diff --git a/LICENSE b/LICENSE
new file mode 100644
index 0000000000000000000000000000000000000000..ecc0896bf1784f869b355fae4b9eab26550a4fef
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 OpenMMLab
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/PowerPaint_v2 b/PowerPaint_v2
new file mode 160000
index 0000000000000000000000000000000000000000..8e039c8f98d8bbdf0c7258104275a4dcf1d1f5fb
--- /dev/null
+++ b/PowerPaint_v2
@@ -0,0 +1 @@
+Subproject commit 8e039c8f98d8bbdf0c7258104275a4dcf1d1f5fb
diff --git a/README.md b/README.md
index 931d2f45a6ecd6c3df134fa4bfd5d6fc2130a8c8..93e288a0f98963db19d285069265a8d6058d5fa9 100644
--- a/README.md
+++ b/README.md
@@ -1,13 +1,97 @@
----
-title: PowerPaint
-emoji: 🐠
-colorFrom: green
-colorTo: gray
-sdk: gradio
-sdk_version: 4.33.0
-app_file: app.py
-pinned: false
-license: mit
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# A Task is Worth One Word: Learning with Task Prompts for High-Quality Versatile Image Inpainting
+
+
+### [Project Page](https://powerpaint.github.io/) | [Paper](https://arxiv.org/abs/2312.03594) | [Online Demo(OpenXlab)](https://openxlab.org.cn/apps/detail/rangoliu/PowerPaint#basic-information)
+
+This README provides a step-by-step guide to download the repository, set up the required virtual environment named "PowerPaint" using conda, and run PowerPaint with or without ControlNet.
+
+**Feel free to try it and give it a star!**:star:
+
+## 🚀 News
+
+**May 22, 2024**:fire:
+
+- We have open-sourced the model weights for PowerPaint v2-1, rectifying some existing issues that were present during the training process of version 2. [](https://huggingface.co/JunhaoZhuang/PowerPaint-v2-1)
+
+**April 7, 2024**:fire:
+
+- We open source the model weights and code for PowerPaint v2. [](https://openxlab.org.cn/models/detail/zhuangjunhao/PowerPaint_v2) [](https://huggingface.co/JunhaoZhuang/PowerPaint_v2)
+
+**April 6, 2024**:
+
+- We have retrained a new PowerPaint, taking inspiration from Brushnet. The [Online Demo](https://openxlab.org.cn/apps/detail/rangoliu/PowerPaint) has been updated accordingly. **We plan to release the model weights and code as open source in the next few days**.
+- Tips: We preserve the cross-attention layer that was deleted by BrushNet for the task prompts input.
+
+| | Object insertion | Object Removal|Shape-guided Object Insertion|Outpainting|
+|-----------------|-----------------|-----------------|-----------------|-----------------|
+| Original Image| ||||
+| Output| | |||
+
+**December 22, 2023**:wrench:
+
+- The logical error in loading ControlNet has been rectified. The `gradio_PowerPaint.py` file and [Online Demo](https://openxlab.org.cn/apps/detail/rangoliu/PowerPaint) have also been updated.
+
+**December 18, 2023**
+
+*Enhanced PowerPaint Model*
+
+- We are delighted to announce the release of more stable model weights. These refined weights can now be accessed on [Hugging Face](https://huggingface.co/JunhaoZhuang/PowerPaint-v1/tree/main). The `gradio_PowerPaint.py` file and [Online Demo](https://openxlab.org.cn/apps/detail/rangoliu/PowerPaint) have also been updated as part of this release.
+
+
+
+________________
+
+
+## Getting Started
+
+```bash
+# Clone the Repository
+git clone https://github.com/zhuang2002/PowerPaint.git
+
+# Navigate to the Repository
+cd projects/powerpaint
+
+# Create Virtual Environment with Conda
+conda create --name PowerPaint python=3.9
+conda activate PowerPaint
+
+# Install Dependencies
+pip install -r requirements.txt
+```
+## PowerPaint v2
+
+```bash
+python gradio_PowerPaint_BrushNet.py
+```
+
+## PowerPaint v1
+
+```bash
+# Create Models Folder
+mkdir models
+
+# Set up Git LFS
+git lfs install
+
+# Clone PowerPaint Model
+git lfs clone https://huggingface.co/JunhaoZhuang/PowerPaint-v1/ ./models
+
+python gradio_PowerPaint.py
+```
+
+This command will launch the Gradio interface for PowerPaint.
+
+Feel free to explore and edit images with PowerPaint!
+
+## BibTeX
+
+```
+@misc{zhuang2023task,
+ title={A Task is Worth One Word: Learning with Task Prompts for High-Quality Versatile Image Inpainting},
+ author={Junhao Zhuang and Yanhong Zeng and Wenran Liu and Chun Yuan and Kai Chen},
+ year={2023},
+ eprint={2312.03594},
+ archivePrefix={arXiv},
+ primaryClass={cs.CV}
+}
+```
diff --git a/app.py b/app.py
new file mode 100644
index 0000000000000000000000000000000000000000..79da621fcc8fa4bca4b69e82c7b750699917604e
--- /dev/null
+++ b/app.py
@@ -0,0 +1,449 @@
+import os
+import random
+
+import gradio as gr
+import numpy as np
+import torch
+from PIL import Image, ImageFilter
+from transformers import CLIPTextModel
+
+from diffusers import UniPCMultistepScheduler
+from model.BrushNet_CA import BrushNetModel
+from model.diffusers_c.models import UNet2DConditionModel
+from pipeline.pipeline_PowerPaint_Brushnet_CA import StableDiffusionPowerPaintBrushNetPipeline
+from utils.utils import TokenizerWrapper, add_tokens
+
+
+base_path = "./PowerPaint_v2"
+os.system("apt install git")
+os.system("apt install git-lfs")
+os.system(f"git lfs clone https://code.openxlab.org.cn/zhuangjunhao/PowerPaint_v2.git {base_path}")
+os.system(f"cd {base_path} && git lfs pull")
+os.system("cd ..")
+torch.set_grad_enabled(False)
+context_prompt = ""
+context_negative_prompt = ""
+base_model_path = "./PowerPaint_v2/realisticVisionV60B1_v51VAE/"
+dtype = torch.float16
+unet = UNet2DConditionModel.from_pretrained(
+ "runwayml/stable-diffusion-v1-5", subfolder="unet", revision=None, torch_dtype=dtype
+)
+text_encoder_brushnet = CLIPTextModel.from_pretrained(
+ "runwayml/stable-diffusion-v1-5", subfolder="text_encoder", revision=None, torch_dtype=dtype
+)
+brushnet = BrushNetModel.from_unet(unet)
+global pipe
+pipe = StableDiffusionPowerPaintBrushNetPipeline.from_pretrained(
+ base_model_path,
+ brushnet=brushnet,
+ text_encoder_brushnet=text_encoder_brushnet,
+ torch_dtype=dtype,
+ low_cpu_mem_usage=False,
+ safety_checker=None,
+)
+pipe.unet = UNet2DConditionModel.from_pretrained(base_model_path, subfolder="unet", revision=None, torch_dtype=dtype)
+pipe.tokenizer = TokenizerWrapper(from_pretrained=base_model_path, subfolder="tokenizer", revision=None)
+add_tokens(
+ tokenizer=pipe.tokenizer,
+ text_encoder=pipe.text_encoder_brushnet,
+ placeholder_tokens=["P_ctxt", "P_shape", "P_obj"],
+ initialize_tokens=["a", "a", "a"],
+ num_vectors_per_token=10,
+)
+from safetensors.torch import load_model
+
+
+load_model(pipe.brushnet, "./PowerPaint_v2/PowerPaint_Brushnet/diffusion_pytorch_model.safetensors")
+
+pipe.text_encoder_brushnet.load_state_dict(
+ torch.load("./PowerPaint_v2/PowerPaint_Brushnet/pytorch_model.bin"), strict=False
+)
+
+pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+
+pipe.enable_model_cpu_offload()
+global current_control
+current_control = "canny"
+# controlnet_conditioning_scale = 0.8
+
+
+def set_seed(seed):
+ torch.manual_seed(seed)
+ torch.cuda.manual_seed(seed)
+ torch.cuda.manual_seed_all(seed)
+ np.random.seed(seed)
+ random.seed(seed)
+
+
+def add_task(control_type):
+ # print(control_type)
+ if control_type == "object-removal":
+ promptA = "P_ctxt"
+ promptB = "P_ctxt"
+ negative_promptA = "P_obj"
+ negative_promptB = "P_obj"
+ elif control_type == "context-aware":
+ promptA = "P_ctxt"
+ promptB = "P_ctxt"
+ negative_promptA = ""
+ negative_promptB = ""
+ elif control_type == "shape-guided":
+ promptA = "P_shape"
+ promptB = "P_ctxt"
+ negative_promptA = "P_shape"
+ negative_promptB = "P_ctxt"
+ elif control_type == "image-outpainting":
+ promptA = "P_ctxt"
+ promptB = "P_ctxt"
+ negative_promptA = "P_obj"
+ negative_promptB = "P_obj"
+ else:
+ promptA = "P_obj"
+ promptB = "P_obj"
+ negative_promptA = "P_obj"
+ negative_promptB = "P_obj"
+
+ return promptA, promptB, negative_promptA, negative_promptB
+
+
+def predict(
+ input_image,
+ prompt,
+ fitting_degree,
+ ddim_steps,
+ scale,
+ seed,
+ negative_prompt,
+ task,
+ vertical_expansion_ratio,
+ horizontal_expansion_ratio,
+):
+ size1, size2 = input_image["image"].convert("RGB").size
+
+ if task != "image-outpainting":
+ if size1 < size2:
+ input_image["image"] = input_image["image"].convert("RGB").resize((640, int(size2 / size1 * 640)))
+ else:
+ input_image["image"] = input_image["image"].convert("RGB").resize((int(size1 / size2 * 640), 640))
+ else:
+ if size1 < size2:
+ input_image["image"] = input_image["image"].convert("RGB").resize((512, int(size2 / size1 * 512)))
+ else:
+ input_image["image"] = input_image["image"].convert("RGB").resize((int(size1 / size2 * 512), 512))
+
+ if task == "image-outpainting" or task == "context-aware":
+ prompt = prompt + " empty scene"
+ if task == "object-removal":
+ prompt = prompt + " empty scene blur"
+
+ if vertical_expansion_ratio != None and horizontal_expansion_ratio != None:
+ o_W, o_H = input_image["image"].convert("RGB").size
+ c_W = int(horizontal_expansion_ratio * o_W)
+ c_H = int(vertical_expansion_ratio * o_H)
+
+ expand_img = np.ones((c_H, c_W, 3), dtype=np.uint8) * 127
+ original_img = np.array(input_image["image"])
+ expand_img[
+ int((c_H - o_H) / 2.0) : int((c_H - o_H) / 2.0) + o_H,
+ int((c_W - o_W) / 2.0) : int((c_W - o_W) / 2.0) + o_W,
+ :,
+ ] = original_img
+
+ blurry_gap = 10
+
+ expand_mask = np.ones((c_H, c_W, 3), dtype=np.uint8) * 255
+ if vertical_expansion_ratio == 1 and horizontal_expansion_ratio != 1:
+ expand_mask[
+ int((c_H - o_H) / 2.0) : int((c_H - o_H) / 2.0) + o_H,
+ int((c_W - o_W) / 2.0) + blurry_gap : int((c_W - o_W) / 2.0) + o_W - blurry_gap,
+ :,
+ ] = 0
+ elif vertical_expansion_ratio != 1 and horizontal_expansion_ratio != 1:
+ expand_mask[
+ int((c_H - o_H) / 2.0) + blurry_gap : int((c_H - o_H) / 2.0) + o_H - blurry_gap,
+ int((c_W - o_W) / 2.0) + blurry_gap : int((c_W - o_W) / 2.0) + o_W - blurry_gap,
+ :,
+ ] = 0
+ elif vertical_expansion_ratio != 1 and horizontal_expansion_ratio == 1:
+ expand_mask[
+ int((c_H - o_H) / 2.0) + blurry_gap : int((c_H - o_H) / 2.0) + o_H - blurry_gap,
+ int((c_W - o_W) / 2.0) : int((c_W - o_W) / 2.0) + o_W,
+ :,
+ ] = 0
+
+ input_image["image"] = Image.fromarray(expand_img)
+ input_image["mask"] = Image.fromarray(expand_mask)
+
+ promptA, promptB, negative_promptA, negative_promptB = add_task(task)
+ img = np.array(input_image["image"].convert("RGB"))
+
+ W = int(np.shape(img)[0] - np.shape(img)[0] % 8)
+ H = int(np.shape(img)[1] - np.shape(img)[1] % 8)
+ input_image["image"] = input_image["image"].resize((H, W))
+ input_image["mask"] = input_image["mask"].resize((H, W))
+
+ np_inpimg = np.array(input_image["image"])
+ np_inmask = np.array(input_image["mask"]) / 255.0
+
+ np_inpimg = np_inpimg * (1 - np_inmask)
+
+ input_image["image"] = Image.fromarray(np_inpimg.astype(np.uint8)).convert("RGB")
+
+ set_seed(seed)
+ global pipe
+ result = pipe(
+ promptA=promptA,
+ promptB=promptB,
+ promptU=prompt,
+ tradoff=fitting_degree,
+ tradoff_nag=fitting_degree,
+ image=input_image["image"].convert("RGB"),
+ mask=input_image["mask"].convert("RGB"),
+ num_inference_steps=ddim_steps,
+ generator=torch.Generator("cuda").manual_seed(seed),
+ brushnet_conditioning_scale=1.0,
+ negative_promptA=negative_promptA,
+ negative_promptB=negative_promptB,
+ negative_promptU=negative_prompt,
+ guidance_scale=scale,
+ width=H,
+ height=W,
+ ).images[0]
+ mask_np = np.array(input_image["mask"].convert("RGB"))
+ red = np.array(result).astype("float") * 1
+ red[:, :, 0] = 180.0
+ red[:, :, 2] = 0
+ red[:, :, 1] = 0
+ result_m = np.array(result)
+ result_m = Image.fromarray(
+ (
+ result_m.astype("float") * (1 - mask_np.astype("float") / 512.0) + mask_np.astype("float") / 512.0 * red
+ ).astype("uint8")
+ )
+ m_img = input_image["mask"].convert("RGB").filter(ImageFilter.GaussianBlur(radius=3))
+ m_img = np.asarray(m_img) / 255.0
+ img_np = np.asarray(input_image["image"].convert("RGB")) / 255.0
+ ours_np = np.asarray(result) / 255.0
+ ours_np = ours_np * m_img + (1 - m_img) * img_np
+ result_paste = Image.fromarray(np.uint8(ours_np * 255))
+
+ dict_res = [input_image["mask"].convert("RGB"), result_m]
+
+ dict_out = [result]
+
+ return dict_out, dict_res
+
+
+def infer(
+ input_image,
+ text_guided_prompt,
+ text_guided_negative_prompt,
+ shape_guided_prompt,
+ shape_guided_negative_prompt,
+ fitting_degree,
+ ddim_steps,
+ scale,
+ seed,
+ task,
+ vertical_expansion_ratio,
+ horizontal_expansion_ratio,
+ outpaint_prompt,
+ outpaint_negative_prompt,
+ removal_prompt,
+ removal_negative_prompt,
+ context_prompt,
+ context_negative_prompt,
+):
+ if task == "text-guided":
+ prompt = text_guided_prompt
+ negative_prompt = text_guided_negative_prompt
+ elif task == "shape-guided":
+ prompt = shape_guided_prompt
+ negative_prompt = shape_guided_negative_prompt
+ elif task == "object-removal":
+ prompt = removal_prompt
+ negative_prompt = removal_negative_prompt
+ elif task == "context-aware":
+ prompt = context_prompt
+ negative_prompt = context_negative_prompt
+ elif task == "image-outpainting":
+ prompt = outpaint_prompt
+ negative_prompt = outpaint_negative_prompt
+ return predict(
+ input_image,
+ prompt,
+ fitting_degree,
+ ddim_steps,
+ scale,
+ seed,
+ negative_prompt,
+ task,
+ vertical_expansion_ratio,
+ horizontal_expansion_ratio,
+ )
+ else:
+ task = "text-guided"
+ prompt = text_guided_prompt
+ negative_prompt = text_guided_negative_prompt
+
+ return predict(input_image, prompt, fitting_degree, ddim_steps, scale, seed, negative_prompt, task, None, None)
+
+
+def select_tab_text_guided():
+ return "text-guided"
+
+
+def select_tab_object_removal():
+ return "object-removal"
+
+
+def select_tab_context_aware():
+ return "context-aware"
+
+
+def select_tab_image_outpainting():
+ return "image-outpainting"
+
+
+def select_tab_shape_guided():
+ return "shape-guided"
+
+
+with gr.Blocks(css="style.css") as demo:
+ with gr.Row():
+ gr.Markdown(
+ "PowerPaint: High-Quality Versatile Image Inpainting
" # noqa
+ )
+ with gr.Row():
+ gr.Markdown(
+ "" # noqa
+ )
+ with gr.Row():
+ gr.Markdown(
+ "**Note:** Due to network-related factors, the page may experience occasional bugs！ If the inpainting results deviate significantly from expectations, consider toggling between task options to refresh the content." # noqa
+ )
+ # Attention: Due to network-related factors, the page may experience occasional bugs. If the inpainting results deviate significantly from expectations, consider toggling between task options to refresh the content.
+ with gr.Row():
+ with gr.Column():
+ gr.Markdown("### Input image and draw mask")
+ input_image = gr.Image(source="upload", tool="sketch", type="pil")
+
+ task = gr.Radio(
+ ["text-guided", "object-removal", "shape-guided", "image-outpainting"], show_label=False, visible=False
+ )
+
+ # Text-guided object inpainting
+ with gr.Tab("Text-guided object inpainting") as tab_text_guided:
+ enable_text_guided = gr.Checkbox(
+ label="Enable text-guided object inpainting", value=True, interactive=False
+ )
+ text_guided_prompt = gr.Textbox(label="Prompt")
+ text_guided_negative_prompt = gr.Textbox(label="negative_prompt")
+ tab_text_guided.select(fn=select_tab_text_guided, inputs=None, outputs=task)
+
+ # Object removal inpainting
+ with gr.Tab("Object removal inpainting") as tab_object_removal:
+ enable_object_removal = gr.Checkbox(
+ label="Enable object removal inpainting",
+ value=True,
+ info="The recommended configuration for the Guidance Scale is 10 or higher. \
+ If undesired objects appear in the masked area, \
+ you can address this by specifically increasing the Guidance Scale.",
+ interactive=False,
+ )
+ removal_prompt = gr.Textbox(label="Prompt")
+ removal_negative_prompt = gr.Textbox(label="negative_prompt")
+ context_prompt = removal_prompt
+ context_negative_prompt = removal_negative_prompt
+ tab_object_removal.select(fn=select_tab_object_removal, inputs=None, outputs=task)
+
+ # Object image outpainting
+ with gr.Tab("Image outpainting") as tab_image_outpainting:
+ enable_object_removal = gr.Checkbox(
+ label="Enable image outpainting",
+ value=True,
+ info="The recommended configuration for the Guidance Scale is 15 or higher. \
+ If unwanted random objects appear in the extended image region, \
+ you can enhance the cleanliness of the extension area by increasing the Guidance Scale.",
+ interactive=False,
+ )
+ outpaint_prompt = gr.Textbox(label="Outpainting_prompt")
+ outpaint_negative_prompt = gr.Textbox(label="Outpainting_negative_prompt")
+ horizontal_expansion_ratio = gr.Slider(
+ label="horizontal expansion ratio",
+ minimum=1,
+ maximum=4,
+ step=0.05,
+ value=1,
+ )
+ vertical_expansion_ratio = gr.Slider(
+ label="vertical expansion ratio",
+ minimum=1,
+ maximum=4,
+ step=0.05,
+ value=1,
+ )
+ tab_image_outpainting.select(fn=select_tab_image_outpainting, inputs=None, outputs=task)
+
+ # Shape-guided object inpainting
+ with gr.Tab("Shape-guided object inpainting") as tab_shape_guided:
+ enable_shape_guided = gr.Checkbox(
+ label="Enable shape-guided object inpainting", value=True, interactive=False
+ )
+ shape_guided_prompt = gr.Textbox(label="shape_guided_prompt")
+ shape_guided_negative_prompt = gr.Textbox(label="shape_guided_negative_prompt")
+ fitting_degree = gr.Slider(
+ label="fitting degree",
+ minimum=0.3,
+ maximum=1,
+ step=0.05,
+ value=1,
+ )
+ tab_shape_guided.select(fn=select_tab_shape_guided, inputs=None, outputs=task)
+
+ run_button = gr.Button(label="Run")
+ with gr.Accordion("Advanced options", open=False):
+ ddim_steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=50, step=1)
+ scale = gr.Slider(label="Guidance Scale", minimum=0.1, maximum=45.0, value=12, step=0.1)
+ seed = gr.Slider(
+ label="Seed",
+ minimum=0,
+ maximum=2147483647,
+ step=1,
+ randomize=True,
+ )
+ with gr.Column():
+ gr.Markdown("### Inpainting result")
+ inpaint_result = gr.Gallery(label="Generated images", show_label=False, columns=2)
+ gr.Markdown("### Mask")
+ gallery = gr.Gallery(label="Generated masks", show_label=False, columns=2)
+
+ run_button.click(
+ fn=infer,
+ inputs=[
+ input_image,
+ text_guided_prompt,
+ text_guided_negative_prompt,
+ shape_guided_prompt,
+ shape_guided_negative_prompt,
+ fitting_degree,
+ ddim_steps,
+ scale,
+ seed,
+ task,
+ vertical_expansion_ratio,
+ horizontal_expansion_ratio,
+ outpaint_prompt,
+ outpaint_negative_prompt,
+ removal_prompt,
+ removal_negative_prompt,
+ context_prompt,
+ context_negative_prompt,
+ ],
+ outputs=[inpaint_result, gallery],
+ )
+
+demo.queue()
+demo.launch(share=False, server_name="0.0.0.0", server_port=7860)
diff --git a/gradio_PowerPaint.py b/gradio_PowerPaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..0b35344a25b56ab5658fd5151deac5f518c19bcb
--- /dev/null
+++ b/gradio_PowerPaint.py
@@ -0,0 +1,573 @@
+import random
+
+import cv2
+import gradio as gr
+import numpy as np
+import torch
+from controlnet_aux import HEDdetector, OpenposeDetector
+from PIL import Image, ImageFilter
+from transformers import DPTFeatureExtractor, DPTForDepthEstimation
+
+from diffusers.pipelines.controlnet.pipeline_controlnet import ControlNetModel
+from pipeline.pipeline_PowerPaint import StableDiffusionInpaintPipeline as Pipeline
+from pipeline.pipeline_PowerPaint_ControlNet import StableDiffusionControlNetInpaintPipeline as controlnetPipeline
+from utils.utils import TokenizerWrapper, add_tokens
+
+
+torch.set_grad_enabled(False)
+
+weight_dtype = torch.float16
+global pipe
+pipe = Pipeline.from_pretrained("runwayml/stable-diffusion-inpainting", torch_dtype=weight_dtype)
+pipe.tokenizer = TokenizerWrapper(
+ from_pretrained="runwayml/stable-diffusion-v1-5", subfolder="tokenizer", revision=None
+)
+
+add_tokens(
+ tokenizer=pipe.tokenizer,
+ text_encoder=pipe.text_encoder,
+ placeholder_tokens=["P_ctxt", "P_shape", "P_obj"],
+ initialize_tokens=["a", "a", "a"],
+ num_vectors_per_token=10,
+)
+
+from safetensors.torch import load_model
+
+
+load_model(pipe.unet, "./models/unet/unet.safetensors")
+load_model(pipe.text_encoder, "./models/unet/text_encoder.safetensors")
+pipe = pipe.to("cuda")
+
+
+depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda")
+feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas")
+openpose = OpenposeDetector.from_pretrained("lllyasviel/ControlNet")
+hed = HEDdetector.from_pretrained("lllyasviel/ControlNet")
+
+global current_control
+current_control = "canny"
+# controlnet_conditioning_scale = 0.8
+
+
+def set_seed(seed):
+ torch.manual_seed(seed)
+ torch.cuda.manual_seed(seed)
+ torch.cuda.manual_seed_all(seed)
+ np.random.seed(seed)
+ random.seed(seed)
+
+
+def get_depth_map(image):
+ image = feature_extractor(images=image, return_tensors="pt").pixel_values.to("cuda")
+ with torch.no_grad(), torch.autocast("cuda"):
+ depth_map = depth_estimator(image).predicted_depth
+
+ depth_map = torch.nn.functional.interpolate(
+ depth_map.unsqueeze(1),
+ size=(1024, 1024),
+ mode="bicubic",
+ align_corners=False,
+ )
+ depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
+ depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
+ depth_map = (depth_map - depth_min) / (depth_max - depth_min)
+ image = torch.cat([depth_map] * 3, dim=1)
+
+ image = image.permute(0, 2, 3, 1).cpu().numpy()[0]
+ image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8))
+ return image
+
+
+def add_task(prompt, negative_prompt, control_type):
+ # print(control_type)
+ if control_type == "object-removal":
+ promptA = "empty scene blur " + prompt + " P_ctxt"
+ promptB = "empty scene blur " + prompt + " P_ctxt"
+ negative_promptA = negative_prompt + " P_obj"
+ negative_promptB = negative_prompt + " P_obj"
+ elif control_type == "shape-guided":
+ promptA = prompt + " P_shape"
+ promptB = prompt + " P_ctxt"
+ negative_promptA = (
+ negative_prompt + ", worst quality, low quality, normal quality, bad quality, blurry P_shape"
+ )
+ negative_promptB = negative_prompt + ", worst quality, low quality, normal quality, bad quality, blurry P_ctxt"
+ elif control_type == "image-outpainting":
+ promptA = "empty scene " + prompt + " P_ctxt"
+ promptB = "empty scene " + prompt + " P_ctxt"
+ negative_promptA = negative_prompt + " P_obj"
+ negative_promptB = negative_prompt + " P_obj"
+ else:
+ promptA = prompt + " P_obj"
+ promptB = prompt + " P_obj"
+ negative_promptA = negative_prompt + ", worst quality, low quality, normal quality, bad quality, blurry, P_obj"
+ negative_promptB = negative_prompt + ", worst quality, low quality, normal quality, bad quality, blurry, P_obj"
+
+ return promptA, promptB, negative_promptA, negative_promptB
+
+
+def predict(
+ input_image,
+ prompt,
+ fitting_degree,
+ ddim_steps,
+ scale,
+ seed,
+ negative_prompt,
+ task,
+ vertical_expansion_ratio,
+ horizontal_expansion_ratio,
+):
+ size1, size2 = input_image["image"].convert("RGB").size
+
+ if task != "image-outpainting":
+ if size1 < size2:
+ input_image["image"] = input_image["image"].convert("RGB").resize((640, int(size2 / size1 * 640)))
+ else:
+ input_image["image"] = input_image["image"].convert("RGB").resize((int(size1 / size2 * 640), 640))
+ else:
+ if size1 < size2:
+ input_image["image"] = input_image["image"].convert("RGB").resize((512, int(size2 / size1 * 512)))
+ else:
+ input_image["image"] = input_image["image"].convert("RGB").resize((int(size1 / size2 * 512), 512))
+
+ if vertical_expansion_ratio != None and horizontal_expansion_ratio != None:
+ o_W, o_H = input_image["image"].convert("RGB").size
+ c_W = int(horizontal_expansion_ratio * o_W)
+ c_H = int(vertical_expansion_ratio * o_H)
+
+ expand_img = np.ones((c_H, c_W, 3), dtype=np.uint8) * 127
+ original_img = np.array(input_image["image"])
+ expand_img[
+ int((c_H - o_H) / 2.0) : int((c_H - o_H) / 2.0) + o_H,
+ int((c_W - o_W) / 2.0) : int((c_W - o_W) / 2.0) + o_W,
+ :,
+ ] = original_img
+
+ blurry_gap = 10
+
+ expand_mask = np.ones((c_H, c_W, 3), dtype=np.uint8) * 255
+ if vertical_expansion_ratio == 1 and horizontal_expansion_ratio != 1:
+ expand_mask[
+ int((c_H - o_H) / 2.0) : int((c_H - o_H) / 2.0) + o_H,
+ int((c_W - o_W) / 2.0) + blurry_gap : int((c_W - o_W) / 2.0) + o_W - blurry_gap,
+ :,
+ ] = 0
+ elif vertical_expansion_ratio != 1 and horizontal_expansion_ratio != 1:
+ expand_mask[
+ int((c_H - o_H) / 2.0) + blurry_gap : int((c_H - o_H) / 2.0) + o_H - blurry_gap,
+ int((c_W - o_W) / 2.0) + blurry_gap : int((c_W - o_W) / 2.0) + o_W - blurry_gap,
+ :,
+ ] = 0
+ elif vertical_expansion_ratio != 1 and horizontal_expansion_ratio == 1:
+ expand_mask[
+ int((c_H - o_H) / 2.0) + blurry_gap : int((c_H - o_H) / 2.0) + o_H - blurry_gap,
+ int((c_W - o_W) / 2.0) : int((c_W - o_W) / 2.0) + o_W,
+ :,
+ ] = 0
+
+ input_image["image"] = Image.fromarray(expand_img)
+ input_image["mask"] = Image.fromarray(expand_mask)
+
+ promptA, promptB, negative_promptA, negative_promptB = add_task(prompt, negative_prompt, task)
+ print(promptA, promptB, negative_promptA, negative_promptB)
+ img = np.array(input_image["image"].convert("RGB"))
+
+ W = int(np.shape(img)[0] - np.shape(img)[0] % 8)
+ H = int(np.shape(img)[1] - np.shape(img)[1] % 8)
+ input_image["image"] = input_image["image"].resize((H, W))
+ input_image["mask"] = input_image["mask"].resize((H, W))
+ set_seed(seed)
+ global pipe
+ result = pipe(
+ promptA=promptA,
+ promptB=promptB,
+ tradoff=fitting_degree,
+ tradoff_nag=fitting_degree,
+ negative_promptA=negative_promptA,
+ negative_promptB=negative_promptB,
+ image=input_image["image"].convert("RGB"),
+ mask_image=input_image["mask"].convert("RGB"),
+ width=H,
+ height=W,
+ guidance_scale=scale,
+ num_inference_steps=ddim_steps,
+ ).images[0]
+ mask_np = np.array(input_image["mask"].convert("RGB"))
+ red = np.array(result).astype("float") * 1
+ red[:, :, 0] = 180.0
+ red[:, :, 2] = 0
+ red[:, :, 1] = 0
+ result_m = np.array(result)
+ result_m = Image.fromarray(
+ (
+ result_m.astype("float") * (1 - mask_np.astype("float") / 512.0) + mask_np.astype("float") / 512.0 * red
+ ).astype("uint8")
+ )
+ m_img = input_image["mask"].convert("RGB").filter(ImageFilter.GaussianBlur(radius=3))
+ m_img = np.asarray(m_img) / 255.0
+ img_np = np.asarray(input_image["image"].convert("RGB")) / 255.0
+ ours_np = np.asarray(result) / 255.0
+ ours_np = ours_np * m_img + (1 - m_img) * img_np
+ result_paste = Image.fromarray(np.uint8(ours_np * 255))
+
+ dict_res = [input_image["mask"].convert("RGB"), result_m]
+
+ dict_out = [input_image["image"].convert("RGB"), result_paste]
+
+ return dict_out, dict_res
+
+
+def predict_controlnet(
+ input_image,
+ input_control_image,
+ control_type,
+ prompt,
+ ddim_steps,
+ scale,
+ seed,
+ negative_prompt,
+ controlnet_conditioning_scale,
+):
+ promptA = prompt + " P_obj"
+ promptB = prompt + " P_obj"
+ negative_promptA = negative_prompt
+ negative_promptB = negative_prompt
+ size1, size2 = input_image["image"].convert("RGB").size
+
+ if size1 < size2:
+ input_image["image"] = input_image["image"].convert("RGB").resize((640, int(size2 / size1 * 640)))
+ else:
+ input_image["image"] = input_image["image"].convert("RGB").resize((int(size1 / size2 * 640), 640))
+ img = np.array(input_image["image"].convert("RGB"))
+ W = int(np.shape(img)[0] - np.shape(img)[0] % 8)
+ H = int(np.shape(img)[1] - np.shape(img)[1] % 8)
+ input_image["image"] = input_image["image"].resize((H, W))
+ input_image["mask"] = input_image["mask"].resize((H, W))
+
+ global current_control
+ global pipe
+
+ base_control = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=weight_dtype)
+ control_pipe = controlnetPipeline(
+ pipe.vae, pipe.text_encoder, pipe.tokenizer, pipe.unet, base_control, pipe.scheduler, None, None, False
+ )
+ control_pipe = control_pipe.to("cuda")
+ current_control = "canny"
+ if current_control != control_type:
+ if control_type == "canny" or control_type is None:
+ control_pipe.controlnet = ControlNetModel.from_pretrained(
+ "lllyasviel/sd-controlnet-canny", torch_dtype=weight_dtype
+ )
+ elif control_type == "pose":
+ control_pipe.controlnet = ControlNetModel.from_pretrained(
+ "lllyasviel/sd-controlnet-openpose", torch_dtype=weight_dtype
+ )
+ elif control_type == "depth":
+ control_pipe.controlnet = ControlNetModel.from_pretrained(
+ "lllyasviel/sd-controlnet-depth", torch_dtype=weight_dtype
+ )
+ else:
+ control_pipe.controlnet = ControlNetModel.from_pretrained(
+ "lllyasviel/sd-controlnet-hed", torch_dtype=weight_dtype
+ )
+ control_pipe = control_pipe.to("cuda")
+ current_control = control_type
+
+ controlnet_image = input_control_image
+ if current_control == "canny":
+ controlnet_image = controlnet_image.resize((H, W))
+ controlnet_image = np.array(controlnet_image)
+ controlnet_image = cv2.Canny(controlnet_image, 100, 200)
+ controlnet_image = controlnet_image[:, :, None]
+ controlnet_image = np.concatenate([controlnet_image, controlnet_image, controlnet_image], axis=2)
+ controlnet_image = Image.fromarray(controlnet_image)
+ elif current_control == "pose":
+ controlnet_image = openpose(controlnet_image)
+ elif current_control == "depth":
+ controlnet_image = controlnet_image.resize((H, W))
+ controlnet_image = get_depth_map(controlnet_image)
+ else:
+ controlnet_image = hed(controlnet_image)
+
+ mask_np = np.array(input_image["mask"].convert("RGB"))
+ controlnet_image = controlnet_image.resize((H, W))
+ set_seed(seed)
+ result = control_pipe(
+ promptA=promptB,
+ promptB=promptA,
+ tradoff=1.0,
+ tradoff_nag=1.0,
+ negative_promptA=negative_promptA,
+ negative_promptB=negative_promptB,
+ image=input_image["image"].convert("RGB"),
+ mask_image=input_image["mask"].convert("RGB"),
+ control_image=controlnet_image,
+ width=H,
+ height=W,
+ guidance_scale=scale,
+ controlnet_conditioning_scale=controlnet_conditioning_scale,
+ num_inference_steps=ddim_steps,
+ ).images[0]
+ red = np.array(result).astype("float") * 1
+ red[:, :, 0] = 180.0
+ red[:, :, 2] = 0
+ red[:, :, 1] = 0
+ result_m = np.array(result)
+ result_m = Image.fromarray(
+ (
+ result_m.astype("float") * (1 - mask_np.astype("float") / 512.0) + mask_np.astype("float") / 512.0 * red
+ ).astype("uint8")
+ )
+
+ mask_np = np.array(input_image["mask"].convert("RGB"))
+ m_img = input_image["mask"].convert("RGB").filter(ImageFilter.GaussianBlur(radius=4))
+ m_img = np.asarray(m_img) / 255.0
+ img_np = np.asarray(input_image["image"].convert("RGB")) / 255.0
+ ours_np = np.asarray(result) / 255.0
+ ours_np = ours_np * m_img + (1 - m_img) * img_np
+ result_paste = Image.fromarray(np.uint8(ours_np * 255))
+ return [input_image["image"].convert("RGB"), result_paste], [controlnet_image, result_m]
+
+
+def infer(
+ input_image,
+ text_guided_prompt,
+ text_guided_negative_prompt,
+ shape_guided_prompt,
+ shape_guided_negative_prompt,
+ fitting_degree,
+ ddim_steps,
+ scale,
+ seed,
+ task,
+ enable_control,
+ input_control_image,
+ control_type,
+ vertical_expansion_ratio,
+ horizontal_expansion_ratio,
+ outpaint_prompt,
+ outpaint_negative_prompt,
+ controlnet_conditioning_scale,
+ removal_prompt,
+ removal_negative_prompt,
+):
+ if task == "text-guided":
+ prompt = text_guided_prompt
+ negative_prompt = text_guided_negative_prompt
+ elif task == "shape-guided":
+ prompt = shape_guided_prompt
+ negative_prompt = shape_guided_negative_prompt
+ elif task == "object-removal":
+ prompt = removal_prompt
+ negative_prompt = removal_negative_prompt
+ elif task == "image-outpainting":
+ prompt = outpaint_prompt
+ negative_prompt = outpaint_negative_prompt
+ return predict(
+ input_image,
+ prompt,
+ fitting_degree,
+ ddim_steps,
+ scale,
+ seed,
+ negative_prompt,
+ task,
+ vertical_expansion_ratio,
+ horizontal_expansion_ratio,
+ )
+ else:
+ task = "text-guided"
+ prompt = text_guided_prompt
+ negative_prompt = text_guided_negative_prompt
+
+ if enable_control and task == "text-guided":
+ return predict_controlnet(
+ input_image,
+ input_control_image,
+ control_type,
+ prompt,
+ ddim_steps,
+ scale,
+ seed,
+ negative_prompt,
+ controlnet_conditioning_scale,
+ )
+ else:
+ return predict(input_image, prompt, fitting_degree, ddim_steps, scale, seed, negative_prompt, task, None, None)
+
+
+def select_tab_text_guided():
+ return "text-guided"
+
+
+def select_tab_object_removal():
+ return "object-removal"
+
+
+def select_tab_image_outpainting():
+ return "image-outpainting"
+
+
+def select_tab_shape_guided():
+ return "shape-guided"
+
+
+with gr.Blocks(css="style.css") as demo:
+ with gr.Row():
+ gr.Markdown(
+ "PowerPaint: High-Quality Versatile Image Inpainting
" # noqa
+ )
+ with gr.Row():
+ gr.Markdown(
+ "" # noqa
+ )
+ with gr.Row():
+ gr.Markdown(
+ "**Note:** Due to network-related factors, the page may experience occasional bugs！ If the inpainting results deviate significantly from expectations, consider toggling between task options to refresh the content." # noqa
+ )
+ # Attention: Due to network-related factors, the page may experience occasional bugs. If the inpainting results deviate significantly from expectations, consider toggling between task options to refresh the content.
+ with gr.Row():
+ with gr.Column():
+ gr.Markdown("### Input image and draw mask")
+ input_image = gr.Image(source="upload", tool="sketch", type="pil")
+
+ task = gr.Radio(
+ ["text-guided", "object-removal", "shape-guided", "image-outpainting"], show_label=False, visible=False
+ )
+
+ # Text-guided object inpainting
+ with gr.Tab("Text-guided object inpainting") as tab_text_guided:
+ enable_text_guided = gr.Checkbox(
+ label="Enable text-guided object inpainting", value=True, interactive=False
+ )
+ text_guided_prompt = gr.Textbox(label="Prompt")
+ text_guided_negative_prompt = gr.Textbox(label="negative_prompt")
+ gr.Markdown("### Controlnet setting")
+ enable_control = gr.Checkbox(
+ label="Enable controlnet", info="Enable this if you want to use controlnet"
+ )
+ controlnet_conditioning_scale = gr.Slider(
+ label="controlnet conditioning scale",
+ minimum=0,
+ maximum=1,
+ step=0.05,
+ value=0.5,
+ )
+ control_type = gr.Radio(["canny", "pose", "depth", "hed"], label="Control type")
+ input_control_image = gr.Image(source="upload", type="pil")
+ tab_text_guided.select(fn=select_tab_text_guided, inputs=None, outputs=task)
+
+ # Object removal inpainting
+ with gr.Tab("Object removal inpainting") as tab_object_removal:
+ enable_object_removal = gr.Checkbox(
+ label="Enable object removal inpainting",
+ value=True,
+ info="The recommended configuration for the Guidance Scale is 10 or higher. \
+ If undesired objects appear in the masked area, \
+ you can address this by specifically increasing the Guidance Scale.",
+ interactive=False,
+ )
+ removal_prompt = gr.Textbox(label="Prompt")
+ removal_negative_prompt = gr.Textbox(label="negative_prompt")
+ tab_object_removal.select(fn=select_tab_object_removal, inputs=None, outputs=task)
+
+ # Object image outpainting
+ with gr.Tab("Image outpainting") as tab_image_outpainting:
+ enable_object_removal = gr.Checkbox(
+ label="Enable image outpainting",
+ value=True,
+ info="The recommended configuration for the Guidance Scale is 10 or higher. \
+ If unwanted random objects appear in the extended image region, \
+ you can enhance the cleanliness of the extension area by increasing the Guidance Scale.",
+ interactive=False,
+ )
+ outpaint_prompt = gr.Textbox(label="Outpainting_prompt")
+ outpaint_negative_prompt = gr.Textbox(label="Outpainting_negative_prompt")
+ horizontal_expansion_ratio = gr.Slider(
+ label="horizontal expansion ratio",
+ minimum=1,
+ maximum=4,
+ step=0.05,
+ value=1,
+ )
+ vertical_expansion_ratio = gr.Slider(
+ label="vertical expansion ratio",
+ minimum=1,
+ maximum=4,
+ step=0.05,
+ value=1,
+ )
+ tab_image_outpainting.select(fn=select_tab_image_outpainting, inputs=None, outputs=task)
+
+ # Shape-guided object inpainting
+ with gr.Tab("Shape-guided object inpainting") as tab_shape_guided:
+ enable_shape_guided = gr.Checkbox(
+ label="Enable shape-guided object inpainting", value=True, interactive=False
+ )
+ shape_guided_prompt = gr.Textbox(label="shape_guided_prompt")
+ shape_guided_negative_prompt = gr.Textbox(label="shape_guided_negative_prompt")
+ fitting_degree = gr.Slider(
+ label="fitting degree",
+ minimum=0,
+ maximum=1,
+ step=0.05,
+ value=1,
+ )
+ tab_shape_guided.select(fn=select_tab_shape_guided, inputs=None, outputs=task)
+
+ run_button = gr.Button(label="Run")
+ with gr.Accordion("Advanced options", open=False):
+ ddim_steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=45, step=1)
+ scale = gr.Slider(
+ label="Guidance Scale",
+ info="For object removal and image outpainting, it is recommended to set the value at 10 or above.",
+ minimum=0.1,
+ maximum=30.0,
+ value=7.5,
+ step=0.1,
+ )
+ seed = gr.Slider(
+ label="Seed",
+ minimum=0,
+ maximum=2147483647,
+ step=1,
+ randomize=True,
+ )
+ with gr.Column():
+ gr.Markdown("### Inpainting result")
+ inpaint_result = gr.Gallery(label="Generated images", show_label=False, columns=2)
+ gr.Markdown("### Mask")
+ gallery = gr.Gallery(label="Generated masks", show_label=False, columns=2)
+
+ run_button.click(
+ fn=infer,
+ inputs=[
+ input_image,
+ text_guided_prompt,
+ text_guided_negative_prompt,
+ shape_guided_prompt,
+ shape_guided_negative_prompt,
+ fitting_degree,
+ ddim_steps,
+ scale,
+ seed,
+ task,
+ enable_control,
+ input_control_image,
+ control_type,
+ vertical_expansion_ratio,
+ horizontal_expansion_ratio,
+ outpaint_prompt,
+ outpaint_negative_prompt,
+ controlnet_conditioning_scale,
+ removal_prompt,
+ removal_negative_prompt,
+ ],
+ outputs=[inpaint_result, gallery],
+ )
+
+demo.queue()
+demo.launch(share=False, server_name="0.0.0.0", server_port=7860)
diff --git a/model/BrushNet_CA.py b/model/BrushNet_CA.py
new file mode 100644
index 0000000000000000000000000000000000000000..6a6e04f7c6debd1acd541e34227748aecb613d66
--- /dev/null
+++ b/model/BrushNet_CA.py
@@ -0,0 +1,960 @@
+import sys
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+
+sys.path.append(".model")
+import torch
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from diffusers.models.embeddings import (
+ TextImageProjection,
+ TextImageTimeEmbedding,
+ TextTimeEmbedding,
+ TimestepEmbedding,
+ Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils import BaseOutput, logging
+from model.diffusers_c.models.unets.unet_2d_blocks import (
+ CrossAttnDownBlock2D,
+ DownBlock2D,
+ get_down_block,
+ get_mid_block,
+ get_up_block,
+)
+from model.diffusers_c.models.unets.unet_2d_condition import UNet2DConditionModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class BrushNetOutput(BaseOutput):
+ """
+ The output of [`BrushNetModel`].
+
+ Args:
+ up_block_res_samples (`tuple[torch.Tensor]`):
+ A tuple of upsample activations at different resolutions for each upsampling block. Each tensor should
+ be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+ used to condition the original UNet's upsampling activations.
+ down_block_res_samples (`tuple[torch.Tensor]`):
+ A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
+ be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+ used to condition the original UNet's downsampling activations.
+ mid_down_block_re_sample (`torch.Tensor`):
+ The activation of the midde block (the lowest sample resolution). Each tensor should be of shape
+ `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
+ Output can be used to condition the original UNet's middle block activation.
+ """
+
+ up_block_res_samples: Tuple[torch.Tensor]
+ down_block_res_samples: Tuple[torch.Tensor]
+ mid_block_res_sample: torch.Tensor
+
+
+class BrushNetModel(ModelMixin, ConfigMixin):
+ """
+ A BrushNet model.
+
+ Args:
+ in_channels (`int`, defaults to 4):
+ The number of channels in the input sample.
+ flip_sin_to_cos (`bool`, defaults to `True`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, defaults to 0):
+ The frequency shift to apply to the time embedding.
+ down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+ The tuple of downsample blocks to use.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+ Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+ `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+ The tuple of upsample blocks to use.
+ only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
+ block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, defaults to 2):
+ The number of layers per block.
+ downsample_padding (`int`, defaults to 1):
+ The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, defaults to 1):
+ The scale factor to use for the mid block.
+ act_fn (`str`, defaults to "silu"):
+ The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use for the normalization. If None, normalization and activation layers is skipped
+ in post-processing.
+ norm_eps (`float`, defaults to 1e-5):
+ The epsilon to use for the normalization.
+ cross_attention_dim (`int`, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ encoder_hid_dim (`int`, *optional*, defaults to None):
+ If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+ dimension to `cross_attention_dim`.
+ encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+ If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+ embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+ attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
+ The dimension of the attention heads.
+ use_linear_projection (`bool`, defaults to `False`):
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None,
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+ addition_embed_type (`str`, *optional*, defaults to `None`):
+ Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+ "text". "text" will use the `TextTimeEmbedding` layer.
+ num_class_embeds (`int`, *optional*, defaults to 0):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+ class conditioning with `class_embed_type` equal to `None`.
+ upcast_attention (`bool`, defaults to `False`):
+ resnet_time_scale_shift (`str`, defaults to `"default"`):
+ Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
+ projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
+ The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
+ `class_embed_type="projection"`.
+ brushnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+ The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+ conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
+ The tuple of output channel for each block in the `conditioning_embedding` layer.
+ global_pool_conditions (`bool`, defaults to `False`):
+ TODO(Patrick) - unused parameter.
+ addition_embed_type_num_heads (`int`, defaults to 64):
+ The number of heads to use for the `TextTimeEmbedding` layer.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 4,
+ conditioning_channels: int = 5,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ),
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+ up_block_types: Tuple[str, ...] = (
+ "UpBlock2D",
+ "CrossAttnUpBlock2D",
+ "CrossAttnUpBlock2D",
+ "CrossAttnUpBlock2D",
+ ),
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+ layers_per_block: int = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: int = 1280,
+ transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+ encoder_hid_dim: Optional[int] = None,
+ encoder_hid_dim_type: Optional[str] = None,
+ attention_head_dim: Union[int, Tuple[int, ...]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
+ use_linear_projection: bool = False,
+ class_embed_type: Optional[str] = None,
+ addition_embed_type: Optional[str] = None,
+ addition_time_embed_dim: Optional[int] = None,
+ num_class_embeds: Optional[int] = None,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ projection_class_embeddings_input_dim: Optional[int] = None,
+ brushnet_conditioning_channel_order: str = "rgb",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ global_pool_conditions: bool = False,
+ addition_embed_type_num_heads: int = 64,
+ ):
+ super().__init__()
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+ # input
+ conv_in_kernel = 3
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in_condition = nn.Conv2d(
+ in_channels + conditioning_channels,
+ block_out_channels[0],
+ kernel_size=conv_in_kernel,
+ padding=conv_in_padding,
+ )
+
+ # time
+ time_embed_dim = block_out_channels[0] * 4
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+ timestep_input_dim = block_out_channels[0]
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ )
+
+ if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+ encoder_hid_dim_type = "text_proj"
+ self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+ logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+ if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+ )
+
+ if encoder_hid_dim_type == "text_proj":
+ self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+ elif encoder_hid_dim_type == "text_image_proj":
+ # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+ self.encoder_hid_proj = TextImageProjection(
+ text_embed_dim=encoder_hid_dim,
+ image_embed_dim=cross_attention_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+
+ elif encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+ )
+ else:
+ self.encoder_hid_proj = None
+
+ # class embedding
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ elif class_embed_type == "projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+ # 2. it projects from an arbitrary input dimension.
+ #
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+ self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ else:
+ self.class_embedding = None
+
+ if addition_embed_type == "text":
+ if encoder_hid_dim is not None:
+ text_time_embedding_from_dim = encoder_hid_dim
+ else:
+ text_time_embedding_from_dim = cross_attention_dim
+
+ self.add_embedding = TextTimeEmbedding(
+ text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+ )
+ elif addition_embed_type == "text_image":
+ # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+ self.add_embedding = TextImageTimeEmbedding(
+ text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+ )
+ elif addition_embed_type == "text_time":
+ self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+ self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+ elif addition_embed_type is not None:
+ raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+ self.down_blocks = nn.ModuleList([])
+ self.brushnet_down_blocks = nn.ModuleList([])
+
+ if isinstance(only_cross_attention, bool):
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+ if isinstance(attention_head_dim, int):
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ # down
+ output_channel = block_out_channels[0]
+
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[i],
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ downsample_padding=downsample_padding,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ self.down_blocks.append(down_block)
+
+ for _ in range(layers_per_block):
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ if not is_final_block:
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ # mid
+ mid_block_channel = block_out_channels[-1]
+
+ brushnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_mid_block = brushnet_block
+
+ self.mid_block = get_mid_block(
+ mid_block_type,
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ in_channels=mid_block_channel,
+ temb_channels=time_embed_dim,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+ only_cross_attention = list(reversed(only_cross_attention))
+
+ output_channel = reversed_block_out_channels[0]
+
+ self.up_blocks = nn.ModuleList([])
+ self.brushnet_up_blocks = nn.ModuleList([])
+
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=layers_per_block + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resolution_idx=i,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=reversed_num_attention_heads[i],
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ for _ in range(layers_per_block + 1):
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_up_blocks.append(brushnet_block)
+
+ if not is_final_block:
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_up_blocks.append(brushnet_block)
+
+ @classmethod
+ def from_unet(
+ cls,
+ unet: UNet2DConditionModel,
+ brushnet_conditioning_channel_order: str = "rgb",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ load_weights_from_unet: bool = True,
+ conditioning_channels: int = 5,
+ ):
+ r"""
+ Instantiate a [`BrushNetModel`] from [`UNet2DConditionModel`].
+
+ Parameters:
+ unet (`UNet2DConditionModel`):
+ The UNet model weights to copy to the [`BrushNetModel`]. All configuration options are also copied
+ where applicable.
+ """
+ transformer_layers_per_block = (
+ unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
+ )
+ encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
+ encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None
+ addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
+ addition_time_embed_dim = (
+ unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
+ )
+
+ brushnet = cls(
+ in_channels=unet.config.in_channels,
+ conditioning_channels=conditioning_channels,
+ flip_sin_to_cos=unet.config.flip_sin_to_cos,
+ freq_shift=unet.config.freq_shift,
+ # down_block_types=['DownBlock2D','DownBlock2D','DownBlock2D','DownBlock2D'],
+ down_block_types=[
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ],
+ # mid_block_type='MidBlock2D',
+ mid_block_type="UNetMidBlock2DCrossAttn",
+ # up_block_types=['UpBlock2D','UpBlock2D','UpBlock2D','UpBlock2D'],
+ up_block_types=["UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"],
+ only_cross_attention=unet.config.only_cross_attention,
+ block_out_channels=unet.config.block_out_channels,
+ layers_per_block=unet.config.layers_per_block,
+ downsample_padding=unet.config.downsample_padding,
+ mid_block_scale_factor=unet.config.mid_block_scale_factor,
+ act_fn=unet.config.act_fn,
+ norm_num_groups=unet.config.norm_num_groups,
+ norm_eps=unet.config.norm_eps,
+ cross_attention_dim=unet.config.cross_attention_dim,
+ transformer_layers_per_block=transformer_layers_per_block,
+ encoder_hid_dim=encoder_hid_dim,
+ encoder_hid_dim_type=encoder_hid_dim_type,
+ attention_head_dim=unet.config.attention_head_dim,
+ num_attention_heads=unet.config.num_attention_heads,
+ use_linear_projection=unet.config.use_linear_projection,
+ class_embed_type=unet.config.class_embed_type,
+ addition_embed_type=addition_embed_type,
+ addition_time_embed_dim=addition_time_embed_dim,
+ num_class_embeds=unet.config.num_class_embeds,
+ upcast_attention=unet.config.upcast_attention,
+ resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+ projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
+ brushnet_conditioning_channel_order=brushnet_conditioning_channel_order,
+ conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+ )
+
+ if load_weights_from_unet:
+ conv_in_condition_weight = torch.zeros_like(brushnet.conv_in_condition.weight)
+ conv_in_condition_weight[:, :4, ...] = unet.conv_in.weight
+ conv_in_condition_weight[:, 4:8, ...] = unet.conv_in.weight
+ brushnet.conv_in_condition.weight = torch.nn.Parameter(conv_in_condition_weight)
+ brushnet.conv_in_condition.bias = unet.conv_in.bias
+
+ brushnet.time_proj.load_state_dict(unet.time_proj.state_dict())
+ brushnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+ if brushnet.class_embedding:
+ brushnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
+
+ brushnet.down_blocks.load_state_dict(unet.down_blocks.state_dict(), strict=False)
+ brushnet.mid_block.load_state_dict(unet.mid_block.state_dict(), strict=False)
+ brushnet.up_blocks.load_state_dict(unet.up_blocks.state_dict(), strict=False)
+
+ return brushnet.to(unet.dtype)
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+ def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+ module.gradient_checkpointing = value
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ brushnet_cond: torch.FloatTensor,
+ conditioning_scale: float = 1.0,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guess_mode: bool = False,
+ return_dict: bool = True,
+ ) -> Union[BrushNetOutput, Tuple[Tuple[torch.FloatTensor, ...], torch.FloatTensor]]:
+ """
+ The [`BrushNetModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor.
+ timestep (`Union[torch.Tensor, float, int]`):
+ The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.Tensor`):
+ The encoder hidden states.
+ brushnet_cond (`torch.FloatTensor`):
+ The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+ conditioning_scale (`float`, defaults to `1.0`):
+ The scale factor for BrushNet outputs.
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+ Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+ timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+ embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ added_cond_kwargs (`dict`):
+ Additional conditions for the Stable Diffusion XL UNet.
+ cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+ A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+ guess_mode (`bool`, defaults to `False`):
+ In this mode, the BrushNet encoder tries its best to recognize the input content of the input even if
+ you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
+ return_dict (`bool`, defaults to `True`):
+ Whether or not to return a [`~models.brushnet.BrushNetOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.brushnet.BrushNetOutput`] **or** `tuple`:
+ If `return_dict` is `True`, a [`~models.brushnet.BrushNetOutput`] is returned, otherwise a tuple is
+ returned where the first element is the sample tensor.
+ """
+ # check channel order
+ channel_order = self.config.brushnet_conditioning_channel_order
+
+ if channel_order == "rgb":
+ # in rgb order by default
+ ...
+ elif channel_order == "bgr":
+ brushnet_cond = torch.flip(brushnet_cond, dims=[1])
+ else:
+ raise ValueError(f"unknown `brushnet_conditioning_channel_order`: {channel_order}")
+
+ # prepare attention_mask
+ if attention_mask is not None:
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+
+ emb = self.time_embedding(t_emb, timestep_cond)
+ aug_emb = None
+
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+ emb = emb + class_emb
+
+ if self.config.addition_embed_type is not None:
+ if self.config.addition_embed_type == "text":
+ aug_emb = self.add_embedding(encoder_hidden_states)
+
+ elif self.config.addition_embed_type == "text_time":
+ if "text_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+ )
+ text_embeds = added_cond_kwargs.get("text_embeds")
+ if "time_ids" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+ )
+ time_ids = added_cond_kwargs.get("time_ids")
+ time_embeds = self.add_time_proj(time_ids.flatten())
+ time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+ add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+ add_embeds = add_embeds.to(emb.dtype)
+ aug_emb = self.add_embedding(add_embeds)
+
+ emb = emb + aug_emb if aug_emb is not None else emb
+
+ # 2. pre-process
+ brushnet_cond = torch.concat([sample, brushnet_cond], 1)
+ sample = self.conv_in_condition(brushnet_cond)
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+ down_block_res_samples += res_samples
+
+ # 4. PaintingNet down blocks
+ brushnet_down_block_res_samples = ()
+ for down_block_res_sample, brushnet_down_block in zip(down_block_res_samples, self.brushnet_down_blocks):
+ down_block_res_sample = brushnet_down_block(down_block_res_sample)
+ brushnet_down_block_res_samples = brushnet_down_block_res_samples + (down_block_res_sample,)
+
+ # 5. mid
+ if self.mid_block is not None:
+ if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = self.mid_block(sample, emb)
+
+ # 6. BrushNet mid blocks
+ brushnet_mid_block_res_sample = self.brushnet_mid_block(sample)
+
+ # 7. up
+ up_block_res_samples = ()
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample, up_res_samples = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ return_res_samples=True,
+ )
+ else:
+ sample, up_res_samples = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ return_res_samples=True,
+ )
+
+ up_block_res_samples += up_res_samples
+
+ # 8. BrushNet up blocks
+ brushnet_up_block_res_samples = ()
+ for up_block_res_sample, brushnet_up_block in zip(up_block_res_samples, self.brushnet_up_blocks):
+ up_block_res_sample = brushnet_up_block(up_block_res_sample)
+ brushnet_up_block_res_samples = brushnet_up_block_res_samples + (up_block_res_sample,)
+
+ # 6. scaling
+ if guess_mode and not self.config.global_pool_conditions:
+ scales = torch.logspace(
+ -1,
+ 0,
+ len(brushnet_down_block_res_samples) + 1 + len(brushnet_up_block_res_samples),
+ device=sample.device,
+ ) # 0.1 to 1.0
+ scales = scales * conditioning_scale
+
+ brushnet_down_block_res_samples = [
+ sample * scale
+ for sample, scale in zip(
+ brushnet_down_block_res_samples, scales[: len(brushnet_down_block_res_samples)]
+ )
+ ]
+ brushnet_mid_block_res_sample = (
+ brushnet_mid_block_res_sample * scales[len(brushnet_down_block_res_samples)]
+ )
+ brushnet_up_block_res_samples = [
+ sample * scale
+ for sample, scale in zip(
+ brushnet_up_block_res_samples, scales[len(brushnet_down_block_res_samples) + 1 :]
+ )
+ ]
+ else:
+ brushnet_down_block_res_samples = [
+ sample * conditioning_scale for sample in brushnet_down_block_res_samples
+ ]
+ brushnet_mid_block_res_sample = brushnet_mid_block_res_sample * conditioning_scale
+ brushnet_up_block_res_samples = [sample * conditioning_scale for sample in brushnet_up_block_res_samples]
+
+ if self.config.global_pool_conditions:
+ brushnet_down_block_res_samples = [
+ torch.mean(sample, dim=(2, 3), keepdim=True) for sample in brushnet_down_block_res_samples
+ ]
+ brushnet_mid_block_res_sample = torch.mean(brushnet_mid_block_res_sample, dim=(2, 3), keepdim=True)
+ brushnet_up_block_res_samples = [
+ torch.mean(sample, dim=(2, 3), keepdim=True) for sample in brushnet_up_block_res_samples
+ ]
+
+ if not return_dict:
+ return (brushnet_down_block_res_samples, brushnet_mid_block_res_sample, brushnet_up_block_res_samples)
+
+ return BrushNetOutput(
+ down_block_res_samples=brushnet_down_block_res_samples,
+ mid_block_res_sample=brushnet_mid_block_res_sample,
+ up_block_res_samples=brushnet_up_block_res_samples,
+ )
+
+
+def zero_module(module):
+ for p in module.parameters():
+ nn.init.zeros_(p)
+ return module
diff --git a/model/__init__.py b/model/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/model/__pycache__/BrushNet_CA.cpython-310.pyc b/model/__pycache__/BrushNet_CA.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..cd407509535380b8aa479a2d95a06295fad78172
Binary files /dev/null and b/model/__pycache__/BrushNet_CA.cpython-310.pyc differ
diff --git a/model/__pycache__/__init__.cpython-310.pyc b/model/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8e134af0343b5fec1f9a3c2549358db097e7bf05
Binary files /dev/null and b/model/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/__init__.py b/model/diffusers_c/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..6c5fd8370d0ed4d30b8e150e3176f11a7ad05c32
--- /dev/null
+++ b/model/diffusers_c/__init__.py
@@ -0,0 +1,789 @@
+__version__ = "0.27.0.dev0"
+
+from typing import TYPE_CHECKING
+
+from .utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ is_flax_available,
+ is_k_diffusion_available,
+ is_librosa_available,
+ is_note_seq_available,
+ is_onnx_available,
+ is_scipy_available,
+ is_torch_available,
+ is_torchsde_available,
+ is_transformers_available,
+)
+
+
+# Lazy Import based on
+# https://github.com/huggingface/transformers/blob/main/src/transformers/__init__.py
+
+# When adding a new object to this init, please add it to `_import_structure`. The `_import_structure` is a dictionary submodule to list of object names,
+# and is used to defer the actual importing for when the objects are requested.
+# This way `import diffusers` provides the names in the namespace without actually importing anything (and especially none of the backends).
+
+_import_structure = {
+ "configuration_utils": ["ConfigMixin"],
+ "models": [],
+ "pipelines": [],
+ "schedulers": [],
+ "utils": [
+ "OptionalDependencyNotAvailable",
+ "is_flax_available",
+ "is_inflect_available",
+ "is_invisible_watermark_available",
+ "is_k_diffusion_available",
+ "is_k_diffusion_version",
+ "is_librosa_available",
+ "is_note_seq_available",
+ "is_onnx_available",
+ "is_scipy_available",
+ "is_torch_available",
+ "is_torchsde_available",
+ "is_transformers_available",
+ "is_transformers_version",
+ "is_unidecode_available",
+ "logging",
+ ],
+}
+
+try:
+ if not is_onnx_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_onnx_objects # noqa F403
+
+ _import_structure["utils.dummy_onnx_objects"] = [
+ name for name in dir(dummy_onnx_objects) if not name.startswith("_")
+ ]
+
+else:
+ _import_structure["pipelines"].extend(["OnnxRuntimeModel"])
+
+try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_pt_objects # noqa F403
+
+ _import_structure["utils.dummy_pt_objects"] = [name for name in dir(dummy_pt_objects) if not name.startswith("_")]
+
+else:
+ _import_structure["models"].extend(
+ [
+ "AsymmetricAutoencoderKL",
+ "AutoencoderKL",
+ "AutoencoderKLTemporalDecoder",
+ "AutoencoderTiny",
+ "ConsistencyDecoderVAE",
+ "BrushNetModel",
+ "ControlNetModel",
+ "I2VGenXLUNet",
+ "Kandinsky3UNet",
+ "ModelMixin",
+ "MotionAdapter",
+ "MultiAdapter",
+ "PriorTransformer",
+ "StableCascadeUNet",
+ "T2IAdapter",
+ "T5FilmDecoder",
+ "Transformer2DModel",
+ "UNet1DModel",
+ "UNet2DConditionModel",
+ "UNet2DModel",
+ "UNet3DConditionModel",
+ "UNetMotionModel",
+ "UNetSpatioTemporalConditionModel",
+ "UVit2DModel",
+ "VQModel",
+ ]
+ )
+
+ _import_structure["optimization"] = [
+ "get_constant_schedule",
+ "get_constant_schedule_with_warmup",
+ "get_cosine_schedule_with_warmup",
+ "get_cosine_with_hard_restarts_schedule_with_warmup",
+ "get_linear_schedule_with_warmup",
+ "get_polynomial_decay_schedule_with_warmup",
+ "get_scheduler",
+ ]
+ _import_structure["pipelines"].extend(
+ [
+ "AudioPipelineOutput",
+ "AutoPipelineForImage2Image",
+ "AutoPipelineForInpainting",
+ "AutoPipelineForText2Image",
+ "ConsistencyModelPipeline",
+ "DanceDiffusionPipeline",
+ "DDIMPipeline",
+ "DDPMPipeline",
+ "DiffusionPipeline",
+ "DiTPipeline",
+ "ImagePipelineOutput",
+ "KarrasVePipeline",
+ "LDMPipeline",
+ "LDMSuperResolutionPipeline",
+ "PNDMPipeline",
+ "RePaintPipeline",
+ "ScoreSdeVePipeline",
+ "StableDiffusionMixin",
+ ]
+ )
+ _import_structure["schedulers"].extend(
+ [
+ "AmusedScheduler",
+ "CMStochasticIterativeScheduler",
+ "DDIMInverseScheduler",
+ "DDIMParallelScheduler",
+ "DDIMScheduler",
+ "DDPMParallelScheduler",
+ "DDPMScheduler",
+ "DDPMWuerstchenScheduler",
+ "DEISMultistepScheduler",
+ "DPMSolverMultistepInverseScheduler",
+ "DPMSolverMultistepScheduler",
+ "DPMSolverSinglestepScheduler",
+ "EDMDPMSolverMultistepScheduler",
+ "EDMEulerScheduler",
+ "EulerAncestralDiscreteScheduler",
+ "EulerDiscreteScheduler",
+ "HeunDiscreteScheduler",
+ "IPNDMScheduler",
+ "KarrasVeScheduler",
+ "KDPM2AncestralDiscreteScheduler",
+ "KDPM2DiscreteScheduler",
+ "LCMScheduler",
+ "PNDMScheduler",
+ "RePaintScheduler",
+ "SASolverScheduler",
+ "SchedulerMixin",
+ "ScoreSdeVeScheduler",
+ "TCDScheduler",
+ "UnCLIPScheduler",
+ "UniPCMultistepScheduler",
+ "VQDiffusionScheduler",
+ ]
+ )
+ _import_structure["training_utils"] = ["EMAModel"]
+
+try:
+ if not (is_torch_available() and is_scipy_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_torch_and_scipy_objects # noqa F403
+
+ _import_structure["utils.dummy_torch_and_scipy_objects"] = [
+ name for name in dir(dummy_torch_and_scipy_objects) if not name.startswith("_")
+ ]
+
+else:
+ _import_structure["schedulers"].extend(["LMSDiscreteScheduler"])
+
+try:
+ if not (is_torch_available() and is_torchsde_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_torch_and_torchsde_objects # noqa F403
+
+ _import_structure["utils.dummy_torch_and_torchsde_objects"] = [
+ name for name in dir(dummy_torch_and_torchsde_objects) if not name.startswith("_")
+ ]
+
+else:
+ _import_structure["schedulers"].extend(["DPMSolverSDEScheduler"])
+
+try:
+ if not (is_torch_available() and is_transformers_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _import_structure["utils.dummy_torch_and_transformers_objects"] = [
+ name for name in dir(dummy_torch_and_transformers_objects) if not name.startswith("_")
+ ]
+
+else:
+ _import_structure["pipelines"].extend(
+ [
+ "AltDiffusionImg2ImgPipeline",
+ "AltDiffusionPipeline",
+ "AmusedImg2ImgPipeline",
+ "AmusedInpaintPipeline",
+ "AmusedPipeline",
+ "AnimateDiffPipeline",
+ "AnimateDiffVideoToVideoPipeline",
+ "AudioLDM2Pipeline",
+ "AudioLDM2ProjectionModel",
+ "AudioLDM2UNet2DConditionModel",
+ "AudioLDMPipeline",
+ "BlipDiffusionControlNetPipeline",
+ "BlipDiffusionPipeline",
+ "CLIPImageProjection",
+ "CycleDiffusionPipeline",
+ "I2VGenXLPipeline",
+ "IFImg2ImgPipeline",
+ "IFImg2ImgSuperResolutionPipeline",
+ "IFInpaintingPipeline",
+ "IFInpaintingSuperResolutionPipeline",
+ "IFPipeline",
+ "IFSuperResolutionPipeline",
+ "ImageTextPipelineOutput",
+ "Kandinsky3Img2ImgPipeline",
+ "Kandinsky3Pipeline",
+ "KandinskyCombinedPipeline",
+ "KandinskyImg2ImgCombinedPipeline",
+ "KandinskyImg2ImgPipeline",
+ "KandinskyInpaintCombinedPipeline",
+ "KandinskyInpaintPipeline",
+ "KandinskyPipeline",
+ "KandinskyPriorPipeline",
+ "KandinskyV22CombinedPipeline",
+ "KandinskyV22ControlnetImg2ImgPipeline",
+ "KandinskyV22ControlnetPipeline",
+ "KandinskyV22Img2ImgCombinedPipeline",
+ "KandinskyV22Img2ImgPipeline",
+ "KandinskyV22InpaintCombinedPipeline",
+ "KandinskyV22InpaintPipeline",
+ "KandinskyV22Pipeline",
+ "KandinskyV22PriorEmb2EmbPipeline",
+ "KandinskyV22PriorPipeline",
+ "LatentConsistencyModelImg2ImgPipeline",
+ "LatentConsistencyModelPipeline",
+ "LDMTextToImagePipeline",
+ "MusicLDMPipeline",
+ "PaintByExamplePipeline",
+ "PIAPipeline",
+ "PixArtAlphaPipeline",
+ "SemanticStableDiffusionPipeline",
+ "ShapEImg2ImgPipeline",
+ "ShapEPipeline",
+ "StableCascadeCombinedPipeline",
+ "StableCascadeDecoderPipeline",
+ "StableCascadePriorPipeline",
+ "StableDiffusionAdapterPipeline",
+ "StableDiffusionAttendAndExcitePipeline",
+ "StableDiffusionBrushNetPipeline",
+ "StableDiffusionBrushNetPowerPaintPipeline",
+ "StableDiffusionControlNetImg2ImgPipeline",
+ "StableDiffusionControlNetInpaintPipeline",
+ "StableDiffusionControlNetPipeline",
+ "StableDiffusionDepth2ImgPipeline",
+ "StableDiffusionDiffEditPipeline",
+ "StableDiffusionGLIGENPipeline",
+ "StableDiffusionGLIGENTextImagePipeline",
+ "StableDiffusionImageVariationPipeline",
+ "StableDiffusionImg2ImgPipeline",
+ "StableDiffusionInpaintPipeline",
+ "StableDiffusionInpaintPipelineLegacy",
+ "StableDiffusionInstructPix2PixPipeline",
+ "StableDiffusionLatentUpscalePipeline",
+ "StableDiffusionLDM3DPipeline",
+ "StableDiffusionModelEditingPipeline",
+ "StableDiffusionPanoramaPipeline",
+ "StableDiffusionParadigmsPipeline",
+ "StableDiffusionPipeline",
+ "StableDiffusionPipelineSafe",
+ "StableDiffusionPix2PixZeroPipeline",
+ "StableDiffusionSAGPipeline",
+ "StableDiffusionUpscalePipeline",
+ "StableDiffusionXLAdapterPipeline",
+ "StableDiffusionXLControlNetImg2ImgPipeline",
+ "StableDiffusionXLControlNetInpaintPipeline",
+ "StableDiffusionXLControlNetPipeline",
+ "StableDiffusionXLImg2ImgPipeline",
+ "StableDiffusionXLInpaintPipeline",
+ "StableDiffusionXLInstructPix2PixPipeline",
+ "StableDiffusionXLPipeline",
+ "StableUnCLIPImg2ImgPipeline",
+ "StableUnCLIPPipeline",
+ "StableVideoDiffusionPipeline",
+ "TextToVideoSDPipeline",
+ "TextToVideoZeroPipeline",
+ "TextToVideoZeroSDXLPipeline",
+ "UnCLIPImageVariationPipeline",
+ "UnCLIPPipeline",
+ "UniDiffuserModel",
+ "UniDiffuserPipeline",
+ "UniDiffuserTextDecoder",
+ "VersatileDiffusionDualGuidedPipeline",
+ "VersatileDiffusionImageVariationPipeline",
+ "VersatileDiffusionPipeline",
+ "VersatileDiffusionTextToImagePipeline",
+ "VideoToVideoSDPipeline",
+ "VQDiffusionPipeline",
+ "WuerstchenCombinedPipeline",
+ "WuerstchenDecoderPipeline",
+ "WuerstchenPriorPipeline",
+ ]
+ )
+
+try:
+ if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_torch_and_transformers_and_k_diffusion_objects # noqa F403
+
+ _import_structure["utils.dummy_torch_and_transformers_and_k_diffusion_objects"] = [
+ name for name in dir(dummy_torch_and_transformers_and_k_diffusion_objects) if not name.startswith("_")
+ ]
+
+else:
+ _import_structure["pipelines"].extend(["StableDiffusionKDiffusionPipeline", "StableDiffusionXLKDiffusionPipeline"])
+
+try:
+ if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_torch_and_transformers_and_onnx_objects # noqa F403
+
+ _import_structure["utils.dummy_torch_and_transformers_and_onnx_objects"] = [
+ name for name in dir(dummy_torch_and_transformers_and_onnx_objects) if not name.startswith("_")
+ ]
+
+else:
+ _import_structure["pipelines"].extend(
+ [
+ "OnnxStableDiffusionImg2ImgPipeline",
+ "OnnxStableDiffusionInpaintPipeline",
+ "OnnxStableDiffusionInpaintPipelineLegacy",
+ "OnnxStableDiffusionPipeline",
+ "OnnxStableDiffusionUpscalePipeline",
+ "StableDiffusionOnnxPipeline",
+ ]
+ )
+
+try:
+ if not (is_torch_available() and is_librosa_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_torch_and_librosa_objects # noqa F403
+
+ _import_structure["utils.dummy_torch_and_librosa_objects"] = [
+ name for name in dir(dummy_torch_and_librosa_objects) if not name.startswith("_")
+ ]
+
+else:
+ _import_structure["pipelines"].extend(["AudioDiffusionPipeline", "Mel"])
+
+try:
+ if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_transformers_and_torch_and_note_seq_objects # noqa F403
+
+ _import_structure["utils.dummy_transformers_and_torch_and_note_seq_objects"] = [
+ name for name in dir(dummy_transformers_and_torch_and_note_seq_objects) if not name.startswith("_")
+ ]
+
+
+else:
+ _import_structure["pipelines"].extend(["SpectrogramDiffusionPipeline"])
+
+try:
+ if not is_flax_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_flax_objects # noqa F403
+
+ _import_structure["utils.dummy_flax_objects"] = [
+ name for name in dir(dummy_flax_objects) if not name.startswith("_")
+ ]
+
+
+else:
+ _import_structure["models.controlnet_flax"] = ["FlaxControlNetModel"]
+ _import_structure["models.modeling_flax_utils"] = ["FlaxModelMixin"]
+ _import_structure["models.unets.unet_2d_condition_flax"] = ["FlaxUNet2DConditionModel"]
+ _import_structure["models.vae_flax"] = ["FlaxAutoencoderKL"]
+ _import_structure["pipelines"].extend(["FlaxDiffusionPipeline"])
+ _import_structure["schedulers"].extend(
+ [
+ "FlaxDDIMScheduler",
+ "FlaxDDPMScheduler",
+ "FlaxDPMSolverMultistepScheduler",
+ "FlaxEulerDiscreteScheduler",
+ "FlaxKarrasVeScheduler",
+ "FlaxLMSDiscreteScheduler",
+ "FlaxPNDMScheduler",
+ "FlaxSchedulerMixin",
+ "FlaxScoreSdeVeScheduler",
+ ]
+ )
+
+
+try:
+ if not (is_flax_available() and is_transformers_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_flax_and_transformers_objects # noqa F403
+
+ _import_structure["utils.dummy_flax_and_transformers_objects"] = [
+ name for name in dir(dummy_flax_and_transformers_objects) if not name.startswith("_")
+ ]
+
+
+else:
+ _import_structure["pipelines"].extend(
+ [
+ "FlaxStableDiffusionControlNetPipeline",
+ "FlaxStableDiffusionImg2ImgPipeline",
+ "FlaxStableDiffusionInpaintPipeline",
+ "FlaxStableDiffusionPipeline",
+ "FlaxStableDiffusionXLPipeline",
+ ]
+ )
+
+try:
+ if not (is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from .utils import dummy_note_seq_objects # noqa F403
+
+ _import_structure["utils.dummy_note_seq_objects"] = [
+ name for name in dir(dummy_note_seq_objects) if not name.startswith("_")
+ ]
+
+
+else:
+ _import_structure["pipelines"].extend(["MidiProcessor"])
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .configuration_utils import ConfigMixin
+
+ try:
+ if not is_onnx_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_onnx_objects import * # noqa F403
+ else:
+ from .pipelines import OnnxRuntimeModel
+
+ try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_pt_objects import * # noqa F403
+ else:
+ from .models import (
+ AsymmetricAutoencoderKL,
+ AutoencoderKL,
+ AutoencoderKLTemporalDecoder,
+ AutoencoderTiny,
+ BrushNetModel,
+ ConsistencyDecoderVAE,
+ ControlNetModel,
+ I2VGenXLUNet,
+ Kandinsky3UNet,
+ ModelMixin,
+ MotionAdapter,
+ MultiAdapter,
+ PriorTransformer,
+ T2IAdapter,
+ T5FilmDecoder,
+ Transformer2DModel,
+ UNet1DModel,
+ UNet2DConditionModel,
+ UNet2DModel,
+ UNet3DConditionModel,
+ UNetMotionModel,
+ UNetSpatioTemporalConditionModel,
+ UVit2DModel,
+ VQModel,
+ )
+ from .optimization import (
+ get_constant_schedule,
+ get_constant_schedule_with_warmup,
+ get_cosine_schedule_with_warmup,
+ get_cosine_with_hard_restarts_schedule_with_warmup,
+ get_linear_schedule_with_warmup,
+ get_polynomial_decay_schedule_with_warmup,
+ get_scheduler,
+ )
+ from .pipelines import (
+ AudioPipelineOutput,
+ AutoPipelineForImage2Image,
+ AutoPipelineForInpainting,
+ AutoPipelineForText2Image,
+ BlipDiffusionControlNetPipeline,
+ BlipDiffusionPipeline,
+ CLIPImageProjection,
+ ConsistencyModelPipeline,
+ DanceDiffusionPipeline,
+ DDIMPipeline,
+ DDPMPipeline,
+ DiffusionPipeline,
+ DiTPipeline,
+ ImagePipelineOutput,
+ KarrasVePipeline,
+ LDMPipeline,
+ LDMSuperResolutionPipeline,
+ PNDMPipeline,
+ RePaintPipeline,
+ ScoreSdeVePipeline,
+ StableDiffusionMixin,
+ )
+ from .schedulers import (
+ AmusedScheduler,
+ CMStochasticIterativeScheduler,
+ DDIMInverseScheduler,
+ DDIMParallelScheduler,
+ DDIMScheduler,
+ DDPMParallelScheduler,
+ DDPMScheduler,
+ DDPMWuerstchenScheduler,
+ DEISMultistepScheduler,
+ DPMSolverMultistepInverseScheduler,
+ DPMSolverMultistepScheduler,
+ DPMSolverSinglestepScheduler,
+ EDMDPMSolverMultistepScheduler,
+ EDMEulerScheduler,
+ EulerAncestralDiscreteScheduler,
+ EulerDiscreteScheduler,
+ HeunDiscreteScheduler,
+ IPNDMScheduler,
+ KarrasVeScheduler,
+ KDPM2AncestralDiscreteScheduler,
+ KDPM2DiscreteScheduler,
+ LCMScheduler,
+ PNDMScheduler,
+ RePaintScheduler,
+ SASolverScheduler,
+ SchedulerMixin,
+ ScoreSdeVeScheduler,
+ TCDScheduler,
+ UnCLIPScheduler,
+ UniPCMultistepScheduler,
+ VQDiffusionScheduler,
+ )
+ from .training_utils import EMAModel
+
+ try:
+ if not (is_torch_available() and is_scipy_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_torch_and_scipy_objects import * # noqa F403
+ else:
+ from .schedulers import LMSDiscreteScheduler
+
+ try:
+ if not (is_torch_available() and is_torchsde_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_torch_and_torchsde_objects import * # noqa F403
+ else:
+ from .schedulers import DPMSolverSDEScheduler
+
+ try:
+ if not (is_torch_available() and is_transformers_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_torch_and_transformers_objects import * # noqa F403
+ else:
+ from .pipelines import (
+ AltDiffusionImg2ImgPipeline,
+ AltDiffusionPipeline,
+ AmusedImg2ImgPipeline,
+ AmusedInpaintPipeline,
+ AmusedPipeline,
+ AnimateDiffPipeline,
+ AnimateDiffVideoToVideoPipeline,
+ AudioLDM2Pipeline,
+ AudioLDM2ProjectionModel,
+ AudioLDM2UNet2DConditionModel,
+ AudioLDMPipeline,
+ CLIPImageProjection,
+ CycleDiffusionPipeline,
+ I2VGenXLPipeline,
+ IFImg2ImgPipeline,
+ IFImg2ImgSuperResolutionPipeline,
+ IFInpaintingPipeline,
+ IFInpaintingSuperResolutionPipeline,
+ IFPipeline,
+ IFSuperResolutionPipeline,
+ ImageTextPipelineOutput,
+ Kandinsky3Img2ImgPipeline,
+ Kandinsky3Pipeline,
+ KandinskyCombinedPipeline,
+ KandinskyImg2ImgCombinedPipeline,
+ KandinskyImg2ImgPipeline,
+ KandinskyInpaintCombinedPipeline,
+ KandinskyInpaintPipeline,
+ KandinskyPipeline,
+ KandinskyPriorPipeline,
+ KandinskyV22CombinedPipeline,
+ KandinskyV22ControlnetImg2ImgPipeline,
+ KandinskyV22ControlnetPipeline,
+ KandinskyV22Img2ImgCombinedPipeline,
+ KandinskyV22Img2ImgPipeline,
+ KandinskyV22InpaintCombinedPipeline,
+ KandinskyV22InpaintPipeline,
+ KandinskyV22Pipeline,
+ KandinskyV22PriorEmb2EmbPipeline,
+ KandinskyV22PriorPipeline,
+ LatentConsistencyModelImg2ImgPipeline,
+ LatentConsistencyModelPipeline,
+ LDMTextToImagePipeline,
+ MusicLDMPipeline,
+ PaintByExamplePipeline,
+ PIAPipeline,
+ PixArtAlphaPipeline,
+ SemanticStableDiffusionPipeline,
+ ShapEImg2ImgPipeline,
+ ShapEPipeline,
+ StableCascadeCombinedPipeline,
+ StableCascadeDecoderPipeline,
+ StableCascadePriorPipeline,
+ StableDiffusionAdapterPipeline,
+ StableDiffusionAttendAndExcitePipeline,
+ StableDiffusionBrushNetPipeline,
+ StableDiffusionBrushNetPowerPaintPipeline,
+ StableDiffusionControlNetImg2ImgPipeline,
+ StableDiffusionControlNetInpaintPipeline,
+ StableDiffusionControlNetPipeline,
+ StableDiffusionDepth2ImgPipeline,
+ StableDiffusionDiffEditPipeline,
+ StableDiffusionGLIGENPipeline,
+ StableDiffusionGLIGENTextImagePipeline,
+ StableDiffusionImageVariationPipeline,
+ StableDiffusionImg2ImgPipeline,
+ StableDiffusionInpaintPipeline,
+ StableDiffusionInpaintPipelineLegacy,
+ StableDiffusionInstructPix2PixPipeline,
+ StableDiffusionLatentUpscalePipeline,
+ StableDiffusionLDM3DPipeline,
+ StableDiffusionModelEditingPipeline,
+ StableDiffusionPanoramaPipeline,
+ StableDiffusionParadigmsPipeline,
+ StableDiffusionPipeline,
+ StableDiffusionPipelineSafe,
+ StableDiffusionPix2PixZeroPipeline,
+ StableDiffusionSAGPipeline,
+ StableDiffusionUpscalePipeline,
+ StableDiffusionXLAdapterPipeline,
+ StableDiffusionXLControlNetImg2ImgPipeline,
+ StableDiffusionXLControlNetInpaintPipeline,
+ StableDiffusionXLControlNetPipeline,
+ StableDiffusionXLImg2ImgPipeline,
+ StableDiffusionXLInpaintPipeline,
+ StableDiffusionXLInstructPix2PixPipeline,
+ StableDiffusionXLPipeline,
+ StableUnCLIPImg2ImgPipeline,
+ StableUnCLIPPipeline,
+ StableVideoDiffusionPipeline,
+ TextToVideoSDPipeline,
+ TextToVideoZeroPipeline,
+ TextToVideoZeroSDXLPipeline,
+ UnCLIPImageVariationPipeline,
+ UnCLIPPipeline,
+ UniDiffuserModel,
+ UniDiffuserPipeline,
+ UniDiffuserTextDecoder,
+ VersatileDiffusionDualGuidedPipeline,
+ VersatileDiffusionImageVariationPipeline,
+ VersatileDiffusionPipeline,
+ VersatileDiffusionTextToImagePipeline,
+ VideoToVideoSDPipeline,
+ VQDiffusionPipeline,
+ WuerstchenCombinedPipeline,
+ WuerstchenDecoderPipeline,
+ WuerstchenPriorPipeline,
+ )
+
+ try:
+ if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_torch_and_transformers_and_k_diffusion_objects import * # noqa F403
+ else:
+ from .pipelines import StableDiffusionKDiffusionPipeline, StableDiffusionXLKDiffusionPipeline
+
+ try:
+ if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_torch_and_transformers_and_onnx_objects import * # noqa F403
+ else:
+ from .pipelines import (
+ OnnxStableDiffusionImg2ImgPipeline,
+ OnnxStableDiffusionInpaintPipeline,
+ OnnxStableDiffusionInpaintPipelineLegacy,
+ OnnxStableDiffusionPipeline,
+ OnnxStableDiffusionUpscalePipeline,
+ StableDiffusionOnnxPipeline,
+ )
+
+ try:
+ if not (is_torch_available() and is_librosa_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_torch_and_librosa_objects import * # noqa F403
+ else:
+ from .pipelines import AudioDiffusionPipeline, Mel
+
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_transformers_and_torch_and_note_seq_objects import * # noqa F403
+ else:
+ from .pipelines import SpectrogramDiffusionPipeline
+
+ try:
+ if not is_flax_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_flax_objects import * # noqa F403
+ else:
+ from .models.controlnet_flax import FlaxControlNetModel
+ from .models.modeling_flax_utils import FlaxModelMixin
+ from .models.unets.unet_2d_condition_flax import FlaxUNet2DConditionModel
+ from .models.vae_flax import FlaxAutoencoderKL
+ from .pipelines import FlaxDiffusionPipeline
+ from .schedulers import (
+ FlaxDDIMScheduler,
+ FlaxDDPMScheduler,
+ FlaxDPMSolverMultistepScheduler,
+ FlaxEulerDiscreteScheduler,
+ FlaxKarrasVeScheduler,
+ FlaxLMSDiscreteScheduler,
+ FlaxPNDMScheduler,
+ FlaxSchedulerMixin,
+ FlaxScoreSdeVeScheduler,
+ )
+
+ try:
+ if not (is_flax_available() and is_transformers_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_flax_and_transformers_objects import * # noqa F403
+ else:
+ from .pipelines import (
+ FlaxStableDiffusionControlNetPipeline,
+ FlaxStableDiffusionImg2ImgPipeline,
+ FlaxStableDiffusionInpaintPipeline,
+ FlaxStableDiffusionPipeline,
+ FlaxStableDiffusionXLPipeline,
+ )
+
+ try:
+ if not (is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from .utils.dummy_note_seq_objects import * # noqa F403
+ else:
+ from .pipelines import MidiProcessor
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ extra_objects={"__version__": __version__},
+ )
diff --git a/model/diffusers_c/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..bc10ae10d6e10b41c4c7f70d4bd05d9637565c9b
Binary files /dev/null and b/model/diffusers_c/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/__pycache__/configuration_utils.cpython-310.pyc b/model/diffusers_c/__pycache__/configuration_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..75100cd830d0fc139ddd1fde17040f314beb6459
Binary files /dev/null and b/model/diffusers_c/__pycache__/configuration_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/__pycache__/dependency_versions_check.cpython-310.pyc b/model/diffusers_c/__pycache__/dependency_versions_check.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ca09ee7e3c3dc0cf1330a3de00313ba0e841c97a
Binary files /dev/null and b/model/diffusers_c/__pycache__/dependency_versions_check.cpython-310.pyc differ
diff --git a/model/diffusers_c/__pycache__/dependency_versions_table.cpython-310.pyc b/model/diffusers_c/__pycache__/dependency_versions_table.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4dc678bcc54ae26d39cb4923c7f93ce18d13b1a4
Binary files /dev/null and b/model/diffusers_c/__pycache__/dependency_versions_table.cpython-310.pyc differ
diff --git a/model/diffusers_c/__pycache__/image_processor.cpython-310.pyc b/model/diffusers_c/__pycache__/image_processor.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e41159780f15b499c4bdf502fe8f66a70318720a
Binary files /dev/null and b/model/diffusers_c/__pycache__/image_processor.cpython-310.pyc differ
diff --git a/model/diffusers_c/commands/__init__.py b/model/diffusers_c/commands/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8208283f6e40a8e46175b1672d6bf44f9d83a02b
--- /dev/null
+++ b/model/diffusers_c/commands/__init__.py
@@ -0,0 +1,27 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from abc import ABC, abstractmethod
+from argparse import ArgumentParser
+
+
+class BaseDiffusersCLICommand(ABC):
+ @staticmethod
+ @abstractmethod
+ def register_subcommand(parser: ArgumentParser):
+ raise NotImplementedError()
+
+ @abstractmethod
+ def run(self):
+ raise NotImplementedError()
diff --git a/model/diffusers_c/commands/diffusers_cli.py b/model/diffusers_c/commands/diffusers_cli.py
new file mode 100644
index 0000000000000000000000000000000000000000..f582c3bcd0df0c5167a8de18123b4474e64bb344
--- /dev/null
+++ b/model/diffusers_c/commands/diffusers_cli.py
@@ -0,0 +1,43 @@
+#!/usr/bin/env python
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from argparse import ArgumentParser
+
+from .env import EnvironmentCommand
+from .fp16_safetensors import FP16SafetensorsCommand
+
+
+def main():
+ parser = ArgumentParser("Diffusers CLI tool", usage="diffusers-cli []")
+ commands_parser = parser.add_subparsers(help="diffusers-cli command helpers")
+
+ # Register commands
+ EnvironmentCommand.register_subcommand(commands_parser)
+ FP16SafetensorsCommand.register_subcommand(commands_parser)
+
+ # Let's go
+ args = parser.parse_args()
+
+ if not hasattr(args, "func"):
+ parser.print_help()
+ exit(1)
+
+ # Run
+ service = args.func(args)
+ service.run()
+
+
+if __name__ == "__main__":
+ main()
diff --git a/model/diffusers_c/commands/env.py b/model/diffusers_c/commands/env.py
new file mode 100644
index 0000000000000000000000000000000000000000..baa69b361f5dc077bb08913f6a00886d0d67f91b
--- /dev/null
+++ b/model/diffusers_c/commands/env.py
@@ -0,0 +1,84 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import platform
+from argparse import ArgumentParser
+
+import huggingface_hub
+
+from .. import __version__ as version
+from ..utils import is_accelerate_available, is_torch_available, is_transformers_available, is_xformers_available
+from . import BaseDiffusersCLICommand
+
+
+def info_command_factory(_):
+ return EnvironmentCommand()
+
+
+class EnvironmentCommand(BaseDiffusersCLICommand):
+ @staticmethod
+ def register_subcommand(parser: ArgumentParser):
+ download_parser = parser.add_parser("env")
+ download_parser.set_defaults(func=info_command_factory)
+
+ def run(self):
+ hub_version = huggingface_hub.__version__
+
+ pt_version = "not installed"
+ pt_cuda_available = "NA"
+ if is_torch_available():
+ import torch
+
+ pt_version = torch.__version__
+ pt_cuda_available = torch.cuda.is_available()
+
+ transformers_version = "not installed"
+ if is_transformers_available():
+ import transformers
+
+ transformers_version = transformers.__version__
+
+ accelerate_version = "not installed"
+ if is_accelerate_available():
+ import accelerate
+
+ accelerate_version = accelerate.__version__
+
+ xformers_version = "not installed"
+ if is_xformers_available():
+ import xformers
+
+ xformers_version = xformers.__version__
+
+ info = {
+ "`diffusers` version": version,
+ "Platform": platform.platform(),
+ "Python version": platform.python_version(),
+ "PyTorch version (GPU?)": f"{pt_version} ({pt_cuda_available})",
+ "Huggingface_hub version": hub_version,
+ "Transformers version": transformers_version,
+ "Accelerate version": accelerate_version,
+ "xFormers version": xformers_version,
+ "Using GPU in script?": "",
+ "Using distributed or parallel set-up in script?": "",
+ }
+
+ print("\nCopy-and-paste the text below in your GitHub issue and FILL OUT the two last points.\n")
+ print(self.format_dict(info))
+
+ return info
+
+ @staticmethod
+ def format_dict(d):
+ return "\n".join([f"- {prop}: {val}" for prop, val in d.items()]) + "\n"
diff --git a/model/diffusers_c/commands/fp16_safetensors.py b/model/diffusers_c/commands/fp16_safetensors.py
new file mode 100644
index 0000000000000000000000000000000000000000..b26b8816bc4cf1a6272d3eebf5ab2be8a5dd865b
--- /dev/null
+++ b/model/diffusers_c/commands/fp16_safetensors.py
@@ -0,0 +1,132 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Usage example:
+ diffusers-cli fp16_safetensors --ckpt_id=openai/shap-e --fp16 --use_safetensors
+"""
+
+import glob
+import json
+import warnings
+from argparse import ArgumentParser, Namespace
+from importlib import import_module
+
+import huggingface_hub
+import torch
+from huggingface_hub import hf_hub_download
+from packaging import version
+
+from ..utils import logging
+from . import BaseDiffusersCLICommand
+
+
+def conversion_command_factory(args: Namespace):
+ if args.use_auth_token:
+ warnings.warn(
+ "The `--use_auth_token` flag is deprecated and will be removed in a future version. Authentication is now"
+ " handled automatically if user is logged in."
+ )
+ return FP16SafetensorsCommand(args.ckpt_id, args.fp16, args.use_safetensors)
+
+
+class FP16SafetensorsCommand(BaseDiffusersCLICommand):
+ @staticmethod
+ def register_subcommand(parser: ArgumentParser):
+ conversion_parser = parser.add_parser("fp16_safetensors")
+ conversion_parser.add_argument(
+ "--ckpt_id",
+ type=str,
+ help="Repo id of the checkpoints on which to run the conversion. Example: 'openai/shap-e'.",
+ )
+ conversion_parser.add_argument(
+ "--fp16", action="store_true", help="If serializing the variables in FP16 precision."
+ )
+ conversion_parser.add_argument(
+ "--use_safetensors", action="store_true", help="If serializing in the safetensors format."
+ )
+ conversion_parser.add_argument(
+ "--use_auth_token",
+ action="store_true",
+ help="When working with checkpoints having private visibility. When used `huggingface-cli login` needs to be run beforehand.",
+ )
+ conversion_parser.set_defaults(func=conversion_command_factory)
+
+ def __init__(self, ckpt_id: str, fp16: bool, use_safetensors: bool):
+ self.logger = logging.get_logger("diffusers-cli/fp16_safetensors")
+ self.ckpt_id = ckpt_id
+ self.local_ckpt_dir = f"/tmp/{ckpt_id}"
+ self.fp16 = fp16
+
+ self.use_safetensors = use_safetensors
+
+ if not self.use_safetensors and not self.fp16:
+ raise NotImplementedError(
+ "When `use_safetensors` and `fp16` both are False, then this command is of no use."
+ )
+
+ def run(self):
+ if version.parse(huggingface_hub.__version__) < version.parse("0.9.0"):
+ raise ImportError(
+ "The huggingface_hub version must be >= 0.9.0 to use this command. Please update your huggingface_hub"
+ " installation."
+ )
+ else:
+ from huggingface_hub import create_commit
+ from huggingface_hub._commit_api import CommitOperationAdd
+
+ model_index = hf_hub_download(repo_id=self.ckpt_id, filename="model_index.json")
+ with open(model_index, "r") as f:
+ pipeline_class_name = json.load(f)["_class_name"]
+ pipeline_class = getattr(import_module("diffusers"), pipeline_class_name)
+ self.logger.info(f"Pipeline class imported: {pipeline_class_name}.")
+
+ # Load the appropriate pipeline. We could have use `DiffusionPipeline`
+ # here, but just to avoid any rough edge cases.
+ pipeline = pipeline_class.from_pretrained(
+ self.ckpt_id, torch_dtype=torch.float16 if self.fp16 else torch.float32
+ )
+ pipeline.save_pretrained(
+ self.local_ckpt_dir,
+ safe_serialization=True if self.use_safetensors else False,
+ variant="fp16" if self.fp16 else None,
+ )
+ self.logger.info(f"Pipeline locally saved to {self.local_ckpt_dir}.")
+
+ # Fetch all the paths.
+ if self.fp16:
+ modified_paths = glob.glob(f"{self.local_ckpt_dir}/*/*.fp16.*")
+ elif self.use_safetensors:
+ modified_paths = glob.glob(f"{self.local_ckpt_dir}/*/*.safetensors")
+
+ # Prepare for the PR.
+ commit_message = f"Serialize variables with FP16: {self.fp16} and safetensors: {self.use_safetensors}."
+ operations = []
+ for path in modified_paths:
+ operations.append(CommitOperationAdd(path_in_repo="/".join(path.split("/")[4:]), path_or_fileobj=path))
+
+ # Open the PR.
+ commit_description = (
+ "Variables converted by the [`diffusers`' `fp16_safetensors`"
+ " CLI](https://github.com/huggingface/diffusers/blob/main/src/diffusers/commands/fp16_safetensors.py)."
+ )
+ hub_pr_url = create_commit(
+ repo_id=self.ckpt_id,
+ operations=operations,
+ commit_message=commit_message,
+ commit_description=commit_description,
+ repo_type="model",
+ create_pr=True,
+ ).pr_url
+ self.logger.info(f"PR created here: {hub_pr_url}.")
diff --git a/model/diffusers_c/configuration_utils.py b/model/diffusers_c/configuration_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..104a0bb0daf8e567914b492686f99c8ee015c311
--- /dev/null
+++ b/model/diffusers_c/configuration_utils.py
@@ -0,0 +1,703 @@
+# Copyright 2024 The HuggingFace Inc. team.
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""ConfigMixin base class and utilities."""
+
+import dataclasses
+import functools
+import importlib
+import inspect
+import json
+import os
+import re
+from collections import OrderedDict
+from pathlib import PosixPath
+from typing import Any, Dict, Tuple, Union
+
+import numpy as np
+from huggingface_hub import create_repo, hf_hub_download
+from huggingface_hub.utils import (
+ EntryNotFoundError,
+ RepositoryNotFoundError,
+ RevisionNotFoundError,
+ validate_hf_hub_args,
+)
+from requests import HTTPError
+
+from . import __version__
+from .utils import (
+ HUGGINGFACE_CO_RESOLVE_ENDPOINT,
+ DummyObject,
+ deprecate,
+ extract_commit_hash,
+ http_user_agent,
+ logging,
+)
+
+
+logger = logging.get_logger(__name__)
+
+_re_configuration_file = re.compile(r"config\.(.*)\.json")
+
+
+class FrozenDict(OrderedDict):
+ def __init__(self, *args, **kwargs):
+ super().__init__(*args, **kwargs)
+
+ for key, value in self.items():
+ setattr(self, key, value)
+
+ self.__frozen = True
+
+ def __delitem__(self, *args, **kwargs):
+ raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.")
+
+ def setdefault(self, *args, **kwargs):
+ raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.")
+
+ def pop(self, *args, **kwargs):
+ raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.")
+
+ def update(self, *args, **kwargs):
+ raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.")
+
+ def __setattr__(self, name, value):
+ if hasattr(self, "__frozen") and self.__frozen:
+ raise Exception(f"You cannot use ``__setattr__`` on a {self.__class__.__name__} instance.")
+ super().__setattr__(name, value)
+
+ def __setitem__(self, name, value):
+ if hasattr(self, "__frozen") and self.__frozen:
+ raise Exception(f"You cannot use ``__setattr__`` on a {self.__class__.__name__} instance.")
+ super().__setitem__(name, value)
+
+
+class ConfigMixin:
+ r"""
+ Base class for all configuration classes. All configuration parameters are stored under `self.config`. Also
+ provides the [`~ConfigMixin.from_config`] and [`~ConfigMixin.save_config`] methods for loading, downloading, and
+ saving classes that inherit from [`ConfigMixin`].
+
+ Class attributes:
+ - **config_name** (`str`) -- A filename under which the config should stored when calling
+ [`~ConfigMixin.save_config`] (should be overridden by parent class).
+ - **ignore_for_config** (`List[str]`) -- A list of attributes that should not be saved in the config (should be
+ overridden by subclass).
+ - **has_compatibles** (`bool`) -- Whether the class has compatible classes (should be overridden by subclass).
+ - **_deprecated_kwargs** (`List[str]`) -- Keyword arguments that are deprecated. Note that the `init` function
+ should only have a `kwargs` argument if at least one argument is deprecated (should be overridden by
+ subclass).
+ """
+
+ config_name = None
+ ignore_for_config = []
+ has_compatibles = False
+
+ _deprecated_kwargs = []
+
+ def register_to_config(self, **kwargs):
+ if self.config_name is None:
+ raise NotImplementedError(f"Make sure that {self.__class__} has defined a class name `config_name`")
+ # Special case for `kwargs` used in deprecation warning added to schedulers
+ # TODO: remove this when we remove the deprecation warning, and the `kwargs` argument,
+ # or solve in a more general way.
+ kwargs.pop("kwargs", None)
+
+ if not hasattr(self, "_internal_dict"):
+ internal_dict = kwargs
+ else:
+ previous_dict = dict(self._internal_dict)
+ internal_dict = {**self._internal_dict, **kwargs}
+ logger.debug(f"Updating config from {previous_dict} to {internal_dict}")
+
+ self._internal_dict = FrozenDict(internal_dict)
+
+ def __getattr__(self, name: str) -> Any:
+ """The only reason we overwrite `getattr` here is to gracefully deprecate accessing
+ config attributes directly. See https://github.com/huggingface/diffusers/pull/3129
+
+ This function is mostly copied from PyTorch's __getattr__ overwrite:
+ https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
+ """
+
+ is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)
+ is_attribute = name in self.__dict__
+
+ if is_in_config and not is_attribute:
+ deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'scheduler.config.{name}'."
+ deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False)
+ return self._internal_dict[name]
+
+ raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
+
+ def save_config(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs):
+ """
+ Save a configuration object to the directory specified in `save_directory` so that it can be reloaded using the
+ [`~ConfigMixin.from_config`] class method.
+
+ Args:
+ save_directory (`str` or `os.PathLike`):
+ Directory where the configuration JSON file is saved (will be created if it does not exist).
+ push_to_hub (`bool`, *optional*, defaults to `False`):
+ Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the
+ repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+ namespace).
+ kwargs (`Dict[str, Any]`, *optional*):
+ Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+ """
+ if os.path.isfile(save_directory):
+ raise AssertionError(f"Provided path ({save_directory}) should be a directory, not a file")
+
+ os.makedirs(save_directory, exist_ok=True)
+
+ # If we save using the predefined names, we can load using `from_config`
+ output_config_file = os.path.join(save_directory, self.config_name)
+
+ self.to_json_file(output_config_file)
+ logger.info(f"Configuration saved in {output_config_file}")
+
+ if push_to_hub:
+ commit_message = kwargs.pop("commit_message", None)
+ private = kwargs.pop("private", False)
+ create_pr = kwargs.pop("create_pr", False)
+ token = kwargs.pop("token", None)
+ repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
+ repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id
+
+ self._upload_folder(
+ save_directory,
+ repo_id,
+ token=token,
+ commit_message=commit_message,
+ create_pr=create_pr,
+ )
+
+ @classmethod
+ def from_config(cls, config: Union[FrozenDict, Dict[str, Any]] = None, return_unused_kwargs=False, **kwargs):
+ r"""
+ Instantiate a Python class from a config dictionary.
+
+ Parameters:
+ config (`Dict[str, Any]`):
+ A config dictionary from which the Python class is instantiated. Make sure to only load configuration
+ files of compatible classes.
+ return_unused_kwargs (`bool`, *optional*, defaults to `False`):
+ Whether kwargs that are not consumed by the Python class should be returned or not.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to update the configuration object (after it is loaded) and initiate the Python class.
+ `**kwargs` are passed directly to the underlying scheduler/model's `__init__` method and eventually
+ overwrite the same named arguments in `config`.
+
+ Returns:
+ [`ModelMixin`] or [`SchedulerMixin`]:
+ A model or scheduler object instantiated from a config dictionary.
+
+ Examples:
+
+ ```python
+ >>> from diffusers import DDPMScheduler, DDIMScheduler, PNDMScheduler
+
+ >>> # Download scheduler from huggingface.co and cache.
+ >>> scheduler = DDPMScheduler.from_pretrained("google/ddpm-cifar10-32")
+
+ >>> # Instantiate DDIM scheduler class with same config as DDPM
+ >>> scheduler = DDIMScheduler.from_config(scheduler.config)
+
+ >>> # Instantiate PNDM scheduler class with same config as DDPM
+ >>> scheduler = PNDMScheduler.from_config(scheduler.config)
+ ```
+ """
+ # <===== TO BE REMOVED WITH DEPRECATION
+ # TODO(Patrick) - make sure to remove the following lines when config=="model_path" is deprecated
+ if "pretrained_model_name_or_path" in kwargs:
+ config = kwargs.pop("pretrained_model_name_or_path")
+
+ if config is None:
+ raise ValueError("Please make sure to provide a config as the first positional argument.")
+ # ======>
+
+ if not isinstance(config, dict):
+ deprecation_message = "It is deprecated to pass a pretrained model name or path to `from_config`."
+ if "Scheduler" in cls.__name__:
+ deprecation_message += (
+ f"If you were trying to load a scheduler, please use {cls}.from_pretrained(...) instead."
+ " Otherwise, please make sure to pass a configuration dictionary instead. This functionality will"
+ " be removed in v1.0.0."
+ )
+ elif "Model" in cls.__name__:
+ deprecation_message += (
+ f"If you were trying to load a model, please use {cls}.load_config(...) followed by"
+ f" {cls}.from_config(...) instead. Otherwise, please make sure to pass a configuration dictionary"
+ " instead. This functionality will be removed in v1.0.0."
+ )
+ deprecate("config-passed-as-path", "1.0.0", deprecation_message, standard_warn=False)
+ config, kwargs = cls.load_config(pretrained_model_name_or_path=config, return_unused_kwargs=True, **kwargs)
+
+ init_dict, unused_kwargs, hidden_dict = cls.extract_init_dict(config, **kwargs)
+
+ # Allow dtype to be specified on initialization
+ if "dtype" in unused_kwargs:
+ init_dict["dtype"] = unused_kwargs.pop("dtype")
+
+ # add possible deprecated kwargs
+ for deprecated_kwarg in cls._deprecated_kwargs:
+ if deprecated_kwarg in unused_kwargs:
+ init_dict[deprecated_kwarg] = unused_kwargs.pop(deprecated_kwarg)
+
+ # Return model and optionally state and/or unused_kwargs
+ model = cls(**init_dict)
+
+ # make sure to also save config parameters that might be used for compatible classes
+ # update _class_name
+ if "_class_name" in hidden_dict:
+ hidden_dict["_class_name"] = cls.__name__
+
+ model.register_to_config(**hidden_dict)
+
+ # add hidden kwargs of compatible classes to unused_kwargs
+ unused_kwargs = {**unused_kwargs, **hidden_dict}
+
+ if return_unused_kwargs:
+ return (model, unused_kwargs)
+ else:
+ return model
+
+ @classmethod
+ def get_config_dict(cls, *args, **kwargs):
+ deprecation_message = (
+ f" The function get_config_dict is deprecated. Please use {cls}.load_config instead. This function will be"
+ " removed in version v1.0.0"
+ )
+ deprecate("get_config_dict", "1.0.0", deprecation_message, standard_warn=False)
+ return cls.load_config(*args, **kwargs)
+
+ @classmethod
+ @validate_hf_hub_args
+ def load_config(
+ cls,
+ pretrained_model_name_or_path: Union[str, os.PathLike],
+ return_unused_kwargs=False,
+ return_commit_hash=False,
+ **kwargs,
+ ) -> Tuple[Dict[str, Any], Dict[str, Any]]:
+ r"""
+ Load a model or scheduler configuration.
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+ the Hub.
+ - A path to a *directory* (for example `./my_model_directory`) containing model weights saved with
+ [`~ConfigMixin.save_config`].
+
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ subfolder (`str`, *optional*, defaults to `""`):
+ The subfolder location of a model file within a larger model repository on the Hub or locally.
+ return_unused_kwargs (`bool`, *optional*, defaults to `False):
+ Whether unused keyword arguments of the config are returned.
+ return_commit_hash (`bool`, *optional*, defaults to `False):
+ Whether the `commit_hash` of the loaded configuration are returned.
+
+ Returns:
+ `dict`:
+ A dictionary of all the parameters stored in a JSON configuration file.
+
+ """
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ token = kwargs.pop("token", None)
+ local_files_only = kwargs.pop("local_files_only", False)
+ revision = kwargs.pop("revision", None)
+ _ = kwargs.pop("mirror", None)
+ subfolder = kwargs.pop("subfolder", None)
+ user_agent = kwargs.pop("user_agent", {})
+
+ user_agent = {**user_agent, "file_type": "config"}
+ user_agent = http_user_agent(user_agent)
+
+ pretrained_model_name_or_path = str(pretrained_model_name_or_path)
+
+ if cls.config_name is None:
+ raise ValueError(
+ "`self.config_name` is not defined. Note that one should not load a config from "
+ "`ConfigMixin`. Please make sure to define `config_name` in a class inheriting from `ConfigMixin`"
+ )
+
+ if os.path.isfile(pretrained_model_name_or_path):
+ config_file = pretrained_model_name_or_path
+ elif os.path.isdir(pretrained_model_name_or_path):
+ if os.path.isfile(os.path.join(pretrained_model_name_or_path, cls.config_name)):
+ # Load from a PyTorch checkpoint
+ config_file = os.path.join(pretrained_model_name_or_path, cls.config_name)
+ elif subfolder is not None and os.path.isfile(
+ os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)
+ ):
+ config_file = os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)
+ else:
+ raise EnvironmentError(
+ f"Error no file named {cls.config_name} found in directory {pretrained_model_name_or_path}."
+ )
+ else:
+ try:
+ # Load from URL or cache if already cached
+ config_file = hf_hub_download(
+ pretrained_model_name_or_path,
+ filename=cls.config_name,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ proxies=proxies,
+ resume_download=resume_download,
+ local_files_only=local_files_only,
+ token=token,
+ user_agent=user_agent,
+ subfolder=subfolder,
+ revision=revision,
+ )
+ except RepositoryNotFoundError:
+ raise EnvironmentError(
+ f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier"
+ " listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a"
+ " token having permission to this repo with `token` or log in with `huggingface-cli login`."
+ )
+ except RevisionNotFoundError:
+ raise EnvironmentError(
+ f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for"
+ " this model name. Check the model page at"
+ f" 'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions."
+ )
+ except EntryNotFoundError:
+ raise EnvironmentError(
+ f"{pretrained_model_name_or_path} does not appear to have a file named {cls.config_name}."
+ )
+ except HTTPError as err:
+ raise EnvironmentError(
+ "There was a specific connection error when trying to load"
+ f" {pretrained_model_name_or_path}:\n{err}"
+ )
+ except ValueError:
+ raise EnvironmentError(
+ f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it"
+ f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a"
+ f" directory containing a {cls.config_name} file.\nCheckout your internet connection or see how to"
+ " run the library in offline mode at"
+ " 'https://huggingface.co/docs/diffusers/installation#offline-mode'."
+ )
+ except EnvironmentError:
+ raise EnvironmentError(
+ f"Can't load config for '{pretrained_model_name_or_path}'. If you were trying to load it from "
+ "'https://huggingface.co/models', make sure you don't have a local directory with the same name. "
+ f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory "
+ f"containing a {cls.config_name} file"
+ )
+
+ try:
+ # Load config dict
+ config_dict = cls._dict_from_json_file(config_file)
+
+ commit_hash = extract_commit_hash(config_file)
+ except (json.JSONDecodeError, UnicodeDecodeError):
+ raise EnvironmentError(f"It looks like the config file at '{config_file}' is not a valid JSON file.")
+
+ if not (return_unused_kwargs or return_commit_hash):
+ return config_dict
+
+ outputs = (config_dict,)
+
+ if return_unused_kwargs:
+ outputs += (kwargs,)
+
+ if return_commit_hash:
+ outputs += (commit_hash,)
+
+ return outputs
+
+ @staticmethod
+ def _get_init_keys(cls):
+ return set(dict(inspect.signature(cls.__init__).parameters).keys())
+
+ @classmethod
+ def extract_init_dict(cls, config_dict, **kwargs):
+ # Skip keys that were not present in the original config, so default __init__ values were used
+ used_defaults = config_dict.get("_use_default_values", [])
+ config_dict = {k: v for k, v in config_dict.items() if k not in used_defaults and k != "_use_default_values"}
+
+ # 0. Copy origin config dict
+ original_dict = dict(config_dict.items())
+
+ # 1. Retrieve expected config attributes from __init__ signature
+ expected_keys = cls._get_init_keys(cls)
+ expected_keys.remove("self")
+ # remove general kwargs if present in dict
+ if "kwargs" in expected_keys:
+ expected_keys.remove("kwargs")
+ # remove flax internal keys
+ if hasattr(cls, "_flax_internal_args"):
+ for arg in cls._flax_internal_args:
+ expected_keys.remove(arg)
+
+ # 2. Remove attributes that cannot be expected from expected config attributes
+ # remove keys to be ignored
+ if len(cls.ignore_for_config) > 0:
+ expected_keys = expected_keys - set(cls.ignore_for_config)
+
+ # load diffusers library to import compatible and original scheduler
+ diffusers_library = importlib.import_module(__name__.split(".")[0])
+
+ if cls.has_compatibles:
+ compatible_classes = [c for c in cls._get_compatibles() if not isinstance(c, DummyObject)]
+ else:
+ compatible_classes = []
+
+ expected_keys_comp_cls = set()
+ for c in compatible_classes:
+ expected_keys_c = cls._get_init_keys(c)
+ expected_keys_comp_cls = expected_keys_comp_cls.union(expected_keys_c)
+ expected_keys_comp_cls = expected_keys_comp_cls - cls._get_init_keys(cls)
+ config_dict = {k: v for k, v in config_dict.items() if k not in expected_keys_comp_cls}
+
+ # remove attributes from orig class that cannot be expected
+ orig_cls_name = config_dict.pop("_class_name", cls.__name__)
+ if (
+ isinstance(orig_cls_name, str)
+ and orig_cls_name != cls.__name__
+ and hasattr(diffusers_library, orig_cls_name)
+ ):
+ orig_cls = getattr(diffusers_library, orig_cls_name)
+ unexpected_keys_from_orig = cls._get_init_keys(orig_cls) - expected_keys
+ config_dict = {k: v for k, v in config_dict.items() if k not in unexpected_keys_from_orig}
+ elif not isinstance(orig_cls_name, str) and not isinstance(orig_cls_name, (list, tuple)):
+ raise ValueError(
+ "Make sure that the `_class_name` is of type string or list of string (for custom pipelines)."
+ )
+
+ # remove private attributes
+ config_dict = {k: v for k, v in config_dict.items() if not k.startswith("_")}
+
+ # 3. Create keyword arguments that will be passed to __init__ from expected keyword arguments
+ init_dict = {}
+ for key in expected_keys:
+ # if config param is passed to kwarg and is present in config dict
+ # it should overwrite existing config dict key
+ if key in kwargs and key in config_dict:
+ config_dict[key] = kwargs.pop(key)
+
+ if key in kwargs:
+ # overwrite key
+ init_dict[key] = kwargs.pop(key)
+ elif key in config_dict:
+ # use value from config dict
+ init_dict[key] = config_dict.pop(key)
+
+ # 4. Give nice warning if unexpected values have been passed
+ if len(config_dict) > 0:
+ logger.warning(
+ f"The config attributes {config_dict} were passed to {cls.__name__}, "
+ "but are not expected and will be ignored. Please verify your "
+ f"{cls.config_name} configuration file."
+ )
+
+ # 5. Give nice info if config attributes are initialized to default because they have not been passed
+ passed_keys = set(init_dict.keys())
+ if len(expected_keys - passed_keys) > 0:
+ logger.info(
+ f"{expected_keys - passed_keys} was not found in config. Values will be initialized to default values."
+ )
+
+ # 6. Define unused keyword arguments
+ unused_kwargs = {**config_dict, **kwargs}
+
+ # 7. Define "hidden" config parameters that were saved for compatible classes
+ hidden_config_dict = {k: v for k, v in original_dict.items() if k not in init_dict}
+
+ return init_dict, unused_kwargs, hidden_config_dict
+
+ @classmethod
+ def _dict_from_json_file(cls, json_file: Union[str, os.PathLike]):
+ with open(json_file, "r", encoding="utf-8") as reader:
+ text = reader.read()
+ return json.loads(text)
+
+ def __repr__(self):
+ return f"{self.__class__.__name__} {self.to_json_string()}"
+
+ @property
+ def config(self) -> Dict[str, Any]:
+ """
+ Returns the config of the class as a frozen dictionary
+
+ Returns:
+ `Dict[str, Any]`: Config of the class.
+ """
+ return self._internal_dict
+
+ def to_json_string(self) -> str:
+ """
+ Serializes the configuration instance to a JSON string.
+
+ Returns:
+ `str`:
+ String containing all the attributes that make up the configuration instance in JSON format.
+ """
+ config_dict = self._internal_dict if hasattr(self, "_internal_dict") else {}
+ config_dict["_class_name"] = self.__class__.__name__
+ config_dict["_diffusers_version"] = __version__
+
+ def to_json_saveable(value):
+ if isinstance(value, np.ndarray):
+ value = value.tolist()
+ elif isinstance(value, PosixPath):
+ value = str(value)
+ return value
+
+ config_dict = {k: to_json_saveable(v) for k, v in config_dict.items()}
+ # Don't save "_ignore_files" or "_use_default_values"
+ config_dict.pop("_ignore_files", None)
+ config_dict.pop("_use_default_values", None)
+
+ return json.dumps(config_dict, indent=2, sort_keys=True) + "\n"
+
+ def to_json_file(self, json_file_path: Union[str, os.PathLike]):
+ """
+ Save the configuration instance's parameters to a JSON file.
+
+ Args:
+ json_file_path (`str` or `os.PathLike`):
+ Path to the JSON file to save a configuration instance's parameters.
+ """
+ with open(json_file_path, "w", encoding="utf-8") as writer:
+ writer.write(self.to_json_string())
+
+
+def register_to_config(init):
+ r"""
+ Decorator to apply on the init of classes inheriting from [`ConfigMixin`] so that all the arguments are
+ automatically sent to `self.register_for_config`. To ignore a specific argument accepted by the init but that
+ shouldn't be registered in the config, use the `ignore_for_config` class variable
+
+ Warning: Once decorated, all private arguments (beginning with an underscore) are trashed and not sent to the init!
+ """
+
+ @functools.wraps(init)
+ def inner_init(self, *args, **kwargs):
+ # Ignore private kwargs in the init.
+ init_kwargs = {k: v for k, v in kwargs.items() if not k.startswith("_")}
+ config_init_kwargs = {k: v for k, v in kwargs.items() if k.startswith("_")}
+ if not isinstance(self, ConfigMixin):
+ raise RuntimeError(
+ f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does "
+ "not inherit from `ConfigMixin`."
+ )
+
+ ignore = getattr(self, "ignore_for_config", [])
+ # Get positional arguments aligned with kwargs
+ new_kwargs = {}
+ signature = inspect.signature(init)
+ parameters = {
+ name: p.default for i, (name, p) in enumerate(signature.parameters.items()) if i > 0 and name not in ignore
+ }
+ for arg, name in zip(args, parameters.keys()):
+ new_kwargs[name] = arg
+
+ # Then add all kwargs
+ new_kwargs.update(
+ {
+ k: init_kwargs.get(k, default)
+ for k, default in parameters.items()
+ if k not in ignore and k not in new_kwargs
+ }
+ )
+
+ # Take note of the parameters that were not present in the loaded config
+ if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0:
+ new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs))
+
+ new_kwargs = {**config_init_kwargs, **new_kwargs}
+ getattr(self, "register_to_config")(**new_kwargs)
+ init(self, *args, **init_kwargs)
+
+ return inner_init
+
+
+def flax_register_to_config(cls):
+ original_init = cls.__init__
+
+ @functools.wraps(original_init)
+ def init(self, *args, **kwargs):
+ if not isinstance(self, ConfigMixin):
+ raise RuntimeError(
+ f"`@register_for_config` was applied to {self.__class__.__name__} init method, but this class does "
+ "not inherit from `ConfigMixin`."
+ )
+
+ # Ignore private kwargs in the init. Retrieve all passed attributes
+ init_kwargs = dict(kwargs.items())
+
+ # Retrieve default values
+ fields = dataclasses.fields(self)
+ default_kwargs = {}
+ for field in fields:
+ # ignore flax specific attributes
+ if field.name in self._flax_internal_args:
+ continue
+ if type(field.default) == dataclasses._MISSING_TYPE:
+ default_kwargs[field.name] = None
+ else:
+ default_kwargs[field.name] = getattr(self, field.name)
+
+ # Make sure init_kwargs override default kwargs
+ new_kwargs = {**default_kwargs, **init_kwargs}
+ # dtype should be part of `init_kwargs`, but not `new_kwargs`
+ if "dtype" in new_kwargs:
+ new_kwargs.pop("dtype")
+
+ # Get positional arguments aligned with kwargs
+ for i, arg in enumerate(args):
+ name = fields[i].name
+ new_kwargs[name] = arg
+
+ # Take note of the parameters that were not present in the loaded config
+ if len(set(new_kwargs.keys()) - set(init_kwargs)) > 0:
+ new_kwargs["_use_default_values"] = list(set(new_kwargs.keys()) - set(init_kwargs))
+
+ getattr(self, "register_to_config")(**new_kwargs)
+ original_init(self, *args, **kwargs)
+
+ cls.__init__ = init
+ return cls
diff --git a/model/diffusers_c/dependency_versions_check.py b/model/diffusers_c/dependency_versions_check.py
new file mode 100644
index 0000000000000000000000000000000000000000..0728b3a7c0932cd06c920947e1ea57f3864f239a
--- /dev/null
+++ b/model/diffusers_c/dependency_versions_check.py
@@ -0,0 +1,34 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .dependency_versions_table import deps
+from .utils.versions import require_version, require_version_core
+
+
+# define which module versions we always want to check at run time
+# (usually the ones defined in `install_requires` in setup.py)
+#
+# order specific notes:
+# - tqdm must be checked before tokenizers
+
+pkgs_to_check_at_runtime = "python requests filelock numpy".split()
+for pkg in pkgs_to_check_at_runtime:
+ if pkg in deps:
+ require_version_core(deps[pkg])
+ else:
+ raise ValueError(f"can't find {pkg} in {deps.keys()}, check dependency_versions_table.py")
+
+
+def dep_version_check(pkg, hint=None):
+ require_version(deps[pkg], hint)
diff --git a/model/diffusers_c/dependency_versions_table.py b/model/diffusers_c/dependency_versions_table.py
new file mode 100644
index 0000000000000000000000000000000000000000..bf7a7d3cc5090661b5d2c6826c2b8e0f722584cc
--- /dev/null
+++ b/model/diffusers_c/dependency_versions_table.py
@@ -0,0 +1,45 @@
+# THIS FILE HAS BEEN AUTOGENERATED. To update:
+# 1. modify the `_deps` dict in setup.py
+# 2. run `make deps_table_update`
+deps = {
+ "Pillow": "Pillow",
+ "accelerate": "accelerate>=0.11.0",
+ "compel": "compel==0.1.8",
+ "datasets": "datasets",
+ "filelock": "filelock",
+ "flax": "flax>=0.4.1",
+ "hf-doc-builder": "hf-doc-builder>=0.3.0",
+ "huggingface-hub": "huggingface-hub",
+ "requests-mock": "requests-mock==1.10.0",
+ "importlib_metadata": "importlib_metadata",
+ "invisible-watermark": "invisible-watermark>=0.2.0",
+ "isort": "isort>=5.5.4",
+ "jax": "jax>=0.4.1",
+ "jaxlib": "jaxlib>=0.4.1",
+ "Jinja2": "Jinja2",
+ "k-diffusion": "k-diffusion>=0.0.12",
+ "torchsde": "torchsde",
+ "note_seq": "note_seq",
+ "librosa": "librosa",
+ "numpy": "numpy",
+ "parameterized": "parameterized",
+ "peft": "peft>=0.6.0",
+ "protobuf": "protobuf>=3.20.3,<4",
+ "pytest": "pytest",
+ "pytest-timeout": "pytest-timeout",
+ "pytest-xdist": "pytest-xdist",
+ "python": "python>=3.8.0",
+ "ruff": "ruff==0.1.5",
+ "safetensors": "safetensors>=0.3.1",
+ "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",
+ "GitPython": "GitPython<3.1.19",
+ "scipy": "scipy",
+ "onnx": "onnx",
+ "regex": "regex!=2019.12.17",
+ "requests": "requests",
+ "tensorboard": "tensorboard",
+ "torch": "torch>=1.4",
+ "torchvision": "torchvision",
+ "transformers": "transformers>=4.25.1",
+ "urllib3": "urllib3<=2.0.0",
+}
diff --git a/model/diffusers_c/experimental/README.md b/model/diffusers_c/experimental/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..c36b510e8efe792a3dfa561fc49f37b35da8327e
--- /dev/null
+++ b/model/diffusers_c/experimental/README.md
@@ -0,0 +1,5 @@
+# 🧨 Diffusers Experimental
+
+We are adding experimental code to support novel applications and usages of the Diffusers library.
+Currently, the following experiments are supported:
+* Reinforcement learning via an implementation of the [Diffuser](https://arxiv.org/abs/2205.09991) model.
diff --git a/model/diffusers_c/experimental/__init__.py b/model/diffusers_c/experimental/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..ebc8155403016dfd8ad7fb78d246f9da9098ac50
--- /dev/null
+++ b/model/diffusers_c/experimental/__init__.py
@@ -0,0 +1 @@
+from .rl import ValueGuidedRLPipeline
diff --git a/model/diffusers_c/experimental/rl/__init__.py b/model/diffusers_c/experimental/rl/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..7b338d3173e12d478b6b6d6fd0e50650a0ab5a4c
--- /dev/null
+++ b/model/diffusers_c/experimental/rl/__init__.py
@@ -0,0 +1 @@
+from .value_guided_sampling import ValueGuidedRLPipeline
diff --git a/model/diffusers_c/experimental/rl/value_guided_sampling.py b/model/diffusers_c/experimental/rl/value_guided_sampling.py
new file mode 100644
index 0000000000000000000000000000000000000000..2f9de857480ec590ae1f795bb2d29bbeccec1331
--- /dev/null
+++ b/model/diffusers_c/experimental/rl/value_guided_sampling.py
@@ -0,0 +1,153 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import torch
+import tqdm
+
+from ...models.unets.unet_1d import UNet1DModel
+from ...pipelines import DiffusionPipeline
+from ...utils.dummy_pt_objects import DDPMScheduler
+from ...utils.torch_utils import randn_tensor
+
+
+class ValueGuidedRLPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for value-guided sampling from a diffusion model trained to predict sequences of states.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ value_function ([`UNet1DModel`]):
+ A specialized UNet for fine-tuning trajectories base on reward.
+ unet ([`UNet1DModel`]):
+ UNet architecture to denoise the encoded trajectories.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded trajectories. Default for this
+ application is [`DDPMScheduler`].
+ env ():
+ An environment following the OpenAI gym API to act in. For now only Hopper has pretrained models.
+ """
+
+ def __init__(
+ self,
+ value_function: UNet1DModel,
+ unet: UNet1DModel,
+ scheduler: DDPMScheduler,
+ env,
+ ):
+ super().__init__()
+
+ self.register_modules(value_function=value_function, unet=unet, scheduler=scheduler, env=env)
+
+ self.data = env.get_dataset()
+ self.means = {}
+ for key in self.data.keys():
+ try:
+ self.means[key] = self.data[key].mean()
+ except: # noqa: E722
+ pass
+ self.stds = {}
+ for key in self.data.keys():
+ try:
+ self.stds[key] = self.data[key].std()
+ except: # noqa: E722
+ pass
+ self.state_dim = env.observation_space.shape[0]
+ self.action_dim = env.action_space.shape[0]
+
+ def normalize(self, x_in, key):
+ return (x_in - self.means[key]) / self.stds[key]
+
+ def de_normalize(self, x_in, key):
+ return x_in * self.stds[key] + self.means[key]
+
+ def to_torch(self, x_in):
+ if isinstance(x_in, dict):
+ return {k: self.to_torch(v) for k, v in x_in.items()}
+ elif torch.is_tensor(x_in):
+ return x_in.to(self.unet.device)
+ return torch.tensor(x_in, device=self.unet.device)
+
+ def reset_x0(self, x_in, cond, act_dim):
+ for key, val in cond.items():
+ x_in[:, key, act_dim:] = val.clone()
+ return x_in
+
+ def run_diffusion(self, x, conditions, n_guide_steps, scale):
+ batch_size = x.shape[0]
+ y = None
+ for i in tqdm.tqdm(self.scheduler.timesteps):
+ # create batch of timesteps to pass into model
+ timesteps = torch.full((batch_size,), i, device=self.unet.device, dtype=torch.long)
+ for _ in range(n_guide_steps):
+ with torch.enable_grad():
+ x.requires_grad_()
+
+ # permute to match dimension for pre-trained models
+ y = self.value_function(x.permute(0, 2, 1), timesteps).sample
+ grad = torch.autograd.grad([y.sum()], [x])[0]
+
+ posterior_variance = self.scheduler._get_variance(i)
+ model_std = torch.exp(0.5 * posterior_variance)
+ grad = model_std * grad
+
+ grad[timesteps < 2] = 0
+ x = x.detach()
+ x = x + scale * grad
+ x = self.reset_x0(x, conditions, self.action_dim)
+
+ prev_x = self.unet(x.permute(0, 2, 1), timesteps).sample.permute(0, 2, 1)
+
+ # TODO: verify deprecation of this kwarg
+ x = self.scheduler.step(prev_x, i, x)["prev_sample"]
+
+ # apply conditions to the trajectory (set the initial state)
+ x = self.reset_x0(x, conditions, self.action_dim)
+ x = self.to_torch(x)
+ return x, y
+
+ def __call__(self, obs, batch_size=64, planning_horizon=32, n_guide_steps=2, scale=0.1):
+ # normalize the observations and create batch dimension
+ obs = self.normalize(obs, "observations")
+ obs = obs[None].repeat(batch_size, axis=0)
+
+ conditions = {0: self.to_torch(obs)}
+ shape = (batch_size, planning_horizon, self.state_dim + self.action_dim)
+
+ # generate initial noise and apply our conditions (to make the trajectories start at current state)
+ x1 = randn_tensor(shape, device=self.unet.device)
+ x = self.reset_x0(x1, conditions, self.action_dim)
+ x = self.to_torch(x)
+
+ # run the diffusion process
+ x, y = self.run_diffusion(x, conditions, n_guide_steps, scale)
+
+ # sort output trajectories by value
+ sorted_idx = y.argsort(0, descending=True).squeeze()
+ sorted_values = x[sorted_idx]
+ actions = sorted_values[:, :, : self.action_dim]
+ actions = actions.detach().cpu().numpy()
+ denorm_actions = self.de_normalize(actions, key="actions")
+
+ # select the action with the highest value
+ if y is not None:
+ selected_index = 0
+ else:
+ # if we didn't run value guiding, select a random action
+ selected_index = np.random.randint(0, batch_size)
+
+ denorm_actions = denorm_actions[selected_index, 0]
+ return denorm_actions
diff --git a/model/diffusers_c/image_processor.py b/model/diffusers_c/image_processor.py
new file mode 100644
index 0000000000000000000000000000000000000000..daeb8fd6fa6de2855da489c89d17abbb14ae4b58
--- /dev/null
+++ b/model/diffusers_c/image_processor.py
@@ -0,0 +1,990 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from PIL import Image, ImageFilter, ImageOps
+
+from .configuration_utils import ConfigMixin, register_to_config
+from .utils import CONFIG_NAME, PIL_INTERPOLATION, deprecate
+
+
+PipelineImageInput = Union[
+ PIL.Image.Image,
+ np.ndarray,
+ torch.FloatTensor,
+ List[PIL.Image.Image],
+ List[np.ndarray],
+ List[torch.FloatTensor],
+]
+
+PipelineDepthInput = PipelineImageInput
+
+
+class VaeImageProcessor(ConfigMixin):
+ """
+ Image processor for VAE.
+
+ Args:
+ do_resize (`bool`, *optional*, defaults to `True`):
+ Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`. Can accept
+ `height` and `width` arguments from [`image_processor.VaeImageProcessor.preprocess`] method.
+ vae_scale_factor (`int`, *optional*, defaults to `8`):
+ VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
+ resample (`str`, *optional*, defaults to `lanczos`):
+ Resampling filter to use when resizing the image.
+ do_normalize (`bool`, *optional*, defaults to `True`):
+ Whether to normalize the image to [-1,1].
+ do_binarize (`bool`, *optional*, defaults to `False`):
+ Whether to binarize the image to 0/1.
+ do_convert_rgb (`bool`, *optional*, defaults to be `False`):
+ Whether to convert the images to RGB format.
+ do_convert_grayscale (`bool`, *optional*, defaults to be `False`):
+ Whether to convert the images to grayscale format.
+ """
+
+ config_name = CONFIG_NAME
+
+ @register_to_config
+ def __init__(
+ self,
+ do_resize: bool = True,
+ vae_scale_factor: int = 8,
+ resample: str = "lanczos",
+ do_normalize: bool = True,
+ do_binarize: bool = False,
+ do_convert_rgb: bool = False,
+ do_convert_grayscale: bool = False,
+ ):
+ super().__init__()
+ if do_convert_rgb and do_convert_grayscale:
+ raise ValueError(
+ "`do_convert_rgb` and `do_convert_grayscale` can not both be set to `True`,"
+ " if you intended to convert the image into RGB format, please set `do_convert_grayscale = False`.",
+ " if you intended to convert the image into grayscale format, please set `do_convert_rgb = False`",
+ )
+ self.config.do_convert_rgb = False
+
+ @staticmethod
+ def numpy_to_pil(images: np.ndarray) -> List[PIL.Image.Image]:
+ """
+ Convert a numpy image or a batch of images to a PIL image.
+ """
+ if images.ndim == 3:
+ images = images[None, ...]
+ images = (images * 255).round().astype("uint8")
+ if images.shape[-1] == 1:
+ # special case for grayscale (single channel) images
+ pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
+ else:
+ pil_images = [Image.fromarray(image) for image in images]
+
+ return pil_images
+
+ @staticmethod
+ def pil_to_numpy(images: Union[List[PIL.Image.Image], PIL.Image.Image]) -> np.ndarray:
+ """
+ Convert a PIL image or a list of PIL images to NumPy arrays.
+ """
+ if not isinstance(images, list):
+ images = [images]
+ images = [np.array(image).astype(np.float32) / 255.0 for image in images]
+ images = np.stack(images, axis=0)
+
+ return images
+
+ @staticmethod
+ def numpy_to_pt(images: np.ndarray) -> torch.FloatTensor:
+ """
+ Convert a NumPy image to a PyTorch tensor.
+ """
+ if images.ndim == 3:
+ images = images[..., None]
+
+ images = torch.from_numpy(images.transpose(0, 3, 1, 2))
+ return images
+
+ @staticmethod
+ def pt_to_numpy(images: torch.FloatTensor) -> np.ndarray:
+ """
+ Convert a PyTorch tensor to a NumPy image.
+ """
+ images = images.cpu().permute(0, 2, 3, 1).float().numpy()
+ return images
+
+ @staticmethod
+ def normalize(images: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
+ """
+ Normalize an image array to [-1,1].
+ """
+ return 2.0 * images - 1.0
+
+ @staticmethod
+ def denormalize(images: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
+ """
+ Denormalize an image array to [0,1].
+ """
+ return (images / 2 + 0.5).clamp(0, 1)
+
+ @staticmethod
+ def convert_to_rgb(image: PIL.Image.Image) -> PIL.Image.Image:
+ """
+ Converts a PIL image to RGB format.
+ """
+ image = image.convert("RGB")
+
+ return image
+
+ @staticmethod
+ def convert_to_grayscale(image: PIL.Image.Image) -> PIL.Image.Image:
+ """
+ Converts a PIL image to grayscale format.
+ """
+ image = image.convert("L")
+
+ return image
+
+ @staticmethod
+ def blur(image: PIL.Image.Image, blur_factor: int = 4) -> PIL.Image.Image:
+ """
+ Applies Gaussian blur to an image.
+ """
+ image = image.filter(ImageFilter.GaussianBlur(blur_factor))
+
+ return image
+
+ @staticmethod
+ def get_crop_region(mask_image: PIL.Image.Image, width: int, height: int, pad=0):
+ """
+ Finds a rectangular region that contains all masked ares in an image, and expands region to match the aspect ratio of the original image;
+ for example, if user drew mask in a 128x32 region, and the dimensions for processing are 512x512, the region will be expanded to 128x128.
+
+ Args:
+ mask_image (PIL.Image.Image): Mask image.
+ width (int): Width of the image to be processed.
+ height (int): Height of the image to be processed.
+ pad (int, optional): Padding to be added to the crop region. Defaults to 0.
+
+ Returns:
+ tuple: (x1, y1, x2, y2) represent a rectangular region that contains all masked ares in an image and matches the original aspect ratio.
+ """
+
+ mask_image = mask_image.convert("L")
+ mask = np.array(mask_image)
+
+ # 1. find a rectangular region that contains all masked ares in an image
+ h, w = mask.shape
+ crop_left = 0
+ for i in range(w):
+ if not (mask[:, i] == 0).all():
+ break
+ crop_left += 1
+
+ crop_right = 0
+ for i in reversed(range(w)):
+ if not (mask[:, i] == 0).all():
+ break
+ crop_right += 1
+
+ crop_top = 0
+ for i in range(h):
+ if not (mask[i] == 0).all():
+ break
+ crop_top += 1
+
+ crop_bottom = 0
+ for i in reversed(range(h)):
+ if not (mask[i] == 0).all():
+ break
+ crop_bottom += 1
+
+ # 2. add padding to the crop region
+ x1, y1, x2, y2 = (
+ int(max(crop_left - pad, 0)),
+ int(max(crop_top - pad, 0)),
+ int(min(w - crop_right + pad, w)),
+ int(min(h - crop_bottom + pad, h)),
+ )
+
+ # 3. expands crop region to match the aspect ratio of the image to be processed
+ ratio_crop_region = (x2 - x1) / (y2 - y1)
+ ratio_processing = width / height
+
+ if ratio_crop_region > ratio_processing:
+ desired_height = (x2 - x1) / ratio_processing
+ desired_height_diff = int(desired_height - (y2 - y1))
+ y1 -= desired_height_diff // 2
+ y2 += desired_height_diff - desired_height_diff // 2
+ if y2 >= mask_image.height:
+ diff = y2 - mask_image.height
+ y2 -= diff
+ y1 -= diff
+ if y1 < 0:
+ y2 -= y1
+ y1 -= y1
+ if y2 >= mask_image.height:
+ y2 = mask_image.height
+ else:
+ desired_width = (y2 - y1) * ratio_processing
+ desired_width_diff = int(desired_width - (x2 - x1))
+ x1 -= desired_width_diff // 2
+ x2 += desired_width_diff - desired_width_diff // 2
+ if x2 >= mask_image.width:
+ diff = x2 - mask_image.width
+ x2 -= diff
+ x1 -= diff
+ if x1 < 0:
+ x2 -= x1
+ x1 -= x1
+ if x2 >= mask_image.width:
+ x2 = mask_image.width
+
+ return x1, y1, x2, y2
+
+ def _resize_and_fill(
+ self,
+ image: PIL.Image.Image,
+ width: int,
+ height: int,
+ ) -> PIL.Image.Image:
+ """
+ Resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image within the dimensions, filling empty with data from image.
+
+ Args:
+ image: The image to resize.
+ width: The width to resize the image to.
+ height: The height to resize the image to.
+ """
+
+ ratio = width / height
+ src_ratio = image.width / image.height
+
+ src_w = width if ratio < src_ratio else image.width * height // image.height
+ src_h = height if ratio >= src_ratio else image.height * width // image.width
+
+ resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION["lanczos"])
+ res = Image.new("RGB", (width, height))
+ res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
+
+ if ratio < src_ratio:
+ fill_height = height // 2 - src_h // 2
+ if fill_height > 0:
+ res.paste(resized.resize((width, fill_height), box=(0, 0, width, 0)), box=(0, 0))
+ res.paste(
+ resized.resize((width, fill_height), box=(0, resized.height, width, resized.height)),
+ box=(0, fill_height + src_h),
+ )
+ elif ratio > src_ratio:
+ fill_width = width // 2 - src_w // 2
+ if fill_width > 0:
+ res.paste(resized.resize((fill_width, height), box=(0, 0, 0, height)), box=(0, 0))
+ res.paste(
+ resized.resize((fill_width, height), box=(resized.width, 0, resized.width, height)),
+ box=(fill_width + src_w, 0),
+ )
+
+ return res
+
+ def _resize_and_crop(
+ self,
+ image: PIL.Image.Image,
+ width: int,
+ height: int,
+ ) -> PIL.Image.Image:
+ """
+ Resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image within the dimensions, cropping the excess.
+
+ Args:
+ image: The image to resize.
+ width: The width to resize the image to.
+ height: The height to resize the image to.
+ """
+ ratio = width / height
+ src_ratio = image.width / image.height
+
+ src_w = width if ratio > src_ratio else image.width * height // image.height
+ src_h = height if ratio <= src_ratio else image.height * width // image.width
+
+ resized = image.resize((src_w, src_h), resample=PIL_INTERPOLATION["lanczos"])
+ res = Image.new("RGB", (width, height))
+ res.paste(resized, box=(width // 2 - src_w // 2, height // 2 - src_h // 2))
+ return res
+
+ def resize(
+ self,
+ image: Union[PIL.Image.Image, np.ndarray, torch.Tensor],
+ height: int,
+ width: int,
+ resize_mode: str = "default", # "default", "fill", "crop"
+ ) -> Union[PIL.Image.Image, np.ndarray, torch.Tensor]:
+ """
+ Resize image.
+
+ Args:
+ image (`PIL.Image.Image`, `np.ndarray` or `torch.Tensor`):
+ The image input, can be a PIL image, numpy array or pytorch tensor.
+ height (`int`):
+ The height to resize to.
+ width (`int`):
+ The width to resize to.
+ resize_mode (`str`, *optional*, defaults to `default`):
+ The resize mode to use, can be one of `default` or `fill`. If `default`, will resize the image to fit
+ within the specified width and height, and it may not maintaining the original aspect ratio.
+ If `fill`, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image
+ within the dimensions, filling empty with data from image.
+ If `crop`, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image
+ within the dimensions, cropping the excess.
+ Note that resize_mode `fill` and `crop` are only supported for PIL image input.
+
+ Returns:
+ `PIL.Image.Image`, `np.ndarray` or `torch.Tensor`:
+ The resized image.
+ """
+ if resize_mode != "default" and not isinstance(image, PIL.Image.Image):
+ raise ValueError(f"Only PIL image input is supported for resize_mode {resize_mode}")
+ if isinstance(image, PIL.Image.Image):
+ if resize_mode == "default":
+ image = image.resize((width, height), resample=PIL_INTERPOLATION[self.config.resample])
+ elif resize_mode == "fill":
+ image = self._resize_and_fill(image, width, height)
+ elif resize_mode == "crop":
+ image = self._resize_and_crop(image, width, height)
+ else:
+ raise ValueError(f"resize_mode {resize_mode} is not supported")
+
+ elif isinstance(image, torch.Tensor):
+ image = torch.nn.functional.interpolate(
+ image,
+ size=(height, width),
+ )
+ elif isinstance(image, np.ndarray):
+ image = self.numpy_to_pt(image)
+ image = torch.nn.functional.interpolate(
+ image,
+ size=(height, width),
+ )
+ image = self.pt_to_numpy(image)
+ return image
+
+ def binarize(self, image: PIL.Image.Image) -> PIL.Image.Image:
+ """
+ Create a mask.
+
+ Args:
+ image (`PIL.Image.Image`):
+ The image input, should be a PIL image.
+
+ Returns:
+ `PIL.Image.Image`:
+ The binarized image. Values less than 0.5 are set to 0, values greater than 0.5 are set to 1.
+ """
+ image[image < 0.5] = 0
+ image[image >= 0.5] = 1
+
+ return image
+
+ def get_default_height_width(
+ self,
+ image: Union[PIL.Image.Image, np.ndarray, torch.Tensor],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ ) -> Tuple[int, int]:
+ """
+ This function return the height and width that are downscaled to the next integer multiple of
+ `vae_scale_factor`.
+
+ Args:
+ image(`PIL.Image.Image`, `np.ndarray` or `torch.Tensor`):
+ The image input, can be a PIL image, numpy array or pytorch tensor. if it is a numpy array, should have
+ shape `[batch, height, width]` or `[batch, height, width, channel]` if it is a pytorch tensor, should
+ have shape `[batch, channel, height, width]`.
+ height (`int`, *optional*, defaults to `None`):
+ The height in preprocessed image. If `None`, will use the height of `image` input.
+ width (`int`, *optional*`, defaults to `None`):
+ The width in preprocessed. If `None`, will use the width of the `image` input.
+ """
+
+ if height is None:
+ if isinstance(image, PIL.Image.Image):
+ height = image.height
+ elif isinstance(image, torch.Tensor):
+ height = image.shape[2]
+ else:
+ height = image.shape[1]
+
+ if width is None:
+ if isinstance(image, PIL.Image.Image):
+ width = image.width
+ elif isinstance(image, torch.Tensor):
+ width = image.shape[3]
+ else:
+ width = image.shape[2]
+
+ width, height = (
+ x - x % self.config.vae_scale_factor for x in (width, height)
+ ) # resize to integer multiple of vae_scale_factor
+
+ return height, width
+
+ def preprocess(
+ self,
+ image: PipelineImageInput,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ resize_mode: str = "default", # "default", "fill", "crop"
+ crops_coords: Optional[Tuple[int, int, int, int]] = None,
+ ) -> torch.Tensor:
+ """
+ Preprocess the image input.
+
+ Args:
+ image (`pipeline_image_input`):
+ The image input, accepted formats are PIL images, NumPy arrays, PyTorch tensors; Also accept list of supported formats.
+ height (`int`, *optional*, defaults to `None`):
+ The height in preprocessed image. If `None`, will use the `get_default_height_width()` to get default height.
+ width (`int`, *optional*`, defaults to `None`):
+ The width in preprocessed. If `None`, will use get_default_height_width()` to get the default width.
+ resize_mode (`str`, *optional*, defaults to `default`):
+ The resize mode, can be one of `default` or `fill`. If `default`, will resize the image to fit
+ within the specified width and height, and it may not maintaining the original aspect ratio.
+ If `fill`, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image
+ within the dimensions, filling empty with data from image.
+ If `crop`, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image
+ within the dimensions, cropping the excess.
+ Note that resize_mode `fill` and `crop` are only supported for PIL image input.
+ crops_coords (`List[Tuple[int, int, int, int]]`, *optional*, defaults to `None`):
+ The crop coordinates for each image in the batch. If `None`, will not crop the image.
+ """
+ supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor)
+
+ # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image
+ if self.config.do_convert_grayscale and isinstance(image, (torch.Tensor, np.ndarray)) and image.ndim == 3:
+ if isinstance(image, torch.Tensor):
+ # if image is a pytorch tensor could have 2 possible shapes:
+ # 1. batch x height x width: we should insert the channel dimension at position 1
+ # 2. channel x height x width: we should insert batch dimension at position 0,
+ # however, since both channel and batch dimension has same size 1, it is same to insert at position 1
+ # for simplicity, we insert a dimension of size 1 at position 1 for both cases
+ image = image.unsqueeze(1)
+ else:
+ # if it is a numpy array, it could have 2 possible shapes:
+ # 1. batch x height x width: insert channel dimension on last position
+ # 2. height x width x channel: insert batch dimension on first position
+ if image.shape[-1] == 1:
+ image = np.expand_dims(image, axis=0)
+ else:
+ image = np.expand_dims(image, axis=-1)
+
+ if isinstance(image, supported_formats):
+ image = [image]
+ elif not (isinstance(image, list) and all(isinstance(i, supported_formats) for i in image)):
+ raise ValueError(
+ f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support {', '.join(supported_formats)}"
+ )
+
+ if isinstance(image[0], PIL.Image.Image):
+ if crops_coords is not None:
+ image = [i.crop(crops_coords) for i in image]
+ if self.config.do_resize:
+ height, width = self.get_default_height_width(image[0], height, width)
+ image = [self.resize(i, height, width, resize_mode=resize_mode) for i in image]
+ if self.config.do_convert_rgb:
+ image = [self.convert_to_rgb(i) for i in image]
+ elif self.config.do_convert_grayscale:
+ image = [self.convert_to_grayscale(i) for i in image]
+ image = self.pil_to_numpy(image) # to np
+ image = self.numpy_to_pt(image) # to pt
+
+ elif isinstance(image[0], np.ndarray):
+ image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
+
+ image = self.numpy_to_pt(image)
+
+ height, width = self.get_default_height_width(image, height, width)
+ if self.config.do_resize:
+ image = self.resize(image, height, width)
+
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
+
+ if self.config.do_convert_grayscale and image.ndim == 3:
+ image = image.unsqueeze(1)
+
+ channel = image.shape[1]
+ # don't need any preprocess if the image is latents
+ if channel == 4:
+ return image
+
+ height, width = self.get_default_height_width(image, height, width)
+ if self.config.do_resize:
+ image = self.resize(image, height, width)
+
+ # expected range [0,1], normalize to [-1,1]
+ do_normalize = self.config.do_normalize
+ if do_normalize and image.min() < 0:
+ warnings.warn(
+ "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] "
+ f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{image.min()},{image.max()}]",
+ FutureWarning,
+ )
+ do_normalize = False
+
+ if do_normalize:
+ image = self.normalize(image)
+
+ if self.config.do_binarize:
+ image = self.binarize(image)
+
+ return image
+
+ def postprocess(
+ self,
+ image: torch.FloatTensor,
+ output_type: str = "pil",
+ do_denormalize: Optional[List[bool]] = None,
+ ) -> Union[PIL.Image.Image, np.ndarray, torch.FloatTensor]:
+ """
+ Postprocess the image output from tensor to `output_type`.
+
+ Args:
+ image (`torch.FloatTensor`):
+ The image input, should be a pytorch tensor with shape `B x C x H x W`.
+ output_type (`str`, *optional*, defaults to `pil`):
+ The output type of the image, can be one of `pil`, `np`, `pt`, `latent`.
+ do_denormalize (`List[bool]`, *optional*, defaults to `None`):
+ Whether to denormalize the image to [0,1]. If `None`, will use the value of `do_normalize` in the
+ `VaeImageProcessor` config.
+
+ Returns:
+ `PIL.Image.Image`, `np.ndarray` or `torch.FloatTensor`:
+ The postprocessed image.
+ """
+ if not isinstance(image, torch.Tensor):
+ raise ValueError(
+ f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor"
+ )
+ if output_type not in ["latent", "pt", "np", "pil"]:
+ deprecation_message = (
+ f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: "
+ "`pil`, `np`, `pt`, `latent`"
+ )
+ deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False)
+ output_type = "np"
+
+ if output_type == "latent":
+ return image
+
+ if do_denormalize is None:
+ do_denormalize = [self.config.do_normalize] * image.shape[0]
+
+ image = torch.stack(
+ [self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
+ )
+
+ if output_type == "pt":
+ return image
+
+ image = self.pt_to_numpy(image)
+
+ if output_type == "np":
+ return image
+
+ if output_type == "pil":
+ return self.numpy_to_pil(image)
+
+ def apply_overlay(
+ self,
+ mask: PIL.Image.Image,
+ init_image: PIL.Image.Image,
+ image: PIL.Image.Image,
+ crop_coords: Optional[Tuple[int, int, int, int]] = None,
+ ) -> PIL.Image.Image:
+ """
+ overlay the inpaint output to the original image
+ """
+
+ width, height = image.width, image.height
+
+ init_image = self.resize(init_image, width=width, height=height)
+ mask = self.resize(mask, width=width, height=height)
+
+ init_image_masked = PIL.Image.new("RGBa", (width, height))
+ init_image_masked.paste(init_image.convert("RGBA").convert("RGBa"), mask=ImageOps.invert(mask.convert("L")))
+ init_image_masked = init_image_masked.convert("RGBA")
+
+ if crop_coords is not None:
+ x, y, x2, y2 = crop_coords
+ w = x2 - x
+ h = y2 - y
+ base_image = PIL.Image.new("RGBA", (width, height))
+ image = self.resize(image, height=h, width=w, resize_mode="crop")
+ base_image.paste(image, (x, y))
+ image = base_image.convert("RGB")
+
+ image = image.convert("RGBA")
+ image.alpha_composite(init_image_masked)
+ image = image.convert("RGB")
+
+ return image
+
+
+class VaeImageProcessorLDM3D(VaeImageProcessor):
+ """
+ Image processor for VAE LDM3D.
+
+ Args:
+ do_resize (`bool`, *optional*, defaults to `True`):
+ Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`.
+ vae_scale_factor (`int`, *optional*, defaults to `8`):
+ VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
+ resample (`str`, *optional*, defaults to `lanczos`):
+ Resampling filter to use when resizing the image.
+ do_normalize (`bool`, *optional*, defaults to `True`):
+ Whether to normalize the image to [-1,1].
+ """
+
+ config_name = CONFIG_NAME
+
+ @register_to_config
+ def __init__(
+ self,
+ do_resize: bool = True,
+ vae_scale_factor: int = 8,
+ resample: str = "lanczos",
+ do_normalize: bool = True,
+ ):
+ super().__init__()
+
+ @staticmethod
+ def numpy_to_pil(images: np.ndarray) -> List[PIL.Image.Image]:
+ """
+ Convert a NumPy image or a batch of images to a PIL image.
+ """
+ if images.ndim == 3:
+ images = images[None, ...]
+ images = (images * 255).round().astype("uint8")
+ if images.shape[-1] == 1:
+ # special case for grayscale (single channel) images
+ pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
+ else:
+ pil_images = [Image.fromarray(image[:, :, :3]) for image in images]
+
+ return pil_images
+
+ @staticmethod
+ def depth_pil_to_numpy(images: Union[List[PIL.Image.Image], PIL.Image.Image]) -> np.ndarray:
+ """
+ Convert a PIL image or a list of PIL images to NumPy arrays.
+ """
+ if not isinstance(images, list):
+ images = [images]
+
+ images = [np.array(image).astype(np.float32) / (2**16 - 1) for image in images]
+ images = np.stack(images, axis=0)
+ return images
+
+ @staticmethod
+ def rgblike_to_depthmap(image: Union[np.ndarray, torch.Tensor]) -> Union[np.ndarray, torch.Tensor]:
+ """
+ Args:
+ image: RGB-like depth image
+
+ Returns: depth map
+
+ """
+ return image[:, :, 1] * 2**8 + image[:, :, 2]
+
+ def numpy_to_depth(self, images: np.ndarray) -> List[PIL.Image.Image]:
+ """
+ Convert a NumPy depth image or a batch of images to a PIL image.
+ """
+ if images.ndim == 3:
+ images = images[None, ...]
+ images_depth = images[:, :, :, 3:]
+ if images.shape[-1] == 6:
+ images_depth = (images_depth * 255).round().astype("uint8")
+ pil_images = [
+ Image.fromarray(self.rgblike_to_depthmap(image_depth), mode="I;16") for image_depth in images_depth
+ ]
+ elif images.shape[-1] == 4:
+ images_depth = (images_depth * 65535.0).astype(np.uint16)
+ pil_images = [Image.fromarray(image_depth, mode="I;16") for image_depth in images_depth]
+ else:
+ raise Exception("Not supported")
+
+ return pil_images
+
+ def postprocess(
+ self,
+ image: torch.FloatTensor,
+ output_type: str = "pil",
+ do_denormalize: Optional[List[bool]] = None,
+ ) -> Union[PIL.Image.Image, np.ndarray, torch.FloatTensor]:
+ """
+ Postprocess the image output from tensor to `output_type`.
+
+ Args:
+ image (`torch.FloatTensor`):
+ The image input, should be a pytorch tensor with shape `B x C x H x W`.
+ output_type (`str`, *optional*, defaults to `pil`):
+ The output type of the image, can be one of `pil`, `np`, `pt`, `latent`.
+ do_denormalize (`List[bool]`, *optional*, defaults to `None`):
+ Whether to denormalize the image to [0,1]. If `None`, will use the value of `do_normalize` in the
+ `VaeImageProcessor` config.
+
+ Returns:
+ `PIL.Image.Image`, `np.ndarray` or `torch.FloatTensor`:
+ The postprocessed image.
+ """
+ if not isinstance(image, torch.Tensor):
+ raise ValueError(
+ f"Input for postprocessing is in incorrect format: {type(image)}. We only support pytorch tensor"
+ )
+ if output_type not in ["latent", "pt", "np", "pil"]:
+ deprecation_message = (
+ f"the output_type {output_type} is outdated and has been set to `np`. Please make sure to set it to one of these instead: "
+ "`pil`, `np`, `pt`, `latent`"
+ )
+ deprecate("Unsupported output_type", "1.0.0", deprecation_message, standard_warn=False)
+ output_type = "np"
+
+ if do_denormalize is None:
+ do_denormalize = [self.config.do_normalize] * image.shape[0]
+
+ image = torch.stack(
+ [self.denormalize(image[i]) if do_denormalize[i] else image[i] for i in range(image.shape[0])]
+ )
+
+ image = self.pt_to_numpy(image)
+
+ if output_type == "np":
+ if image.shape[-1] == 6:
+ image_depth = np.stack([self.rgblike_to_depthmap(im[:, :, 3:]) for im in image], axis=0)
+ else:
+ image_depth = image[:, :, :, 3:]
+ return image[:, :, :, :3], image_depth
+
+ if output_type == "pil":
+ return self.numpy_to_pil(image), self.numpy_to_depth(image)
+ else:
+ raise Exception(f"This type {output_type} is not supported")
+
+ def preprocess(
+ self,
+ rgb: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
+ depth: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ target_res: Optional[int] = None,
+ ) -> torch.Tensor:
+ """
+ Preprocess the image input. Accepted formats are PIL images, NumPy arrays or PyTorch tensors.
+ """
+ supported_formats = (PIL.Image.Image, np.ndarray, torch.Tensor)
+
+ # Expand the missing dimension for 3-dimensional pytorch tensor or numpy array that represents grayscale image
+ if self.config.do_convert_grayscale and isinstance(rgb, (torch.Tensor, np.ndarray)) and rgb.ndim == 3:
+ raise Exception("This is not yet supported")
+
+ if isinstance(rgb, supported_formats):
+ rgb = [rgb]
+ depth = [depth]
+ elif not (isinstance(rgb, list) and all(isinstance(i, supported_formats) for i in rgb)):
+ raise ValueError(
+ f"Input is in incorrect format: {[type(i) for i in rgb]}. Currently, we only support {', '.join(supported_formats)}"
+ )
+
+ if isinstance(rgb[0], PIL.Image.Image):
+ if self.config.do_convert_rgb:
+ raise Exception("This is not yet supported")
+ # rgb = [self.convert_to_rgb(i) for i in rgb]
+ # depth = [self.convert_to_depth(i) for i in depth] #TODO define convert_to_depth
+ if self.config.do_resize or target_res:
+ height, width = self.get_default_height_width(rgb[0], height, width) if not target_res else target_res
+ rgb = [self.resize(i, height, width) for i in rgb]
+ depth = [self.resize(i, height, width) for i in depth]
+ rgb = self.pil_to_numpy(rgb) # to np
+ rgb = self.numpy_to_pt(rgb) # to pt
+
+ depth = self.depth_pil_to_numpy(depth) # to np
+ depth = self.numpy_to_pt(depth) # to pt
+
+ elif isinstance(rgb[0], np.ndarray):
+ rgb = np.concatenate(rgb, axis=0) if rgb[0].ndim == 4 else np.stack(rgb, axis=0)
+ rgb = self.numpy_to_pt(rgb)
+ height, width = self.get_default_height_width(rgb, height, width)
+ if self.config.do_resize:
+ rgb = self.resize(rgb, height, width)
+
+ depth = np.concatenate(depth, axis=0) if rgb[0].ndim == 4 else np.stack(depth, axis=0)
+ depth = self.numpy_to_pt(depth)
+ height, width = self.get_default_height_width(depth, height, width)
+ if self.config.do_resize:
+ depth = self.resize(depth, height, width)
+
+ elif isinstance(rgb[0], torch.Tensor):
+ raise Exception("This is not yet supported")
+ # rgb = torch.cat(rgb, axis=0) if rgb[0].ndim == 4 else torch.stack(rgb, axis=0)
+
+ # if self.config.do_convert_grayscale and rgb.ndim == 3:
+ # rgb = rgb.unsqueeze(1)
+
+ # channel = rgb.shape[1]
+
+ # height, width = self.get_default_height_width(rgb, height, width)
+ # if self.config.do_resize:
+ # rgb = self.resize(rgb, height, width)
+
+ # depth = torch.cat(depth, axis=0) if depth[0].ndim == 4 else torch.stack(depth, axis=0)
+
+ # if self.config.do_convert_grayscale and depth.ndim == 3:
+ # depth = depth.unsqueeze(1)
+
+ # channel = depth.shape[1]
+ # # don't need any preprocess if the image is latents
+ # if depth == 4:
+ # return rgb, depth
+
+ # height, width = self.get_default_height_width(depth, height, width)
+ # if self.config.do_resize:
+ # depth = self.resize(depth, height, width)
+ # expected range [0,1], normalize to [-1,1]
+ do_normalize = self.config.do_normalize
+ if rgb.min() < 0 and do_normalize:
+ warnings.warn(
+ "Passing `image` as torch tensor with value range in [-1,1] is deprecated. The expected value range for image tensor is [0,1] "
+ f"when passing as pytorch tensor or numpy Array. You passed `image` with value range [{rgb.min()},{rgb.max()}]",
+ FutureWarning,
+ )
+ do_normalize = False
+
+ if do_normalize:
+ rgb = self.normalize(rgb)
+ depth = self.normalize(depth)
+
+ if self.config.do_binarize:
+ rgb = self.binarize(rgb)
+ depth = self.binarize(depth)
+
+ return rgb, depth
+
+
+class IPAdapterMaskProcessor(VaeImageProcessor):
+ """
+ Image processor for IP Adapter image masks.
+
+ Args:
+ do_resize (`bool`, *optional*, defaults to `True`):
+ Whether to downscale the image's (height, width) dimensions to multiples of `vae_scale_factor`.
+ vae_scale_factor (`int`, *optional*, defaults to `8`):
+ VAE scale factor. If `do_resize` is `True`, the image is automatically resized to multiples of this factor.
+ resample (`str`, *optional*, defaults to `lanczos`):
+ Resampling filter to use when resizing the image.
+ do_normalize (`bool`, *optional*, defaults to `False`):
+ Whether to normalize the image to [-1,1].
+ do_binarize (`bool`, *optional*, defaults to `True`):
+ Whether to binarize the image to 0/1.
+ do_convert_grayscale (`bool`, *optional*, defaults to be `True`):
+ Whether to convert the images to grayscale format.
+
+ """
+
+ config_name = CONFIG_NAME
+
+ @register_to_config
+ def __init__(
+ self,
+ do_resize: bool = True,
+ vae_scale_factor: int = 8,
+ resample: str = "lanczos",
+ do_normalize: bool = False,
+ do_binarize: bool = True,
+ do_convert_grayscale: bool = True,
+ ):
+ super().__init__(
+ do_resize=do_resize,
+ vae_scale_factor=vae_scale_factor,
+ resample=resample,
+ do_normalize=do_normalize,
+ do_binarize=do_binarize,
+ do_convert_grayscale=do_convert_grayscale,
+ )
+
+ @staticmethod
+ def downsample(mask: torch.FloatTensor, batch_size: int, num_queries: int, value_embed_dim: int):
+ """
+ Downsamples the provided mask tensor to match the expected dimensions for scaled dot-product attention.
+ If the aspect ratio of the mask does not match the aspect ratio of the output image, a warning is issued.
+
+ Args:
+ mask (`torch.FloatTensor`):
+ The input mask tensor generated with `IPAdapterMaskProcessor.preprocess()`.
+ batch_size (`int`):
+ The batch size.
+ num_queries (`int`):
+ The number of queries.
+ value_embed_dim (`int`):
+ The dimensionality of the value embeddings.
+
+ Returns:
+ `torch.FloatTensor`:
+ The downsampled mask tensor.
+
+ """
+ o_h = mask.shape[1]
+ o_w = mask.shape[2]
+ ratio = o_w / o_h
+ mask_h = int(math.sqrt(num_queries / ratio))
+ mask_h = int(mask_h) + int((num_queries % int(mask_h)) != 0)
+ mask_w = num_queries // mask_h
+
+ mask_downsample = F.interpolate(mask.unsqueeze(0), size=(mask_h, mask_w), mode="bicubic").squeeze(0)
+
+ # Repeat batch_size times
+ if mask_downsample.shape[0] < batch_size:
+ mask_downsample = mask_downsample.repeat(batch_size, 1, 1)
+
+ mask_downsample = mask_downsample.view(mask_downsample.shape[0], -1)
+
+ downsampled_area = mask_h * mask_w
+ # If the output image and the mask do not have the same aspect ratio, tensor shapes will not match
+ # Pad tensor if downsampled_mask.shape[1] is smaller than num_queries
+ if downsampled_area < num_queries:
+ warnings.warn(
+ "The aspect ratio of the mask does not match the aspect ratio of the output image. "
+ "Please update your masks or adjust the output size for optimal performance.",
+ UserWarning,
+ )
+ mask_downsample = F.pad(mask_downsample, (0, num_queries - mask_downsample.shape[1]), value=0.0)
+ # Discard last embeddings if downsampled_mask.shape[1] is bigger than num_queries
+ if downsampled_area > num_queries:
+ warnings.warn(
+ "The aspect ratio of the mask does not match the aspect ratio of the output image. "
+ "Please update your masks or adjust the output size for optimal performance.",
+ UserWarning,
+ )
+ mask_downsample = mask_downsample[:, :num_queries]
+
+ # Repeat last dimension to match SDPA output shape
+ mask_downsample = mask_downsample.view(mask_downsample.shape[0], mask_downsample.shape[1], 1).repeat(
+ 1, 1, value_embed_dim
+ )
+
+ return mask_downsample
diff --git a/model/diffusers_c/loaders/__init__.py b/model/diffusers_c/loaders/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..4da047435d8ef91ca2a22613d24ee6617c3a4f59
--- /dev/null
+++ b/model/diffusers_c/loaders/__init__.py
@@ -0,0 +1,88 @@
+from typing import TYPE_CHECKING
+
+from ..utils import DIFFUSERS_SLOW_IMPORT, _LazyModule, deprecate
+from ..utils.import_utils import is_peft_available, is_torch_available, is_transformers_available
+
+
+def text_encoder_lora_state_dict(text_encoder):
+ deprecate(
+ "text_encoder_load_state_dict in `models`",
+ "0.27.0",
+ "`text_encoder_lora_state_dict` is deprecated and will be removed in 0.27.0. Make sure to retrieve the weights using `get_peft_model`. See https://huggingface.co/docs/peft/v0.6.2/en/quicktour#peftmodel for more information.",
+ )
+ state_dict = {}
+
+ for name, module in text_encoder_attn_modules(text_encoder):
+ for k, v in module.q_proj.lora_linear_layer.state_dict().items():
+ state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v
+
+ for k, v in module.k_proj.lora_linear_layer.state_dict().items():
+ state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v
+
+ for k, v in module.v_proj.lora_linear_layer.state_dict().items():
+ state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v
+
+ for k, v in module.out_proj.lora_linear_layer.state_dict().items():
+ state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v
+
+ return state_dict
+
+
+if is_transformers_available():
+
+ def text_encoder_attn_modules(text_encoder):
+ deprecate(
+ "text_encoder_attn_modules in `models`",
+ "0.27.0",
+ "`text_encoder_lora_state_dict` is deprecated and will be removed in 0.27.0. Make sure to retrieve the weights using `get_peft_model`. See https://huggingface.co/docs/peft/v0.6.2/en/quicktour#peftmodel for more information.",
+ )
+ from transformers import CLIPTextModel, CLIPTextModelWithProjection
+
+ attn_modules = []
+
+ if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+ for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+ name = f"text_model.encoder.layers.{i}.self_attn"
+ mod = layer.self_attn
+ attn_modules.append((name, mod))
+ else:
+ raise ValueError(f"do not know how to get attention modules for: {text_encoder.__class__.__name__}")
+
+ return attn_modules
+
+
+_import_structure = {}
+
+if is_torch_available():
+ _import_structure["autoencoder"] = ["FromOriginalVAEMixin"]
+
+ _import_structure["controlnet"] = ["FromOriginalControlNetMixin"]
+ _import_structure["unet"] = ["UNet2DConditionLoadersMixin"]
+ _import_structure["utils"] = ["AttnProcsLayers"]
+ if is_transformers_available():
+ _import_structure["single_file"] = ["FromSingleFileMixin"]
+ _import_structure["lora"] = ["LoraLoaderMixin", "StableDiffusionXLLoraLoaderMixin"]
+ _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
+ _import_structure["ip_adapter"] = ["IPAdapterMixin"]
+
+_import_structure["peft"] = ["PeftAdapterMixin"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ if is_torch_available():
+ from .autoencoder import FromOriginalVAEMixin
+ from .controlnet import FromOriginalControlNetMixin
+ from .unet import UNet2DConditionLoadersMixin
+ from .utils import AttnProcsLayers
+
+ if is_transformers_available():
+ from .ip_adapter import IPAdapterMixin
+ from .lora import LoraLoaderMixin, StableDiffusionXLLoraLoaderMixin
+ from .single_file import FromSingleFileMixin
+ from .textual_inversion import TextualInversionLoaderMixin
+
+ from .peft import PeftAdapterMixin
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
diff --git a/model/diffusers_c/loaders/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/loaders/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..933ba001138c6829f2dd66f66afe451347244fcc
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d9607d4eb49520df8d63a8a8e86fa8105f6800b6
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/autoencoder.cpython-310.pyc b/model/diffusers_c/loaders/__pycache__/autoencoder.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9196f33e27b6d72e153a0bc809a8472565071d73
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/autoencoder.cpython-310.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/autoencoder.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/autoencoder.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f000f24849c30b9537d707f615213628347513b6
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/autoencoder.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/controlnet.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/controlnet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4abce221cd67568d91fafd5f621ed0e2b2f5475c
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/controlnet.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/ip_adapter.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/ip_adapter.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..22bfbaa501e83197b9437e772447b3ea3ba90a71
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/ip_adapter.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/lora.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/lora.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0f02e1391f5e603a9129c1c00834ae0d038bfa2b
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/lora.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/lora_conversion_utils.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/lora_conversion_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..08441c7ec9b6c7b4a66221cb649ed3539df00670
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/lora_conversion_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/peft.cpython-310.pyc b/model/diffusers_c/loaders/__pycache__/peft.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c539014da50e9caa053a7912515cb162d046619c
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/peft.cpython-310.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/peft.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/peft.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..72d7cd809a3fdb67d0a71667862a499ba7dc2573
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/peft.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/single_file.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/single_file.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..58060716f77f45bb86e339e687da94806f848075
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/single_file.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/single_file_utils.cpython-310.pyc b/model/diffusers_c/loaders/__pycache__/single_file_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a12b343f21b13ba7132df839603ec2c960234885
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/single_file_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/single_file_utils.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/single_file_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d2073968d55eecdcc2d959457a20fd75423247a4
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/single_file_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/textual_inversion.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/textual_inversion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7968b6c7055fd083b666709f146ee5aec5e7a6e2
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/textual_inversion.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/unet.cpython-310.pyc b/model/diffusers_c/loaders/__pycache__/unet.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b31898d7332ebf2cadd427a64af496bd93d12f58
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/unet.cpython-310.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/unet.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/unet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d562165120ea6c78f71b6fd4de37be97d2a52584
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/unet.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/utils.cpython-310.pyc b/model/diffusers_c/loaders/__pycache__/utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..db06e3f5ac8302e781dcf517e6d8441572b29466
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/loaders/__pycache__/utils.cpython-39.pyc b/model/diffusers_c/loaders/__pycache__/utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8e51d4477317ef27eb8d512d6e2589d23ec40c54
Binary files /dev/null and b/model/diffusers_c/loaders/__pycache__/utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/loaders/autoencoder.py b/model/diffusers_c/loaders/autoencoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..b91d27f7d63e3904cc2faaadb90865ca115d28b6
--- /dev/null
+++ b/model/diffusers_c/loaders/autoencoder.py
@@ -0,0 +1,146 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from huggingface_hub.utils import validate_hf_hub_args
+
+from .single_file_utils import (
+ create_diffusers_vae_model_from_ldm,
+ fetch_ldm_config_and_checkpoint,
+)
+
+
+class FromOriginalVAEMixin:
+ """
+ Load pretrained AutoencoderKL weights saved in the `.ckpt` or `.safetensors` format into a [`AutoencoderKL`].
+ """
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+ r"""
+ Instantiate a [`AutoencoderKL`] from pretrained ControlNet weights saved in the original `.ckpt` or
+ `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ Parameters:
+ pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+ - A link to the `.ckpt` file (for example
+ `"https://huggingface.co//blob/main/.ckpt"`) on the Hub.
+ - A path to a *file* containing all pipeline weights.
+ config_file (`str`, *optional*):
+ Filepath to the configuration YAML file associated with the model. If not provided it will default to:
+ https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
+ dtype is automatically derived from the model's weights.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to True, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ image_size (`int`, *optional*, defaults to 512):
+ The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+ Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+ scaling_factor (`float`, *optional*, defaults to 0.18215):
+ The component-wise standard deviation of the trained latent space computed using the first batch of the
+ training set. This is used to scale the latent space to have unit variance when training the diffusion
+ model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+ diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z
+ = 1 / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution
+ Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to overwrite load and saveable variables (for example the pipeline components of the
+ specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+ method. See example below for more information.
+
+
+
+ Make sure to pass both `image_size` and `scaling_factor` to `from_single_file()` if you're loading
+ a VAE from SDXL or a Stable Diffusion v2 model or higher.
+
+
+
+ Examples:
+
+ ```py
+ from diffusers import AutoencoderKL
+
+ url = "https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/vae-ft-mse-840000-ema-pruned.safetensors" # can also be local file
+ model = AutoencoderKL.from_single_file(url)
+ ```
+ """
+
+ original_config_file = kwargs.pop("original_config_file", None)
+ config_file = kwargs.pop("config_file", None)
+ resume_download = kwargs.pop("resume_download", False)
+ force_download = kwargs.pop("force_download", False)
+ proxies = kwargs.pop("proxies", None)
+ token = kwargs.pop("token", None)
+ cache_dir = kwargs.pop("cache_dir", None)
+ local_files_only = kwargs.pop("local_files_only", None)
+ revision = kwargs.pop("revision", None)
+ torch_dtype = kwargs.pop("torch_dtype", None)
+
+ class_name = cls.__name__
+
+ if (config_file is not None) and (original_config_file is not None):
+ raise ValueError(
+ "You cannot pass both `config_file` and `original_config_file` to `from_single_file`. Please use only one of these arguments."
+ )
+
+ original_config_file = original_config_file or config_file
+ original_config, checkpoint = fetch_ldm_config_and_checkpoint(
+ pretrained_model_link_or_path=pretrained_model_link_or_path,
+ class_name=class_name,
+ original_config_file=original_config_file,
+ resume_download=resume_download,
+ force_download=force_download,
+ proxies=proxies,
+ token=token,
+ revision=revision,
+ local_files_only=local_files_only,
+ cache_dir=cache_dir,
+ )
+
+ image_size = kwargs.pop("image_size", None)
+ scaling_factor = kwargs.pop("scaling_factor", None)
+ component = create_diffusers_vae_model_from_ldm(
+ class_name,
+ original_config,
+ checkpoint,
+ image_size=image_size,
+ scaling_factor=scaling_factor,
+ torch_dtype=torch_dtype,
+ )
+ vae = component["vae"]
+ if torch_dtype is not None:
+ vae = vae.to(torch_dtype)
+
+ return vae
diff --git a/model/diffusers_c/loaders/controlnet.py b/model/diffusers_c/loaders/controlnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..d323f60aa7ee32bfb1cdbe348696539d19097f45
--- /dev/null
+++ b/model/diffusers_c/loaders/controlnet.py
@@ -0,0 +1,136 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from huggingface_hub.utils import validate_hf_hub_args
+
+from .single_file_utils import (
+ create_diffusers_controlnet_model_from_ldm,
+ fetch_ldm_config_and_checkpoint,
+)
+
+
+class FromOriginalControlNetMixin:
+ """
+ Load pretrained ControlNet weights saved in the `.ckpt` or `.safetensors` format into a [`ControlNetModel`].
+ """
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+ r"""
+ Instantiate a [`ControlNetModel`] from pretrained ControlNet weights saved in the original `.ckpt` or
+ `.safetensors` format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ Parameters:
+ pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+ - A link to the `.ckpt` file (for example
+ `"https://huggingface.co//blob/main/.ckpt"`) on the Hub.
+ - A path to a *file* containing all pipeline weights.
+ config_file (`str`, *optional*):
+ Filepath to the configuration YAML file associated with the model. If not provided it will default to:
+ https://raw.githubusercontent.com/lllyasviel/ControlNet/main/models/cldm_v15.yaml
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
+ dtype is automatically derived from the model's weights.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to True, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ image_size (`int`, *optional*, defaults to 512):
+ The image size the model was trained on. Use 512 for all Stable Diffusion v1 models and the Stable
+ Diffusion v2 base model. Use 768 for Stable Diffusion v2.
+ upcast_attention (`bool`, *optional*, defaults to `None`):
+ Whether the attention computation should always be upcasted.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to overwrite load and saveable variables (for example the pipeline components of the
+ specific pipeline class). The overwritten components are directly passed to the pipelines `__init__`
+ method. See example below for more information.
+
+ Examples:
+
+ ```py
+ from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
+
+ url = "https://huggingface.co/lllyasviel/ControlNet-v1-1/blob/main/control_v11p_sd15_canny.pth" # can also be a local path
+ model = ControlNetModel.from_single_file(url)
+
+ url = "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned.safetensors" # can also be a local path
+ pipe = StableDiffusionControlNetPipeline.from_single_file(url, controlnet=controlnet)
+ ```
+ """
+ original_config_file = kwargs.pop("original_config_file", None)
+ config_file = kwargs.pop("config_file", None)
+ resume_download = kwargs.pop("resume_download", False)
+ force_download = kwargs.pop("force_download", False)
+ proxies = kwargs.pop("proxies", None)
+ token = kwargs.pop("token", None)
+ cache_dir = kwargs.pop("cache_dir", None)
+ local_files_only = kwargs.pop("local_files_only", None)
+ revision = kwargs.pop("revision", None)
+ torch_dtype = kwargs.pop("torch_dtype", None)
+
+ class_name = cls.__name__
+ if (config_file is not None) and (original_config_file is not None):
+ raise ValueError(
+ "You cannot pass both `config_file` and `original_config_file` to `from_single_file`. Please use only one of these arguments."
+ )
+
+ original_config_file = config_file or original_config_file
+ original_config, checkpoint = fetch_ldm_config_and_checkpoint(
+ pretrained_model_link_or_path=pretrained_model_link_or_path,
+ class_name=class_name,
+ original_config_file=original_config_file,
+ resume_download=resume_download,
+ force_download=force_download,
+ proxies=proxies,
+ token=token,
+ revision=revision,
+ local_files_only=local_files_only,
+ cache_dir=cache_dir,
+ )
+
+ upcast_attention = kwargs.pop("upcast_attention", False)
+ image_size = kwargs.pop("image_size", None)
+
+ component = create_diffusers_controlnet_model_from_ldm(
+ class_name,
+ original_config,
+ checkpoint,
+ upcast_attention=upcast_attention,
+ image_size=image_size,
+ torch_dtype=torch_dtype,
+ )
+ controlnet = component["controlnet"]
+ if torch_dtype is not None:
+ controlnet = controlnet.to(torch_dtype)
+
+ return controlnet
diff --git a/model/diffusers_c/loaders/ip_adapter.py b/model/diffusers_c/loaders/ip_adapter.py
new file mode 100644
index 0000000000000000000000000000000000000000..93959b9f0a6df6b74b4374519ae736e528e1e1e9
--- /dev/null
+++ b/model/diffusers_c/loaders/ip_adapter.py
@@ -0,0 +1,281 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from pathlib import Path
+from typing import Dict, List, Optional, Union
+
+import torch
+from huggingface_hub.utils import validate_hf_hub_args
+from safetensors import safe_open
+
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
+from ..utils import (
+ _get_model_file,
+ is_accelerate_available,
+ is_torch_version,
+ is_transformers_available,
+ logging,
+)
+
+
+if is_transformers_available():
+ from transformers import (
+ CLIPImageProcessor,
+ CLIPVisionModelWithProjection,
+ )
+
+ from ..models.attention_processor import (
+ IPAdapterAttnProcessor,
+ IPAdapterAttnProcessor2_0,
+ )
+
+logger = logging.get_logger(__name__)
+
+
+class IPAdapterMixin:
+ """Mixin for handling IP Adapters."""
+
+ @validate_hf_hub_args
+ def load_ip_adapter(
+ self,
+ pretrained_model_name_or_path_or_dict: Union[str, List[str], Dict[str, torch.Tensor]],
+ subfolder: Union[str, List[str]],
+ weight_name: Union[str, List[str]],
+ image_encoder_folder: Optional[str] = "image_encoder",
+ **kwargs,
+ ):
+ """
+ Parameters:
+ pretrained_model_name_or_path_or_dict (`str` or `List[str]` or `os.PathLike` or `List[os.PathLike]` or `dict` or `List[dict]`):
+ Can be either:
+
+ - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+ the Hub.
+ - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+ with [`ModelMixin.save_pretrained`].
+ - A [torch state
+ dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+ subfolder (`str` or `List[str]`):
+ The subfolder location of a model file within a larger model repository on the Hub or locally.
+ If a list is passed, it should have the same length as `weight_name`.
+ weight_name (`str` or `List[str]`):
+ The name of the weight file to load. If a list is passed, it should have the same length as
+ `weight_name`.
+ image_encoder_folder (`str`, *optional*, defaults to `image_encoder`):
+ The subfolder location of the image encoder within a larger model repository on the Hub or locally.
+ Pass `None` to not load the image encoder. If the image encoder is located in a folder inside `subfolder`,
+ you only need to pass the name of the folder that contains image encoder weights, e.g. `image_encoder_folder="image_encoder"`.
+ If the image encoder is located in a folder other than `subfolder`, you should pass the path to the folder that contains image encoder weights,
+ for example, `image_encoder_folder="different_subfolder/image_encoder"`.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ """
+
+ # handle the list inputs for multiple IP Adapters
+ if not isinstance(weight_name, list):
+ weight_name = [weight_name]
+
+ if not isinstance(pretrained_model_name_or_path_or_dict, list):
+ pretrained_model_name_or_path_or_dict = [pretrained_model_name_or_path_or_dict]
+ if len(pretrained_model_name_or_path_or_dict) == 1:
+ pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict * len(weight_name)
+
+ if not isinstance(subfolder, list):
+ subfolder = [subfolder]
+ if len(subfolder) == 1:
+ subfolder = subfolder * len(weight_name)
+
+ if len(weight_name) != len(pretrained_model_name_or_path_or_dict):
+ raise ValueError("`weight_name` and `pretrained_model_name_or_path_or_dict` must have the same length.")
+
+ if len(weight_name) != len(subfolder):
+ raise ValueError("`weight_name` and `subfolder` must have the same length.")
+
+ # Load the main state dict first.
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ local_files_only = kwargs.pop("local_files_only", None)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+
+ if low_cpu_mem_usage and not is_accelerate_available():
+ low_cpu_mem_usage = False
+ logger.warning(
+ "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+ " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+ " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+ " install accelerate\n```\n."
+ )
+
+ if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+ raise NotImplementedError(
+ "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+ " `low_cpu_mem_usage=False`."
+ )
+
+ user_agent = {
+ "file_type": "attn_procs_weights",
+ "framework": "pytorch",
+ }
+ state_dicts = []
+ for pretrained_model_name_or_path_or_dict, weight_name, subfolder in zip(
+ pretrained_model_name_or_path_or_dict, weight_name, subfolder
+ ):
+ if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+ model_file = _get_model_file(
+ pretrained_model_name_or_path_or_dict,
+ weights_name=weight_name,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ )
+ if weight_name.endswith(".safetensors"):
+ state_dict = {"image_proj": {}, "ip_adapter": {}}
+ with safe_open(model_file, framework="pt", device="cpu") as f:
+ for key in f.keys():
+ if key.startswith("image_proj."):
+ state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+ elif key.startswith("ip_adapter."):
+ state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+ else:
+ state_dict = torch.load(model_file, map_location="cpu")
+ else:
+ state_dict = pretrained_model_name_or_path_or_dict
+
+ keys = list(state_dict.keys())
+ if keys != ["image_proj", "ip_adapter"]:
+ raise ValueError("Required keys are (`image_proj` and `ip_adapter`) missing from the state dict.")
+
+ state_dicts.append(state_dict)
+
+ # load CLIP image encoder here if it has not been registered to the pipeline yet
+ if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is None:
+ if image_encoder_folder is not None:
+ if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+ logger.info(f"loading image_encoder from {pretrained_model_name_or_path_or_dict}")
+ if image_encoder_folder.count("/") == 0:
+ image_encoder_subfolder = Path(subfolder, image_encoder_folder).as_posix()
+ else:
+ image_encoder_subfolder = Path(image_encoder_folder).as_posix()
+
+ image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+ pretrained_model_name_or_path_or_dict,
+ subfolder=image_encoder_subfolder,
+ low_cpu_mem_usage=low_cpu_mem_usage,
+ ).to(self.device, dtype=self.dtype)
+ self.register_modules(image_encoder=image_encoder)
+ else:
+ raise ValueError(
+ "`image_encoder` cannot be loaded because `pretrained_model_name_or_path_or_dict` is a state dict."
+ )
+ else:
+ logger.warning(
+ "image_encoder is not loaded since `image_encoder_folder=None` passed. You will not be able to use `ip_adapter_image` when calling the pipeline with IP-Adapter."
+ "Use `ip_adapter_image_embeds` to pass pre-generated image embedding instead."
+ )
+
+ # create feature extractor if it has not been registered to the pipeline yet
+ if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is None:
+ feature_extractor = CLIPImageProcessor()
+ self.register_modules(feature_extractor=feature_extractor)
+
+ # load ip-adapter into unet
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ unet._load_ip_adapter_weights(state_dicts, low_cpu_mem_usage=low_cpu_mem_usage)
+
+ def set_ip_adapter_scale(self, scale):
+ """
+ Sets the conditioning scale between text and image.
+
+ Example:
+
+ ```py
+ pipeline.set_ip_adapter_scale(0.5)
+ ```
+ """
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ for attn_processor in unet.attn_processors.values():
+ if isinstance(attn_processor, (IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0)):
+ if not isinstance(scale, list):
+ scale = [scale] * len(attn_processor.scale)
+ if len(attn_processor.scale) != len(scale):
+ raise ValueError(
+ f"`scale` should be a list of same length as the number if ip-adapters "
+ f"Expected {len(attn_processor.scale)} but got {len(scale)}."
+ )
+ attn_processor.scale = scale
+
+ def unload_ip_adapter(self):
+ """
+ Unloads the IP Adapter weights
+
+ Examples:
+
+ ```python
+ >>> # Assuming `pipeline` is already loaded with the IP Adapter weights.
+ >>> pipeline.unload_ip_adapter()
+ >>> ...
+ ```
+ """
+ # remove CLIP image encoder
+ if hasattr(self, "image_encoder") and getattr(self, "image_encoder", None) is not None:
+ self.image_encoder = None
+ self.register_to_config(image_encoder=[None, None])
+
+ # remove feature extractor only when safety_checker is None as safety_checker uses
+ # the feature_extractor later
+ if not hasattr(self, "safety_checker"):
+ if hasattr(self, "feature_extractor") and getattr(self, "feature_extractor", None) is not None:
+ self.feature_extractor = None
+ self.register_to_config(feature_extractor=[None, None])
+
+ # remove hidden encoder
+ self.unet.encoder_hid_proj = None
+ self.config.encoder_hid_dim_type = None
+
+ # restore original Unet attention processors layers
+ self.unet.set_default_attn_processor()
diff --git a/model/diffusers_c/loaders/lora.py b/model/diffusers_c/loaders/lora.py
new file mode 100644
index 0000000000000000000000000000000000000000..964087e0e06a70839c86e17fd771a412c02fb4fa
--- /dev/null
+++ b/model/diffusers_c/loaders/lora.py
@@ -0,0 +1,1349 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import inspect
+import os
+from pathlib import Path
+from typing import Callable, Dict, List, Optional, Union
+
+import safetensors
+import torch
+from huggingface_hub import model_info
+from huggingface_hub.constants import HF_HUB_OFFLINE
+from huggingface_hub.utils import validate_hf_hub_args
+from packaging import version
+from torch import nn
+
+from .. import __version__
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
+from ..utils import (
+ USE_PEFT_BACKEND,
+ _get_model_file,
+ convert_state_dict_to_diffusers,
+ convert_state_dict_to_peft,
+ convert_unet_state_dict_to_peft,
+ delete_adapter_layers,
+ get_adapter_name,
+ get_peft_kwargs,
+ is_accelerate_available,
+ is_transformers_available,
+ logging,
+ recurse_remove_peft_layers,
+ scale_lora_layers,
+ set_adapter_layers,
+ set_weights_and_activate_adapters,
+)
+from .lora_conversion_utils import _convert_kohya_lora_to_diffusers, _maybe_map_sgm_blocks_to_diffusers
+
+
+if is_transformers_available():
+ from transformers import PreTrainedModel
+
+ from ..models.lora import text_encoder_attn_modules, text_encoder_mlp_modules
+
+if is_accelerate_available():
+ from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
+TRANSFORMER_NAME = "transformer"
+
+LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
+LORA_DEPRECATION_MESSAGE = "You are using an old version of LoRA backend. This will be deprecated in the next releases in favor of PEFT make sure to install the latest PEFT and transformers packages in the future."
+
+
+class LoraLoaderMixin:
+ r"""
+ Load LoRA layers into [`UNet2DConditionModel`] and
+ [`CLIPTextModel`](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel).
+ """
+
+ text_encoder_name = TEXT_ENCODER_NAME
+ unet_name = UNET_NAME
+ transformer_name = TRANSFORMER_NAME
+ num_fused_loras = 0
+
+ def load_lora_weights(
+ self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
+ ):
+ """
+ Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
+ `self.text_encoder`.
+
+ All kwargs are forwarded to `self.lora_state_dict`.
+
+ See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+
+ See [`~loaders.LoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is loaded into
+ `self.unet`.
+
+ See [`~loaders.LoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state dict is loaded
+ into `self.text_encoder`.
+
+ Parameters:
+ pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+ See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+ kwargs (`dict`, *optional*):
+ See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+ adapter_name (`str`, *optional*):
+ Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+ `default_{i}` where i is the total number of adapters being loaded.
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ # if a dict is passed, copy it instead of modifying it inplace
+ if isinstance(pretrained_model_name_or_path_or_dict, dict):
+ pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+
+ # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+ state_dict, network_alphas = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
+ is_correct_format = all("lora" in key for key in state_dict.keys())
+ if not is_correct_format:
+ raise ValueError("Invalid LoRA checkpoint.")
+
+ low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+
+ self.load_lora_into_unet(
+ state_dict,
+ network_alphas=network_alphas,
+ unet=getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet,
+ low_cpu_mem_usage=low_cpu_mem_usage,
+ adapter_name=adapter_name,
+ _pipeline=self,
+ )
+ self.load_lora_into_text_encoder(
+ state_dict,
+ network_alphas=network_alphas,
+ text_encoder=getattr(self, self.text_encoder_name)
+ if not hasattr(self, "text_encoder")
+ else self.text_encoder,
+ lora_scale=self.lora_scale,
+ low_cpu_mem_usage=low_cpu_mem_usage,
+ adapter_name=adapter_name,
+ _pipeline=self,
+ )
+
+ @classmethod
+ @validate_hf_hub_args
+ def lora_state_dict(
+ cls,
+ pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+ **kwargs,
+ ):
+ r"""
+ Return state dict for lora weights and the network alphas.
+
+
+
+ We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+
+ This function is experimental and might change in the future.
+
+
+
+ Parameters:
+ pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+ Can be either:
+
+ - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+ the Hub.
+ - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+ with [`ModelMixin.save_pretrained`].
+ - A [torch state
+ dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ subfolder (`str`, *optional*, defaults to `""`):
+ The subfolder location of a model file within a larger model repository on the Hub or locally.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+
+ """
+ # Load the main state dict first which has the LoRA layers for either of
+ # UNet and text encoder or both.
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ local_files_only = kwargs.pop("local_files_only", None)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ subfolder = kwargs.pop("subfolder", None)
+ weight_name = kwargs.pop("weight_name", None)
+ unet_config = kwargs.pop("unet_config", None)
+ use_safetensors = kwargs.pop("use_safetensors", None)
+
+ allow_pickle = False
+ if use_safetensors is None:
+ use_safetensors = True
+ allow_pickle = True
+
+ user_agent = {
+ "file_type": "attn_procs_weights",
+ "framework": "pytorch",
+ }
+
+ model_file = None
+ if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+ # Let's first try to load .safetensors weights
+ if (use_safetensors and weight_name is None) or (
+ weight_name is not None and weight_name.endswith(".safetensors")
+ ):
+ try:
+ # Here we're relaxing the loading check to enable more Inference API
+ # friendliness where sometimes, it's not at all possible to automatically
+ # determine `weight_name`.
+ if weight_name is None:
+ weight_name = cls._best_guess_weight_name(
+ pretrained_model_name_or_path_or_dict,
+ file_extension=".safetensors",
+ local_files_only=local_files_only,
+ )
+ model_file = _get_model_file(
+ pretrained_model_name_or_path_or_dict,
+ weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ )
+ state_dict = safetensors.torch.load_file(model_file, device="cpu")
+ except (IOError, safetensors.SafetensorError) as e:
+ if not allow_pickle:
+ raise e
+ # try loading non-safetensors weights
+ model_file = None
+ pass
+
+ if model_file is None:
+ if weight_name is None:
+ weight_name = cls._best_guess_weight_name(
+ pretrained_model_name_or_path_or_dict, file_extension=".bin", local_files_only=local_files_only
+ )
+ model_file = _get_model_file(
+ pretrained_model_name_or_path_or_dict,
+ weights_name=weight_name or LORA_WEIGHT_NAME,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ )
+ state_dict = torch.load(model_file, map_location="cpu")
+ else:
+ state_dict = pretrained_model_name_or_path_or_dict
+
+ network_alphas = None
+ # TODO: replace it with a method from `state_dict_utils`
+ if all(
+ (
+ k.startswith("lora_te_")
+ or k.startswith("lora_unet_")
+ or k.startswith("lora_te1_")
+ or k.startswith("lora_te2_")
+ )
+ for k in state_dict.keys()
+ ):
+ # Map SDXL blocks correctly.
+ if unet_config is not None:
+ # use unet config to remap block numbers
+ state_dict = _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config)
+ state_dict, network_alphas = _convert_kohya_lora_to_diffusers(state_dict)
+
+ return state_dict, network_alphas
+
+ @classmethod
+ def _best_guess_weight_name(
+ cls, pretrained_model_name_or_path_or_dict, file_extension=".safetensors", local_files_only=False
+ ):
+ if local_files_only or HF_HUB_OFFLINE:
+ raise ValueError("When using the offline mode, you must specify a `weight_name`.")
+
+ targeted_files = []
+
+ if os.path.isfile(pretrained_model_name_or_path_or_dict):
+ return
+ elif os.path.isdir(pretrained_model_name_or_path_or_dict):
+ targeted_files = [
+ f for f in os.listdir(pretrained_model_name_or_path_or_dict) if f.endswith(file_extension)
+ ]
+ else:
+ files_in_repo = model_info(pretrained_model_name_or_path_or_dict).siblings
+ targeted_files = [f.rfilename for f in files_in_repo if f.rfilename.endswith(file_extension)]
+ if len(targeted_files) == 0:
+ return
+
+ # "scheduler" does not correspond to a LoRA checkpoint.
+ # "optimizer" does not correspond to a LoRA checkpoint
+ # only top-level checkpoints are considered and not the other ones, hence "checkpoint".
+ unallowed_substrings = {"scheduler", "optimizer", "checkpoint"}
+ targeted_files = list(
+ filter(lambda x: all(substring not in x for substring in unallowed_substrings), targeted_files)
+ )
+
+ if any(f.endswith(LORA_WEIGHT_NAME) for f in targeted_files):
+ targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME), targeted_files))
+ elif any(f.endswith(LORA_WEIGHT_NAME_SAFE) for f in targeted_files):
+ targeted_files = list(filter(lambda x: x.endswith(LORA_WEIGHT_NAME_SAFE), targeted_files))
+
+ if len(targeted_files) > 1:
+ raise ValueError(
+ f"Provided path contains more than one weights file in the {file_extension} format. Either specify `weight_name` in `load_lora_weights` or make sure there's only one `.safetensors` or `.bin` file in {pretrained_model_name_or_path_or_dict}."
+ )
+ weight_name = targeted_files[0]
+ return weight_name
+
+ @classmethod
+ def _optionally_disable_offloading(cls, _pipeline):
+ """
+ Optionally removes offloading in case the pipeline has been already sequentially offloaded to CPU.
+
+ Args:
+ _pipeline (`DiffusionPipeline`):
+ The pipeline to disable offloading for.
+
+ Returns:
+ tuple:
+ A tuple indicating if `is_model_cpu_offload` or `is_sequential_cpu_offload` is True.
+ """
+ is_model_cpu_offload = False
+ is_sequential_cpu_offload = False
+
+ if _pipeline is not None:
+ for _, component in _pipeline.components.items():
+ if isinstance(component, nn.Module) and hasattr(component, "_hf_hook"):
+ if not is_model_cpu_offload:
+ is_model_cpu_offload = isinstance(component._hf_hook, CpuOffload)
+ if not is_sequential_cpu_offload:
+ is_sequential_cpu_offload = isinstance(component._hf_hook, AlignDevicesHook)
+
+ logger.info(
+ "Accelerate hooks detected. Since you have called `load_lora_weights()`, the previous hooks will be first removed. Then the LoRA parameters will be loaded and the hooks will be applied again."
+ )
+ remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+ return (is_model_cpu_offload, is_sequential_cpu_offload)
+
+ @classmethod
+ def load_lora_into_unet(
+ cls, state_dict, network_alphas, unet, low_cpu_mem_usage=None, adapter_name=None, _pipeline=None
+ ):
+ """
+ This will load the LoRA layers specified in `state_dict` into `unet`.
+
+ Parameters:
+ state_dict (`dict`):
+ A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+ into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+ encoder lora layers.
+ network_alphas (`Dict[str, float]`):
+ See `LoRALinearLayer` for more details.
+ unet (`UNet2DConditionModel`):
+ The UNet model to load the LoRA layers into.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ adapter_name (`str`, *optional*):
+ Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+ `default_{i}` where i is the total number of adapters being loaded.
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+
+ low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
+ # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
+ # then the `state_dict` keys should have `cls.unet_name` and/or `cls.text_encoder_name` as
+ # their prefixes.
+ keys = list(state_dict.keys())
+
+ if all(key.startswith(cls.unet_name) or key.startswith(cls.text_encoder_name) for key in keys):
+ # Load the layers corresponding to UNet.
+ logger.info(f"Loading {cls.unet_name}.")
+
+ unet_keys = [k for k in keys if k.startswith(cls.unet_name)]
+ state_dict = {k.replace(f"{cls.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
+
+ if network_alphas is not None:
+ alpha_keys = [k for k in network_alphas.keys() if k.startswith(cls.unet_name)]
+ network_alphas = {
+ k.replace(f"{cls.unet_name}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+ }
+
+ else:
+ # Otherwise, we're dealing with the old format. This means the `state_dict` should only
+ # contain the module names of the `unet` as its keys WITHOUT any prefix.
+ if not USE_PEFT_BACKEND:
+ warn_message = "You have saved the LoRA weights using the old format. To convert the old LoRA weights to the new format, you can first load them in a dictionary and then create a new dictionary like the following: `new_state_dict = {f'unet.{module_name}': params for module_name, params in old_state_dict.items()}`."
+ logger.warn(warn_message)
+
+ if len(state_dict.keys()) > 0:
+ if adapter_name in getattr(unet, "peft_config", {}):
+ raise ValueError(
+ f"Adapter name {adapter_name} already in use in the Unet - please select a new adapter name."
+ )
+
+ state_dict = convert_unet_state_dict_to_peft(state_dict)
+
+ if network_alphas is not None:
+ # The alphas state dict have the same structure as Unet, thus we convert it to peft format using
+ # `convert_unet_state_dict_to_peft` method.
+ network_alphas = convert_unet_state_dict_to_peft(network_alphas)
+
+ rank = {}
+ for key, val in state_dict.items():
+ if "lora_B" in key:
+ rank[key] = val.shape[1]
+
+ lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict, is_unet=True)
+ lora_config = LoraConfig(**lora_config_kwargs)
+
+ # adapter_name
+ if adapter_name is None:
+ adapter_name = get_adapter_name(unet)
+
+ # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
+ # otherwise loading LoRA weights will lead to an error
+ is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
+
+ inject_adapter_in_model(lora_config, unet, adapter_name=adapter_name)
+ incompatible_keys = set_peft_model_state_dict(unet, state_dict, adapter_name)
+
+ if incompatible_keys is not None:
+ # check only for unexpected keys
+ unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+ if unexpected_keys:
+ logger.warning(
+ f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+ f" {unexpected_keys}. "
+ )
+
+ # Offload back.
+ if is_model_cpu_offload:
+ _pipeline.enable_model_cpu_offload()
+ elif is_sequential_cpu_offload:
+ _pipeline.enable_sequential_cpu_offload()
+ # Unsafe code />
+
+ unet.load_attn_procs(
+ state_dict, network_alphas=network_alphas, low_cpu_mem_usage=low_cpu_mem_usage, _pipeline=_pipeline
+ )
+
+ @classmethod
+ def load_lora_into_text_encoder(
+ cls,
+ state_dict,
+ network_alphas,
+ text_encoder,
+ prefix=None,
+ lora_scale=1.0,
+ low_cpu_mem_usage=None,
+ adapter_name=None,
+ _pipeline=None,
+ ):
+ """
+ This will load the LoRA layers specified in `state_dict` into `text_encoder`
+
+ Parameters:
+ state_dict (`dict`):
+ A standard state dict containing the lora layer parameters. The key should be prefixed with an
+ additional `text_encoder` to distinguish between unet lora layers.
+ network_alphas (`Dict[str, float]`):
+ See `LoRALinearLayer` for more details.
+ text_encoder (`CLIPTextModel`):
+ The text encoder model to load the LoRA layers into.
+ prefix (`str`):
+ Expected prefix of the `text_encoder` in the `state_dict`.
+ lora_scale (`float`):
+ How much to scale the output of the lora linear layer before it is added with the output of the regular
+ lora layer.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ adapter_name (`str`, *optional*):
+ Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+ `default_{i}` where i is the total number of adapters being loaded.
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ from peft import LoraConfig
+
+ low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
+
+ # If the serialization format is new (introduced in https://github.com/huggingface/diffusers/pull/2918),
+ # then the `state_dict` keys should have `self.unet_name` and/or `self.text_encoder_name` as
+ # their prefixes.
+ keys = list(state_dict.keys())
+ prefix = cls.text_encoder_name if prefix is None else prefix
+
+ # Safe prefix to check with.
+ if any(cls.text_encoder_name in key for key in keys):
+ # Load the layers corresponding to text encoder and make necessary adjustments.
+ text_encoder_keys = [k for k in keys if k.startswith(prefix) and k.split(".")[0] == prefix]
+ text_encoder_lora_state_dict = {
+ k.replace(f"{prefix}.", ""): v for k, v in state_dict.items() if k in text_encoder_keys
+ }
+
+ if len(text_encoder_lora_state_dict) > 0:
+ logger.info(f"Loading {prefix}.")
+ rank = {}
+ text_encoder_lora_state_dict = convert_state_dict_to_diffusers(text_encoder_lora_state_dict)
+
+ # convert state dict
+ text_encoder_lora_state_dict = convert_state_dict_to_peft(text_encoder_lora_state_dict)
+
+ for name, _ in text_encoder_attn_modules(text_encoder):
+ rank_key = f"{name}.out_proj.lora_B.weight"
+ rank[rank_key] = text_encoder_lora_state_dict[rank_key].shape[1]
+
+ patch_mlp = any(".mlp." in key for key in text_encoder_lora_state_dict.keys())
+ if patch_mlp:
+ for name, _ in text_encoder_mlp_modules(text_encoder):
+ rank_key_fc1 = f"{name}.fc1.lora_B.weight"
+ rank_key_fc2 = f"{name}.fc2.lora_B.weight"
+
+ rank[rank_key_fc1] = text_encoder_lora_state_dict[rank_key_fc1].shape[1]
+ rank[rank_key_fc2] = text_encoder_lora_state_dict[rank_key_fc2].shape[1]
+
+ if network_alphas is not None:
+ alpha_keys = [
+ k for k in network_alphas.keys() if k.startswith(prefix) and k.split(".")[0] == prefix
+ ]
+ network_alphas = {
+ k.replace(f"{prefix}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+ }
+
+ lora_config_kwargs = get_peft_kwargs(rank, network_alphas, text_encoder_lora_state_dict, is_unet=False)
+ lora_config = LoraConfig(**lora_config_kwargs)
+
+ # adapter_name
+ if adapter_name is None:
+ adapter_name = get_adapter_name(text_encoder)
+
+ is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
+
+ # inject LoRA layers and load the state dict
+ # in transformers we automatically check whether the adapter name is already in use or not
+ text_encoder.load_adapter(
+ adapter_name=adapter_name,
+ adapter_state_dict=text_encoder_lora_state_dict,
+ peft_config=lora_config,
+ )
+
+ # scale LoRA layers with `lora_scale`
+ scale_lora_layers(text_encoder, weight=lora_scale)
+
+ text_encoder.to(device=text_encoder.device, dtype=text_encoder.dtype)
+
+ # Offload back.
+ if is_model_cpu_offload:
+ _pipeline.enable_model_cpu_offload()
+ elif is_sequential_cpu_offload:
+ _pipeline.enable_sequential_cpu_offload()
+ # Unsafe code />
+
+ @classmethod
+ def load_lora_into_transformer(
+ cls, state_dict, network_alphas, transformer, low_cpu_mem_usage=None, adapter_name=None, _pipeline=None
+ ):
+ """
+ This will load the LoRA layers specified in `state_dict` into `transformer`.
+
+ Parameters:
+ state_dict (`dict`):
+ A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+ into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+ encoder lora layers.
+ network_alphas (`Dict[str, float]`):
+ See `LoRALinearLayer` for more details.
+ unet (`UNet2DConditionModel`):
+ The UNet model to load the LoRA layers into.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ adapter_name (`str`, *optional*):
+ Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+ `default_{i}` where i is the total number of adapters being loaded.
+ """
+ from peft import LoraConfig, inject_adapter_in_model, set_peft_model_state_dict
+
+ low_cpu_mem_usage = low_cpu_mem_usage if low_cpu_mem_usage is not None else _LOW_CPU_MEM_USAGE_DEFAULT
+
+ keys = list(state_dict.keys())
+
+ transformer_keys = [k for k in keys if k.startswith(cls.transformer_name)]
+ state_dict = {
+ k.replace(f"{cls.transformer_name}.", ""): v for k, v in state_dict.items() if k in transformer_keys
+ }
+
+ if network_alphas is not None:
+ alpha_keys = [k for k in network_alphas.keys() if k.startswith(cls.transformer_name)]
+ network_alphas = {
+ k.replace(f"{cls.transformer_name}.", ""): v for k, v in network_alphas.items() if k in alpha_keys
+ }
+
+ if len(state_dict.keys()) > 0:
+ if adapter_name in getattr(transformer, "peft_config", {}):
+ raise ValueError(
+ f"Adapter name {adapter_name} already in use in the transformer - please select a new adapter name."
+ )
+
+ rank = {}
+ for key, val in state_dict.items():
+ if "lora_B" in key:
+ rank[key] = val.shape[1]
+
+ lora_config_kwargs = get_peft_kwargs(rank, network_alphas, state_dict)
+ lora_config = LoraConfig(**lora_config_kwargs)
+
+ # adapter_name
+ if adapter_name is None:
+ adapter_name = get_adapter_name(transformer)
+
+ # In case the pipeline has been already offloaded to CPU - temporarily remove the hooks
+ # otherwise loading LoRA weights will lead to an error
+ is_model_cpu_offload, is_sequential_cpu_offload = cls._optionally_disable_offloading(_pipeline)
+
+ inject_adapter_in_model(lora_config, transformer, adapter_name=adapter_name)
+ incompatible_keys = set_peft_model_state_dict(transformer, state_dict, adapter_name)
+
+ if incompatible_keys is not None:
+ # check only for unexpected keys
+ unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+ if unexpected_keys:
+ logger.warning(
+ f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+ f" {unexpected_keys}. "
+ )
+
+ # Offload back.
+ if is_model_cpu_offload:
+ _pipeline.enable_model_cpu_offload()
+ elif is_sequential_cpu_offload:
+ _pipeline.enable_sequential_cpu_offload()
+ # Unsafe code />
+
+ @property
+ def lora_scale(self) -> float:
+ # property function that returns the lora scale which can be set at run time by the pipeline.
+ # if _lora_scale has not been set, return 1
+ return self._lora_scale if hasattr(self, "_lora_scale") else 1.0
+
+ def _remove_text_encoder_monkey_patch(self):
+ remove_method = recurse_remove_peft_layers
+ if hasattr(self, "text_encoder"):
+ remove_method(self.text_encoder)
+ # In case text encoder have no Lora attached
+ if getattr(self.text_encoder, "peft_config", None) is not None:
+ del self.text_encoder.peft_config
+ self.text_encoder._hf_peft_config_loaded = None
+
+ if hasattr(self, "text_encoder_2"):
+ remove_method(self.text_encoder_2)
+ if getattr(self.text_encoder_2, "peft_config", None) is not None:
+ del self.text_encoder_2.peft_config
+ self.text_encoder_2._hf_peft_config_loaded = None
+
+ @classmethod
+ def save_lora_weights(
+ cls,
+ save_directory: Union[str, os.PathLike],
+ unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+ text_encoder_lora_layers: Dict[str, torch.nn.Module] = None,
+ transformer_lora_layers: Dict[str, torch.nn.Module] = None,
+ is_main_process: bool = True,
+ weight_name: str = None,
+ save_function: Callable = None,
+ safe_serialization: bool = True,
+ ):
+ r"""
+ Save the LoRA parameters corresponding to the UNet and text encoder.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to save LoRA parameters to. Will be created if it doesn't exist.
+ unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+ State dict of the LoRA layers corresponding to the `unet`.
+ text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+ State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
+ encoder LoRA state dict because it comes from 🤗 Transformers.
+ is_main_process (`bool`, *optional*, defaults to `True`):
+ Whether the process calling this is the main process or not. Useful during distributed training and you
+ need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+ process to avoid race conditions.
+ save_function (`Callable`):
+ The function to use to save the state dictionary. Useful during distributed training when you need to
+ replace `torch.save` with another method. Can be configured with the environment variable
+ `DIFFUSERS_SAVE_MODE`.
+ safe_serialization (`bool`, *optional*, defaults to `True`):
+ Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+ """
+ state_dict = {}
+
+ def pack_weights(layers, prefix):
+ layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
+ layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
+ return layers_state_dict
+
+ if not (unet_lora_layers or text_encoder_lora_layers or transformer_lora_layers):
+ raise ValueError(
+ "You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers`, or `transformer_lora_layers`."
+ )
+
+ if unet_lora_layers:
+ state_dict.update(pack_weights(unet_lora_layers, cls.unet_name))
+
+ if text_encoder_lora_layers:
+ state_dict.update(pack_weights(text_encoder_lora_layers, cls.text_encoder_name))
+
+ if transformer_lora_layers:
+ state_dict.update(pack_weights(transformer_lora_layers, "transformer"))
+
+ # Save the model
+ cls.write_lora_layers(
+ state_dict=state_dict,
+ save_directory=save_directory,
+ is_main_process=is_main_process,
+ weight_name=weight_name,
+ save_function=save_function,
+ safe_serialization=safe_serialization,
+ )
+
+ @staticmethod
+ def write_lora_layers(
+ state_dict: Dict[str, torch.Tensor],
+ save_directory: str,
+ is_main_process: bool,
+ weight_name: str,
+ save_function: Callable,
+ safe_serialization: bool,
+ ):
+ if os.path.isfile(save_directory):
+ logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+ return
+
+ if save_function is None:
+ if safe_serialization:
+
+ def save_function(weights, filename):
+ return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
+
+ else:
+ save_function = torch.save
+
+ os.makedirs(save_directory, exist_ok=True)
+
+ if weight_name is None:
+ if safe_serialization:
+ weight_name = LORA_WEIGHT_NAME_SAFE
+ else:
+ weight_name = LORA_WEIGHT_NAME
+
+ save_path = Path(save_directory, weight_name).as_posix()
+ save_function(state_dict, save_path)
+ logger.info(f"Model weights saved in {save_path}")
+
+ def unload_lora_weights(self):
+ """
+ Unloads the LoRA parameters.
+
+ Examples:
+
+ ```python
+ >>> # Assuming `pipeline` is already loaded with the LoRA parameters.
+ >>> pipeline.unload_lora_weights()
+ >>> ...
+ ```
+ """
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+
+ if not USE_PEFT_BACKEND:
+ if version.parse(__version__) > version.parse("0.23"):
+ logger.warning(
+ "You are using `unload_lora_weights` to disable and unload lora weights. If you want to iteratively enable and disable adapter weights,"
+ "you can use `pipe.enable_lora()` or `pipe.disable_lora()`. After installing the latest version of PEFT."
+ )
+
+ for _, module in unet.named_modules():
+ if hasattr(module, "set_lora_layer"):
+ module.set_lora_layer(None)
+ else:
+ recurse_remove_peft_layers(unet)
+ if hasattr(unet, "peft_config"):
+ del unet.peft_config
+
+ # Safe to call the following regardless of LoRA.
+ self._remove_text_encoder_monkey_patch()
+
+ def fuse_lora(
+ self,
+ fuse_unet: bool = True,
+ fuse_text_encoder: bool = True,
+ lora_scale: float = 1.0,
+ safe_fusing: bool = False,
+ adapter_names: Optional[List[str]] = None,
+ ):
+ r"""
+ Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+
+
+
+ This is an experimental API.
+
+
+
+ Args:
+ fuse_unet (`bool`, defaults to `True`): Whether to fuse the UNet LoRA parameters.
+ fuse_text_encoder (`bool`, defaults to `True`):
+ Whether to fuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+ LoRA parameters then it won't have any effect.
+ lora_scale (`float`, defaults to 1.0):
+ Controls how much to influence the outputs with the LoRA parameters.
+ safe_fusing (`bool`, defaults to `False`):
+ Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
+ adapter_names (`List[str]`, *optional*):
+ Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
+
+ Example:
+
+ ```py
+ from diffusers import DiffusionPipeline
+ import torch
+
+ pipeline = DiffusionPipeline.from_pretrained(
+ "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+ ).to("cuda")
+ pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+ pipeline.fuse_lora(lora_scale=0.7)
+ ```
+ """
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ if fuse_unet or fuse_text_encoder:
+ self.num_fused_loras += 1
+ if self.num_fused_loras > 1:
+ logger.warn(
+ "The current API is supported for operating with a single LoRA file. You are trying to load and fuse more than one LoRA which is not well-supported.",
+ )
+
+ if fuse_unet:
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ unet.fuse_lora(lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names)
+
+ def fuse_text_encoder_lora(text_encoder, lora_scale=1.0, safe_fusing=False, adapter_names=None):
+ merge_kwargs = {"safe_merge": safe_fusing}
+
+ for module in text_encoder.modules():
+ if isinstance(module, BaseTunerLayer):
+ if lora_scale != 1.0:
+ module.scale_layer(lora_scale)
+
+ # For BC with previous PEFT versions, we need to check the signature
+ # of the `merge` method to see if it supports the `adapter_names` argument.
+ supported_merge_kwargs = list(inspect.signature(module.merge).parameters)
+ if "adapter_names" in supported_merge_kwargs:
+ merge_kwargs["adapter_names"] = adapter_names
+ elif "adapter_names" not in supported_merge_kwargs and adapter_names is not None:
+ raise ValueError(
+ "The `adapter_names` argument is not supported with your PEFT version. "
+ "Please upgrade to the latest version of PEFT. `pip install -U peft`"
+ )
+
+ module.merge(**merge_kwargs)
+
+ if fuse_text_encoder:
+ if hasattr(self, "text_encoder"):
+ fuse_text_encoder_lora(self.text_encoder, lora_scale, safe_fusing, adapter_names=adapter_names)
+ if hasattr(self, "text_encoder_2"):
+ fuse_text_encoder_lora(self.text_encoder_2, lora_scale, safe_fusing, adapter_names=adapter_names)
+
+ def unfuse_lora(self, unfuse_unet: bool = True, unfuse_text_encoder: bool = True):
+ r"""
+ Reverses the effect of
+ [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.fuse_lora).
+
+
+
+ This is an experimental API.
+
+
+
+ Args:
+ unfuse_unet (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
+ unfuse_text_encoder (`bool`, defaults to `True`):
+ Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+ LoRA parameters then it won't have any effect.
+ """
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ if unfuse_unet:
+ for module in unet.modules():
+ if isinstance(module, BaseTunerLayer):
+ module.unmerge()
+
+ def unfuse_text_encoder_lora(text_encoder):
+ for module in text_encoder.modules():
+ if isinstance(module, BaseTunerLayer):
+ module.unmerge()
+
+ if unfuse_text_encoder:
+ if hasattr(self, "text_encoder"):
+ unfuse_text_encoder_lora(self.text_encoder)
+ if hasattr(self, "text_encoder_2"):
+ unfuse_text_encoder_lora(self.text_encoder_2)
+
+ self.num_fused_loras -= 1
+
+ def set_adapters_for_text_encoder(
+ self,
+ adapter_names: Union[List[str], str],
+ text_encoder: Optional["PreTrainedModel"] = None, # noqa: F821
+ text_encoder_weights: List[float] = None,
+ ):
+ """
+ Sets the adapter layers for the text encoder.
+
+ Args:
+ adapter_names (`List[str]` or `str`):
+ The names of the adapters to use.
+ text_encoder (`torch.nn.Module`, *optional*):
+ The text encoder module to set the adapter layers for. If `None`, it will try to get the `text_encoder`
+ attribute.
+ text_encoder_weights (`List[float]`, *optional*):
+ The weights to use for the text encoder. If `None`, the weights are set to `1.0` for all the adapters.
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ def process_weights(adapter_names, weights):
+ if weights is None:
+ weights = [1.0] * len(adapter_names)
+ elif isinstance(weights, float):
+ weights = [weights]
+
+ if len(adapter_names) != len(weights):
+ raise ValueError(
+ f"Length of adapter names {len(adapter_names)} is not equal to the length of the weights {len(weights)}"
+ )
+ return weights
+
+ adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+ text_encoder_weights = process_weights(adapter_names, text_encoder_weights)
+ text_encoder = text_encoder or getattr(self, "text_encoder", None)
+ if text_encoder is None:
+ raise ValueError(
+ "The pipeline does not have a default `pipe.text_encoder` class. Please make sure to pass a `text_encoder` instead."
+ )
+ set_weights_and_activate_adapters(text_encoder, adapter_names, text_encoder_weights)
+
+ def disable_lora_for_text_encoder(self, text_encoder: Optional["PreTrainedModel"] = None):
+ """
+ Disables the LoRA layers for the text encoder.
+
+ Args:
+ text_encoder (`torch.nn.Module`, *optional*):
+ The text encoder module to disable the LoRA layers for. If `None`, it will try to get the
+ `text_encoder` attribute.
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ text_encoder = text_encoder or getattr(self, "text_encoder", None)
+ if text_encoder is None:
+ raise ValueError("Text Encoder not found.")
+ set_adapter_layers(text_encoder, enabled=False)
+
+ def enable_lora_for_text_encoder(self, text_encoder: Optional["PreTrainedModel"] = None):
+ """
+ Enables the LoRA layers for the text encoder.
+
+ Args:
+ text_encoder (`torch.nn.Module`, *optional*):
+ The text encoder module to enable the LoRA layers for. If `None`, it will try to get the `text_encoder`
+ attribute.
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+ text_encoder = text_encoder or getattr(self, "text_encoder", None)
+ if text_encoder is None:
+ raise ValueError("Text Encoder not found.")
+ set_adapter_layers(self.text_encoder, enabled=True)
+
+ def set_adapters(
+ self,
+ adapter_names: Union[List[str], str],
+ adapter_weights: Optional[List[float]] = None,
+ ):
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ # Handle the UNET
+ unet.set_adapters(adapter_names, adapter_weights)
+
+ # Handle the Text Encoder
+ if hasattr(self, "text_encoder"):
+ self.set_adapters_for_text_encoder(adapter_names, self.text_encoder, adapter_weights)
+ if hasattr(self, "text_encoder_2"):
+ self.set_adapters_for_text_encoder(adapter_names, self.text_encoder_2, adapter_weights)
+
+ def disable_lora(self):
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ # Disable unet adapters
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ unet.disable_lora()
+
+ # Disable text encoder adapters
+ if hasattr(self, "text_encoder"):
+ self.disable_lora_for_text_encoder(self.text_encoder)
+ if hasattr(self, "text_encoder_2"):
+ self.disable_lora_for_text_encoder(self.text_encoder_2)
+
+ def enable_lora(self):
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ # Enable unet adapters
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ unet.enable_lora()
+
+ # Enable text encoder adapters
+ if hasattr(self, "text_encoder"):
+ self.enable_lora_for_text_encoder(self.text_encoder)
+ if hasattr(self, "text_encoder_2"):
+ self.enable_lora_for_text_encoder(self.text_encoder_2)
+
+ def delete_adapters(self, adapter_names: Union[List[str], str]):
+ """
+ Args:
+ Deletes the LoRA layers of `adapter_name` for the unet and text-encoder(s).
+ adapter_names (`Union[List[str], str]`):
+ The names of the adapter to delete. Can be a single string or a list of strings
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ if isinstance(adapter_names, str):
+ adapter_names = [adapter_names]
+
+ # Delete unet adapters
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ unet.delete_adapters(adapter_names)
+
+ for adapter_name in adapter_names:
+ # Delete text encoder adapters
+ if hasattr(self, "text_encoder"):
+ delete_adapter_layers(self.text_encoder, adapter_name)
+ if hasattr(self, "text_encoder_2"):
+ delete_adapter_layers(self.text_encoder_2, adapter_name)
+
+ def get_active_adapters(self) -> List[str]:
+ """
+ Gets the list of the current active adapters.
+
+ Example:
+
+ ```python
+ from diffusers import DiffusionPipeline
+
+ pipeline = DiffusionPipeline.from_pretrained(
+ "stabilityai/stable-diffusion-xl-base-1.0",
+ ).to("cuda")
+ pipeline.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
+ pipeline.get_active_adapters()
+ ```
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError(
+ "PEFT backend is required for this method. Please install the latest version of PEFT `pip install -U peft`"
+ )
+
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ active_adapters = []
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ for module in unet.modules():
+ if isinstance(module, BaseTunerLayer):
+ active_adapters = module.active_adapters
+ break
+
+ return active_adapters
+
+ def get_list_adapters(self) -> Dict[str, List[str]]:
+ """
+ Gets the current list of all available adapters in the pipeline.
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError(
+ "PEFT backend is required for this method. Please install the latest version of PEFT `pip install -U peft`"
+ )
+
+ set_adapters = {}
+
+ if hasattr(self, "text_encoder") and hasattr(self.text_encoder, "peft_config"):
+ set_adapters["text_encoder"] = list(self.text_encoder.peft_config.keys())
+
+ if hasattr(self, "text_encoder_2") and hasattr(self.text_encoder_2, "peft_config"):
+ set_adapters["text_encoder_2"] = list(self.text_encoder_2.peft_config.keys())
+
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ if hasattr(self, self.unet_name) and hasattr(unet, "peft_config"):
+ set_adapters[self.unet_name] = list(self.unet.peft_config.keys())
+
+ return set_adapters
+
+ def set_lora_device(self, adapter_names: List[str], device: Union[torch.device, str, int]) -> None:
+ """
+ Moves the LoRAs listed in `adapter_names` to a target device. Useful for offloading the LoRA to the CPU in case
+ you want to load multiple adapters and free some GPU memory.
+
+ Args:
+ adapter_names (`List[str]`):
+ List of adapters to send device to.
+ device (`Union[torch.device, str, int]`):
+ Device to send the adapters to. Can be either a torch device, a str or an integer.
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ # Handle the UNET
+ unet = getattr(self, self.unet_name) if not hasattr(self, "unet") else self.unet
+ for unet_module in unet.modules():
+ if isinstance(unet_module, BaseTunerLayer):
+ for adapter_name in adapter_names:
+ unet_module.lora_A[adapter_name].to(device)
+ unet_module.lora_B[adapter_name].to(device)
+
+ # Handle the text encoder
+ modules_to_process = []
+ if hasattr(self, "text_encoder"):
+ modules_to_process.append(self.text_encoder)
+
+ if hasattr(self, "text_encoder_2"):
+ modules_to_process.append(self.text_encoder_2)
+
+ for text_encoder in modules_to_process:
+ # loop over submodules
+ for text_encoder_module in text_encoder.modules():
+ if isinstance(text_encoder_module, BaseTunerLayer):
+ for adapter_name in adapter_names:
+ text_encoder_module.lora_A[adapter_name].to(device)
+ text_encoder_module.lora_B[adapter_name].to(device)
+
+
+class StableDiffusionXLLoraLoaderMixin(LoraLoaderMixin):
+ """This class overrides `LoraLoaderMixin` with LoRA loading/saving code that's specific to SDXL"""
+
+ # Override to properly handle the loading and unloading of the additional text encoder.
+ def load_lora_weights(
+ self,
+ pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+ adapter_name: Optional[str] = None,
+ **kwargs,
+ ):
+ """
+ Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.unet` and
+ `self.text_encoder`.
+
+ All kwargs are forwarded to `self.lora_state_dict`.
+
+ See [`~loaders.LoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+
+ See [`~loaders.LoraLoaderMixin.load_lora_into_unet`] for more details on how the state dict is loaded into
+ `self.unet`.
+
+ See [`~loaders.LoraLoaderMixin.load_lora_into_text_encoder`] for more details on how the state dict is loaded
+ into `self.text_encoder`.
+
+ Parameters:
+ pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+ See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+ adapter_name (`str`, *optional*):
+ Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+ `default_{i}` where i is the total number of adapters being loaded.
+ kwargs (`dict`, *optional*):
+ See [`~loaders.LoraLoaderMixin.lora_state_dict`].
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ # We could have accessed the unet config from `lora_state_dict()` too. We pass
+ # it here explicitly to be able to tell that it's coming from an SDXL
+ # pipeline.
+
+ # if a dict is passed, copy it instead of modifying it inplace
+ if isinstance(pretrained_model_name_or_path_or_dict, dict):
+ pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+
+ # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+ state_dict, network_alphas = self.lora_state_dict(
+ pretrained_model_name_or_path_or_dict,
+ unet_config=self.unet.config,
+ **kwargs,
+ )
+ is_correct_format = all("lora" in key for key in state_dict.keys())
+ if not is_correct_format:
+ raise ValueError("Invalid LoRA checkpoint.")
+
+ self.load_lora_into_unet(
+ state_dict, network_alphas=network_alphas, unet=self.unet, adapter_name=adapter_name, _pipeline=self
+ )
+ text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k}
+ if len(text_encoder_state_dict) > 0:
+ self.load_lora_into_text_encoder(
+ text_encoder_state_dict,
+ network_alphas=network_alphas,
+ text_encoder=self.text_encoder,
+ prefix="text_encoder",
+ lora_scale=self.lora_scale,
+ adapter_name=adapter_name,
+ _pipeline=self,
+ )
+
+ text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k}
+ if len(text_encoder_2_state_dict) > 0:
+ self.load_lora_into_text_encoder(
+ text_encoder_2_state_dict,
+ network_alphas=network_alphas,
+ text_encoder=self.text_encoder_2,
+ prefix="text_encoder_2",
+ lora_scale=self.lora_scale,
+ adapter_name=adapter_name,
+ _pipeline=self,
+ )
+
+ @classmethod
+ def save_lora_weights(
+ cls,
+ save_directory: Union[str, os.PathLike],
+ unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+ text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+ text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+ is_main_process: bool = True,
+ weight_name: str = None,
+ save_function: Callable = None,
+ safe_serialization: bool = True,
+ ):
+ r"""
+ Save the LoRA parameters corresponding to the UNet and text encoder.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to save LoRA parameters to. Will be created if it doesn't exist.
+ unet_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+ State dict of the LoRA layers corresponding to the `unet`.
+ text_encoder_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+ State dict of the LoRA layers corresponding to the `text_encoder`. Must explicitly pass the text
+ encoder LoRA state dict because it comes from 🤗 Transformers.
+ is_main_process (`bool`, *optional*, defaults to `True`):
+ Whether the process calling this is the main process or not. Useful during distributed training and you
+ need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+ process to avoid race conditions.
+ save_function (`Callable`):
+ The function to use to save the state dictionary. Useful during distributed training when you need to
+ replace `torch.save` with another method. Can be configured with the environment variable
+ `DIFFUSERS_SAVE_MODE`.
+ safe_serialization (`bool`, *optional*, defaults to `True`):
+ Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+ """
+ state_dict = {}
+
+ def pack_weights(layers, prefix):
+ layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers
+ layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()}
+ return layers_state_dict
+
+ if not (unet_lora_layers or text_encoder_lora_layers or text_encoder_2_lora_layers):
+ raise ValueError(
+ "You must pass at least one of `unet_lora_layers`, `text_encoder_lora_layers` or `text_encoder_2_lora_layers`."
+ )
+
+ if unet_lora_layers:
+ state_dict.update(pack_weights(unet_lora_layers, "unet"))
+
+ if text_encoder_lora_layers and text_encoder_2_lora_layers:
+ state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder"))
+ state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2"))
+
+ cls.write_lora_layers(
+ state_dict=state_dict,
+ save_directory=save_directory,
+ is_main_process=is_main_process,
+ weight_name=weight_name,
+ save_function=save_function,
+ safe_serialization=safe_serialization,
+ )
+
+ def _remove_text_encoder_monkey_patch(self):
+ recurse_remove_peft_layers(self.text_encoder)
+ # TODO: @younesbelkada handle this in transformers side
+ if getattr(self.text_encoder, "peft_config", None) is not None:
+ del self.text_encoder.peft_config
+ self.text_encoder._hf_peft_config_loaded = None
+
+ recurse_remove_peft_layers(self.text_encoder_2)
+ if getattr(self.text_encoder_2, "peft_config", None) is not None:
+ del self.text_encoder_2.peft_config
+ self.text_encoder_2._hf_peft_config_loaded = None
diff --git a/model/diffusers_c/loaders/lora_conversion_utils.py b/model/diffusers_c/loaders/lora_conversion_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..e968ef9628c2800cfa2ef6268b94cf7ca62089fe
--- /dev/null
+++ b/model/diffusers_c/loaders/lora_conversion_utils.py
@@ -0,0 +1,284 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import re
+
+from ..utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+def _maybe_map_sgm_blocks_to_diffusers(state_dict, unet_config, delimiter="_", block_slice_pos=5):
+ # 1. get all state_dict_keys
+ all_keys = list(state_dict.keys())
+ sgm_patterns = ["input_blocks", "middle_block", "output_blocks"]
+
+ # 2. check if needs remapping, if not return original dict
+ is_in_sgm_format = False
+ for key in all_keys:
+ if any(p in key for p in sgm_patterns):
+ is_in_sgm_format = True
+ break
+
+ if not is_in_sgm_format:
+ return state_dict
+
+ # 3. Else remap from SGM patterns
+ new_state_dict = {}
+ inner_block_map = ["resnets", "attentions", "upsamplers"]
+
+ # Retrieves # of down, mid and up blocks
+ input_block_ids, middle_block_ids, output_block_ids = set(), set(), set()
+
+ for layer in all_keys:
+ if "text" in layer:
+ new_state_dict[layer] = state_dict.pop(layer)
+ else:
+ layer_id = int(layer.split(delimiter)[:block_slice_pos][-1])
+ if sgm_patterns[0] in layer:
+ input_block_ids.add(layer_id)
+ elif sgm_patterns[1] in layer:
+ middle_block_ids.add(layer_id)
+ elif sgm_patterns[2] in layer:
+ output_block_ids.add(layer_id)
+ else:
+ raise ValueError(f"Checkpoint not supported because layer {layer} not supported.")
+
+ input_blocks = {
+ layer_id: [key for key in state_dict if f"input_blocks{delimiter}{layer_id}" in key]
+ for layer_id in input_block_ids
+ }
+ middle_blocks = {
+ layer_id: [key for key in state_dict if f"middle_block{delimiter}{layer_id}" in key]
+ for layer_id in middle_block_ids
+ }
+ output_blocks = {
+ layer_id: [key for key in state_dict if f"output_blocks{delimiter}{layer_id}" in key]
+ for layer_id in output_block_ids
+ }
+
+ # Rename keys accordingly
+ for i in input_block_ids:
+ block_id = (i - 1) // (unet_config.layers_per_block + 1)
+ layer_in_block_id = (i - 1) % (unet_config.layers_per_block + 1)
+
+ for key in input_blocks[i]:
+ inner_block_id = int(key.split(delimiter)[block_slice_pos])
+ inner_block_key = inner_block_map[inner_block_id] if "op" not in key else "downsamplers"
+ inner_layers_in_block = str(layer_in_block_id) if "op" not in key else "0"
+ new_key = delimiter.join(
+ key.split(delimiter)[: block_slice_pos - 1]
+ + [str(block_id), inner_block_key, inner_layers_in_block]
+ + key.split(delimiter)[block_slice_pos + 1 :]
+ )
+ new_state_dict[new_key] = state_dict.pop(key)
+
+ for i in middle_block_ids:
+ key_part = None
+ if i == 0:
+ key_part = [inner_block_map[0], "0"]
+ elif i == 1:
+ key_part = [inner_block_map[1], "0"]
+ elif i == 2:
+ key_part = [inner_block_map[0], "1"]
+ else:
+ raise ValueError(f"Invalid middle block id {i}.")
+
+ for key in middle_blocks[i]:
+ new_key = delimiter.join(
+ key.split(delimiter)[: block_slice_pos - 1] + key_part + key.split(delimiter)[block_slice_pos:]
+ )
+ new_state_dict[new_key] = state_dict.pop(key)
+
+ for i in output_block_ids:
+ block_id = i // (unet_config.layers_per_block + 1)
+ layer_in_block_id = i % (unet_config.layers_per_block + 1)
+
+ for key in output_blocks[i]:
+ inner_block_id = int(key.split(delimiter)[block_slice_pos])
+ inner_block_key = inner_block_map[inner_block_id]
+ inner_layers_in_block = str(layer_in_block_id) if inner_block_id < 2 else "0"
+ new_key = delimiter.join(
+ key.split(delimiter)[: block_slice_pos - 1]
+ + [str(block_id), inner_block_key, inner_layers_in_block]
+ + key.split(delimiter)[block_slice_pos + 1 :]
+ )
+ new_state_dict[new_key] = state_dict.pop(key)
+
+ if len(state_dict) > 0:
+ raise ValueError("At this point all state dict entries have to be converted.")
+
+ return new_state_dict
+
+
+def _convert_kohya_lora_to_diffusers(state_dict, unet_name="unet", text_encoder_name="text_encoder"):
+ unet_state_dict = {}
+ te_state_dict = {}
+ te2_state_dict = {}
+ network_alphas = {}
+
+ # every down weight has a corresponding up weight and potentially an alpha weight
+ lora_keys = [k for k in state_dict.keys() if k.endswith("lora_down.weight")]
+ for key in lora_keys:
+ lora_name = key.split(".")[0]
+ lora_name_up = lora_name + ".lora_up.weight"
+ lora_name_alpha = lora_name + ".alpha"
+
+ if lora_name.startswith("lora_unet_"):
+ diffusers_name = key.replace("lora_unet_", "").replace("_", ".")
+
+ if "input.blocks" in diffusers_name:
+ diffusers_name = diffusers_name.replace("input.blocks", "down_blocks")
+ else:
+ diffusers_name = diffusers_name.replace("down.blocks", "down_blocks")
+
+ if "middle.block" in diffusers_name:
+ diffusers_name = diffusers_name.replace("middle.block", "mid_block")
+ else:
+ diffusers_name = diffusers_name.replace("mid.block", "mid_block")
+ if "output.blocks" in diffusers_name:
+ diffusers_name = diffusers_name.replace("output.blocks", "up_blocks")
+ else:
+ diffusers_name = diffusers_name.replace("up.blocks", "up_blocks")
+
+ diffusers_name = diffusers_name.replace("transformer.blocks", "transformer_blocks")
+ diffusers_name = diffusers_name.replace("to.q.lora", "to_q_lora")
+ diffusers_name = diffusers_name.replace("to.k.lora", "to_k_lora")
+ diffusers_name = diffusers_name.replace("to.v.lora", "to_v_lora")
+ diffusers_name = diffusers_name.replace("to.out.0.lora", "to_out_lora")
+ diffusers_name = diffusers_name.replace("proj.in", "proj_in")
+ diffusers_name = diffusers_name.replace("proj.out", "proj_out")
+ diffusers_name = diffusers_name.replace("emb.layers", "time_emb_proj")
+
+ # SDXL specificity.
+ if "emb" in diffusers_name and "time.emb.proj" not in diffusers_name:
+ pattern = r"\.\d+(?=\D*$)"
+ diffusers_name = re.sub(pattern, "", diffusers_name, count=1)
+ if ".in." in diffusers_name:
+ diffusers_name = diffusers_name.replace("in.layers.2", "conv1")
+ if ".out." in diffusers_name:
+ diffusers_name = diffusers_name.replace("out.layers.3", "conv2")
+ if "downsamplers" in diffusers_name or "upsamplers" in diffusers_name:
+ diffusers_name = diffusers_name.replace("op", "conv")
+ if "skip" in diffusers_name:
+ diffusers_name = diffusers_name.replace("skip.connection", "conv_shortcut")
+
+ # LyCORIS specificity.
+ if "time.emb.proj" in diffusers_name:
+ diffusers_name = diffusers_name.replace("time.emb.proj", "time_emb_proj")
+ if "conv.shortcut" in diffusers_name:
+ diffusers_name = diffusers_name.replace("conv.shortcut", "conv_shortcut")
+
+ # General coverage.
+ if "transformer_blocks" in diffusers_name:
+ if "attn1" in diffusers_name or "attn2" in diffusers_name:
+ diffusers_name = diffusers_name.replace("attn1", "attn1.processor")
+ diffusers_name = diffusers_name.replace("attn2", "attn2.processor")
+ unet_state_dict[diffusers_name] = state_dict.pop(key)
+ unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+ elif "ff" in diffusers_name:
+ unet_state_dict[diffusers_name] = state_dict.pop(key)
+ unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+ elif any(key in diffusers_name for key in ("proj_in", "proj_out")):
+ unet_state_dict[diffusers_name] = state_dict.pop(key)
+ unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+ else:
+ unet_state_dict[diffusers_name] = state_dict.pop(key)
+ unet_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+ elif lora_name.startswith("lora_te_"):
+ diffusers_name = key.replace("lora_te_", "").replace("_", ".")
+ diffusers_name = diffusers_name.replace("text.model", "text_model")
+ diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+ diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+ diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+ diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+ diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+ if "self_attn" in diffusers_name:
+ te_state_dict[diffusers_name] = state_dict.pop(key)
+ te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+ elif "mlp" in diffusers_name:
+ # Be aware that this is the new diffusers convention and the rest of the code might
+ # not utilize it yet.
+ diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+ te_state_dict[diffusers_name] = state_dict.pop(key)
+ te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+ # (sayakpaul): Duplicate code. Needs to be cleaned.
+ elif lora_name.startswith("lora_te1_"):
+ diffusers_name = key.replace("lora_te1_", "").replace("_", ".")
+ diffusers_name = diffusers_name.replace("text.model", "text_model")
+ diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+ diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+ diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+ diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+ diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+ if "self_attn" in diffusers_name:
+ te_state_dict[diffusers_name] = state_dict.pop(key)
+ te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+ elif "mlp" in diffusers_name:
+ # Be aware that this is the new diffusers convention and the rest of the code might
+ # not utilize it yet.
+ diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+ te_state_dict[diffusers_name] = state_dict.pop(key)
+ te_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+ # (sayakpaul): Duplicate code. Needs to be cleaned.
+ elif lora_name.startswith("lora_te2_"):
+ diffusers_name = key.replace("lora_te2_", "").replace("_", ".")
+ diffusers_name = diffusers_name.replace("text.model", "text_model")
+ diffusers_name = diffusers_name.replace("self.attn", "self_attn")
+ diffusers_name = diffusers_name.replace("q.proj.lora", "to_q_lora")
+ diffusers_name = diffusers_name.replace("k.proj.lora", "to_k_lora")
+ diffusers_name = diffusers_name.replace("v.proj.lora", "to_v_lora")
+ diffusers_name = diffusers_name.replace("out.proj.lora", "to_out_lora")
+ if "self_attn" in diffusers_name:
+ te2_state_dict[diffusers_name] = state_dict.pop(key)
+ te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+ elif "mlp" in diffusers_name:
+ # Be aware that this is the new diffusers convention and the rest of the code might
+ # not utilize it yet.
+ diffusers_name = diffusers_name.replace(".lora.", ".lora_linear_layer.")
+ te2_state_dict[diffusers_name] = state_dict.pop(key)
+ te2_state_dict[diffusers_name.replace(".down.", ".up.")] = state_dict.pop(lora_name_up)
+
+ # Rename the alphas so that they can be mapped appropriately.
+ if lora_name_alpha in state_dict:
+ alpha = state_dict.pop(lora_name_alpha).item()
+ if lora_name_alpha.startswith("lora_unet_"):
+ prefix = "unet."
+ elif lora_name_alpha.startswith(("lora_te_", "lora_te1_")):
+ prefix = "text_encoder."
+ else:
+ prefix = "text_encoder_2."
+ new_name = prefix + diffusers_name.split(".lora.")[0] + ".alpha"
+ network_alphas.update({new_name: alpha})
+
+ if len(state_dict) > 0:
+ raise ValueError(f"The following keys have not been correctly be renamed: \n\n {', '.join(state_dict.keys())}")
+
+ logger.info("Kohya-style checkpoint detected.")
+ unet_state_dict = {f"{unet_name}.{module_name}": params for module_name, params in unet_state_dict.items()}
+ te_state_dict = {f"{text_encoder_name}.{module_name}": params for module_name, params in te_state_dict.items()}
+ te2_state_dict = (
+ {f"text_encoder_2.{module_name}": params for module_name, params in te2_state_dict.items()}
+ if len(te2_state_dict) > 0
+ else None
+ )
+ if te2_state_dict is not None:
+ te_state_dict.update(te2_state_dict)
+
+ new_state_dict = {**unet_state_dict, **te_state_dict}
+ return new_state_dict, network_alphas
diff --git a/model/diffusers_c/loaders/peft.py b/model/diffusers_c/loaders/peft.py
new file mode 100644
index 0000000000000000000000000000000000000000..8acbe1ea8dedecca362fc309114c74f40dac3d59
--- /dev/null
+++ b/model/diffusers_c/loaders/peft.py
@@ -0,0 +1,185 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import List, Union
+
+from ..utils import MIN_PEFT_VERSION, check_peft_version, is_peft_available
+
+
+class PeftAdapterMixin:
+ """
+ A class containing all functions for loading and using adapters weights that are supported in PEFT library. For
+ more details about adapters and injecting them in a transformer-based model, check out the PEFT [documentation](https://huggingface.co/docs/peft/index).
+
+ Install the latest version of PEFT, and use this mixin to:
+
+ - Attach new adapters in the model.
+ - Attach multiple adapters and iteratively activate/deactivate them.
+ - Activate/deactivate all adapters from the model.
+ - Get a list of the active adapters.
+ """
+
+ _hf_peft_config_loaded = False
+
+ def add_adapter(self, adapter_config, adapter_name: str = "default") -> None:
+ r"""
+ Adds a new adapter to the current model for training. If no adapter name is passed, a default name is assigned
+ to the adapter to follow the convention of the PEFT library.
+
+ If you are not familiar with adapters and PEFT methods, we invite you to read more about them in the PEFT
+ [documentation](https://huggingface.co/docs/peft).
+
+ Args:
+ adapter_config (`[~peft.PeftConfig]`):
+ The configuration of the adapter to add; supported adapters are non-prefix tuning and adaption prompt
+ methods.
+ adapter_name (`str`, *optional*, defaults to `"default"`):
+ The name of the adapter to add. If no name is passed, a default name is assigned to the adapter.
+ """
+ check_peft_version(min_version=MIN_PEFT_VERSION)
+
+ if not is_peft_available():
+ raise ImportError("PEFT is not available. Please install PEFT to use this function: `pip install peft`.")
+
+ from peft import PeftConfig, inject_adapter_in_model
+
+ if not self._hf_peft_config_loaded:
+ self._hf_peft_config_loaded = True
+ elif adapter_name in self.peft_config:
+ raise ValueError(f"Adapter with name {adapter_name} already exists. Please use a different name.")
+
+ if not isinstance(adapter_config, PeftConfig):
+ raise ValueError(
+ f"adapter_config should be an instance of PeftConfig. Got {type(adapter_config)} instead."
+ )
+
+ # Unlike transformers, here we don't need to retrieve the name_or_path of the unet as the loading logic is
+ # handled by the `load_lora_layers` or `LoraLoaderMixin`. Therefore we set it to `None` here.
+ adapter_config.base_model_name_or_path = None
+ inject_adapter_in_model(adapter_config, self, adapter_name)
+ self.set_adapter(adapter_name)
+
+ def set_adapter(self, adapter_name: Union[str, List[str]]) -> None:
+ """
+ Sets a specific adapter by forcing the model to only use that adapter and disables the other adapters.
+
+ If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT
+ [documentation](https://huggingface.co/docs/peft).
+
+ Args:
+ adapter_name (Union[str, List[str]])):
+ The list of adapters to set or the adapter name in the case of a single adapter.
+ """
+ check_peft_version(min_version=MIN_PEFT_VERSION)
+
+ if not self._hf_peft_config_loaded:
+ raise ValueError("No adapter loaded. Please load an adapter first.")
+
+ if isinstance(adapter_name, str):
+ adapter_name = [adapter_name]
+
+ missing = set(adapter_name) - set(self.peft_config)
+ if len(missing) > 0:
+ raise ValueError(
+ f"Following adapter(s) could not be found: {', '.join(missing)}. Make sure you are passing the correct adapter name(s)."
+ f" current loaded adapters are: {list(self.peft_config.keys())}"
+ )
+
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ _adapters_has_been_set = False
+
+ for _, module in self.named_modules():
+ if isinstance(module, BaseTunerLayer):
+ if hasattr(module, "set_adapter"):
+ module.set_adapter(adapter_name)
+ # Previous versions of PEFT does not support multi-adapter inference
+ elif not hasattr(module, "set_adapter") and len(adapter_name) != 1:
+ raise ValueError(
+ "You are trying to set multiple adapters and you have a PEFT version that does not support multi-adapter inference. Please upgrade to the latest version of PEFT."
+ " `pip install -U peft` or `pip install -U git+https://github.com/huggingface/peft.git`"
+ )
+ else:
+ module.active_adapter = adapter_name
+ _adapters_has_been_set = True
+
+ if not _adapters_has_been_set:
+ raise ValueError(
+ "Did not succeeded in setting the adapter. Please make sure you are using a model that supports adapters."
+ )
+
+ def disable_adapters(self) -> None:
+ r"""
+ Disable all adapters attached to the model and fallback to inference with the base model only.
+
+ If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT
+ [documentation](https://huggingface.co/docs/peft).
+ """
+ check_peft_version(min_version=MIN_PEFT_VERSION)
+
+ if not self._hf_peft_config_loaded:
+ raise ValueError("No adapter loaded. Please load an adapter first.")
+
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ for _, module in self.named_modules():
+ if isinstance(module, BaseTunerLayer):
+ if hasattr(module, "enable_adapters"):
+ module.enable_adapters(enabled=False)
+ else:
+ # support for older PEFT versions
+ module.disable_adapters = True
+
+ def enable_adapters(self) -> None:
+ """
+ Enable adapters that are attached to the model. The model uses `self.active_adapters()` to retrieve the
+ list of adapters to enable.
+
+ If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT
+ [documentation](https://huggingface.co/docs/peft).
+ """
+ check_peft_version(min_version=MIN_PEFT_VERSION)
+
+ if not self._hf_peft_config_loaded:
+ raise ValueError("No adapter loaded. Please load an adapter first.")
+
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ for _, module in self.named_modules():
+ if isinstance(module, BaseTunerLayer):
+ if hasattr(module, "enable_adapters"):
+ module.enable_adapters(enabled=True)
+ else:
+ # support for older PEFT versions
+ module.disable_adapters = False
+
+ def active_adapters(self) -> List[str]:
+ """
+ Gets the current list of active adapters of the model.
+
+ If you are not familiar with adapters and PEFT methods, we invite you to read more about them on the PEFT
+ [documentation](https://huggingface.co/docs/peft).
+ """
+ check_peft_version(min_version=MIN_PEFT_VERSION)
+
+ if not is_peft_available():
+ raise ImportError("PEFT is not available. Please install PEFT to use this function: `pip install peft`.")
+
+ if not self._hf_peft_config_loaded:
+ raise ValueError("No adapter loaded. Please load an adapter first.")
+
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ for _, module in self.named_modules():
+ if isinstance(module, BaseTunerLayer):
+ return module.active_adapter
diff --git a/model/diffusers_c/loaders/single_file.py b/model/diffusers_c/loaders/single_file.py
new file mode 100644
index 0000000000000000000000000000000000000000..69cce72468ce02afd0c1293a2413b87cdb9be8b3
--- /dev/null
+++ b/model/diffusers_c/loaders/single_file.py
@@ -0,0 +1,313 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from huggingface_hub.utils import validate_hf_hub_args
+
+from ..utils import is_transformers_available, logging
+from .single_file_utils import (
+ create_diffusers_unet_model_from_ldm,
+ create_diffusers_vae_model_from_ldm,
+ create_scheduler_from_ldm,
+ create_text_encoders_and_tokenizers_from_ldm,
+ fetch_ldm_config_and_checkpoint,
+ infer_model_type,
+)
+
+
+logger = logging.get_logger(__name__)
+
+# Pipelines that support the SDXL Refiner checkpoint
+REFINER_PIPELINES = [
+ "StableDiffusionXLImg2ImgPipeline",
+ "StableDiffusionXLInpaintPipeline",
+ "StableDiffusionXLControlNetImg2ImgPipeline",
+]
+
+if is_transformers_available():
+ from transformers import AutoFeatureExtractor
+
+
+def build_sub_model_components(
+ pipeline_components,
+ pipeline_class_name,
+ component_name,
+ original_config,
+ checkpoint,
+ local_files_only=False,
+ load_safety_checker=False,
+ model_type=None,
+ image_size=None,
+ torch_dtype=None,
+ **kwargs,
+):
+ if component_name in pipeline_components:
+ return {}
+
+ if component_name == "unet":
+ num_in_channels = kwargs.pop("num_in_channels", None)
+ unet_components = create_diffusers_unet_model_from_ldm(
+ pipeline_class_name,
+ original_config,
+ checkpoint,
+ num_in_channels=num_in_channels,
+ image_size=image_size,
+ torch_dtype=torch_dtype,
+ model_type=model_type,
+ )
+ return unet_components
+
+ if component_name == "vae":
+ scaling_factor = kwargs.get("scaling_factor", None)
+ vae_components = create_diffusers_vae_model_from_ldm(
+ pipeline_class_name,
+ original_config,
+ checkpoint,
+ image_size,
+ scaling_factor,
+ torch_dtype,
+ model_type=model_type,
+ )
+ return vae_components
+
+ if component_name == "scheduler":
+ scheduler_type = kwargs.get("scheduler_type", "ddim")
+ prediction_type = kwargs.get("prediction_type", None)
+
+ scheduler_components = create_scheduler_from_ldm(
+ pipeline_class_name,
+ original_config,
+ checkpoint,
+ scheduler_type=scheduler_type,
+ prediction_type=prediction_type,
+ model_type=model_type,
+ )
+
+ return scheduler_components
+
+ if component_name in ["text_encoder", "text_encoder_2", "tokenizer", "tokenizer_2"]:
+ text_encoder_components = create_text_encoders_and_tokenizers_from_ldm(
+ original_config,
+ checkpoint,
+ model_type=model_type,
+ local_files_only=local_files_only,
+ torch_dtype=torch_dtype,
+ )
+ return text_encoder_components
+
+ if component_name == "safety_checker":
+ if load_safety_checker:
+ from ..pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+ safety_checker = StableDiffusionSafetyChecker.from_pretrained(
+ "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only, torch_dtype=torch_dtype
+ )
+ else:
+ safety_checker = None
+ return {"safety_checker": safety_checker}
+
+ if component_name == "feature_extractor":
+ if load_safety_checker:
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
+ "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only
+ )
+ else:
+ feature_extractor = None
+ return {"feature_extractor": feature_extractor}
+
+ return
+
+
+def set_additional_components(
+ pipeline_class_name,
+ original_config,
+ checkpoint=None,
+ model_type=None,
+):
+ components = {}
+ if pipeline_class_name in REFINER_PIPELINES:
+ model_type = infer_model_type(original_config, checkpoint=checkpoint, model_type=model_type)
+ is_refiner = model_type == "SDXL-Refiner"
+ components.update(
+ {
+ "requires_aesthetics_score": is_refiner,
+ "force_zeros_for_empty_prompt": False if is_refiner else True,
+ }
+ )
+
+ return components
+
+
+class FromSingleFileMixin:
+ """
+ Load model weights saved in the `.ckpt` format into a [`DiffusionPipeline`].
+ """
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_single_file(cls, pretrained_model_link_or_path, **kwargs):
+ r"""
+ Instantiate a [`DiffusionPipeline`] from pretrained pipeline weights saved in the `.ckpt` or `.safetensors`
+ format. The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ Parameters:
+ pretrained_model_link_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+ - A link to the `.ckpt` file (for example
+ `"https://huggingface.co//blob/main/.ckpt"`) on the Hub.
+ - A path to a *file* containing all pipeline weights.
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model with another dtype.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ original_config_file (`str`, *optional*):
+ The path to the original config file that was used to train the model. If not provided, the config file
+ will be inferred from the checkpoint file.
+ model_type (`str`, *optional*):
+ The type of model to load. If not provided, the model type will be inferred from the checkpoint file.
+ image_size (`int`, *optional*):
+ The size of the image output. It's used to configure the `sample_size` parameter of the UNet and VAE model.
+ load_safety_checker (`bool`, *optional*, defaults to `False`):
+ Whether to load the safety checker model or not. By default, the safety checker is not loaded unless a `safety_checker` component is passed to the `kwargs`.
+ num_in_channels (`int`, *optional*):
+ Specify the number of input channels for the UNet model. Read more about how to configure UNet model with this parameter
+ [here](https://huggingface.co/docs/diffusers/training/adapt_a_model#configure-unet2dconditionmodel-parameters).
+ scaling_factor (`float`, *optional*):
+ The scaling factor to use for the VAE model. If not provided, it is inferred from the config file first.
+ If the scaling factor is not found in the config file, the default value 0.18215 is used.
+ scheduler_type (`str`, *optional*):
+ The type of scheduler to load. If not provided, the scheduler type will be inferred from the checkpoint file.
+ prediction_type (`str`, *optional*):
+ The type of prediction to load. If not provided, the prediction type will be inferred from the checkpoint file.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
+ class). The overwritten components are passed directly to the pipelines `__init__` method. See example
+ below for more information.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import StableDiffusionPipeline
+
+ >>> # Download pipeline from huggingface.co and cache.
+ >>> pipeline = StableDiffusionPipeline.from_single_file(
+ ... "https://huggingface.co/WarriorMama777/OrangeMixs/blob/main/Models/AbyssOrangeMix/AbyssOrangeMix.safetensors"
+ ... )
+
+ >>> # Download pipeline from local file
+ >>> # file is downloaded under ./v1-5-pruned-emaonly.ckpt
+ >>> pipeline = StableDiffusionPipeline.from_single_file("./v1-5-pruned-emaonly")
+
+ >>> # Enable float16 and move to GPU
+ >>> pipeline = StableDiffusionPipeline.from_single_file(
+ ... "https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/v1-5-pruned-emaonly.ckpt",
+ ... torch_dtype=torch.float16,
+ ... )
+ >>> pipeline.to("cuda")
+ ```
+ """
+ original_config_file = kwargs.pop("original_config_file", None)
+ resume_download = kwargs.pop("resume_download", False)
+ force_download = kwargs.pop("force_download", False)
+ proxies = kwargs.pop("proxies", None)
+ token = kwargs.pop("token", None)
+ cache_dir = kwargs.pop("cache_dir", None)
+ local_files_only = kwargs.pop("local_files_only", False)
+ revision = kwargs.pop("revision", None)
+ torch_dtype = kwargs.pop("torch_dtype", None)
+
+ class_name = cls.__name__
+
+ original_config, checkpoint = fetch_ldm_config_and_checkpoint(
+ pretrained_model_link_or_path=pretrained_model_link_or_path,
+ class_name=class_name,
+ original_config_file=original_config_file,
+ resume_download=resume_download,
+ force_download=force_download,
+ proxies=proxies,
+ token=token,
+ revision=revision,
+ local_files_only=local_files_only,
+ cache_dir=cache_dir,
+ )
+
+ from ..pipelines.pipeline_utils import _get_pipeline_class
+
+ pipeline_class = _get_pipeline_class(
+ cls,
+ config=None,
+ cache_dir=cache_dir,
+ )
+
+ expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class)
+ passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
+ passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
+
+ model_type = kwargs.pop("model_type", None)
+ image_size = kwargs.pop("image_size", None)
+ load_safety_checker = (kwargs.pop("load_safety_checker", False)) or (
+ passed_class_obj.get("safety_checker", None) is not None
+ )
+
+ init_kwargs = {}
+ for name in expected_modules:
+ if name in passed_class_obj:
+ init_kwargs[name] = passed_class_obj[name]
+ else:
+ components = build_sub_model_components(
+ init_kwargs,
+ class_name,
+ name,
+ original_config,
+ checkpoint,
+ model_type=model_type,
+ image_size=image_size,
+ load_safety_checker=load_safety_checker,
+ local_files_only=local_files_only,
+ torch_dtype=torch_dtype,
+ **kwargs,
+ )
+ if not components:
+ continue
+ init_kwargs.update(components)
+
+ additional_components = set_additional_components(class_name, original_config, model_type=model_type)
+ if additional_components:
+ init_kwargs.update(additional_components)
+
+ init_kwargs.update(passed_pipe_kwargs)
+ pipe = pipeline_class(**init_kwargs)
+
+ if torch_dtype is not None:
+ pipe.to(dtype=torch_dtype)
+
+ return pipe
diff --git a/model/diffusers_c/loaders/single_file_utils.py b/model/diffusers_c/loaders/single_file_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..13aaca88e80fef5408a668ed53d550ca8aca90db
--- /dev/null
+++ b/model/diffusers_c/loaders/single_file_utils.py
@@ -0,0 +1,1512 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Conversion script for the Stable Diffusion checkpoints."""
+
+import os
+import re
+from contextlib import nullcontext
+from io import BytesIO
+from urllib.parse import urlparse
+
+import requests
+import yaml
+
+from ..models.modeling_utils import load_state_dict
+from ..schedulers import (
+ DDIMScheduler,
+ DDPMScheduler,
+ DPMSolverMultistepScheduler,
+ EDMDPMSolverMultistepScheduler,
+ EulerAncestralDiscreteScheduler,
+ EulerDiscreteScheduler,
+ HeunDiscreteScheduler,
+ LMSDiscreteScheduler,
+ PNDMScheduler,
+)
+from ..utils import is_accelerate_available, is_transformers_available, logging
+from ..utils.hub_utils import _get_model_file
+
+
+if is_transformers_available():
+ from transformers import (
+ CLIPTextConfig,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ )
+
+if is_accelerate_available():
+ from accelerate import init_empty_weights
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+CONFIG_URLS = {
+ "v1": "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml",
+ "v2": "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/v2-inference-v.yaml",
+ "xl": "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml",
+ "xl_refiner": "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_refiner.yaml",
+ "upscale": "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/x4-upscaling.yaml",
+ "controlnet": "https://raw.githubusercontent.com/lllyasviel/ControlNet/main/models/cldm_v15.yaml",
+}
+
+CHECKPOINT_KEY_NAMES = {
+ "v2": "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight",
+ "xl_base": "conditioner.embedders.1.model.transformer.resblocks.9.mlp.c_proj.bias",
+ "xl_refiner": "conditioner.embedders.0.model.transformer.resblocks.9.mlp.c_proj.bias",
+}
+
+SCHEDULER_DEFAULT_CONFIG = {
+ "beta_schedule": "scaled_linear",
+ "beta_start": 0.00085,
+ "beta_end": 0.012,
+ "interpolation_type": "linear",
+ "num_train_timesteps": 1000,
+ "prediction_type": "epsilon",
+ "sample_max_value": 1.0,
+ "set_alpha_to_one": False,
+ "skip_prk_steps": True,
+ "steps_offset": 1,
+ "timestep_spacing": "leading",
+}
+
+DIFFUSERS_TO_LDM_MAPPING = {
+ "unet": {
+ "layers": {
+ "time_embedding.linear_1.weight": "time_embed.0.weight",
+ "time_embedding.linear_1.bias": "time_embed.0.bias",
+ "time_embedding.linear_2.weight": "time_embed.2.weight",
+ "time_embedding.linear_2.bias": "time_embed.2.bias",
+ "conv_in.weight": "input_blocks.0.0.weight",
+ "conv_in.bias": "input_blocks.0.0.bias",
+ "conv_norm_out.weight": "out.0.weight",
+ "conv_norm_out.bias": "out.0.bias",
+ "conv_out.weight": "out.2.weight",
+ "conv_out.bias": "out.2.bias",
+ },
+ "class_embed_type": {
+ "class_embedding.linear_1.weight": "label_emb.0.0.weight",
+ "class_embedding.linear_1.bias": "label_emb.0.0.bias",
+ "class_embedding.linear_2.weight": "label_emb.0.2.weight",
+ "class_embedding.linear_2.bias": "label_emb.0.2.bias",
+ },
+ "addition_embed_type": {
+ "add_embedding.linear_1.weight": "label_emb.0.0.weight",
+ "add_embedding.linear_1.bias": "label_emb.0.0.bias",
+ "add_embedding.linear_2.weight": "label_emb.0.2.weight",
+ "add_embedding.linear_2.bias": "label_emb.0.2.bias",
+ },
+ },
+ "controlnet": {
+ "layers": {
+ "time_embedding.linear_1.weight": "time_embed.0.weight",
+ "time_embedding.linear_1.bias": "time_embed.0.bias",
+ "time_embedding.linear_2.weight": "time_embed.2.weight",
+ "time_embedding.linear_2.bias": "time_embed.2.bias",
+ "conv_in.weight": "input_blocks.0.0.weight",
+ "conv_in.bias": "input_blocks.0.0.bias",
+ "controlnet_cond_embedding.conv_in.weight": "input_hint_block.0.weight",
+ "controlnet_cond_embedding.conv_in.bias": "input_hint_block.0.bias",
+ "controlnet_cond_embedding.conv_out.weight": "input_hint_block.14.weight",
+ "controlnet_cond_embedding.conv_out.bias": "input_hint_block.14.bias",
+ },
+ "class_embed_type": {
+ "class_embedding.linear_1.weight": "label_emb.0.0.weight",
+ "class_embedding.linear_1.bias": "label_emb.0.0.bias",
+ "class_embedding.linear_2.weight": "label_emb.0.2.weight",
+ "class_embedding.linear_2.bias": "label_emb.0.2.bias",
+ },
+ "addition_embed_type": {
+ "add_embedding.linear_1.weight": "label_emb.0.0.weight",
+ "add_embedding.linear_1.bias": "label_emb.0.0.bias",
+ "add_embedding.linear_2.weight": "label_emb.0.2.weight",
+ "add_embedding.linear_2.bias": "label_emb.0.2.bias",
+ },
+ },
+ "vae": {
+ "encoder.conv_in.weight": "encoder.conv_in.weight",
+ "encoder.conv_in.bias": "encoder.conv_in.bias",
+ "encoder.conv_out.weight": "encoder.conv_out.weight",
+ "encoder.conv_out.bias": "encoder.conv_out.bias",
+ "encoder.conv_norm_out.weight": "encoder.norm_out.weight",
+ "encoder.conv_norm_out.bias": "encoder.norm_out.bias",
+ "decoder.conv_in.weight": "decoder.conv_in.weight",
+ "decoder.conv_in.bias": "decoder.conv_in.bias",
+ "decoder.conv_out.weight": "decoder.conv_out.weight",
+ "decoder.conv_out.bias": "decoder.conv_out.bias",
+ "decoder.conv_norm_out.weight": "decoder.norm_out.weight",
+ "decoder.conv_norm_out.bias": "decoder.norm_out.bias",
+ "quant_conv.weight": "quant_conv.weight",
+ "quant_conv.bias": "quant_conv.bias",
+ "post_quant_conv.weight": "post_quant_conv.weight",
+ "post_quant_conv.bias": "post_quant_conv.bias",
+ },
+ "openclip": {
+ "layers": {
+ "text_model.embeddings.position_embedding.weight": "positional_embedding",
+ "text_model.embeddings.token_embedding.weight": "token_embedding.weight",
+ "text_model.final_layer_norm.weight": "ln_final.weight",
+ "text_model.final_layer_norm.bias": "ln_final.bias",
+ "text_projection.weight": "text_projection",
+ },
+ "transformer": {
+ "text_model.encoder.layers.": "resblocks.",
+ "layer_norm1": "ln_1",
+ "layer_norm2": "ln_2",
+ ".fc1.": ".c_fc.",
+ ".fc2.": ".c_proj.",
+ ".self_attn": ".attn",
+ "transformer.text_model.final_layer_norm.": "ln_final.",
+ "transformer.text_model.embeddings.token_embedding.weight": "token_embedding.weight",
+ "transformer.text_model.embeddings.position_embedding.weight": "positional_embedding",
+ },
+ },
+}
+
+LDM_VAE_KEY = "first_stage_model."
+LDM_VAE_DEFAULT_SCALING_FACTOR = 0.18215
+PLAYGROUND_VAE_SCALING_FACTOR = 0.5
+LDM_UNET_KEY = "model.diffusion_model."
+LDM_CONTROLNET_KEY = "control_model."
+LDM_CLIP_PREFIX_TO_REMOVE = ["cond_stage_model.transformer.", "conditioner.embedders.0.transformer."]
+LDM_OPEN_CLIP_TEXT_PROJECTION_DIM = 1024
+
+SD_2_TEXT_ENCODER_KEYS_TO_IGNORE = [
+ "cond_stage_model.model.transformer.resblocks.23.attn.in_proj_bias",
+ "cond_stage_model.model.transformer.resblocks.23.attn.in_proj_weight",
+ "cond_stage_model.model.transformer.resblocks.23.attn.out_proj.bias",
+ "cond_stage_model.model.transformer.resblocks.23.attn.out_proj.weight",
+ "cond_stage_model.model.transformer.resblocks.23.ln_1.bias",
+ "cond_stage_model.model.transformer.resblocks.23.ln_1.weight",
+ "cond_stage_model.model.transformer.resblocks.23.ln_2.bias",
+ "cond_stage_model.model.transformer.resblocks.23.ln_2.weight",
+ "cond_stage_model.model.transformer.resblocks.23.mlp.c_fc.bias",
+ "cond_stage_model.model.transformer.resblocks.23.mlp.c_fc.weight",
+ "cond_stage_model.model.transformer.resblocks.23.mlp.c_proj.bias",
+ "cond_stage_model.model.transformer.resblocks.23.mlp.c_proj.weight",
+ "cond_stage_model.model.text_projection",
+]
+
+
+VALID_URL_PREFIXES = ["https://huggingface.co/", "huggingface.co/", "hf.co/", "https://hf.co/"]
+
+
+def _extract_repo_id_and_weights_name(pretrained_model_name_or_path):
+ pattern = r"([^/]+)/([^/]+)/(?:blob/main/)?(.+)"
+ weights_name = None
+ repo_id = (None,)
+ for prefix in VALID_URL_PREFIXES:
+ pretrained_model_name_or_path = pretrained_model_name_or_path.replace(prefix, "")
+ match = re.match(pattern, pretrained_model_name_or_path)
+ if not match:
+ return repo_id, weights_name
+
+ repo_id = f"{match.group(1)}/{match.group(2)}"
+ weights_name = match.group(3)
+
+ return repo_id, weights_name
+
+
+def fetch_ldm_config_and_checkpoint(
+ pretrained_model_link_or_path,
+ class_name,
+ original_config_file=None,
+ resume_download=False,
+ force_download=False,
+ proxies=None,
+ token=None,
+ cache_dir=None,
+ local_files_only=None,
+ revision=None,
+):
+ if os.path.isfile(pretrained_model_link_or_path):
+ checkpoint = load_state_dict(pretrained_model_link_or_path)
+
+ else:
+ repo_id, weights_name = _extract_repo_id_and_weights_name(pretrained_model_link_or_path)
+ checkpoint_path = _get_model_file(
+ repo_id,
+ weights_name=weights_name,
+ force_download=force_download,
+ cache_dir=cache_dir,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ )
+ checkpoint = load_state_dict(checkpoint_path)
+
+ # some checkpoints contain the model state dict under a "state_dict" key
+ while "state_dict" in checkpoint:
+ checkpoint = checkpoint["state_dict"]
+
+ original_config = fetch_original_config(class_name, checkpoint, original_config_file)
+
+ return original_config, checkpoint
+
+
+def infer_original_config_file(class_name, checkpoint):
+ if CHECKPOINT_KEY_NAMES["v2"] in checkpoint and checkpoint[CHECKPOINT_KEY_NAMES["v2"]].shape[-1] == 1024:
+ config_url = CONFIG_URLS["v2"]
+
+ elif CHECKPOINT_KEY_NAMES["xl_base"] in checkpoint:
+ config_url = CONFIG_URLS["xl"]
+
+ elif CHECKPOINT_KEY_NAMES["xl_refiner"] in checkpoint:
+ config_url = CONFIG_URLS["xl_refiner"]
+
+ elif class_name == "StableDiffusionUpscalePipeline":
+ config_url = CONFIG_URLS["upscale"]
+
+ elif class_name == "ControlNetModel":
+ config_url = CONFIG_URLS["controlnet"]
+
+ else:
+ config_url = CONFIG_URLS["v1"]
+
+ original_config_file = BytesIO(requests.get(config_url).content)
+
+ return original_config_file
+
+
+def fetch_original_config(pipeline_class_name, checkpoint, original_config_file=None):
+ def is_valid_url(url):
+ result = urlparse(url)
+ if result.scheme and result.netloc:
+ return True
+
+ return False
+
+ if original_config_file is None:
+ original_config_file = infer_original_config_file(pipeline_class_name, checkpoint)
+
+ elif os.path.isfile(original_config_file):
+ with open(original_config_file, "r") as fp:
+ original_config_file = fp.read()
+
+ elif is_valid_url(original_config_file):
+ original_config_file = BytesIO(requests.get(original_config_file).content)
+
+ else:
+ raise ValueError("Invalid `original_config_file` provided. Please set it to a valid file path or URL.")
+
+ original_config = yaml.safe_load(original_config_file)
+
+ return original_config
+
+
+def infer_model_type(original_config, checkpoint=None, model_type=None):
+ if model_type is not None:
+ return model_type
+
+ has_cond_stage_config = (
+ "cond_stage_config" in original_config["model"]["params"]
+ and original_config["model"]["params"]["cond_stage_config"] is not None
+ )
+ has_network_config = (
+ "network_config" in original_config["model"]["params"]
+ and original_config["model"]["params"]["network_config"] is not None
+ )
+
+ if has_cond_stage_config:
+ model_type = original_config["model"]["params"]["cond_stage_config"]["target"].split(".")[-1]
+
+ elif has_network_config:
+ context_dim = original_config["model"]["params"]["network_config"]["params"]["context_dim"]
+ if "edm_mean" in checkpoint and "edm_std" in checkpoint:
+ model_type = "Playground"
+ elif context_dim == 2048:
+ model_type = "SDXL"
+ else:
+ model_type = "SDXL-Refiner"
+ else:
+ raise ValueError("Unable to infer model type from config")
+
+ logger.debug(f"No `model_type` given, `model_type` inferred as: {model_type}")
+
+ return model_type
+
+
+def get_default_scheduler_config():
+ return SCHEDULER_DEFAULT_CONFIG
+
+
+def set_image_size(pipeline_class_name, original_config, checkpoint, image_size=None, model_type=None):
+ if image_size:
+ return image_size
+
+ global_step = checkpoint["global_step"] if "global_step" in checkpoint else None
+ model_type = infer_model_type(original_config, checkpoint, model_type)
+
+ if pipeline_class_name == "StableDiffusionUpscalePipeline":
+ image_size = original_config["model"]["params"]["unet_config"]["params"]["image_size"]
+ return image_size
+
+ elif model_type in ["SDXL", "SDXL-Refiner", "Playground"]:
+ image_size = 1024
+ return image_size
+
+ elif (
+ "parameterization" in original_config["model"]["params"]
+ and original_config["model"]["params"]["parameterization"] == "v"
+ ):
+ # NOTE: For stable diffusion 2 base one has to pass `image_size==512`
+ # as it relies on a brittle global step parameter here
+ image_size = 512 if global_step == 875000 else 768
+ return image_size
+
+ else:
+ image_size = 512
+ return image_size
+
+
+# Copied from diffusers.pipelines.stable_diffusion.convert_from_ckpt.conv_attn_to_linear
+def conv_attn_to_linear(checkpoint):
+ keys = list(checkpoint.keys())
+ attn_keys = ["query.weight", "key.weight", "value.weight"]
+ for key in keys:
+ if ".".join(key.split(".")[-2:]) in attn_keys:
+ if checkpoint[key].ndim > 2:
+ checkpoint[key] = checkpoint[key][:, :, 0, 0]
+ elif "proj_attn.weight" in key:
+ if checkpoint[key].ndim > 2:
+ checkpoint[key] = checkpoint[key][:, :, 0]
+
+
+def create_unet_diffusers_config(original_config, image_size: int):
+ """
+ Creates a config for the diffusers based on the config of the LDM model.
+ """
+ if (
+ "unet_config" in original_config["model"]["params"]
+ and original_config["model"]["params"]["unet_config"] is not None
+ ):
+ unet_params = original_config["model"]["params"]["unet_config"]["params"]
+ else:
+ unet_params = original_config["model"]["params"]["network_config"]["params"]
+
+ vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"]
+ block_out_channels = [unet_params["model_channels"] * mult for mult in unet_params["channel_mult"]]
+
+ down_block_types = []
+ resolution = 1
+ for i in range(len(block_out_channels)):
+ block_type = "CrossAttnDownBlock2D" if resolution in unet_params["attention_resolutions"] else "DownBlock2D"
+ down_block_types.append(block_type)
+ if i != len(block_out_channels) - 1:
+ resolution *= 2
+
+ up_block_types = []
+ for i in range(len(block_out_channels)):
+ block_type = "CrossAttnUpBlock2D" if resolution in unet_params["attention_resolutions"] else "UpBlock2D"
+ up_block_types.append(block_type)
+ resolution //= 2
+
+ if unet_params["transformer_depth"] is not None:
+ transformer_layers_per_block = (
+ unet_params["transformer_depth"]
+ if isinstance(unet_params["transformer_depth"], int)
+ else list(unet_params["transformer_depth"])
+ )
+ else:
+ transformer_layers_per_block = 1
+
+ vae_scale_factor = 2 ** (len(vae_params["ch_mult"]) - 1)
+
+ head_dim = unet_params["num_heads"] if "num_heads" in unet_params else None
+ use_linear_projection = (
+ unet_params["use_linear_in_transformer"] if "use_linear_in_transformer" in unet_params else False
+ )
+ if use_linear_projection:
+ # stable diffusion 2-base-512 and 2-768
+ if head_dim is None:
+ head_dim_mult = unet_params["model_channels"] // unet_params["num_head_channels"]
+ head_dim = [head_dim_mult * c for c in list(unet_params["channel_mult"])]
+
+ class_embed_type = None
+ addition_embed_type = None
+ addition_time_embed_dim = None
+ projection_class_embeddings_input_dim = None
+ context_dim = None
+
+ if unet_params["context_dim"] is not None:
+ context_dim = (
+ unet_params["context_dim"]
+ if isinstance(unet_params["context_dim"], int)
+ else unet_params["context_dim"][0]
+ )
+
+ if "num_classes" in unet_params:
+ if unet_params["num_classes"] == "sequential":
+ if context_dim in [2048, 1280]:
+ # SDXL
+ addition_embed_type = "text_time"
+ addition_time_embed_dim = 256
+ else:
+ class_embed_type = "projection"
+ assert "adm_in_channels" in unet_params
+ projection_class_embeddings_input_dim = unet_params["adm_in_channels"]
+
+ config = {
+ "sample_size": image_size // vae_scale_factor,
+ "in_channels": unet_params["in_channels"],
+ "down_block_types": down_block_types,
+ "block_out_channels": block_out_channels,
+ "layers_per_block": unet_params["num_res_blocks"],
+ "cross_attention_dim": context_dim,
+ "attention_head_dim": head_dim,
+ "use_linear_projection": use_linear_projection,
+ "class_embed_type": class_embed_type,
+ "addition_embed_type": addition_embed_type,
+ "addition_time_embed_dim": addition_time_embed_dim,
+ "projection_class_embeddings_input_dim": projection_class_embeddings_input_dim,
+ "transformer_layers_per_block": transformer_layers_per_block,
+ }
+
+ if "disable_self_attentions" in unet_params:
+ config["only_cross_attention"] = unet_params["disable_self_attentions"]
+
+ if "num_classes" in unet_params and isinstance(unet_params["num_classes"], int):
+ config["num_class_embeds"] = unet_params["num_classes"]
+
+ config["out_channels"] = unet_params["out_channels"]
+ config["up_block_types"] = up_block_types
+
+ return config
+
+
+def create_controlnet_diffusers_config(original_config, image_size: int):
+ unet_params = original_config["model"]["params"]["control_stage_config"]["params"]
+ diffusers_unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
+
+ controlnet_config = {
+ "conditioning_channels": unet_params["hint_channels"],
+ "in_channels": diffusers_unet_config["in_channels"],
+ "down_block_types": diffusers_unet_config["down_block_types"],
+ "block_out_channels": diffusers_unet_config["block_out_channels"],
+ "layers_per_block": diffusers_unet_config["layers_per_block"],
+ "cross_attention_dim": diffusers_unet_config["cross_attention_dim"],
+ "attention_head_dim": diffusers_unet_config["attention_head_dim"],
+ "use_linear_projection": diffusers_unet_config["use_linear_projection"],
+ "class_embed_type": diffusers_unet_config["class_embed_type"],
+ "addition_embed_type": diffusers_unet_config["addition_embed_type"],
+ "addition_time_embed_dim": diffusers_unet_config["addition_time_embed_dim"],
+ "projection_class_embeddings_input_dim": diffusers_unet_config["projection_class_embeddings_input_dim"],
+ "transformer_layers_per_block": diffusers_unet_config["transformer_layers_per_block"],
+ }
+
+ return controlnet_config
+
+
+def create_vae_diffusers_config(original_config, image_size, scaling_factor=None, latents_mean=None, latents_std=None):
+ """
+ Creates a config for the diffusers based on the config of the LDM model.
+ """
+ vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"]
+ if (scaling_factor is None) and (latents_mean is not None) and (latents_std is not None):
+ scaling_factor = PLAYGROUND_VAE_SCALING_FACTOR
+ elif (scaling_factor is None) and ("scale_factor" in original_config["model"]["params"]):
+ scaling_factor = original_config["model"]["params"]["scale_factor"]
+ elif scaling_factor is None:
+ scaling_factor = LDM_VAE_DEFAULT_SCALING_FACTOR
+
+ block_out_channels = [vae_params["ch"] * mult for mult in vae_params["ch_mult"]]
+ down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels)
+ up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels)
+
+ config = {
+ "sample_size": image_size,
+ "in_channels": vae_params["in_channels"],
+ "out_channels": vae_params["out_ch"],
+ "down_block_types": down_block_types,
+ "up_block_types": up_block_types,
+ "block_out_channels": block_out_channels,
+ "latent_channels": vae_params["z_channels"],
+ "layers_per_block": vae_params["num_res_blocks"],
+ "scaling_factor": scaling_factor,
+ }
+ if latents_mean is not None and latents_std is not None:
+ config.update({"latents_mean": latents_mean, "latents_std": latents_std})
+
+ return config
+
+
+def update_unet_resnet_ldm_to_diffusers(ldm_keys, new_checkpoint, checkpoint, mapping=None):
+ for ldm_key in ldm_keys:
+ diffusers_key = (
+ ldm_key.replace("in_layers.0", "norm1")
+ .replace("in_layers.2", "conv1")
+ .replace("out_layers.0", "norm2")
+ .replace("out_layers.3", "conv2")
+ .replace("emb_layers.1", "time_emb_proj")
+ .replace("skip_connection", "conv_shortcut")
+ )
+ if mapping:
+ diffusers_key = diffusers_key.replace(mapping["old"], mapping["new"])
+ new_checkpoint[diffusers_key] = checkpoint.pop(ldm_key)
+
+
+def update_unet_attention_ldm_to_diffusers(ldm_keys, new_checkpoint, checkpoint, mapping):
+ for ldm_key in ldm_keys:
+ diffusers_key = ldm_key.replace(mapping["old"], mapping["new"])
+ new_checkpoint[diffusers_key] = checkpoint.pop(ldm_key)
+
+
+def convert_ldm_unet_checkpoint(checkpoint, config, extract_ema=False):
+ """
+ Takes a state dict and a config, and returns a converted checkpoint.
+ """
+ # extract state_dict for UNet
+ unet_state_dict = {}
+ keys = list(checkpoint.keys())
+ unet_key = LDM_UNET_KEY
+
+ # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
+ if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema:
+ logger.warning("Checkpoint has both EMA and non-EMA weights.")
+ logger.warning(
+ "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
+ " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
+ )
+ for key in keys:
+ if key.startswith("model.diffusion_model"):
+ flat_ema_key = "model_ema." + "".join(key.split(".")[1:])
+ unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key)
+ else:
+ if sum(k.startswith("model_ema") for k in keys) > 100:
+ logger.warning(
+ "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
+ " weights (usually better for inference), please make sure to add the `--extract_ema` flag."
+ )
+ for key in keys:
+ if key.startswith(unet_key):
+ unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key)
+
+ new_checkpoint = {}
+ ldm_unet_keys = DIFFUSERS_TO_LDM_MAPPING["unet"]["layers"]
+ for diffusers_key, ldm_key in ldm_unet_keys.items():
+ if ldm_key not in unet_state_dict:
+ continue
+ new_checkpoint[diffusers_key] = unet_state_dict[ldm_key]
+
+ if ("class_embed_type" in config) and (config["class_embed_type"] in ["timestep", "projection"]):
+ class_embed_keys = DIFFUSERS_TO_LDM_MAPPING["unet"]["class_embed_type"]
+ for diffusers_key, ldm_key in class_embed_keys.items():
+ new_checkpoint[diffusers_key] = unet_state_dict[ldm_key]
+
+ if ("addition_embed_type" in config) and (config["addition_embed_type"] == "text_time"):
+ addition_embed_keys = DIFFUSERS_TO_LDM_MAPPING["unet"]["addition_embed_type"]
+ for diffusers_key, ldm_key in addition_embed_keys.items():
+ new_checkpoint[diffusers_key] = unet_state_dict[ldm_key]
+
+ # Relevant to StableDiffusionUpscalePipeline
+ if "num_class_embeds" in config:
+ if (config["num_class_embeds"] is not None) and ("label_emb.weight" in unet_state_dict):
+ new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"]
+
+ # Retrieves the keys for the input blocks only
+ num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer})
+ input_blocks = {
+ layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key]
+ for layer_id in range(num_input_blocks)
+ }
+
+ # Retrieves the keys for the middle blocks only
+ num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer})
+ middle_blocks = {
+ layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key]
+ for layer_id in range(num_middle_blocks)
+ }
+
+ # Retrieves the keys for the output blocks only
+ num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer})
+ output_blocks = {
+ layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key]
+ for layer_id in range(num_output_blocks)
+ }
+
+ # Down blocks
+ for i in range(1, num_input_blocks):
+ block_id = (i - 1) // (config["layers_per_block"] + 1)
+ layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)
+
+ resnets = [
+ key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
+ ]
+ update_unet_resnet_ldm_to_diffusers(
+ resnets,
+ new_checkpoint,
+ unet_state_dict,
+ {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"},
+ )
+
+ if f"input_blocks.{i}.0.op.weight" in unet_state_dict:
+ new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop(
+ f"input_blocks.{i}.0.op.weight"
+ )
+ new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(
+ f"input_blocks.{i}.0.op.bias"
+ )
+
+ attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
+ if attentions:
+ update_unet_attention_ldm_to_diffusers(
+ attentions,
+ new_checkpoint,
+ unet_state_dict,
+ {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"},
+ )
+
+ # Mid blocks
+ resnet_0 = middle_blocks[0]
+ attentions = middle_blocks[1]
+ resnet_1 = middle_blocks[2]
+
+ update_unet_resnet_ldm_to_diffusers(
+ resnet_0, new_checkpoint, unet_state_dict, mapping={"old": "middle_block.0", "new": "mid_block.resnets.0"}
+ )
+ update_unet_resnet_ldm_to_diffusers(
+ resnet_1, new_checkpoint, unet_state_dict, mapping={"old": "middle_block.2", "new": "mid_block.resnets.1"}
+ )
+ update_unet_attention_ldm_to_diffusers(
+ attentions, new_checkpoint, unet_state_dict, mapping={"old": "middle_block.1", "new": "mid_block.attentions.0"}
+ )
+
+ # Up Blocks
+ for i in range(num_output_blocks):
+ block_id = i // (config["layers_per_block"] + 1)
+ layer_in_block_id = i % (config["layers_per_block"] + 1)
+
+ resnets = [
+ key for key in output_blocks[i] if f"output_blocks.{i}.0" in key and f"output_blocks.{i}.0.op" not in key
+ ]
+ update_unet_resnet_ldm_to_diffusers(
+ resnets,
+ new_checkpoint,
+ unet_state_dict,
+ {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"},
+ )
+
+ attentions = [
+ key for key in output_blocks[i] if f"output_blocks.{i}.1" in key and f"output_blocks.{i}.1.conv" not in key
+ ]
+ if attentions:
+ update_unet_attention_ldm_to_diffusers(
+ attentions,
+ new_checkpoint,
+ unet_state_dict,
+ {"old": f"output_blocks.{i}.1", "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}"},
+ )
+
+ if f"output_blocks.{i}.1.conv.weight" in unet_state_dict:
+ new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
+ f"output_blocks.{i}.1.conv.weight"
+ ]
+ new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
+ f"output_blocks.{i}.1.conv.bias"
+ ]
+ if f"output_blocks.{i}.2.conv.weight" in unet_state_dict:
+ new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
+ f"output_blocks.{i}.2.conv.weight"
+ ]
+ new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
+ f"output_blocks.{i}.2.conv.bias"
+ ]
+
+ return new_checkpoint
+
+
+def convert_controlnet_checkpoint(
+ checkpoint,
+ config,
+):
+ # Some controlnet ckpt files are distributed independently from the rest of the
+ # model components i.e. https://huggingface.co/thibaud/controlnet-sd21/
+ if "time_embed.0.weight" in checkpoint:
+ controlnet_state_dict = checkpoint
+
+ else:
+ controlnet_state_dict = {}
+ keys = list(checkpoint.keys())
+ controlnet_key = LDM_CONTROLNET_KEY
+ for key in keys:
+ if key.startswith(controlnet_key):
+ controlnet_state_dict[key.replace(controlnet_key, "")] = checkpoint.pop(key)
+
+ new_checkpoint = {}
+ ldm_controlnet_keys = DIFFUSERS_TO_LDM_MAPPING["controlnet"]["layers"]
+ for diffusers_key, ldm_key in ldm_controlnet_keys.items():
+ if ldm_key not in controlnet_state_dict:
+ continue
+ new_checkpoint[diffusers_key] = controlnet_state_dict[ldm_key]
+
+ # Retrieves the keys for the input blocks only
+ num_input_blocks = len(
+ {".".join(layer.split(".")[:2]) for layer in controlnet_state_dict if "input_blocks" in layer}
+ )
+ input_blocks = {
+ layer_id: [key for key in controlnet_state_dict if f"input_blocks.{layer_id}" in key]
+ for layer_id in range(num_input_blocks)
+ }
+
+ # Down blocks
+ for i in range(1, num_input_blocks):
+ block_id = (i - 1) // (config["layers_per_block"] + 1)
+ layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)
+
+ resnets = [
+ key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
+ ]
+ update_unet_resnet_ldm_to_diffusers(
+ resnets,
+ new_checkpoint,
+ controlnet_state_dict,
+ {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"},
+ )
+
+ if f"input_blocks.{i}.0.op.weight" in controlnet_state_dict:
+ new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = controlnet_state_dict.pop(
+ f"input_blocks.{i}.0.op.weight"
+ )
+ new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = controlnet_state_dict.pop(
+ f"input_blocks.{i}.0.op.bias"
+ )
+
+ attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
+ if attentions:
+ update_unet_attention_ldm_to_diffusers(
+ attentions,
+ new_checkpoint,
+ controlnet_state_dict,
+ {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"},
+ )
+
+ # controlnet down blocks
+ for i in range(num_input_blocks):
+ new_checkpoint[f"controlnet_down_blocks.{i}.weight"] = controlnet_state_dict.pop(f"zero_convs.{i}.0.weight")
+ new_checkpoint[f"controlnet_down_blocks.{i}.bias"] = controlnet_state_dict.pop(f"zero_convs.{i}.0.bias")
+
+ # Retrieves the keys for the middle blocks only
+ num_middle_blocks = len(
+ {".".join(layer.split(".")[:2]) for layer in controlnet_state_dict if "middle_block" in layer}
+ )
+ middle_blocks = {
+ layer_id: [key for key in controlnet_state_dict if f"middle_block.{layer_id}" in key]
+ for layer_id in range(num_middle_blocks)
+ }
+ if middle_blocks:
+ resnet_0 = middle_blocks[0]
+ attentions = middle_blocks[1]
+ resnet_1 = middle_blocks[2]
+
+ update_unet_resnet_ldm_to_diffusers(
+ resnet_0,
+ new_checkpoint,
+ controlnet_state_dict,
+ mapping={"old": "middle_block.0", "new": "mid_block.resnets.0"},
+ )
+ update_unet_resnet_ldm_to_diffusers(
+ resnet_1,
+ new_checkpoint,
+ controlnet_state_dict,
+ mapping={"old": "middle_block.2", "new": "mid_block.resnets.1"},
+ )
+ update_unet_attention_ldm_to_diffusers(
+ attentions,
+ new_checkpoint,
+ controlnet_state_dict,
+ mapping={"old": "middle_block.1", "new": "mid_block.attentions.0"},
+ )
+
+ # mid block
+ new_checkpoint["controlnet_mid_block.weight"] = controlnet_state_dict.pop("middle_block_out.0.weight")
+ new_checkpoint["controlnet_mid_block.bias"] = controlnet_state_dict.pop("middle_block_out.0.bias")
+
+ # controlnet cond embedding blocks
+ cond_embedding_blocks = {
+ ".".join(layer.split(".")[:2])
+ for layer in controlnet_state_dict
+ if "input_hint_block" in layer and ("input_hint_block.0" not in layer) and ("input_hint_block.14" not in layer)
+ }
+ num_cond_embedding_blocks = len(cond_embedding_blocks)
+
+ for idx in range(1, num_cond_embedding_blocks + 1):
+ diffusers_idx = idx - 1
+ cond_block_id = 2 * idx
+
+ new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_idx}.weight"] = controlnet_state_dict.pop(
+ f"input_hint_block.{cond_block_id}.weight"
+ )
+ new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_idx}.bias"] = controlnet_state_dict.pop(
+ f"input_hint_block.{cond_block_id}.bias"
+ )
+
+ return new_checkpoint
+
+
+def create_diffusers_controlnet_model_from_ldm(
+ pipeline_class_name, original_config, checkpoint, upcast_attention=False, image_size=None, torch_dtype=None
+):
+ # import here to avoid circular imports
+ from ..models import ControlNetModel
+
+ image_size = set_image_size(pipeline_class_name, original_config, checkpoint, image_size=image_size)
+
+ diffusers_config = create_controlnet_diffusers_config(original_config, image_size=image_size)
+ diffusers_config["upcast_attention"] = upcast_attention
+
+ diffusers_format_controlnet_checkpoint = convert_controlnet_checkpoint(checkpoint, diffusers_config)
+
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+ with ctx():
+ controlnet = ControlNetModel(**diffusers_config)
+
+ if is_accelerate_available():
+ from ..models.modeling_utils import load_model_dict_into_meta
+
+ unexpected_keys = load_model_dict_into_meta(
+ controlnet, diffusers_format_controlnet_checkpoint, dtype=torch_dtype
+ )
+ if controlnet._keys_to_ignore_on_load_unexpected is not None:
+ for pat in controlnet._keys_to_ignore_on_load_unexpected:
+ unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+ if len(unexpected_keys) > 0:
+ logger.warn(
+ f"Some weights of the model checkpoint were not used when initializing {controlnet.__name__}: \n {[', '.join(unexpected_keys)]}"
+ )
+ else:
+ controlnet.load_state_dict(diffusers_format_controlnet_checkpoint)
+
+ if torch_dtype is not None:
+ controlnet = controlnet.to(torch_dtype)
+
+ return {"controlnet": controlnet}
+
+
+def update_vae_resnet_ldm_to_diffusers(keys, new_checkpoint, checkpoint, mapping):
+ for ldm_key in keys:
+ diffusers_key = ldm_key.replace(mapping["old"], mapping["new"]).replace("nin_shortcut", "conv_shortcut")
+ new_checkpoint[diffusers_key] = checkpoint.pop(ldm_key)
+
+
+def update_vae_attentions_ldm_to_diffusers(keys, new_checkpoint, checkpoint, mapping):
+ for ldm_key in keys:
+ diffusers_key = (
+ ldm_key.replace(mapping["old"], mapping["new"])
+ .replace("norm.weight", "group_norm.weight")
+ .replace("norm.bias", "group_norm.bias")
+ .replace("q.weight", "to_q.weight")
+ .replace("q.bias", "to_q.bias")
+ .replace("k.weight", "to_k.weight")
+ .replace("k.bias", "to_k.bias")
+ .replace("v.weight", "to_v.weight")
+ .replace("v.bias", "to_v.bias")
+ .replace("proj_out.weight", "to_out.0.weight")
+ .replace("proj_out.bias", "to_out.0.bias")
+ )
+ new_checkpoint[diffusers_key] = checkpoint.pop(ldm_key)
+
+ # proj_attn.weight has to be converted from conv 1D to linear
+ shape = new_checkpoint[diffusers_key].shape
+
+ if len(shape) == 3:
+ new_checkpoint[diffusers_key] = new_checkpoint[diffusers_key][:, :, 0]
+ elif len(shape) == 4:
+ new_checkpoint[diffusers_key] = new_checkpoint[diffusers_key][:, :, 0, 0]
+
+
+def convert_ldm_vae_checkpoint(checkpoint, config):
+ # extract state dict for VAE
+ # remove the LDM_VAE_KEY prefix from the ldm checkpoint keys so that it is easier to map them to diffusers keys
+ vae_state_dict = {}
+ keys = list(checkpoint.keys())
+ vae_key = LDM_VAE_KEY if any(k.startswith(LDM_VAE_KEY) for k in keys) else ""
+ for key in keys:
+ if key.startswith(vae_key):
+ vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key)
+
+ new_checkpoint = {}
+ vae_diffusers_ldm_map = DIFFUSERS_TO_LDM_MAPPING["vae"]
+ for diffusers_key, ldm_key in vae_diffusers_ldm_map.items():
+ if ldm_key not in vae_state_dict:
+ continue
+ new_checkpoint[diffusers_key] = vae_state_dict[ldm_key]
+
+ # Retrieves the keys for the encoder down blocks only
+ num_down_blocks = len(config["down_block_types"])
+ down_blocks = {
+ layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)
+ }
+
+ for i in range(num_down_blocks):
+ resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
+ update_vae_resnet_ldm_to_diffusers(
+ resnets,
+ new_checkpoint,
+ vae_state_dict,
+ mapping={"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"},
+ )
+ if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
+ new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(
+ f"encoder.down.{i}.downsample.conv.weight"
+ )
+ new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop(
+ f"encoder.down.{i}.downsample.conv.bias"
+ )
+
+ mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
+ num_mid_res_blocks = 2
+ for i in range(1, num_mid_res_blocks + 1):
+ resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]
+ update_vae_resnet_ldm_to_diffusers(
+ resnets,
+ new_checkpoint,
+ vae_state_dict,
+ mapping={"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"},
+ )
+
+ mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
+ update_vae_attentions_ldm_to_diffusers(
+ mid_attentions, new_checkpoint, vae_state_dict, mapping={"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+ )
+
+ # Retrieves the keys for the decoder up blocks only
+ num_up_blocks = len(config["up_block_types"])
+ up_blocks = {
+ layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)
+ }
+
+ for i in range(num_up_blocks):
+ block_id = num_up_blocks - 1 - i
+ resnets = [
+ key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
+ ]
+ update_vae_resnet_ldm_to_diffusers(
+ resnets,
+ new_checkpoint,
+ vae_state_dict,
+ mapping={"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"},
+ )
+ if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
+ new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
+ f"decoder.up.{block_id}.upsample.conv.weight"
+ ]
+ new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[
+ f"decoder.up.{block_id}.upsample.conv.bias"
+ ]
+
+ mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
+ num_mid_res_blocks = 2
+ for i in range(1, num_mid_res_blocks + 1):
+ resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]
+ update_vae_resnet_ldm_to_diffusers(
+ resnets,
+ new_checkpoint,
+ vae_state_dict,
+ mapping={"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"},
+ )
+
+ mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key]
+ update_vae_attentions_ldm_to_diffusers(
+ mid_attentions, new_checkpoint, vae_state_dict, mapping={"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+ )
+ conv_attn_to_linear(new_checkpoint)
+
+ return new_checkpoint
+
+
+def create_text_encoder_from_ldm_clip_checkpoint(config_name, checkpoint, local_files_only=False, torch_dtype=None):
+ try:
+ config = CLIPTextConfig.from_pretrained(config_name, local_files_only=local_files_only)
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the configuration in the following path: 'openai/clip-vit-large-patch14'."
+ )
+
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+ with ctx():
+ text_model = CLIPTextModel(config)
+
+ keys = list(checkpoint.keys())
+ text_model_dict = {}
+
+ remove_prefixes = LDM_CLIP_PREFIX_TO_REMOVE
+
+ for key in keys:
+ for prefix in remove_prefixes:
+ if key.startswith(prefix):
+ diffusers_key = key.replace(prefix, "")
+ text_model_dict[diffusers_key] = checkpoint[key]
+
+ if is_accelerate_available():
+ from ..models.modeling_utils import load_model_dict_into_meta
+
+ unexpected_keys = load_model_dict_into_meta(text_model, text_model_dict, dtype=torch_dtype)
+ if text_model._keys_to_ignore_on_load_unexpected is not None:
+ for pat in text_model._keys_to_ignore_on_load_unexpected:
+ unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+ if len(unexpected_keys) > 0:
+ logger.warn(
+ f"Some weights of the model checkpoint were not used when initializing {text_model.__class__.__name__}: \n {[', '.join(unexpected_keys)]}"
+ )
+ else:
+ if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)):
+ text_model_dict.pop("text_model.embeddings.position_ids", None)
+
+ text_model.load_state_dict(text_model_dict)
+
+ if torch_dtype is not None:
+ text_model = text_model.to(torch_dtype)
+
+ return text_model
+
+
+def create_text_encoder_from_open_clip_checkpoint(
+ config_name,
+ checkpoint,
+ prefix="cond_stage_model.model.",
+ has_projection=False,
+ local_files_only=False,
+ torch_dtype=None,
+ **config_kwargs,
+):
+ try:
+ config = CLIPTextConfig.from_pretrained(config_name, **config_kwargs, local_files_only=local_files_only)
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the configuration in the following path: '{config_name}'."
+ )
+
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+ with ctx():
+ text_model = CLIPTextModelWithProjection(config) if has_projection else CLIPTextModel(config)
+
+ text_model_dict = {}
+ text_proj_key = prefix + "text_projection"
+ text_proj_dim = (
+ int(checkpoint[text_proj_key].shape[0]) if text_proj_key in checkpoint else LDM_OPEN_CLIP_TEXT_PROJECTION_DIM
+ )
+ text_model_dict["text_model.embeddings.position_ids"] = text_model.text_model.embeddings.get_buffer("position_ids")
+
+ keys = list(checkpoint.keys())
+ keys_to_ignore = SD_2_TEXT_ENCODER_KEYS_TO_IGNORE
+
+ openclip_diffusers_ldm_map = DIFFUSERS_TO_LDM_MAPPING["openclip"]["layers"]
+ for diffusers_key, ldm_key in openclip_diffusers_ldm_map.items():
+ ldm_key = prefix + ldm_key
+ if ldm_key not in checkpoint:
+ continue
+ if ldm_key in keys_to_ignore:
+ continue
+ if ldm_key.endswith("text_projection"):
+ text_model_dict[diffusers_key] = checkpoint[ldm_key].T.contiguous()
+ else:
+ text_model_dict[diffusers_key] = checkpoint[ldm_key]
+
+ for key in keys:
+ if key in keys_to_ignore:
+ continue
+
+ if not key.startswith(prefix + "transformer."):
+ continue
+
+ diffusers_key = key.replace(prefix + "transformer.", "")
+ transformer_diffusers_to_ldm_map = DIFFUSERS_TO_LDM_MAPPING["openclip"]["transformer"]
+ for new_key, old_key in transformer_diffusers_to_ldm_map.items():
+ diffusers_key = (
+ diffusers_key.replace(old_key, new_key).replace(".in_proj_weight", "").replace(".in_proj_bias", "")
+ )
+
+ if key.endswith(".in_proj_weight"):
+ weight_value = checkpoint[key]
+
+ text_model_dict[diffusers_key + ".q_proj.weight"] = weight_value[:text_proj_dim, :]
+ text_model_dict[diffusers_key + ".k_proj.weight"] = weight_value[text_proj_dim : text_proj_dim * 2, :]
+ text_model_dict[diffusers_key + ".v_proj.weight"] = weight_value[text_proj_dim * 2 :, :]
+
+ elif key.endswith(".in_proj_bias"):
+ weight_value = checkpoint[key]
+ text_model_dict[diffusers_key + ".q_proj.bias"] = weight_value[:text_proj_dim]
+ text_model_dict[diffusers_key + ".k_proj.bias"] = weight_value[text_proj_dim : text_proj_dim * 2]
+ text_model_dict[diffusers_key + ".v_proj.bias"] = weight_value[text_proj_dim * 2 :]
+ else:
+ text_model_dict[diffusers_key] = checkpoint[key]
+
+ if is_accelerate_available():
+ from ..models.modeling_utils import load_model_dict_into_meta
+
+ unexpected_keys = load_model_dict_into_meta(text_model, text_model_dict, dtype=torch_dtype)
+ if text_model._keys_to_ignore_on_load_unexpected is not None:
+ for pat in text_model._keys_to_ignore_on_load_unexpected:
+ unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+ if len(unexpected_keys) > 0:
+ logger.warn(
+ f"Some weights of the model checkpoint were not used when initializing {text_model.__class__.__name__}: \n {[', '.join(unexpected_keys)]}"
+ )
+
+ else:
+ if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)):
+ text_model_dict.pop("text_model.embeddings.position_ids", None)
+
+ text_model.load_state_dict(text_model_dict)
+
+ if torch_dtype is not None:
+ text_model = text_model.to(torch_dtype)
+
+ return text_model
+
+
+def create_diffusers_unet_model_from_ldm(
+ pipeline_class_name,
+ original_config,
+ checkpoint,
+ num_in_channels=None,
+ upcast_attention=False,
+ extract_ema=False,
+ image_size=None,
+ torch_dtype=None,
+ model_type=None,
+):
+ from ..models import UNet2DConditionModel
+
+ if num_in_channels is None:
+ if pipeline_class_name in [
+ "StableDiffusionInpaintPipeline",
+ "StableDiffusionControlNetInpaintPipeline",
+ "StableDiffusionXLInpaintPipeline",
+ "StableDiffusionXLControlNetInpaintPipeline",
+ ]:
+ num_in_channels = 9
+
+ elif pipeline_class_name == "StableDiffusionUpscalePipeline":
+ num_in_channels = 7
+
+ else:
+ num_in_channels = 4
+
+ image_size = set_image_size(
+ pipeline_class_name, original_config, checkpoint, image_size=image_size, model_type=model_type
+ )
+ unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
+ unet_config["in_channels"] = num_in_channels
+ unet_config["upcast_attention"] = upcast_attention
+
+ diffusers_format_unet_checkpoint = convert_ldm_unet_checkpoint(checkpoint, unet_config, extract_ema=extract_ema)
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+
+ with ctx():
+ unet = UNet2DConditionModel(**unet_config)
+
+ if is_accelerate_available():
+ from ..models.modeling_utils import load_model_dict_into_meta
+
+ unexpected_keys = load_model_dict_into_meta(unet, diffusers_format_unet_checkpoint, dtype=torch_dtype)
+ if unet._keys_to_ignore_on_load_unexpected is not None:
+ for pat in unet._keys_to_ignore_on_load_unexpected:
+ unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+ if len(unexpected_keys) > 0:
+ logger.warn(
+ f"Some weights of the model checkpoint were not used when initializing {unet.__name__}: \n {[', '.join(unexpected_keys)]}"
+ )
+ else:
+ unet.load_state_dict(diffusers_format_unet_checkpoint)
+
+ if torch_dtype is not None:
+ unet = unet.to(torch_dtype)
+
+ return {"unet": unet}
+
+
+def create_diffusers_vae_model_from_ldm(
+ pipeline_class_name,
+ original_config,
+ checkpoint,
+ image_size=None,
+ scaling_factor=None,
+ torch_dtype=None,
+ model_type=None,
+):
+ # import here to avoid circular imports
+ from ..models import AutoencoderKL
+
+ image_size = set_image_size(
+ pipeline_class_name, original_config, checkpoint, image_size=image_size, model_type=model_type
+ )
+ model_type = infer_model_type(original_config, checkpoint, model_type)
+
+ if model_type == "Playground":
+ edm_mean = (
+ checkpoint["edm_mean"].to(dtype=torch_dtype).tolist() if torch_dtype else checkpoint["edm_mean"].tolist()
+ )
+ edm_std = (
+ checkpoint["edm_std"].to(dtype=torch_dtype).tolist() if torch_dtype else checkpoint["edm_std"].tolist()
+ )
+ else:
+ edm_mean = None
+ edm_std = None
+
+ vae_config = create_vae_diffusers_config(
+ original_config,
+ image_size=image_size,
+ scaling_factor=scaling_factor,
+ latents_mean=edm_mean,
+ latents_std=edm_std,
+ )
+ diffusers_format_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+
+ with ctx():
+ vae = AutoencoderKL(**vae_config)
+
+ if is_accelerate_available():
+ from ..models.modeling_utils import load_model_dict_into_meta
+
+ unexpected_keys = load_model_dict_into_meta(vae, diffusers_format_vae_checkpoint, dtype=torch_dtype)
+ if vae._keys_to_ignore_on_load_unexpected is not None:
+ for pat in vae._keys_to_ignore_on_load_unexpected:
+ unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+ if len(unexpected_keys) > 0:
+ logger.warn(
+ f"Some weights of the model checkpoint were not used when initializing {vae.__name__}: \n {[', '.join(unexpected_keys)]}"
+ )
+ else:
+ vae.load_state_dict(diffusers_format_vae_checkpoint)
+
+ if torch_dtype is not None:
+ vae = vae.to(torch_dtype)
+
+ return {"vae": vae}
+
+
+def create_text_encoders_and_tokenizers_from_ldm(
+ original_config,
+ checkpoint,
+ model_type=None,
+ local_files_only=False,
+ torch_dtype=None,
+):
+ model_type = infer_model_type(original_config, checkpoint=checkpoint, model_type=model_type)
+
+ if model_type == "FrozenOpenCLIPEmbedder":
+ config_name = "stabilityai/stable-diffusion-2"
+ config_kwargs = {"subfolder": "text_encoder"}
+
+ try:
+ text_encoder = create_text_encoder_from_open_clip_checkpoint(
+ config_name, checkpoint, local_files_only=local_files_only, torch_dtype=torch_dtype, **config_kwargs
+ )
+ tokenizer = CLIPTokenizer.from_pretrained(
+ config_name, subfolder="tokenizer", local_files_only=local_files_only
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the text_encoder in the following path: '{config_name}'."
+ )
+ else:
+ return {"text_encoder": text_encoder, "tokenizer": tokenizer}
+
+ elif model_type == "FrozenCLIPEmbedder":
+ try:
+ config_name = "openai/clip-vit-large-patch14"
+ text_encoder = create_text_encoder_from_ldm_clip_checkpoint(
+ config_name,
+ checkpoint,
+ local_files_only=local_files_only,
+ torch_dtype=torch_dtype,
+ )
+ tokenizer = CLIPTokenizer.from_pretrained(config_name, local_files_only=local_files_only)
+
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: '{config_name}'."
+ )
+ else:
+ return {"text_encoder": text_encoder, "tokenizer": tokenizer}
+
+ elif model_type == "SDXL-Refiner":
+ config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
+ config_kwargs = {"projection_dim": 1280}
+ prefix = "conditioner.embedders.0.model."
+
+ try:
+ tokenizer_2 = CLIPTokenizer.from_pretrained(config_name, pad_token="!", local_files_only=local_files_only)
+ text_encoder_2 = create_text_encoder_from_open_clip_checkpoint(
+ config_name,
+ checkpoint,
+ prefix=prefix,
+ has_projection=True,
+ local_files_only=local_files_only,
+ torch_dtype=torch_dtype,
+ **config_kwargs,
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the text_encoder_2 and tokenizer_2 in the following path: {config_name} with `pad_token` set to '!'."
+ )
+
+ else:
+ return {
+ "text_encoder": None,
+ "tokenizer": None,
+ "tokenizer_2": tokenizer_2,
+ "text_encoder_2": text_encoder_2,
+ }
+
+ elif model_type in ["SDXL", "Playground"]:
+ try:
+ config_name = "openai/clip-vit-large-patch14"
+ tokenizer = CLIPTokenizer.from_pretrained(config_name, local_files_only=local_files_only)
+ text_encoder = create_text_encoder_from_ldm_clip_checkpoint(
+ config_name, checkpoint, local_files_only=local_files_only, torch_dtype=torch_dtype
+ )
+
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the text_encoder and tokenizer in the following path: 'openai/clip-vit-large-patch14'."
+ )
+
+ try:
+ config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
+ config_kwargs = {"projection_dim": 1280}
+ prefix = "conditioner.embedders.1.model."
+ tokenizer_2 = CLIPTokenizer.from_pretrained(config_name, pad_token="!", local_files_only=local_files_only)
+ text_encoder_2 = create_text_encoder_from_open_clip_checkpoint(
+ config_name,
+ checkpoint,
+ prefix=prefix,
+ has_projection=True,
+ local_files_only=local_files_only,
+ torch_dtype=torch_dtype,
+ **config_kwargs,
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the text_encoder_2 and tokenizer_2 in the following path: {config_name} with `pad_token` set to '!'."
+ )
+
+ return {
+ "tokenizer": tokenizer,
+ "text_encoder": text_encoder,
+ "tokenizer_2": tokenizer_2,
+ "text_encoder_2": text_encoder_2,
+ }
+
+ return
+
+
+def create_scheduler_from_ldm(
+ pipeline_class_name,
+ original_config,
+ checkpoint,
+ prediction_type=None,
+ scheduler_type="ddim",
+ model_type=None,
+):
+ scheduler_config = get_default_scheduler_config()
+ model_type = infer_model_type(original_config, checkpoint=checkpoint, model_type=model_type)
+
+ global_step = checkpoint["global_step"] if "global_step" in checkpoint else None
+
+ num_train_timesteps = getattr(original_config["model"]["params"], "timesteps", None) or 1000
+ scheduler_config["num_train_timesteps"] = num_train_timesteps
+
+ if (
+ "parameterization" in original_config["model"]["params"]
+ and original_config["model"]["params"]["parameterization"] == "v"
+ ):
+ if prediction_type is None:
+ # NOTE: For stable diffusion 2 base it is recommended to pass `prediction_type=="epsilon"`
+ # as it relies on a brittle global step parameter here
+ prediction_type = "epsilon" if global_step == 875000 else "v_prediction"
+
+ else:
+ prediction_type = prediction_type or "epsilon"
+
+ scheduler_config["prediction_type"] = prediction_type
+
+ if model_type in ["SDXL", "SDXL-Refiner"]:
+ scheduler_type = "euler"
+ elif model_type == "Playground":
+ scheduler_type = "edm_dpm_solver_multistep"
+ else:
+ beta_start = original_config["model"]["params"].get("linear_start", 0.02)
+ beta_end = original_config["model"]["params"].get("linear_end", 0.085)
+ scheduler_config["beta_start"] = beta_start
+ scheduler_config["beta_end"] = beta_end
+ scheduler_config["beta_schedule"] = "scaled_linear"
+ scheduler_config["clip_sample"] = False
+ scheduler_config["set_alpha_to_one"] = False
+
+ if scheduler_type == "pndm":
+ scheduler_config["skip_prk_steps"] = True
+ scheduler = PNDMScheduler.from_config(scheduler_config)
+
+ elif scheduler_type == "lms":
+ scheduler = LMSDiscreteScheduler.from_config(scheduler_config)
+
+ elif scheduler_type == "heun":
+ scheduler = HeunDiscreteScheduler.from_config(scheduler_config)
+
+ elif scheduler_type == "euler":
+ scheduler = EulerDiscreteScheduler.from_config(scheduler_config)
+
+ elif scheduler_type == "euler-ancestral":
+ scheduler = EulerAncestralDiscreteScheduler.from_config(scheduler_config)
+
+ elif scheduler_type == "dpm":
+ scheduler = DPMSolverMultistepScheduler.from_config(scheduler_config)
+
+ elif scheduler_type == "ddim":
+ scheduler = DDIMScheduler.from_config(scheduler_config)
+
+ elif scheduler_type == "edm_dpm_solver_multistep":
+ scheduler_config = {
+ "algorithm_type": "dpmsolver++",
+ "dynamic_thresholding_ratio": 0.995,
+ "euler_at_final": False,
+ "final_sigmas_type": "zero",
+ "lower_order_final": True,
+ "num_train_timesteps": 1000,
+ "prediction_type": "epsilon",
+ "rho": 7.0,
+ "sample_max_value": 1.0,
+ "sigma_data": 0.5,
+ "sigma_max": 80.0,
+ "sigma_min": 0.002,
+ "solver_order": 2,
+ "solver_type": "midpoint",
+ "thresholding": False,
+ }
+ scheduler = EDMDPMSolverMultistepScheduler(**scheduler_config)
+
+ else:
+ raise ValueError(f"Scheduler of type {scheduler_type} doesn't exist!")
+
+ if pipeline_class_name == "StableDiffusionUpscalePipeline":
+ scheduler = DDIMScheduler.from_pretrained("stabilityai/stable-diffusion-x4-upscaler", subfolder="scheduler")
+ low_res_scheduler = DDPMScheduler.from_pretrained(
+ "stabilityai/stable-diffusion-x4-upscaler", subfolder="low_res_scheduler"
+ )
+
+ return {
+ "scheduler": scheduler,
+ "low_res_scheduler": low_res_scheduler,
+ }
+
+ return {"scheduler": scheduler}
diff --git a/model/diffusers_c/loaders/textual_inversion.py b/model/diffusers_c/loaders/textual_inversion.py
new file mode 100644
index 0000000000000000000000000000000000000000..aaaf4b68bb5f03a2723ea689095a874d580dffb6
--- /dev/null
+++ b/model/diffusers_c/loaders/textual_inversion.py
@@ -0,0 +1,562 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, List, Optional, Union
+
+import safetensors
+import torch
+from huggingface_hub.utils import validate_hf_hub_args
+from torch import nn
+
+from ..utils import _get_model_file, is_accelerate_available, is_transformers_available, logging
+
+
+if is_transformers_available():
+ from transformers import PreTrainedModel, PreTrainedTokenizer
+
+if is_accelerate_available():
+ from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+TEXT_INVERSION_NAME = "learned_embeds.bin"
+TEXT_INVERSION_NAME_SAFE = "learned_embeds.safetensors"
+
+
+@validate_hf_hub_args
+def load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs):
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ local_files_only = kwargs.pop("local_files_only", None)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ subfolder = kwargs.pop("subfolder", None)
+ weight_name = kwargs.pop("weight_name", None)
+ use_safetensors = kwargs.pop("use_safetensors", None)
+
+ allow_pickle = False
+ if use_safetensors is None:
+ use_safetensors = True
+ allow_pickle = True
+
+ user_agent = {
+ "file_type": "text_inversion",
+ "framework": "pytorch",
+ }
+ state_dicts = []
+ for pretrained_model_name_or_path in pretrained_model_name_or_paths:
+ if not isinstance(pretrained_model_name_or_path, (dict, torch.Tensor)):
+ # 3.1. Load textual inversion file
+ model_file = None
+
+ # Let's first try to load .safetensors weights
+ if (use_safetensors and weight_name is None) or (
+ weight_name is not None and weight_name.endswith(".safetensors")
+ ):
+ try:
+ model_file = _get_model_file(
+ pretrained_model_name_or_path,
+ weights_name=weight_name or TEXT_INVERSION_NAME_SAFE,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ )
+ state_dict = safetensors.torch.load_file(model_file, device="cpu")
+ except Exception as e:
+ if not allow_pickle:
+ raise e
+
+ model_file = None
+
+ if model_file is None:
+ model_file = _get_model_file(
+ pretrained_model_name_or_path,
+ weights_name=weight_name or TEXT_INVERSION_NAME,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ )
+ state_dict = torch.load(model_file, map_location="cpu")
+ else:
+ state_dict = pretrained_model_name_or_path
+
+ state_dicts.append(state_dict)
+
+ return state_dicts
+
+
+class TextualInversionLoaderMixin:
+ r"""
+ Load Textual Inversion tokens and embeddings to the tokenizer and text encoder.
+ """
+
+ def maybe_convert_prompt(self, prompt: Union[str, List[str]], tokenizer: "PreTrainedTokenizer"): # noqa: F821
+ r"""
+ Processes prompts that include a special token corresponding to a multi-vector textual inversion embedding to
+ be replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+ inversion token or if the textual inversion token is a single vector, the input prompt is returned.
+
+ Parameters:
+ prompt (`str` or list of `str`):
+ The prompt or prompts to guide the image generation.
+ tokenizer (`PreTrainedTokenizer`):
+ The tokenizer responsible for encoding the prompt into input tokens.
+
+ Returns:
+ `str` or list of `str`: The converted prompt
+ """
+ if not isinstance(prompt, List):
+ prompts = [prompt]
+ else:
+ prompts = prompt
+
+ prompts = [self._maybe_convert_prompt(p, tokenizer) for p in prompts]
+
+ if not isinstance(prompt, List):
+ return prompts[0]
+
+ return prompts
+
+ def _maybe_convert_prompt(self, prompt: str, tokenizer: "PreTrainedTokenizer"): # noqa: F821
+ r"""
+ Maybe convert a prompt into a "multi vector"-compatible prompt. If the prompt includes a token that corresponds
+ to a multi-vector textual inversion embedding, this function will process the prompt so that the special token
+ is replaced with multiple special tokens each corresponding to one of the vectors. If the prompt has no textual
+ inversion token or a textual inversion token that is a single vector, the input prompt is simply returned.
+
+ Parameters:
+ prompt (`str`):
+ The prompt to guide the image generation.
+ tokenizer (`PreTrainedTokenizer`):
+ The tokenizer responsible for encoding the prompt into input tokens.
+
+ Returns:
+ `str`: The converted prompt
+ """
+ tokens = tokenizer.tokenize(prompt)
+ unique_tokens = set(tokens)
+ for token in unique_tokens:
+ if token in tokenizer.added_tokens_encoder:
+ replacement = token
+ i = 1
+ while f"{token}_{i}" in tokenizer.added_tokens_encoder:
+ replacement += f" {token}_{i}"
+ i += 1
+
+ prompt = prompt.replace(token, replacement)
+
+ return prompt
+
+ def _check_text_inv_inputs(self, tokenizer, text_encoder, pretrained_model_name_or_paths, tokens):
+ if tokenizer is None:
+ raise ValueError(
+ f"{self.__class__.__name__} requires `self.tokenizer` or passing a `tokenizer` of type `PreTrainedTokenizer` for calling"
+ f" `{self.load_textual_inversion.__name__}`"
+ )
+
+ if text_encoder is None:
+ raise ValueError(
+ f"{self.__class__.__name__} requires `self.text_encoder` or passing a `text_encoder` of type `PreTrainedModel` for calling"
+ f" `{self.load_textual_inversion.__name__}`"
+ )
+
+ if len(pretrained_model_name_or_paths) > 1 and len(pretrained_model_name_or_paths) != len(tokens):
+ raise ValueError(
+ f"You have passed a list of models of length {len(pretrained_model_name_or_paths)}, and list of tokens of length {len(tokens)} "
+ f"Make sure both lists have the same length."
+ )
+
+ valid_tokens = [t for t in tokens if t is not None]
+ if len(set(valid_tokens)) < len(valid_tokens):
+ raise ValueError(f"You have passed a list of tokens that contains duplicates: {tokens}")
+
+ @staticmethod
+ def _retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer):
+ all_tokens = []
+ all_embeddings = []
+ for state_dict, token in zip(state_dicts, tokens):
+ if isinstance(state_dict, torch.Tensor):
+ if token is None:
+ raise ValueError(
+ "You are trying to load a textual inversion embedding that has been saved as a PyTorch tensor. Make sure to pass the name of the corresponding token in this case: `token=...`."
+ )
+ loaded_token = token
+ embedding = state_dict
+ elif len(state_dict) == 1:
+ # diffusers
+ loaded_token, embedding = next(iter(state_dict.items()))
+ elif "string_to_param" in state_dict:
+ # A1111
+ loaded_token = state_dict["name"]
+ embedding = state_dict["string_to_param"]["*"]
+ else:
+ raise ValueError(
+ f"Loaded state dictionary is incorrect: {state_dict}. \n\n"
+ "Please verify that the loaded state dictionary of the textual embedding either only has a single key or includes the `string_to_param`"
+ " input key."
+ )
+
+ if token is not None and loaded_token != token:
+ logger.info(f"The loaded token: {loaded_token} is overwritten by the passed token {token}.")
+ else:
+ token = loaded_token
+
+ if token in tokenizer.get_vocab():
+ raise ValueError(
+ f"Token {token} already in tokenizer vocabulary. Please choose a different token name or remove {token} and embedding from the tokenizer and text encoder."
+ )
+
+ all_tokens.append(token)
+ all_embeddings.append(embedding)
+
+ return all_tokens, all_embeddings
+
+ @staticmethod
+ def _extend_tokens_and_embeddings(tokens, embeddings, tokenizer):
+ all_tokens = []
+ all_embeddings = []
+
+ for embedding, token in zip(embeddings, tokens):
+ if f"{token}_1" in tokenizer.get_vocab():
+ multi_vector_tokens = [token]
+ i = 1
+ while f"{token}_{i}" in tokenizer.added_tokens_encoder:
+ multi_vector_tokens.append(f"{token}_{i}")
+ i += 1
+
+ raise ValueError(
+ f"Multi-vector Token {multi_vector_tokens} already in tokenizer vocabulary. Please choose a different token name or remove the {multi_vector_tokens} and embedding from the tokenizer and text encoder."
+ )
+
+ is_multi_vector = len(embedding.shape) > 1 and embedding.shape[0] > 1
+ if is_multi_vector:
+ all_tokens += [token] + [f"{token}_{i}" for i in range(1, embedding.shape[0])]
+ all_embeddings += [e for e in embedding] # noqa: C416
+ else:
+ all_tokens += [token]
+ all_embeddings += [embedding[0]] if len(embedding.shape) > 1 else [embedding]
+
+ return all_tokens, all_embeddings
+
+ @validate_hf_hub_args
+ def load_textual_inversion(
+ self,
+ pretrained_model_name_or_path: Union[str, List[str], Dict[str, torch.Tensor], List[Dict[str, torch.Tensor]]],
+ token: Optional[Union[str, List[str]]] = None,
+ tokenizer: Optional["PreTrainedTokenizer"] = None, # noqa: F821
+ text_encoder: Optional["PreTrainedModel"] = None, # noqa: F821
+ **kwargs,
+ ):
+ r"""
+ Load Textual Inversion embeddings into the text encoder of [`StableDiffusionPipeline`] (both 🤗 Diffusers and
+ Automatic1111 formats are supported).
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike` or `List[str or os.PathLike]` or `Dict` or `List[Dict]`):
+ Can be either one of the following or a list of them:
+
+ - A string, the *model id* (for example `sd-concepts-library/low-poly-hd-logos-icons`) of a
+ pretrained model hosted on the Hub.
+ - A path to a *directory* (for example `./my_text_inversion_directory/`) containing the textual
+ inversion weights.
+ - A path to a *file* (for example `./my_text_inversions.pt`) containing textual inversion weights.
+ - A [torch state
+ dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+ token (`str` or `List[str]`, *optional*):
+ Override the token to use for the textual inversion weights. If `pretrained_model_name_or_path` is a
+ list, then `token` must also be a list of equal length.
+ text_encoder ([`~transformers.CLIPTextModel`], *optional*):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ If not specified, function will take self.tokenizer.
+ tokenizer ([`~transformers.CLIPTokenizer`], *optional*):
+ A `CLIPTokenizer` to tokenize text. If not specified, function will take self.tokenizer.
+ weight_name (`str`, *optional*):
+ Name of a custom weight file. This should be used when:
+
+ - The saved textual inversion file is in 🤗 Diffusers format, but was saved under a specific weight
+ name such as `text_inv.bin`.
+ - The saved textual inversion file is in the Automatic1111 format.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ subfolder (`str`, *optional*, defaults to `""`):
+ The subfolder location of a model file within a larger model repository on the Hub or locally.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+
+ Example:
+
+ To load a Textual Inversion embedding vector in 🤗 Diffusers format:
+
+ ```py
+ from diffusers import StableDiffusionPipeline
+ import torch
+
+ model_id = "runwayml/stable-diffusion-v1-5"
+ pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
+
+ pipe.load_textual_inversion("sd-concepts-library/cat-toy")
+
+ prompt = "A backpack"
+
+ image = pipe(prompt, num_inference_steps=50).images[0]
+ image.save("cat-backpack.png")
+ ```
+
+ To load a Textual Inversion embedding vector in Automatic1111 format, make sure to download the vector first
+ (for example from [civitAI](https://civitai.com/models/3036?modelVersionId=9857)) and then load the vector
+ locally:
+
+ ```py
+ from diffusers import StableDiffusionPipeline
+ import torch
+
+ model_id = "runwayml/stable-diffusion-v1-5"
+ pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to("cuda")
+
+ pipe.load_textual_inversion("./charturnerv2.pt", token="charturnerv2")
+
+ prompt = "charturnerv2, multiple views of the same character in the same outfit, a character turnaround of a woman wearing a black jacket and red shirt, best quality, intricate details."
+
+ image = pipe(prompt, num_inference_steps=50).images[0]
+ image.save("character.png")
+ ```
+
+ """
+ # 1. Set correct tokenizer and text encoder
+ tokenizer = tokenizer or getattr(self, "tokenizer", None)
+ text_encoder = text_encoder or getattr(self, "text_encoder", None)
+
+ # 2. Normalize inputs
+ pretrained_model_name_or_paths = (
+ [pretrained_model_name_or_path]
+ if not isinstance(pretrained_model_name_or_path, list)
+ else pretrained_model_name_or_path
+ )
+ tokens = [token] if not isinstance(token, list) else token
+ if tokens[0] is None:
+ tokens = tokens * len(pretrained_model_name_or_paths)
+
+ # 3. Check inputs
+ self._check_text_inv_inputs(tokenizer, text_encoder, pretrained_model_name_or_paths, tokens)
+
+ # 4. Load state dicts of textual embeddings
+ state_dicts = load_textual_inversion_state_dicts(pretrained_model_name_or_paths, **kwargs)
+
+ # 4.1 Handle the special case when state_dict is a tensor that contains n embeddings for n tokens
+ if len(tokens) > 1 and len(state_dicts) == 1:
+ if isinstance(state_dicts[0], torch.Tensor):
+ state_dicts = list(state_dicts[0])
+ if len(tokens) != len(state_dicts):
+ raise ValueError(
+ f"You have passed a state_dict contains {len(state_dicts)} embeddings, and list of tokens of length {len(tokens)} "
+ f"Make sure both have the same length."
+ )
+
+ # 4. Retrieve tokens and embeddings
+ tokens, embeddings = self._retrieve_tokens_and_embeddings(tokens, state_dicts, tokenizer)
+
+ # 5. Extend tokens and embeddings for multi vector
+ tokens, embeddings = self._extend_tokens_and_embeddings(tokens, embeddings, tokenizer)
+
+ # 6. Make sure all embeddings have the correct size
+ expected_emb_dim = text_encoder.get_input_embeddings().weight.shape[-1]
+ if any(expected_emb_dim != emb.shape[-1] for emb in embeddings):
+ raise ValueError(
+ "Loaded embeddings are of incorrect shape. Expected each textual inversion embedding "
+ "to be of shape {input_embeddings.shape[-1]}, but are {embeddings.shape[-1]} "
+ )
+
+ # 7. Now we can be sure that loading the embedding matrix works
+ # < Unsafe code:
+
+ # 7.1 Offload all hooks in case the pipeline was cpu offloaded before make sure, we offload and onload again
+ is_model_cpu_offload = False
+ is_sequential_cpu_offload = False
+ for _, component in self.components.items():
+ if isinstance(component, nn.Module):
+ if hasattr(component, "_hf_hook"):
+ is_model_cpu_offload = isinstance(getattr(component, "_hf_hook"), CpuOffload)
+ is_sequential_cpu_offload = isinstance(getattr(component, "_hf_hook"), AlignDevicesHook)
+ logger.info(
+ "Accelerate hooks detected. Since you have called `load_textual_inversion()`, the previous hooks will be first removed. Then the textual inversion parameters will be loaded and the hooks will be applied again."
+ )
+ remove_hook_from_module(component, recurse=is_sequential_cpu_offload)
+
+ # 7.2 save expected device and dtype
+ device = text_encoder.device
+ dtype = text_encoder.dtype
+
+ # 7.3 Increase token embedding matrix
+ text_encoder.resize_token_embeddings(len(tokenizer) + len(tokens))
+ input_embeddings = text_encoder.get_input_embeddings().weight
+
+ # 7.4 Load token and embedding
+ for token, embedding in zip(tokens, embeddings):
+ # add tokens and get ids
+ tokenizer.add_tokens(token)
+ token_id = tokenizer.convert_tokens_to_ids(token)
+ input_embeddings.data[token_id] = embedding
+ logger.info(f"Loaded textual inversion embedding for {token}.")
+
+ input_embeddings.to(dtype=dtype, device=device)
+
+ # 7.5 Offload the model again
+ if is_model_cpu_offload:
+ self.enable_model_cpu_offload()
+ elif is_sequential_cpu_offload:
+ self.enable_sequential_cpu_offload()
+
+ # / Unsafe Code >
+
+ def unload_textual_inversion(
+ self,
+ tokens: Optional[Union[str, List[str]]] = None,
+ tokenizer: Optional["PreTrainedTokenizer"] = None,
+ text_encoder: Optional["PreTrainedModel"] = None,
+ ):
+ r"""
+ Unload Textual Inversion embeddings from the text encoder of [`StableDiffusionPipeline`]
+
+ Example:
+ ```py
+ from diffusers import AutoPipelineForText2Image
+ import torch
+
+ pipeline = AutoPipelineForText2Image.from_pretrained("runwayml/stable-diffusion-v1-5")
+
+ # Example 1
+ pipeline.load_textual_inversion("sd-concepts-library/gta5-artwork")
+ pipeline.load_textual_inversion("sd-concepts-library/moeb-style")
+
+ # Remove all token embeddings
+ pipeline.unload_textual_inversion()
+
+ # Example 2
+ pipeline.load_textual_inversion("sd-concepts-library/moeb-style")
+ pipeline.load_textual_inversion("sd-concepts-library/gta5-artwork")
+
+ # Remove just one token
+ pipeline.unload_textual_inversion("")
+
+ # Example 3: unload from SDXL
+ pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0")
+ embedding_path = hf_hub_download(repo_id="linoyts/web_y2k", filename="web_y2k_emb.safetensors", repo_type="model")
+
+ # load embeddings to the text encoders
+ state_dict = load_file(embedding_path)
+
+ # load embeddings of text_encoder 1 (CLIP ViT-L/14)
+ pipeline.load_textual_inversion(state_dict["clip_l"], token=["", ""], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
+ # load embeddings of text_encoder 2 (CLIP ViT-G/14)
+ pipeline.load_textual_inversion(state_dict["clip_g"], token=["", ""], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
+
+ # Unload explicitly from both text encoders abd tokenizers
+ pipeline.unload_textual_inversion(tokens=["", ""], text_encoder=pipeline.text_encoder, tokenizer=pipeline.tokenizer)
+ pipeline.unload_textual_inversion(tokens=["", ""], text_encoder=pipeline.text_encoder_2, tokenizer=pipeline.tokenizer_2)
+
+ ```
+ """
+
+ tokenizer = tokenizer or getattr(self, "tokenizer", None)
+ text_encoder = text_encoder or getattr(self, "text_encoder", None)
+
+ # Get textual inversion tokens and ids
+ token_ids = []
+ last_special_token_id = None
+
+ if tokens:
+ if isinstance(tokens, str):
+ tokens = [tokens]
+ for added_token_id, added_token in tokenizer.added_tokens_decoder.items():
+ if not added_token.special:
+ if added_token.content in tokens:
+ token_ids.append(added_token_id)
+ else:
+ last_special_token_id = added_token_id
+ if len(token_ids) == 0:
+ raise ValueError("No tokens to remove found")
+ else:
+ tokens = []
+ for added_token_id, added_token in tokenizer.added_tokens_decoder.items():
+ if not added_token.special:
+ token_ids.append(added_token_id)
+ tokens.append(added_token.content)
+ else:
+ last_special_token_id = added_token_id
+
+ # Delete from tokenizer
+ for token_id, token_to_remove in zip(token_ids, tokens):
+ del tokenizer._added_tokens_decoder[token_id]
+ del tokenizer._added_tokens_encoder[token_to_remove]
+
+ # Make all token ids sequential in tokenizer
+ key_id = 1
+ for token_id in tokenizer.added_tokens_decoder:
+ if token_id > last_special_token_id and token_id > last_special_token_id + key_id:
+ token = tokenizer._added_tokens_decoder[token_id]
+ tokenizer._added_tokens_decoder[last_special_token_id + key_id] = token
+ del tokenizer._added_tokens_decoder[token_id]
+ tokenizer._added_tokens_encoder[token.content] = last_special_token_id + key_id
+ key_id += 1
+ tokenizer._update_trie()
+
+ # Delete from text encoder
+ text_embedding_dim = text_encoder.get_input_embeddings().embedding_dim
+ temp_text_embedding_weights = text_encoder.get_input_embeddings().weight
+ text_embedding_weights = temp_text_embedding_weights[: last_special_token_id + 1]
+ to_append = []
+ for i in range(last_special_token_id + 1, temp_text_embedding_weights.shape[0]):
+ if i not in token_ids:
+ to_append.append(temp_text_embedding_weights[i].unsqueeze(0))
+ if len(to_append) > 0:
+ to_append = torch.cat(to_append, dim=0)
+ text_embedding_weights = torch.cat([text_embedding_weights, to_append], dim=0)
+ text_embeddings_filtered = nn.Embedding(text_embedding_weights.shape[0], text_embedding_dim)
+ text_embeddings_filtered.weight.data = text_embedding_weights
+ text_encoder.set_input_embeddings(text_embeddings_filtered)
diff --git a/model/diffusers_c/loaders/unet.py b/model/diffusers_c/loaders/unet.py
new file mode 100644
index 0000000000000000000000000000000000000000..9d8e2666c518b0ff3aa53931d29ab6348b3f94c9
--- /dev/null
+++ b/model/diffusers_c/loaders/unet.py
@@ -0,0 +1,898 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import inspect
+import os
+from collections import defaultdict
+from contextlib import nullcontext
+from functools import partial
+from pathlib import Path
+from typing import Callable, Dict, List, Optional, Union
+
+import safetensors
+import torch
+import torch.nn.functional as F
+from huggingface_hub.utils import validate_hf_hub_args
+from torch import nn
+
+from ..models.embeddings import (
+ ImageProjection,
+ IPAdapterFullImageProjection,
+ IPAdapterPlusImageProjection,
+ MultiIPAdapterImageProjection,
+)
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
+from ..utils import (
+ USE_PEFT_BACKEND,
+ _get_model_file,
+ delete_adapter_layers,
+ is_accelerate_available,
+ is_torch_version,
+ logging,
+ set_adapter_layers,
+ set_weights_and_activate_adapters,
+)
+from .utils import AttnProcsLayers
+
+
+if is_accelerate_available():
+ from accelerate import init_empty_weights
+ from accelerate.hooks import AlignDevicesHook, CpuOffload, remove_hook_from_module
+
+logger = logging.get_logger(__name__)
+
+
+TEXT_ENCODER_NAME = "text_encoder"
+UNET_NAME = "unet"
+
+LORA_WEIGHT_NAME = "pytorch_lora_weights.bin"
+LORA_WEIGHT_NAME_SAFE = "pytorch_lora_weights.safetensors"
+
+CUSTOM_DIFFUSION_WEIGHT_NAME = "pytorch_custom_diffusion_weights.bin"
+CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE = "pytorch_custom_diffusion_weights.safetensors"
+
+
+class UNet2DConditionLoadersMixin:
+ """
+ Load LoRA layers into a [`UNet2DCondtionModel`].
+ """
+
+ text_encoder_name = TEXT_ENCODER_NAME
+ unet_name = UNET_NAME
+
+ @validate_hf_hub_args
+ def load_attn_procs(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs):
+ r"""
+ Load pretrained attention processor layers into [`UNet2DConditionModel`]. Attention processor layers have to be
+ defined in
+ [`attention_processor.py`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py)
+ and be a `torch.nn.Module` class.
+
+ Parameters:
+ pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+ Can be either:
+
+ - A string, the model id (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+ the Hub.
+ - A path to a directory (for example `./my_model_directory`) containing the model weights saved
+ with [`ModelMixin.save_pretrained`].
+ - A [torch state
+ dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ subfolder (`str`, *optional*, defaults to `""`):
+ The subfolder location of a model file within a larger model repository on the Hub or locally.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+
+ Example:
+
+ ```py
+ from diffusers import AutoPipelineForText2Image
+ import torch
+
+ pipeline = AutoPipelineForText2Image.from_pretrained(
+ "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+ ).to("cuda")
+ pipeline.unet.load_attn_procs(
+ "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+ )
+ ```
+ """
+ from ..models.attention_processor import CustomDiffusionAttnProcessor
+ from ..models.lora import LoRACompatibleConv, LoRACompatibleLinear, LoRAConv2dLayer, LoRALinearLayer
+
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ local_files_only = kwargs.pop("local_files_only", None)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ subfolder = kwargs.pop("subfolder", None)
+ weight_name = kwargs.pop("weight_name", None)
+ use_safetensors = kwargs.pop("use_safetensors", None)
+ low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+ # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
+ # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+ network_alphas = kwargs.pop("network_alphas", None)
+
+ _pipeline = kwargs.pop("_pipeline", None)
+
+ is_network_alphas_none = network_alphas is None
+
+ allow_pickle = False
+
+ if use_safetensors is None:
+ use_safetensors = True
+ allow_pickle = True
+
+ user_agent = {
+ "file_type": "attn_procs_weights",
+ "framework": "pytorch",
+ }
+
+ model_file = None
+ if not isinstance(pretrained_model_name_or_path_or_dict, dict):
+ # Let's first try to load .safetensors weights
+ if (use_safetensors and weight_name is None) or (
+ weight_name is not None and weight_name.endswith(".safetensors")
+ ):
+ try:
+ model_file = _get_model_file(
+ pretrained_model_name_or_path_or_dict,
+ weights_name=weight_name or LORA_WEIGHT_NAME_SAFE,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ )
+ state_dict = safetensors.torch.load_file(model_file, device="cpu")
+ except IOError as e:
+ if not allow_pickle:
+ raise e
+ # try loading non-safetensors weights
+ pass
+ if model_file is None:
+ model_file = _get_model_file(
+ pretrained_model_name_or_path_or_dict,
+ weights_name=weight_name or LORA_WEIGHT_NAME,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ )
+ state_dict = torch.load(model_file, map_location="cpu")
+ else:
+ state_dict = pretrained_model_name_or_path_or_dict
+
+ # fill attn processors
+ lora_layers_list = []
+
+ is_lora = all(("lora" in k or k.endswith(".alpha")) for k in state_dict.keys()) and not USE_PEFT_BACKEND
+ is_custom_diffusion = any("custom_diffusion" in k for k in state_dict.keys())
+
+ if is_lora:
+ # correct keys
+ state_dict, network_alphas = self.convert_state_dict_legacy_attn_format(state_dict, network_alphas)
+
+ if network_alphas is not None:
+ network_alphas_keys = list(network_alphas.keys())
+ used_network_alphas_keys = set()
+
+ lora_grouped_dict = defaultdict(dict)
+ mapped_network_alphas = {}
+
+ all_keys = list(state_dict.keys())
+ for key in all_keys:
+ value = state_dict.pop(key)
+ attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+ lora_grouped_dict[attn_processor_key][sub_key] = value
+
+ # Create another `mapped_network_alphas` dictionary so that we can properly map them.
+ if network_alphas is not None:
+ for k in network_alphas_keys:
+ if k.replace(".alpha", "") in key:
+ mapped_network_alphas.update({attn_processor_key: network_alphas.get(k)})
+ used_network_alphas_keys.add(k)
+
+ if not is_network_alphas_none:
+ if len(set(network_alphas_keys) - used_network_alphas_keys) > 0:
+ raise ValueError(
+ f"The `network_alphas` has to be empty at this point but has the following keys \n\n {', '.join(network_alphas.keys())}"
+ )
+
+ if len(state_dict) > 0:
+ raise ValueError(
+ f"The `state_dict` has to be empty at this point but has the following keys \n\n {', '.join(state_dict.keys())}"
+ )
+
+ for key, value_dict in lora_grouped_dict.items():
+ attn_processor = self
+ for sub_key in key.split("."):
+ attn_processor = getattr(attn_processor, sub_key)
+
+ # Process non-attention layers, which don't have to_{k,v,q,out_proj}_lora layers
+ # or add_{k,v,q,out_proj}_proj_lora layers.
+ rank = value_dict["lora.down.weight"].shape[0]
+
+ if isinstance(attn_processor, LoRACompatibleConv):
+ in_features = attn_processor.in_channels
+ out_features = attn_processor.out_channels
+ kernel_size = attn_processor.kernel_size
+
+ ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+ with ctx():
+ lora = LoRAConv2dLayer(
+ in_features=in_features,
+ out_features=out_features,
+ rank=rank,
+ kernel_size=kernel_size,
+ stride=attn_processor.stride,
+ padding=attn_processor.padding,
+ network_alpha=mapped_network_alphas.get(key),
+ )
+ elif isinstance(attn_processor, LoRACompatibleLinear):
+ ctx = init_empty_weights if low_cpu_mem_usage else nullcontext
+ with ctx():
+ lora = LoRALinearLayer(
+ attn_processor.in_features,
+ attn_processor.out_features,
+ rank,
+ mapped_network_alphas.get(key),
+ )
+ else:
+ raise ValueError(f"Module {key} is not a LoRACompatibleConv or LoRACompatibleLinear module.")
+
+ value_dict = {k.replace("lora.", ""): v for k, v in value_dict.items()}
+ lora_layers_list.append((attn_processor, lora))
+
+ if low_cpu_mem_usage:
+ device = next(iter(value_dict.values())).device
+ dtype = next(iter(value_dict.values())).dtype
+ load_model_dict_into_meta(lora, value_dict, device=device, dtype=dtype)
+ else:
+ lora.load_state_dict(value_dict)
+
+ elif is_custom_diffusion:
+ attn_processors = {}
+ custom_diffusion_grouped_dict = defaultdict(dict)
+ for key, value in state_dict.items():
+ if len(value) == 0:
+ custom_diffusion_grouped_dict[key] = {}
+ else:
+ if "to_out" in key:
+ attn_processor_key, sub_key = ".".join(key.split(".")[:-3]), ".".join(key.split(".")[-3:])
+ else:
+ attn_processor_key, sub_key = ".".join(key.split(".")[:-2]), ".".join(key.split(".")[-2:])
+ custom_diffusion_grouped_dict[attn_processor_key][sub_key] = value
+
+ for key, value_dict in custom_diffusion_grouped_dict.items():
+ if len(value_dict) == 0:
+ attn_processors[key] = CustomDiffusionAttnProcessor(
+ train_kv=False, train_q_out=False, hidden_size=None, cross_attention_dim=None
+ )
+ else:
+ cross_attention_dim = value_dict["to_k_custom_diffusion.weight"].shape[1]
+ hidden_size = value_dict["to_k_custom_diffusion.weight"].shape[0]
+ train_q_out = True if "to_q_custom_diffusion.weight" in value_dict else False
+ attn_processors[key] = CustomDiffusionAttnProcessor(
+ train_kv=True,
+ train_q_out=train_q_out,
+ hidden_size=hidden_size,
+ cross_attention_dim=cross_attention_dim,
+ )
+ attn_processors[key].load_state_dict(value_dict)
+ elif USE_PEFT_BACKEND:
+ # In that case we have nothing to do as loading the adapter weights is already handled above by `set_peft_model_state_dict`
+ # on the Unet
+ pass
+ else:
+ raise ValueError(
+ f"{model_file} does not seem to be in the correct format expected by LoRA or Custom Diffusion training."
+ )
+
+ #
+
+ def convert_state_dict_legacy_attn_format(self, state_dict, network_alphas):
+ is_new_lora_format = all(
+ key.startswith(self.unet_name) or key.startswith(self.text_encoder_name) for key in state_dict.keys()
+ )
+ if is_new_lora_format:
+ # Strip the `"unet"` prefix.
+ is_text_encoder_present = any(key.startswith(self.text_encoder_name) for key in state_dict.keys())
+ if is_text_encoder_present:
+ warn_message = "The state_dict contains LoRA params corresponding to the text encoder which are not being used here. To use both UNet and text encoder related LoRA params, use [`pipe.load_lora_weights()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraLoaderMixin.load_lora_weights)."
+ logger.warn(warn_message)
+ unet_keys = [k for k in state_dict.keys() if k.startswith(self.unet_name)]
+ state_dict = {k.replace(f"{self.unet_name}.", ""): v for k, v in state_dict.items() if k in unet_keys}
+
+ # change processor format to 'pure' LoRACompatibleLinear format
+ if any("processor" in k.split(".") for k in state_dict.keys()):
+
+ def format_to_lora_compatible(key):
+ if "processor" not in key.split("."):
+ return key
+ return key.replace(".processor", "").replace("to_out_lora", "to_out.0.lora").replace("_lora", ".lora")
+
+ state_dict = {format_to_lora_compatible(k): v for k, v in state_dict.items()}
+
+ if network_alphas is not None:
+ network_alphas = {format_to_lora_compatible(k): v for k, v in network_alphas.items()}
+ return state_dict, network_alphas
+
+ def save_attn_procs(
+ self,
+ save_directory: Union[str, os.PathLike],
+ is_main_process: bool = True,
+ weight_name: str = None,
+ save_function: Callable = None,
+ safe_serialization: bool = True,
+ **kwargs,
+ ):
+ r"""
+ Save attention processor layers to a directory so that it can be reloaded with the
+ [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`] method.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to save an attention processor to (will be created if it doesn't exist).
+ is_main_process (`bool`, *optional*, defaults to `True`):
+ Whether the process calling this is the main process or not. Useful during distributed training and you
+ need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+ process to avoid race conditions.
+ save_function (`Callable`):
+ The function to use to save the state dictionary. Useful during distributed training when you need to
+ replace `torch.save` with another method. Can be configured with the environment variable
+ `DIFFUSERS_SAVE_MODE`.
+ safe_serialization (`bool`, *optional*, defaults to `True`):
+ Whether to save the model using `safetensors` or with `pickle`.
+
+ Example:
+
+ ```py
+ import torch
+ from diffusers import DiffusionPipeline
+
+ pipeline = DiffusionPipeline.from_pretrained(
+ "CompVis/stable-diffusion-v1-4",
+ torch_dtype=torch.float16,
+ ).to("cuda")
+ pipeline.unet.load_attn_procs("path-to-save-model", weight_name="pytorch_custom_diffusion_weights.bin")
+ pipeline.unet.save_attn_procs("path-to-save-model", weight_name="pytorch_custom_diffusion_weights.bin")
+ ```
+ """
+ from ..models.attention_processor import (
+ CustomDiffusionAttnProcessor,
+ CustomDiffusionAttnProcessor2_0,
+ CustomDiffusionXFormersAttnProcessor,
+ )
+
+ if os.path.isfile(save_directory):
+ logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+ return
+
+ if save_function is None:
+ if safe_serialization:
+
+ def save_function(weights, filename):
+ return safetensors.torch.save_file(weights, filename, metadata={"format": "pt"})
+
+ else:
+ save_function = torch.save
+
+ os.makedirs(save_directory, exist_ok=True)
+
+ is_custom_diffusion = any(
+ isinstance(
+ x,
+ (CustomDiffusionAttnProcessor, CustomDiffusionAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor),
+ )
+ for (_, x) in self.attn_processors.items()
+ )
+ if is_custom_diffusion:
+ model_to_save = AttnProcsLayers(
+ {
+ y: x
+ for (y, x) in self.attn_processors.items()
+ if isinstance(
+ x,
+ (
+ CustomDiffusionAttnProcessor,
+ CustomDiffusionAttnProcessor2_0,
+ CustomDiffusionXFormersAttnProcessor,
+ ),
+ )
+ }
+ )
+ state_dict = model_to_save.state_dict()
+ for name, attn in self.attn_processors.items():
+ if len(attn.state_dict()) == 0:
+ state_dict[name] = {}
+ else:
+ model_to_save = AttnProcsLayers(self.attn_processors)
+ state_dict = model_to_save.state_dict()
+
+ if weight_name is None:
+ if safe_serialization:
+ weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME_SAFE if is_custom_diffusion else LORA_WEIGHT_NAME_SAFE
+ else:
+ weight_name = CUSTOM_DIFFUSION_WEIGHT_NAME if is_custom_diffusion else LORA_WEIGHT_NAME
+
+ # Save the model
+ save_path = Path(save_directory, weight_name).as_posix()
+ save_function(state_dict, save_path)
+ logger.info(f"Model weights saved in {save_path}")
+
+ def fuse_lora(self, lora_scale=1.0, safe_fusing=False, adapter_names=None):
+ self.lora_scale = lora_scale
+ self._safe_fusing = safe_fusing
+ self.apply(partial(self._fuse_lora_apply, adapter_names=adapter_names))
+
+ def _fuse_lora_apply(self, module, adapter_names=None):
+ if not USE_PEFT_BACKEND:
+ if hasattr(module, "_fuse_lora"):
+ module._fuse_lora(self.lora_scale, self._safe_fusing)
+
+ if adapter_names is not None:
+ raise ValueError(
+ "The `adapter_names` argument is not supported in your environment. Please switch"
+ " to PEFT backend to use this argument by installing latest PEFT and transformers."
+ " `pip install -U peft transformers`"
+ )
+ else:
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ merge_kwargs = {"safe_merge": self._safe_fusing}
+
+ if isinstance(module, BaseTunerLayer):
+ if self.lora_scale != 1.0:
+ module.scale_layer(self.lora_scale)
+
+ # For BC with prevous PEFT versions, we need to check the signature
+ # of the `merge` method to see if it supports the `adapter_names` argument.
+ supported_merge_kwargs = list(inspect.signature(module.merge).parameters)
+ if "adapter_names" in supported_merge_kwargs:
+ merge_kwargs["adapter_names"] = adapter_names
+ elif "adapter_names" not in supported_merge_kwargs and adapter_names is not None:
+ raise ValueError(
+ "The `adapter_names` argument is not supported with your PEFT version. Please upgrade"
+ " to the latest version of PEFT. `pip install -U peft`"
+ )
+
+ module.merge(**merge_kwargs)
+
+ def unfuse_lora(self):
+ self.apply(self._unfuse_lora_apply)
+
+ def _unfuse_lora_apply(self, module):
+ if not USE_PEFT_BACKEND:
+ if hasattr(module, "_unfuse_lora"):
+ module._unfuse_lora()
+ else:
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ if isinstance(module, BaseTunerLayer):
+ module.unmerge()
+
+ def set_adapters(
+ self,
+ adapter_names: Union[List[str], str],
+ weights: Optional[Union[List[float], float]] = None,
+ ):
+ """
+ Set the currently active adapters for use in the UNet.
+
+ Args:
+ adapter_names (`List[str]` or `str`):
+ The names of the adapters to use.
+ adapter_weights (`Union[List[float], float]`, *optional*):
+ The adapter(s) weights to use with the UNet. If `None`, the weights are set to `1.0` for all the
+ adapters.
+
+ Example:
+
+ ```py
+ from diffusers import AutoPipelineForText2Image
+ import torch
+
+ pipeline = AutoPipelineForText2Image.from_pretrained(
+ "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+ ).to("cuda")
+ pipeline.load_lora_weights(
+ "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+ )
+ pipeline.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+ pipeline.set_adapters(["cinematic", "pixel"], adapter_weights=[0.5, 0.5])
+ ```
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for `set_adapters()`.")
+
+ adapter_names = [adapter_names] if isinstance(adapter_names, str) else adapter_names
+
+ if weights is None:
+ weights = [1.0] * len(adapter_names)
+ elif isinstance(weights, float):
+ weights = [weights] * len(adapter_names)
+
+ if len(adapter_names) != len(weights):
+ raise ValueError(
+ f"Length of adapter names {len(adapter_names)} is not equal to the length of their weights {len(weights)}."
+ )
+
+ set_weights_and_activate_adapters(self, adapter_names, weights)
+
+ def disable_lora(self):
+ """
+ Disable the UNet's active LoRA layers.
+
+ Example:
+
+ ```py
+ from diffusers import AutoPipelineForText2Image
+ import torch
+
+ pipeline = AutoPipelineForText2Image.from_pretrained(
+ "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+ ).to("cuda")
+ pipeline.load_lora_weights(
+ "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+ )
+ pipeline.disable_lora()
+ ```
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+ set_adapter_layers(self, enabled=False)
+
+ def enable_lora(self):
+ """
+ Enable the UNet's active LoRA layers.
+
+ Example:
+
+ ```py
+ from diffusers import AutoPipelineForText2Image
+ import torch
+
+ pipeline = AutoPipelineForText2Image.from_pretrained(
+ "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+ ).to("cuda")
+ pipeline.load_lora_weights(
+ "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_name="cinematic"
+ )
+ pipeline.enable_lora()
+ ```
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+ set_adapter_layers(self, enabled=True)
+
+ def delete_adapters(self, adapter_names: Union[List[str], str]):
+ """
+ Delete an adapter's LoRA layers from the UNet.
+
+ Args:
+ adapter_names (`Union[List[str], str]`):
+ The names (single string or list of strings) of the adapter to delete.
+
+ Example:
+
+ ```py
+ from diffusers import AutoPipelineForText2Image
+ import torch
+
+ pipeline = AutoPipelineForText2Image.from_pretrained(
+ "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+ ).to("cuda")
+ pipeline.load_lora_weights(
+ "jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", adapter_names="cinematic"
+ )
+ pipeline.delete_adapters("cinematic")
+ ```
+ """
+ if not USE_PEFT_BACKEND:
+ raise ValueError("PEFT backend is required for this method.")
+
+ if isinstance(adapter_names, str):
+ adapter_names = [adapter_names]
+
+ for adapter_name in adapter_names:
+ delete_adapter_layers(self, adapter_name)
+
+ # Pop also the corresponding adapter from the config
+ if hasattr(self, "peft_config"):
+ self.peft_config.pop(adapter_name, None)
+
+ def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict, low_cpu_mem_usage=False):
+ if low_cpu_mem_usage:
+ if is_accelerate_available():
+ from accelerate import init_empty_weights
+
+ else:
+ low_cpu_mem_usage = False
+ logger.warning(
+ "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+ " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+ " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+ " install accelerate\n```\n."
+ )
+
+ if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+ raise NotImplementedError(
+ "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+ " `low_cpu_mem_usage=False`."
+ )
+
+ updated_state_dict = {}
+ image_projection = None
+ init_context = init_empty_weights if low_cpu_mem_usage else nullcontext
+
+ if "proj.weight" in state_dict:
+ # IP-Adapter
+ num_image_text_embeds = 4
+ clip_embeddings_dim = state_dict["proj.weight"].shape[-1]
+ cross_attention_dim = state_dict["proj.weight"].shape[0] // 4
+
+ with init_context():
+ image_projection = ImageProjection(
+ cross_attention_dim=cross_attention_dim,
+ image_embed_dim=clip_embeddings_dim,
+ num_image_text_embeds=num_image_text_embeds,
+ )
+
+ for key, value in state_dict.items():
+ diffusers_name = key.replace("proj", "image_embeds")
+ updated_state_dict[diffusers_name] = value
+
+ elif "proj.3.weight" in state_dict:
+ # IP-Adapter Full
+ clip_embeddings_dim = state_dict["proj.0.weight"].shape[0]
+ cross_attention_dim = state_dict["proj.3.weight"].shape[0]
+
+ with init_context():
+ image_projection = IPAdapterFullImageProjection(
+ cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim
+ )
+
+ for key, value in state_dict.items():
+ diffusers_name = key.replace("proj.0", "ff.net.0.proj")
+ diffusers_name = diffusers_name.replace("proj.2", "ff.net.2")
+ diffusers_name = diffusers_name.replace("proj.3", "norm")
+ updated_state_dict[diffusers_name] = value
+
+ else:
+ # IP-Adapter Plus
+ num_image_text_embeds = state_dict["latents"].shape[1]
+ embed_dims = state_dict["proj_in.weight"].shape[1]
+ output_dims = state_dict["proj_out.weight"].shape[0]
+ hidden_dims = state_dict["latents"].shape[2]
+ heads = state_dict["layers.0.0.to_q.weight"].shape[0] // 64
+
+ with init_context():
+ image_projection = IPAdapterPlusImageProjection(
+ embed_dims=embed_dims,
+ output_dims=output_dims,
+ hidden_dims=hidden_dims,
+ heads=heads,
+ num_queries=num_image_text_embeds,
+ )
+
+ for key, value in state_dict.items():
+ diffusers_name = key.replace("0.to", "2.to")
+ diffusers_name = diffusers_name.replace("1.0.weight", "3.0.weight")
+ diffusers_name = diffusers_name.replace("1.0.bias", "3.0.bias")
+ diffusers_name = diffusers_name.replace("1.1.weight", "3.1.net.0.proj.weight")
+ diffusers_name = diffusers_name.replace("1.3.weight", "3.1.net.2.weight")
+
+ if "norm1" in diffusers_name:
+ updated_state_dict[diffusers_name.replace("0.norm1", "0")] = value
+ elif "norm2" in diffusers_name:
+ updated_state_dict[diffusers_name.replace("0.norm2", "1")] = value
+ elif "to_kv" in diffusers_name:
+ v_chunk = value.chunk(2, dim=0)
+ updated_state_dict[diffusers_name.replace("to_kv", "to_k")] = v_chunk[0]
+ updated_state_dict[diffusers_name.replace("to_kv", "to_v")] = v_chunk[1]
+ elif "to_out" in diffusers_name:
+ updated_state_dict[diffusers_name.replace("to_out", "to_out.0")] = value
+ else:
+ updated_state_dict[diffusers_name] = value
+
+ if not low_cpu_mem_usage:
+ image_projection.load_state_dict(updated_state_dict)
+ else:
+ load_model_dict_into_meta(image_projection, updated_state_dict, device=self.device, dtype=self.dtype)
+
+ return image_projection
+
+ def _convert_ip_adapter_attn_to_diffusers(self, state_dicts, low_cpu_mem_usage=False):
+ from ..models.attention_processor import (
+ AttnProcessor,
+ AttnProcessor2_0,
+ IPAdapterAttnProcessor,
+ IPAdapterAttnProcessor2_0,
+ )
+
+ if low_cpu_mem_usage:
+ if is_accelerate_available():
+ from accelerate import init_empty_weights
+
+ else:
+ low_cpu_mem_usage = False
+ logger.warning(
+ "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+ " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+ " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+ " install accelerate\n```\n."
+ )
+
+ if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+ raise NotImplementedError(
+ "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+ " `low_cpu_mem_usage=False`."
+ )
+
+ # set ip-adapter cross-attention processors & load state_dict
+ attn_procs = {}
+ key_id = 1
+ init_context = init_empty_weights if low_cpu_mem_usage else nullcontext
+ for name in self.attn_processors.keys():
+ cross_attention_dim = None if name.endswith("attn1.processor") else self.config.cross_attention_dim
+ if name.startswith("mid_block"):
+ hidden_size = self.config.block_out_channels[-1]
+ elif name.startswith("up_blocks"):
+ block_id = int(name[len("up_blocks.")])
+ hidden_size = list(reversed(self.config.block_out_channels))[block_id]
+ elif name.startswith("down_blocks"):
+ block_id = int(name[len("down_blocks.")])
+ hidden_size = self.config.block_out_channels[block_id]
+
+ if cross_attention_dim is None or "motion_modules" in name:
+ attn_processor_class = (
+ AttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else AttnProcessor
+ )
+ attn_procs[name] = attn_processor_class()
+ else:
+ attn_processor_class = (
+ IPAdapterAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else IPAdapterAttnProcessor
+ )
+ num_image_text_embeds = []
+ for state_dict in state_dicts:
+ if "proj.weight" in state_dict["image_proj"]:
+ # IP-Adapter
+ num_image_text_embeds += [4]
+ elif "proj.3.weight" in state_dict["image_proj"]:
+ # IP-Adapter Full Face
+ num_image_text_embeds += [257] # 256 CLIP tokens + 1 CLS token
+ else:
+ # IP-Adapter Plus
+ num_image_text_embeds += [state_dict["image_proj"]["latents"].shape[1]]
+
+ with init_context():
+ attn_procs[name] = attn_processor_class(
+ hidden_size=hidden_size,
+ cross_attention_dim=cross_attention_dim,
+ scale=1.0,
+ num_tokens=num_image_text_embeds,
+ )
+
+ value_dict = {}
+ for i, state_dict in enumerate(state_dicts):
+ value_dict.update({f"to_k_ip.{i}.weight": state_dict["ip_adapter"][f"{key_id}.to_k_ip.weight"]})
+ value_dict.update({f"to_v_ip.{i}.weight": state_dict["ip_adapter"][f"{key_id}.to_v_ip.weight"]})
+
+ if not low_cpu_mem_usage:
+ attn_procs[name].load_state_dict(value_dict)
+ else:
+ device = next(iter(value_dict.values())).device
+ dtype = next(iter(value_dict.values())).dtype
+ load_model_dict_into_meta(attn_procs[name], value_dict, device=device, dtype=dtype)
+
+ key_id += 2
+
+ return attn_procs
+
+ def _load_ip_adapter_weights(self, state_dicts, low_cpu_mem_usage=False):
+ if not isinstance(state_dicts, list):
+ state_dicts = [state_dicts]
+ # Set encoder_hid_proj after loading ip_adapter weights,
+ # because `IPAdapterPlusImageProjection` also has `attn_processors`.
+ self.encoder_hid_proj = None
+
+ attn_procs = self._convert_ip_adapter_attn_to_diffusers(state_dicts, low_cpu_mem_usage=low_cpu_mem_usage)
+ self.set_attn_processor(attn_procs)
+
+ # convert IP-Adapter Image Projection layers to diffusers
+ image_projection_layers = []
+ for state_dict in state_dicts:
+ image_projection_layer = self._convert_ip_adapter_image_proj_to_diffusers(
+ state_dict["image_proj"], low_cpu_mem_usage=low_cpu_mem_usage
+ )
+ image_projection_layers.append(image_projection_layer)
+
+ self.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
+ self.config.encoder_hid_dim_type = "ip_image_proj"
+
+ self.to(dtype=self.dtype, device=self.device)
diff --git a/model/diffusers_c/loaders/utils.py b/model/diffusers_c/loaders/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..142d72bf6b77edf4af72ee0d30d3d190cd4b3eef
--- /dev/null
+++ b/model/diffusers_c/loaders/utils.py
@@ -0,0 +1,59 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Dict
+
+import torch
+
+
+class AttnProcsLayers(torch.nn.Module):
+ def __init__(self, state_dict: Dict[str, torch.Tensor]):
+ super().__init__()
+ self.layers = torch.nn.ModuleList(state_dict.values())
+ self.mapping = dict(enumerate(state_dict.keys()))
+ self.rev_mapping = {v: k for k, v in enumerate(state_dict.keys())}
+
+ # .processor for unet, .self_attn for text encoder
+ self.split_keys = [".processor", ".self_attn"]
+
+ # we add a hook to state_dict() and load_state_dict() so that the
+ # naming fits with `unet.attn_processors`
+ def map_to(module, state_dict, *args, **kwargs):
+ new_state_dict = {}
+ for key, value in state_dict.items():
+ num = int(key.split(".")[1]) # 0 is always "layers"
+ new_key = key.replace(f"layers.{num}", module.mapping[num])
+ new_state_dict[new_key] = value
+
+ return new_state_dict
+
+ def remap_key(key, state_dict):
+ for k in self.split_keys:
+ if k in key:
+ return key.split(k)[0] + k
+
+ raise ValueError(
+ f"There seems to be a problem with the state_dict: {set(state_dict.keys())}. {key} has to have one of {self.split_keys}."
+ )
+
+ def map_from(module, state_dict, *args, **kwargs):
+ all_keys = list(state_dict.keys())
+ for key in all_keys:
+ replace_key = remap_key(key, state_dict)
+ new_key = key.replace(replace_key, f"layers.{module.rev_mapping[replace_key]}")
+ state_dict[new_key] = state_dict[key]
+ del state_dict[key]
+
+ self._register_state_dict_hook(map_to)
+ self._register_load_state_dict_pre_hook(map_from, with_module=True)
diff --git a/model/diffusers_c/models/README.md b/model/diffusers_c/models/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..dc57ab4578f259d7f9b5230372ad36bb6bd27353
--- /dev/null
+++ b/model/diffusers_c/models/README.md
@@ -0,0 +1,3 @@
+# Models
+
+For more detail on the models, please refer to the [docs](https://huggingface.co/docs/diffusers/api/models/overview).
diff --git a/model/diffusers_c/models/__init__.py b/model/diffusers_c/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..29475b3954c7cc4bbee97cbb91efc8528843b0d6
--- /dev/null
+++ b/model/diffusers_c/models/__init__.py
@@ -0,0 +1,105 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ..utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ _LazyModule,
+ is_flax_available,
+ is_torch_available,
+)
+
+
+_import_structure = {}
+
+if is_torch_available():
+ _import_structure["adapter"] = ["MultiAdapter", "T2IAdapter"]
+ _import_structure["autoencoders.autoencoder_asym_kl"] = ["AsymmetricAutoencoderKL"]
+ _import_structure["autoencoders.autoencoder_kl"] = ["AutoencoderKL"]
+ _import_structure["autoencoders.autoencoder_kl_temporal_decoder"] = ["AutoencoderKLTemporalDecoder"]
+ _import_structure["autoencoders.autoencoder_tiny"] = ["AutoencoderTiny"]
+ _import_structure["autoencoders.consistency_decoder_vae"] = ["ConsistencyDecoderVAE"]
+ _import_structure["brushnet"] = ["BrushNetModel"]
+ _import_structure["controlnet"] = ["ControlNetModel"]
+ _import_structure["dual_transformer_2d"] = ["DualTransformer2DModel"]
+ _import_structure["embeddings"] = ["ImageProjection"]
+ _import_structure["modeling_utils"] = ["ModelMixin"]
+ _import_structure["transformers.prior_transformer"] = ["PriorTransformer"]
+ _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
+ _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
+ _import_structure["transformers.transformer_temporal"] = ["TransformerTemporalModel"]
+ _import_structure["unets.unet_1d"] = ["UNet1DModel"]
+ _import_structure["unets.unet_2d"] = ["UNet2DModel"]
+ _import_structure["unets.unet_2d_condition"] = ["UNet2DConditionModel"]
+ _import_structure["unets.unet_3d_condition"] = ["UNet3DConditionModel"]
+ _import_structure["unets.unet_i2vgen_xl"] = ["I2VGenXLUNet"]
+ _import_structure["unets.unet_kandinsky3"] = ["Kandinsky3UNet"]
+ _import_structure["unets.unet_motion_model"] = ["MotionAdapter", "UNetMotionModel"]
+ _import_structure["unets.unet_spatio_temporal_condition"] = ["UNetSpatioTemporalConditionModel"]
+ _import_structure["unets.unet_stable_cascade"] = ["StableCascadeUNet"]
+ _import_structure["unets.uvit_2d"] = ["UVit2DModel"]
+ _import_structure["vq_model"] = ["VQModel"]
+
+if is_flax_available():
+ _import_structure["controlnet_flax"] = ["FlaxControlNetModel"]
+ _import_structure["unets.unet_2d_condition_flax"] = ["FlaxUNet2DConditionModel"]
+ _import_structure["vae_flax"] = ["FlaxAutoencoderKL"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ if is_torch_available():
+ from .adapter import MultiAdapter, T2IAdapter
+ from .autoencoders import (
+ AsymmetricAutoencoderKL,
+ AutoencoderKL,
+ AutoencoderKLTemporalDecoder,
+ AutoencoderTiny,
+ ConsistencyDecoderVAE,
+ )
+ from .brushnet import BrushNetModel
+ from .controlnet import ControlNetModel
+ from .embeddings import ImageProjection
+ from .modeling_utils import ModelMixin
+ from .transformers import (
+ DualTransformer2DModel,
+ PriorTransformer,
+ T5FilmDecoder,
+ Transformer2DModel,
+ TransformerTemporalModel,
+ )
+ from .unets import (
+ I2VGenXLUNet,
+ Kandinsky3UNet,
+ MotionAdapter,
+ StableCascadeUNet,
+ UNet1DModel,
+ UNet2DConditionModel,
+ UNet2DModel,
+ UNet3DConditionModel,
+ UNetMotionModel,
+ UNetSpatioTemporalConditionModel,
+ UVit2DModel,
+ )
+ from .vq_model import VQModel
+
+ if is_flax_available():
+ from .controlnet_flax import FlaxControlNetModel
+ from .unets import FlaxUNet2DConditionModel
+ from .vae_flax import FlaxAutoencoderKL
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
diff --git a/model/diffusers_c/models/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/models/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d390689c3c94d16a8684fa785f67f4d18a027f55
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/models/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6f738e06827e572170d68ad8e759957a839174e3
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/activations.cpython-310.pyc b/model/diffusers_c/models/__pycache__/activations.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1a3a7379734892fc51e3952f5aa881b5ebf27522
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/activations.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/activations.cpython-39.pyc b/model/diffusers_c/models/__pycache__/activations.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..857bfe20c4457f5aaeb079a373b1f4f151afe94a
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/activations.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/adapter.cpython-39.pyc b/model/diffusers_c/models/__pycache__/adapter.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0ea858d80ccd9c48ab3a011f0d07ae8275c499ca
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/adapter.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/attention.cpython-310.pyc b/model/diffusers_c/models/__pycache__/attention.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c15f0d7a61d7032fb5932c7c12b4585f835b381e
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/attention.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/attention.cpython-39.pyc b/model/diffusers_c/models/__pycache__/attention.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..66bfb0958f3642f7d51d12eb6e1b563173b0614c
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/attention.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/attention_flax.cpython-310.pyc b/model/diffusers_c/models/__pycache__/attention_flax.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d16f3cdfdd6af8db84fc7d982f6bf5efb5619cb1
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/attention_flax.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/attention_processor.cpython-310.pyc b/model/diffusers_c/models/__pycache__/attention_processor.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1c94f6e02f355d34147778e2bacbd52df187b7ba
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/attention_processor.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/attention_processor.cpython-39.pyc b/model/diffusers_c/models/__pycache__/attention_processor.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..66e6f0d0a6123e56faf924f9ad38e0e342828ae9
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/attention_processor.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/brushnet.cpython-39.pyc b/model/diffusers_c/models/__pycache__/brushnet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2d5bcc2d9d089e29f4d9a3256d0bc39baf7b5da5
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/brushnet.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/controlnet.cpython-39.pyc b/model/diffusers_c/models/__pycache__/controlnet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..dcc482e38e761ed31c820eae1ca68665bd632212
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/controlnet.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/downsampling.cpython-310.pyc b/model/diffusers_c/models/__pycache__/downsampling.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..69d9ae49ab34e0a93953da41f1e1179382459bd3
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/downsampling.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/downsampling.cpython-39.pyc b/model/diffusers_c/models/__pycache__/downsampling.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..65e4303e90f53da672a89fe660f1bac1945b1c4c
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/downsampling.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/dual_transformer_2d.cpython-39.pyc b/model/diffusers_c/models/__pycache__/dual_transformer_2d.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..be52c4d90c43878f81660ee0720afc6dd440141a
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/dual_transformer_2d.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/embeddings.cpython-310.pyc b/model/diffusers_c/models/__pycache__/embeddings.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0235411d2762cb5872229c4d7d4b336aab189770
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/embeddings.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/embeddings.cpython-39.pyc b/model/diffusers_c/models/__pycache__/embeddings.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7ed8bef7938e1dce260ccf693caca73ad21be3fc
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/embeddings.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/embeddings_flax.cpython-310.pyc b/model/diffusers_c/models/__pycache__/embeddings_flax.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2b190c178310f82e1aad8bf2bd2b0da17722d84a
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/embeddings_flax.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/lora.cpython-310.pyc b/model/diffusers_c/models/__pycache__/lora.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a26a717e4f61e5387c44e7216d2c0c31414ecb52
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/lora.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/lora.cpython-39.pyc b/model/diffusers_c/models/__pycache__/lora.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b11e519f1e73849d81c322b0685b9ef32f381295
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/lora.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/modeling_flax_pytorch_utils.cpython-310.pyc b/model/diffusers_c/models/__pycache__/modeling_flax_pytorch_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..76ac3bb8fda22bf52c1bc15a8bb1a573bb612360
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/modeling_flax_pytorch_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/modeling_flax_utils.cpython-310.pyc b/model/diffusers_c/models/__pycache__/modeling_flax_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..28dbac4575bffdaf2ab6a6bf4c03364ac7a813a5
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/modeling_flax_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/modeling_outputs.cpython-310.pyc b/model/diffusers_c/models/__pycache__/modeling_outputs.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9695101c25dc66685e77dcd577b1a8f08de8b295
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/modeling_outputs.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/modeling_outputs.cpython-39.pyc b/model/diffusers_c/models/__pycache__/modeling_outputs.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..84fa2ff1af647f84282b581fb20632af619d5873
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/modeling_outputs.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/modeling_utils.cpython-310.pyc b/model/diffusers_c/models/__pycache__/modeling_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6a494153cb74d1144f741fa20b3fbf9d8e9df6df
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/modeling_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/modeling_utils.cpython-39.pyc b/model/diffusers_c/models/__pycache__/modeling_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4c0c166a74286077974718ec98f1767f35721420
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/modeling_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/normalization.cpython-310.pyc b/model/diffusers_c/models/__pycache__/normalization.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6ab92e997b94619eccc0fa8ff9305534971868ab
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/normalization.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/normalization.cpython-39.pyc b/model/diffusers_c/models/__pycache__/normalization.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..301d8989068556969eda8c2d4a436df348b47e24
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/normalization.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/resnet.cpython-310.pyc b/model/diffusers_c/models/__pycache__/resnet.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3d93b0b6e7311eafa8daa92bc47d460cb091f468
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/resnet.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/resnet.cpython-39.pyc b/model/diffusers_c/models/__pycache__/resnet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5f6762cc7b70849653fe296ac54c8918265f7ab3
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/resnet.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/resnet_flax.cpython-310.pyc b/model/diffusers_c/models/__pycache__/resnet_flax.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c95d97dc684040d7e6511fb9cd8867be92e29978
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/resnet_flax.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/unet_2d_condition.cpython-39.pyc b/model/diffusers_c/models/__pycache__/unet_2d_condition.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0cb29dda8e79312cfd7ab8086022b985fd0eb8a9
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/unet_2d_condition.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/upsampling.cpython-310.pyc b/model/diffusers_c/models/__pycache__/upsampling.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..692f0d91d5027f6f6aaa6c08d2f938a2a135f1ce
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/upsampling.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/upsampling.cpython-39.pyc b/model/diffusers_c/models/__pycache__/upsampling.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..515464c08484c67579d5173b3bdbfbb2deee9a4b
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/upsampling.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/__pycache__/vq_model.cpython-39.pyc b/model/diffusers_c/models/__pycache__/vq_model.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..62b04bdc3a72851feca3b35901f924c14f62be00
Binary files /dev/null and b/model/diffusers_c/models/__pycache__/vq_model.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/activations.py b/model/diffusers_c/models/activations.py
new file mode 100644
index 0000000000000000000000000000000000000000..bf79853871e5c19073f5f756ca29c6c53c0b4b9c
--- /dev/null
+++ b/model/diffusers_c/models/activations.py
@@ -0,0 +1,122 @@
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..utils import USE_PEFT_BACKEND
+from .lora import LoRACompatibleLinear
+
+
+ACTIVATION_FUNCTIONS = {
+ "swish": nn.SiLU(),
+ "silu": nn.SiLU(),
+ "mish": nn.Mish(),
+ "gelu": nn.GELU(),
+ "relu": nn.ReLU(),
+}
+
+
+def get_activation(act_fn: str) -> nn.Module:
+ """Helper function to get activation function from string.
+
+ Args:
+ act_fn (str): Name of activation function.
+
+ Returns:
+ nn.Module: Activation function.
+ """
+
+ act_fn = act_fn.lower()
+ if act_fn in ACTIVATION_FUNCTIONS:
+ return ACTIVATION_FUNCTIONS[act_fn]
+ else:
+ raise ValueError(f"Unsupported activation function: {act_fn}")
+
+
+class GELU(nn.Module):
+ r"""
+ GELU activation function with tanh approximation support with `approximate="tanh"`.
+
+ Parameters:
+ dim_in (`int`): The number of channels in the input.
+ dim_out (`int`): The number of channels in the output.
+ approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation.
+ bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+ """
+
+ def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True):
+ super().__init__()
+ self.proj = nn.Linear(dim_in, dim_out, bias=bias)
+ self.approximate = approximate
+
+ def gelu(self, gate: torch.Tensor) -> torch.Tensor:
+ if gate.device.type != "mps":
+ return F.gelu(gate, approximate=self.approximate)
+ # mps: gelu is not implemented for float16
+ return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype)
+
+ def forward(self, hidden_states):
+ hidden_states = self.proj(hidden_states)
+ hidden_states = self.gelu(hidden_states)
+ return hidden_states
+
+
+class GEGLU(nn.Module):
+ r"""
+ A [variant](https://arxiv.org/abs/2002.05202) of the gated linear unit activation function.
+
+ Parameters:
+ dim_in (`int`): The number of channels in the input.
+ dim_out (`int`): The number of channels in the output.
+ bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+ """
+
+ def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
+ super().__init__()
+ linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+ self.proj = linear_cls(dim_in, dim_out * 2, bias=bias)
+
+ def gelu(self, gate: torch.Tensor) -> torch.Tensor:
+ if gate.device.type != "mps":
+ return F.gelu(gate)
+ # mps: gelu is not implemented for float16
+ return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype)
+
+ def forward(self, hidden_states, scale: float = 1.0):
+ args = () if USE_PEFT_BACKEND else (scale,)
+ hidden_states, gate = self.proj(hidden_states, *args).chunk(2, dim=-1)
+ return hidden_states * self.gelu(gate)
+
+
+class ApproximateGELU(nn.Module):
+ r"""
+ The approximate form of the Gaussian Error Linear Unit (GELU). For more details, see section 2 of this
+ [paper](https://arxiv.org/abs/1606.08415).
+
+ Parameters:
+ dim_in (`int`): The number of channels in the input.
+ dim_out (`int`): The number of channels in the output.
+ bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+ """
+
+ def __init__(self, dim_in: int, dim_out: int, bias: bool = True):
+ super().__init__()
+ self.proj = nn.Linear(dim_in, dim_out, bias=bias)
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ x = self.proj(x)
+ return x * torch.sigmoid(1.702 * x)
diff --git a/model/diffusers_c/models/adapter.py b/model/diffusers_c/models/adapter.py
new file mode 100644
index 0000000000000000000000000000000000000000..0f4b2ec033717a8d4eb10fe34226d88e966d1ec5
--- /dev/null
+++ b/model/diffusers_c/models/adapter.py
@@ -0,0 +1,584 @@
+# Copyright 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import os
+from typing import Callable, List, Optional, Union
+
+import torch
+import torch.nn as nn
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import logging
+from .modeling_utils import ModelMixin
+
+
+logger = logging.get_logger(__name__)
+
+
+class MultiAdapter(ModelMixin):
+ r"""
+ MultiAdapter is a wrapper model that contains multiple adapter models and merges their outputs according to
+ user-assigned weighting.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
+ implements for all the model (such as downloading or saving, etc.)
+
+ Parameters:
+ adapters (`List[T2IAdapter]`, *optional*, defaults to None):
+ A list of `T2IAdapter` model instances.
+ """
+
+ def __init__(self, adapters: List["T2IAdapter"]):
+ super(MultiAdapter, self).__init__()
+
+ self.num_adapter = len(adapters)
+ self.adapters = nn.ModuleList(adapters)
+
+ if len(adapters) == 0:
+ raise ValueError("Expecting at least one adapter")
+
+ if len(adapters) == 1:
+ raise ValueError("For a single adapter, please use the `T2IAdapter` class instead of `MultiAdapter`")
+
+ # The outputs from each adapter are added together with a weight.
+ # This means that the change in dimensions from downsampling must
+ # be the same for all adapters. Inductively, it also means the
+ # downscale_factor and total_downscale_factor must be the same for all
+ # adapters.
+ first_adapter_total_downscale_factor = adapters[0].total_downscale_factor
+ first_adapter_downscale_factor = adapters[0].downscale_factor
+ for idx in range(1, len(adapters)):
+ if (
+ adapters[idx].total_downscale_factor != first_adapter_total_downscale_factor
+ or adapters[idx].downscale_factor != first_adapter_downscale_factor
+ ):
+ raise ValueError(
+ f"Expecting all adapters to have the same downscaling behavior, but got:\n"
+ f"adapters[0].total_downscale_factor={first_adapter_total_downscale_factor}\n"
+ f"adapters[0].downscale_factor={first_adapter_downscale_factor}\n"
+ f"adapter[`{idx}`].total_downscale_factor={adapters[idx].total_downscale_factor}\n"
+ f"adapter[`{idx}`].downscale_factor={adapters[idx].downscale_factor}"
+ )
+
+ self.total_downscale_factor = first_adapter_total_downscale_factor
+ self.downscale_factor = first_adapter_downscale_factor
+
+ def forward(self, xs: torch.Tensor, adapter_weights: Optional[List[float]] = None) -> List[torch.Tensor]:
+ r"""
+ Args:
+ xs (`torch.Tensor`):
+ (batch, channel, height, width) input images for multiple adapter models concated along dimension 1,
+ `channel` should equal to `num_adapter` * "number of channel of image".
+ adapter_weights (`List[float]`, *optional*, defaults to None):
+ List of floats representing the weight which will be multiply to each adapter's output before adding
+ them together.
+ """
+ if adapter_weights is None:
+ adapter_weights = torch.tensor([1 / self.num_adapter] * self.num_adapter)
+ else:
+ adapter_weights = torch.tensor(adapter_weights)
+
+ accume_state = None
+ for x, w, adapter in zip(xs, adapter_weights, self.adapters):
+ features = adapter(x)
+ if accume_state is None:
+ accume_state = features
+ for i in range(len(accume_state)):
+ accume_state[i] = w * accume_state[i]
+ else:
+ for i in range(len(features)):
+ accume_state[i] += w * features[i]
+ return accume_state
+
+ def save_pretrained(
+ self,
+ save_directory: Union[str, os.PathLike],
+ is_main_process: bool = True,
+ save_function: Callable = None,
+ safe_serialization: bool = True,
+ variant: Optional[str] = None,
+ ):
+ """
+ Save a model and its configuration file to a directory, so that it can be re-loaded using the
+ `[`~models.adapter.MultiAdapter.from_pretrained`]` class method.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to which to save. Will be created if it doesn't exist.
+ is_main_process (`bool`, *optional*, defaults to `True`):
+ Whether the process calling this is the main process or not. Useful when in distributed training like
+ TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on
+ the main process to avoid race conditions.
+ save_function (`Callable`):
+ The function to use to save the state dictionary. Useful on distributed training like TPUs when one
+ need to replace `torch.save` by another method. Can be configured with the environment variable
+ `DIFFUSERS_SAVE_MODE`.
+ safe_serialization (`bool`, *optional*, defaults to `True`):
+ Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+ variant (`str`, *optional*):
+ If specified, weights are saved in the format pytorch_model..bin.
+ """
+ idx = 0
+ model_path_to_save = save_directory
+ for adapter in self.adapters:
+ adapter.save_pretrained(
+ model_path_to_save,
+ is_main_process=is_main_process,
+ save_function=save_function,
+ safe_serialization=safe_serialization,
+ variant=variant,
+ )
+
+ idx += 1
+ model_path_to_save = model_path_to_save + f"_{idx}"
+
+ @classmethod
+ def from_pretrained(cls, pretrained_model_path: Optional[Union[str, os.PathLike]], **kwargs):
+ r"""
+ Instantiate a pretrained MultiAdapter model from multiple pre-trained adapter models.
+
+ The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
+ the model, you should first set it back in training mode with `model.train()`.
+
+ The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
+ pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
+ task.
+
+ The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
+ weights are discarded.
+
+ Parameters:
+ pretrained_model_path (`os.PathLike`):
+ A path to a *directory* containing model weights saved using
+ [`~diffusers.models.adapter.MultiAdapter.save_pretrained`], e.g., `./my_model_directory/adapter`.
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
+ will be automatically derived from the model's weights.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+ A map that specifies where each submodule should go. It doesn't need to be refined to each
+ parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
+ same device.
+
+ To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
+ more information about each option see [designing a device
+ map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+ max_memory (`Dict`, *optional*):
+ A dictionary device identifier to maximum memory. Will default to the maximum memory available for each
+ GPU and the available CPU RAM if unset.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
+ also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
+ model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
+ setting this argument to `True` will raise an error.
+ variant (`str`, *optional*):
+ If specified load weights from `variant` filename, *e.g.* pytorch_model..bin. `variant` is
+ ignored when using `from_flax`.
+ use_safetensors (`bool`, *optional*, defaults to `None`):
+ If set to `None`, the `safetensors` weights will be downloaded if they're available **and** if the
+ `safetensors` library is installed. If set to `True`, the model will be forcibly loaded from
+ `safetensors` weights. If set to `False`, loading will *not* use `safetensors`.
+ """
+ idx = 0
+ adapters = []
+
+ # load adapter and append to list until no adapter directory exists anymore
+ # first adapter has to be saved under `./mydirectory/adapter` to be compliant with `DiffusionPipeline.from_pretrained`
+ # second, third, ... adapters have to be saved under `./mydirectory/adapter_1`, `./mydirectory/adapter_2`, ...
+ model_path_to_load = pretrained_model_path
+ while os.path.isdir(model_path_to_load):
+ adapter = T2IAdapter.from_pretrained(model_path_to_load, **kwargs)
+ adapters.append(adapter)
+
+ idx += 1
+ model_path_to_load = pretrained_model_path + f"_{idx}"
+
+ logger.info(f"{len(adapters)} adapters loaded from {pretrained_model_path}.")
+
+ if len(adapters) == 0:
+ raise ValueError(
+ f"No T2IAdapters found under {os.path.dirname(pretrained_model_path)}. Expected at least {pretrained_model_path + '_0'}."
+ )
+
+ return cls(adapters)
+
+
+class T2IAdapter(ModelMixin, ConfigMixin):
+ r"""
+ A simple ResNet-like model that accepts images containing control signals such as keyposes and depth. The model
+ generates multiple feature maps that are used as additional conditioning in [`UNet2DConditionModel`]. The model's
+ architecture follows the original implementation of
+ [Adapter](https://github.com/TencentARC/T2I-Adapter/blob/686de4681515662c0ac2ffa07bf5dda83af1038a/ldm/modules/encoders/adapter.py#L97)
+ and
+ [AdapterLight](https://github.com/TencentARC/T2I-Adapter/blob/686de4681515662c0ac2ffa07bf5dda83af1038a/ldm/modules/encoders/adapter.py#L235).
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
+ implements for all the model (such as downloading or saving, etc.)
+
+ Parameters:
+ in_channels (`int`, *optional*, defaults to 3):
+ Number of channels of Aapter's input(*control image*). Set this parameter to 1 if you're using gray scale
+ image as *control image*.
+ channels (`List[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The number of channel of each downsample block's output hidden state. The `len(block_out_channels)` will
+ also determine the number of downsample blocks in the Adapter.
+ num_res_blocks (`int`, *optional*, defaults to 2):
+ Number of ResNet blocks in each downsample block.
+ downscale_factor (`int`, *optional*, defaults to 8):
+ A factor that determines the total downscale factor of the Adapter.
+ adapter_type (`str`, *optional*, defaults to `full_adapter`):
+ The type of Adapter to use. Choose either `full_adapter` or `full_adapter_xl` or `light_adapter`.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 3,
+ channels: List[int] = [320, 640, 1280, 1280],
+ num_res_blocks: int = 2,
+ downscale_factor: int = 8,
+ adapter_type: str = "full_adapter",
+ ):
+ super().__init__()
+
+ if adapter_type == "full_adapter":
+ self.adapter = FullAdapter(in_channels, channels, num_res_blocks, downscale_factor)
+ elif adapter_type == "full_adapter_xl":
+ self.adapter = FullAdapterXL(in_channels, channels, num_res_blocks, downscale_factor)
+ elif adapter_type == "light_adapter":
+ self.adapter = LightAdapter(in_channels, channels, num_res_blocks, downscale_factor)
+ else:
+ raise ValueError(
+ f"Unsupported adapter_type: '{adapter_type}'. Choose either 'full_adapter' or "
+ "'full_adapter_xl' or 'light_adapter'."
+ )
+
+ def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+ r"""
+ This function processes the input tensor `x` through the adapter model and returns a list of feature tensors,
+ each representing information extracted at a different scale from the input. The length of the list is
+ determined by the number of downsample blocks in the Adapter, as specified by the `channels` and
+ `num_res_blocks` parameters during initialization.
+ """
+ return self.adapter(x)
+
+ @property
+ def total_downscale_factor(self):
+ return self.adapter.total_downscale_factor
+
+ @property
+ def downscale_factor(self):
+ """The downscale factor applied in the T2I-Adapter's initial pixel unshuffle operation. If an input image's dimensions are
+ not evenly divisible by the downscale_factor then an exception will be raised.
+ """
+ return self.adapter.unshuffle.downscale_factor
+
+
+# full adapter
+
+
+class FullAdapter(nn.Module):
+ r"""
+ See [`T2IAdapter`] for more information.
+ """
+
+ def __init__(
+ self,
+ in_channels: int = 3,
+ channels: List[int] = [320, 640, 1280, 1280],
+ num_res_blocks: int = 2,
+ downscale_factor: int = 8,
+ ):
+ super().__init__()
+
+ in_channels = in_channels * downscale_factor**2
+
+ self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+ self.conv_in = nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1)
+
+ self.body = nn.ModuleList(
+ [
+ AdapterBlock(channels[0], channels[0], num_res_blocks),
+ *[
+ AdapterBlock(channels[i - 1], channels[i], num_res_blocks, down=True)
+ for i in range(1, len(channels))
+ ],
+ ]
+ )
+
+ self.total_downscale_factor = downscale_factor * 2 ** (len(channels) - 1)
+
+ def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+ r"""
+ This method processes the input tensor `x` through the FullAdapter model and performs operations including
+ pixel unshuffling, convolution, and a stack of AdapterBlocks. It returns a list of feature tensors, each
+ capturing information at a different stage of processing within the FullAdapter model. The number of feature
+ tensors in the list is determined by the number of downsample blocks specified during initialization.
+ """
+ x = self.unshuffle(x)
+ x = self.conv_in(x)
+
+ features = []
+
+ for block in self.body:
+ x = block(x)
+ features.append(x)
+
+ return features
+
+
+class FullAdapterXL(nn.Module):
+ r"""
+ See [`T2IAdapter`] for more information.
+ """
+
+ def __init__(
+ self,
+ in_channels: int = 3,
+ channels: List[int] = [320, 640, 1280, 1280],
+ num_res_blocks: int = 2,
+ downscale_factor: int = 16,
+ ):
+ super().__init__()
+
+ in_channels = in_channels * downscale_factor**2
+
+ self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+ self.conv_in = nn.Conv2d(in_channels, channels[0], kernel_size=3, padding=1)
+
+ self.body = []
+ # blocks to extract XL features with dimensions of [320, 64, 64], [640, 64, 64], [1280, 32, 32], [1280, 32, 32]
+ for i in range(len(channels)):
+ if i == 1:
+ self.body.append(AdapterBlock(channels[i - 1], channels[i], num_res_blocks))
+ elif i == 2:
+ self.body.append(AdapterBlock(channels[i - 1], channels[i], num_res_blocks, down=True))
+ else:
+ self.body.append(AdapterBlock(channels[i], channels[i], num_res_blocks))
+
+ self.body = nn.ModuleList(self.body)
+ # XL has only one downsampling AdapterBlock.
+ self.total_downscale_factor = downscale_factor * 2
+
+ def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+ r"""
+ This method takes the tensor x as input and processes it through FullAdapterXL model. It consists of operations
+ including unshuffling pixels, applying convolution layer and appending each block into list of feature tensors.
+ """
+ x = self.unshuffle(x)
+ x = self.conv_in(x)
+
+ features = []
+
+ for block in self.body:
+ x = block(x)
+ features.append(x)
+
+ return features
+
+
+class AdapterBlock(nn.Module):
+ r"""
+ An AdapterBlock is a helper model that contains multiple ResNet-like blocks. It is used in the `FullAdapter` and
+ `FullAdapterXL` models.
+
+ Parameters:
+ in_channels (`int`):
+ Number of channels of AdapterBlock's input.
+ out_channels (`int`):
+ Number of channels of AdapterBlock's output.
+ num_res_blocks (`int`):
+ Number of ResNet blocks in the AdapterBlock.
+ down (`bool`, *optional*, defaults to `False`):
+ Whether to perform downsampling on AdapterBlock's input.
+ """
+
+ def __init__(self, in_channels: int, out_channels: int, num_res_blocks: int, down: bool = False):
+ super().__init__()
+
+ self.downsample = None
+ if down:
+ self.downsample = nn.AvgPool2d(kernel_size=2, stride=2, ceil_mode=True)
+
+ self.in_conv = None
+ if in_channels != out_channels:
+ self.in_conv = nn.Conv2d(in_channels, out_channels, kernel_size=1)
+
+ self.resnets = nn.Sequential(
+ *[AdapterResnetBlock(out_channels) for _ in range(num_res_blocks)],
+ )
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ r"""
+ This method takes tensor x as input and performs operations downsampling and convolutional layers if the
+ self.downsample and self.in_conv properties of AdapterBlock model are specified. Then it applies a series of
+ residual blocks to the input tensor.
+ """
+ if self.downsample is not None:
+ x = self.downsample(x)
+
+ if self.in_conv is not None:
+ x = self.in_conv(x)
+
+ x = self.resnets(x)
+
+ return x
+
+
+class AdapterResnetBlock(nn.Module):
+ r"""
+ An `AdapterResnetBlock` is a helper model that implements a ResNet-like block.
+
+ Parameters:
+ channels (`int`):
+ Number of channels of AdapterResnetBlock's input and output.
+ """
+
+ def __init__(self, channels: int):
+ super().__init__()
+ self.block1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
+ self.act = nn.ReLU()
+ self.block2 = nn.Conv2d(channels, channels, kernel_size=1)
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ r"""
+ This method takes input tensor x and applies a convolutional layer, ReLU activation, and another convolutional
+ layer on the input tensor. It returns addition with the input tensor.
+ """
+
+ h = self.act(self.block1(x))
+ h = self.block2(h)
+
+ return h + x
+
+
+# light adapter
+
+
+class LightAdapter(nn.Module):
+ r"""
+ See [`T2IAdapter`] for more information.
+ """
+
+ def __init__(
+ self,
+ in_channels: int = 3,
+ channels: List[int] = [320, 640, 1280],
+ num_res_blocks: int = 4,
+ downscale_factor: int = 8,
+ ):
+ super().__init__()
+
+ in_channels = in_channels * downscale_factor**2
+
+ self.unshuffle = nn.PixelUnshuffle(downscale_factor)
+
+ self.body = nn.ModuleList(
+ [
+ LightAdapterBlock(in_channels, channels[0], num_res_blocks),
+ *[
+ LightAdapterBlock(channels[i], channels[i + 1], num_res_blocks, down=True)
+ for i in range(len(channels) - 1)
+ ],
+ LightAdapterBlock(channels[-1], channels[-1], num_res_blocks, down=True),
+ ]
+ )
+
+ self.total_downscale_factor = downscale_factor * (2 ** len(channels))
+
+ def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+ r"""
+ This method takes the input tensor x and performs downscaling and appends it in list of feature tensors. Each
+ feature tensor corresponds to a different level of processing within the LightAdapter.
+ """
+ x = self.unshuffle(x)
+
+ features = []
+
+ for block in self.body:
+ x = block(x)
+ features.append(x)
+
+ return features
+
+
+class LightAdapterBlock(nn.Module):
+ r"""
+ A `LightAdapterBlock` is a helper model that contains multiple `LightAdapterResnetBlocks`. It is used in the
+ `LightAdapter` model.
+
+ Parameters:
+ in_channels (`int`):
+ Number of channels of LightAdapterBlock's input.
+ out_channels (`int`):
+ Number of channels of LightAdapterBlock's output.
+ num_res_blocks (`int`):
+ Number of LightAdapterResnetBlocks in the LightAdapterBlock.
+ down (`bool`, *optional*, defaults to `False`):
+ Whether to perform downsampling on LightAdapterBlock's input.
+ """
+
+ def __init__(self, in_channels: int, out_channels: int, num_res_blocks: int, down: bool = False):
+ super().__init__()
+ mid_channels = out_channels // 4
+
+ self.downsample = None
+ if down:
+ self.downsample = nn.AvgPool2d(kernel_size=2, stride=2, ceil_mode=True)
+
+ self.in_conv = nn.Conv2d(in_channels, mid_channels, kernel_size=1)
+ self.resnets = nn.Sequential(*[LightAdapterResnetBlock(mid_channels) for _ in range(num_res_blocks)])
+ self.out_conv = nn.Conv2d(mid_channels, out_channels, kernel_size=1)
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ r"""
+ This method takes tensor x as input and performs downsampling if required. Then it applies in convolution
+ layer, a sequence of residual blocks, and out convolutional layer.
+ """
+ if self.downsample is not None:
+ x = self.downsample(x)
+
+ x = self.in_conv(x)
+ x = self.resnets(x)
+ x = self.out_conv(x)
+
+ return x
+
+
+class LightAdapterResnetBlock(nn.Module):
+ """
+ A `LightAdapterResnetBlock` is a helper model that implements a ResNet-like block with a slightly different
+ architecture than `AdapterResnetBlock`.
+
+ Parameters:
+ channels (`int`):
+ Number of channels of LightAdapterResnetBlock's input and output.
+ """
+
+ def __init__(self, channels: int):
+ super().__init__()
+ self.block1 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
+ self.act = nn.ReLU()
+ self.block2 = nn.Conv2d(channels, channels, kernel_size=3, padding=1)
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ r"""
+ This function takes input tensor x and processes it through one convolutional layer, ReLU activation, and
+ another convolutional layer and adds it to input tensor.
+ """
+
+ h = self.act(self.block1(x))
+ h = self.block2(h)
+
+ return h + x
diff --git a/model/diffusers_c/models/attention.py b/model/diffusers_c/models/attention.py
new file mode 100644
index 0000000000000000000000000000000000000000..c65e6717959cc8e720ba3036755abadf6d9d343a
--- /dev/null
+++ b/model/diffusers_c/models/attention.py
@@ -0,0 +1,675 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..utils import USE_PEFT_BACKEND
+from ..utils.torch_utils import maybe_allow_in_graph
+from .activations import GEGLU, GELU, ApproximateGELU
+from .attention_processor import Attention
+from .embeddings import SinusoidalPositionalEmbedding
+from .lora import LoRACompatibleLinear
+from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm
+
+
+def _chunked_feed_forward(
+ ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None
+):
+ # "feed_forward_chunk_size" can be used to save memory
+ if hidden_states.shape[chunk_dim] % chunk_size != 0:
+ raise ValueError(
+ f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+ )
+
+ num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+ if lora_scale is None:
+ ff_output = torch.cat(
+ [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+ dim=chunk_dim,
+ )
+ else:
+ # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete
+ ff_output = torch.cat(
+ [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+ dim=chunk_dim,
+ )
+
+ return ff_output
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+ r"""
+ A gated self-attention dense layer that combines visual features and object features.
+
+ Parameters:
+ query_dim (`int`): The number of channels in the query.
+ context_dim (`int`): The number of channels in the context.
+ n_heads (`int`): The number of heads to use for attention.
+ d_head (`int`): The number of channels in each head.
+ """
+
+ def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+ super().__init__()
+
+ # we need a linear projection since we need cat visual feature and obj feature
+ self.linear = nn.Linear(context_dim, query_dim)
+
+ self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+ self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+ self.norm1 = nn.LayerNorm(query_dim)
+ self.norm2 = nn.LayerNorm(query_dim)
+
+ self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+ self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+ self.enabled = True
+
+ def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+ if not self.enabled:
+ return x
+
+ n_visual = x.shape[1]
+ objs = self.linear(objs)
+
+ x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+ x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+ return x
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+ r"""
+ A basic Transformer block.
+
+ Parameters:
+ dim (`int`): The number of channels in the input and output.
+ num_attention_heads (`int`): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`): The number of channels in each head.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+ num_embeds_ada_norm (:
+ obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+ attention_bias (:
+ obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+ only_cross_attention (`bool`, *optional*):
+ Whether to use only cross-attention layers. In this case two cross attention layers are used.
+ double_self_attention (`bool`, *optional*):
+ Whether to use two self-attention layers. In this case no cross attention layers are used.
+ upcast_attention (`bool`, *optional*):
+ Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+ norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+ Whether to use learnable elementwise affine parameters for normalization.
+ norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+ The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+ final_dropout (`bool` *optional*, defaults to False):
+ Whether to apply a final dropout after the last feed-forward layer.
+ attention_type (`str`, *optional*, defaults to `"default"`):
+ The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+ positional_embeddings (`str`, *optional*, defaults to `None`):
+ The type of positional embeddings to apply to.
+ num_positional_embeddings (`int`, *optional*, defaults to `None`):
+ The maximum number of positional embeddings to apply.
+ """
+
+ def __init__(
+ self,
+ dim: int,
+ num_attention_heads: int,
+ attention_head_dim: int,
+ dropout=0.0,
+ cross_attention_dim: Optional[int] = None,
+ activation_fn: str = "geglu",
+ num_embeds_ada_norm: Optional[int] = None,
+ attention_bias: bool = False,
+ only_cross_attention: bool = False,
+ double_self_attention: bool = False,
+ upcast_attention: bool = False,
+ norm_elementwise_affine: bool = True,
+ norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+ norm_eps: float = 1e-5,
+ final_dropout: bool = False,
+ attention_type: str = "default",
+ positional_embeddings: Optional[str] = None,
+ num_positional_embeddings: Optional[int] = None,
+ ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
+ ada_norm_bias: Optional[int] = None,
+ ff_inner_dim: Optional[int] = None,
+ ff_bias: bool = True,
+ attention_out_bias: bool = True,
+ ):
+ super().__init__()
+ self.only_cross_attention = only_cross_attention
+
+ # We keep these boolean flags for backward-compatibility.
+ self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+ self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+ self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+ self.use_layer_norm = norm_type == "layer_norm"
+ self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+ if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+ raise ValueError(
+ f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+ f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+ )
+
+ self.norm_type = norm_type
+ self.num_embeds_ada_norm = num_embeds_ada_norm
+
+ if positional_embeddings and (num_positional_embeddings is None):
+ raise ValueError(
+ "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+ )
+
+ if positional_embeddings == "sinusoidal":
+ self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+ else:
+ self.pos_embed = None
+
+ # Define 3 blocks. Each block has its own normalization layer.
+ # 1. Self-Attn
+ if norm_type == "ada_norm":
+ self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+ elif norm_type == "ada_norm_zero":
+ self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+ elif norm_type == "ada_norm_continuous":
+ self.norm1 = AdaLayerNormContinuous(
+ dim,
+ ada_norm_continous_conditioning_embedding_dim,
+ norm_elementwise_affine,
+ norm_eps,
+ ada_norm_bias,
+ "rms_norm",
+ )
+ else:
+ self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+ self.attn1 = Attention(
+ query_dim=dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+ upcast_attention=upcast_attention,
+ out_bias=attention_out_bias,
+ )
+
+ # 2. Cross-Attn
+ if cross_attention_dim is not None or double_self_attention:
+ # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+ # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+ # the second cross attention block.
+ if norm_type == "ada_norm":
+ self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+ elif norm_type == "ada_norm_continuous":
+ self.norm2 = AdaLayerNormContinuous(
+ dim,
+ ada_norm_continous_conditioning_embedding_dim,
+ norm_elementwise_affine,
+ norm_eps,
+ ada_norm_bias,
+ "rms_norm",
+ )
+ else:
+ self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+ self.attn2 = Attention(
+ query_dim=dim,
+ cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ upcast_attention=upcast_attention,
+ out_bias=attention_out_bias,
+ ) # is self-attn if encoder_hidden_states is none
+ else:
+ self.norm2 = None
+ self.attn2 = None
+
+ # 3. Feed-forward
+ if norm_type == "ada_norm_continuous":
+ self.norm3 = AdaLayerNormContinuous(
+ dim,
+ ada_norm_continous_conditioning_embedding_dim,
+ norm_elementwise_affine,
+ norm_eps,
+ ada_norm_bias,
+ "layer_norm",
+ )
+
+ elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]:
+ self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+ elif norm_type == "layer_norm_i2vgen":
+ self.norm3 = None
+
+ self.ff = FeedForward(
+ dim,
+ dropout=dropout,
+ activation_fn=activation_fn,
+ final_dropout=final_dropout,
+ inner_dim=ff_inner_dim,
+ bias=ff_bias,
+ )
+
+ # 4. Fuser
+ if attention_type == "gated" or attention_type == "gated-text-image":
+ self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+ # 5. Scale-shift for PixArt-Alpha.
+ if norm_type == "ada_norm_single":
+ self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+ # let chunk size default to None
+ self._chunk_size = None
+ self._chunk_dim = 0
+
+ def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+ # Sets chunk feed-forward
+ self._chunk_size = chunk_size
+ self._chunk_dim = dim
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ timestep: Optional[torch.LongTensor] = None,
+ cross_attention_kwargs: Dict[str, Any] = None,
+ class_labels: Optional[torch.LongTensor] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ ) -> torch.FloatTensor:
+ # Notice that normalization is always applied before the real computation in the following blocks.
+ # 0. Self-Attention
+ batch_size = hidden_states.shape[0]
+
+ if self.norm_type == "ada_norm":
+ norm_hidden_states = self.norm1(hidden_states, timestep)
+ elif self.norm_type == "ada_norm_zero":
+ norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+ hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+ )
+ elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
+ norm_hidden_states = self.norm1(hidden_states)
+ elif self.norm_type == "ada_norm_continuous":
+ norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
+ elif self.norm_type == "ada_norm_single":
+ shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+ self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+ ).chunk(6, dim=1)
+ norm_hidden_states = self.norm1(hidden_states)
+ norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+ norm_hidden_states = norm_hidden_states.squeeze(1)
+ else:
+ raise ValueError("Incorrect norm used")
+
+ if self.pos_embed is not None:
+ norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+ # 1. Retrieve lora scale.
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+ # 2. Prepare GLIGEN inputs
+ cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+ gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+ attn_output = self.attn1(
+ norm_hidden_states,
+ encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+ attention_mask=attention_mask,
+ **cross_attention_kwargs,
+ )
+ if self.norm_type == "ada_norm_zero":
+ attn_output = gate_msa.unsqueeze(1) * attn_output
+ elif self.norm_type == "ada_norm_single":
+ attn_output = gate_msa * attn_output
+
+ hidden_states = attn_output + hidden_states
+ if hidden_states.ndim == 4:
+ hidden_states = hidden_states.squeeze(1)
+
+ # 2.5 GLIGEN Control
+ if gligen_kwargs is not None:
+ hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+ # 3. Cross-Attention
+ if self.attn2 is not None:
+ if self.norm_type == "ada_norm":
+ norm_hidden_states = self.norm2(hidden_states, timestep)
+ elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
+ norm_hidden_states = self.norm2(hidden_states)
+ elif self.norm_type == "ada_norm_single":
+ # For PixArt norm2 isn't applied here:
+ # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+ norm_hidden_states = hidden_states
+ elif self.norm_type == "ada_norm_continuous":
+ norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
+ else:
+ raise ValueError("Incorrect norm")
+
+ if self.pos_embed is not None and self.norm_type != "ada_norm_single":
+ norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+ attn_output = self.attn2(
+ norm_hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=encoder_attention_mask,
+ **cross_attention_kwargs,
+ )
+ hidden_states = attn_output + hidden_states
+
+ # 4. Feed-forward
+ # i2vgen doesn't have this norm 🤷♂️
+ if self.norm_type == "ada_norm_continuous":
+ norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+ elif not self.norm_type == "ada_norm_single":
+ norm_hidden_states = self.norm3(hidden_states)
+
+ if self.norm_type == "ada_norm_zero":
+ norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+ if self.norm_type == "ada_norm_single":
+ norm_hidden_states = self.norm2(hidden_states)
+ norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+ if self._chunk_size is not None:
+ # "feed_forward_chunk_size" can be used to save memory
+ ff_output = _chunked_feed_forward(
+ self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale
+ )
+ else:
+ ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+ if self.norm_type == "ada_norm_zero":
+ ff_output = gate_mlp.unsqueeze(1) * ff_output
+ elif self.norm_type == "ada_norm_single":
+ ff_output = gate_mlp * ff_output
+
+ hidden_states = ff_output + hidden_states
+ if hidden_states.ndim == 4:
+ hidden_states = hidden_states.squeeze(1)
+
+ return hidden_states
+
+
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Module):
+ r"""
+ A basic Transformer block for video like data.
+
+ Parameters:
+ dim (`int`): The number of channels in the input and output.
+ time_mix_inner_dim (`int`): The number of channels for temporal attention.
+ num_attention_heads (`int`): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`): The number of channels in each head.
+ cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+ """
+
+ def __init__(
+ self,
+ dim: int,
+ time_mix_inner_dim: int,
+ num_attention_heads: int,
+ attention_head_dim: int,
+ cross_attention_dim: Optional[int] = None,
+ ):
+ super().__init__()
+ self.is_res = dim == time_mix_inner_dim
+
+ self.norm_in = nn.LayerNorm(dim)
+
+ # Define 3 blocks. Each block has its own normalization layer.
+ # 1. Self-Attn
+ self.ff_in = FeedForward(
+ dim,
+ dim_out=time_mix_inner_dim,
+ activation_fn="geglu",
+ )
+
+ self.norm1 = nn.LayerNorm(time_mix_inner_dim)
+ self.attn1 = Attention(
+ query_dim=time_mix_inner_dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ cross_attention_dim=None,
+ )
+
+ # 2. Cross-Attn
+ if cross_attention_dim is not None:
+ # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+ # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+ # the second cross attention block.
+ self.norm2 = nn.LayerNorm(time_mix_inner_dim)
+ self.attn2 = Attention(
+ query_dim=time_mix_inner_dim,
+ cross_attention_dim=cross_attention_dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ ) # is self-attn if encoder_hidden_states is none
+ else:
+ self.norm2 = None
+ self.attn2 = None
+
+ # 3. Feed-forward
+ self.norm3 = nn.LayerNorm(time_mix_inner_dim)
+ self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
+
+ # let chunk size default to None
+ self._chunk_size = None
+ self._chunk_dim = None
+
+ def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
+ # Sets chunk feed-forward
+ self._chunk_size = chunk_size
+ # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
+ self._chunk_dim = 1
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ num_frames: int,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ # Notice that normalization is always applied before the real computation in the following blocks.
+ # 0. Self-Attention
+ batch_size = hidden_states.shape[0]
+
+ batch_frames, seq_length, channels = hidden_states.shape
+ batch_size = batch_frames // num_frames
+
+ hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
+ hidden_states = hidden_states.permute(0, 2, 1, 3)
+ hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)
+
+ residual = hidden_states
+ hidden_states = self.norm_in(hidden_states)
+
+ if self._chunk_size is not None:
+ hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
+ else:
+ hidden_states = self.ff_in(hidden_states)
+
+ if self.is_res:
+ hidden_states = hidden_states + residual
+
+ norm_hidden_states = self.norm1(hidden_states)
+ attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+ hidden_states = attn_output + hidden_states
+
+ # 3. Cross-Attention
+ if self.attn2 is not None:
+ norm_hidden_states = self.norm2(hidden_states)
+ attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+ hidden_states = attn_output + hidden_states
+
+ # 4. Feed-forward
+ norm_hidden_states = self.norm3(hidden_states)
+
+ if self._chunk_size is not None:
+ ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+ else:
+ ff_output = self.ff(norm_hidden_states)
+
+ if self.is_res:
+ hidden_states = ff_output + hidden_states
+ else:
+ hidden_states = ff_output
+
+ hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
+ hidden_states = hidden_states.permute(0, 2, 1, 3)
+ hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)
+
+ return hidden_states
+
+
+class SkipFFTransformerBlock(nn.Module):
+ def __init__(
+ self,
+ dim: int,
+ num_attention_heads: int,
+ attention_head_dim: int,
+ kv_input_dim: int,
+ kv_input_dim_proj_use_bias: bool,
+ dropout=0.0,
+ cross_attention_dim: Optional[int] = None,
+ attention_bias: bool = False,
+ attention_out_bias: bool = True,
+ ):
+ super().__init__()
+ if kv_input_dim != dim:
+ self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
+ else:
+ self.kv_mapper = None
+
+ self.norm1 = RMSNorm(dim, 1e-06)
+
+ self.attn1 = Attention(
+ query_dim=dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ cross_attention_dim=cross_attention_dim,
+ out_bias=attention_out_bias,
+ )
+
+ self.norm2 = RMSNorm(dim, 1e-06)
+
+ self.attn2 = Attention(
+ query_dim=dim,
+ cross_attention_dim=cross_attention_dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ out_bias=attention_out_bias,
+ )
+
+ def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
+ cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+ if self.kv_mapper is not None:
+ encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))
+
+ norm_hidden_states = self.norm1(hidden_states)
+
+ attn_output = self.attn1(
+ norm_hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ **cross_attention_kwargs,
+ )
+
+ hidden_states = attn_output + hidden_states
+
+ norm_hidden_states = self.norm2(hidden_states)
+
+ attn_output = self.attn2(
+ norm_hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ **cross_attention_kwargs,
+ )
+
+ hidden_states = attn_output + hidden_states
+
+ return hidden_states
+
+
+class FeedForward(nn.Module):
+ r"""
+ A feed-forward layer.
+
+ Parameters:
+ dim (`int`): The number of channels in the input.
+ dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+ mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+ final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+ bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+ """
+
+ def __init__(
+ self,
+ dim: int,
+ dim_out: Optional[int] = None,
+ mult: int = 4,
+ dropout: float = 0.0,
+ activation_fn: str = "geglu",
+ final_dropout: bool = False,
+ inner_dim=None,
+ bias: bool = True,
+ ):
+ super().__init__()
+ if inner_dim is None:
+ inner_dim = int(dim * mult)
+ dim_out = dim_out if dim_out is not None else dim
+ linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+ if activation_fn == "gelu":
+ act_fn = GELU(dim, inner_dim, bias=bias)
+ if activation_fn == "gelu-approximate":
+ act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
+ elif activation_fn == "geglu":
+ act_fn = GEGLU(dim, inner_dim, bias=bias)
+ elif activation_fn == "geglu-approximate":
+ act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
+
+ self.net = nn.ModuleList([])
+ # project in
+ self.net.append(act_fn)
+ # project dropout
+ self.net.append(nn.Dropout(dropout))
+ # project out
+ self.net.append(linear_cls(inner_dim, dim_out, bias=bias))
+ # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+ if final_dropout:
+ self.net.append(nn.Dropout(dropout))
+
+ def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+ compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
+ for module in self.net:
+ if isinstance(module, compatible_cls):
+ hidden_states = module(hidden_states, scale)
+ else:
+ hidden_states = module(hidden_states)
+ return hidden_states
diff --git a/model/diffusers_c/models/attention_flax.py b/model/diffusers_c/models/attention_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..25ae5d0a5d6344d566a731c0b1e18f54e4d42581
--- /dev/null
+++ b/model/diffusers_c/models/attention_flax.py
@@ -0,0 +1,494 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import functools
+import math
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+
+
+def _query_chunk_attention(query, key, value, precision, key_chunk_size: int = 4096):
+ """Multi-head dot product attention with a limited number of queries."""
+ num_kv, num_heads, k_features = key.shape[-3:]
+ v_features = value.shape[-1]
+ key_chunk_size = min(key_chunk_size, num_kv)
+ query = query / jnp.sqrt(k_features)
+
+ @functools.partial(jax.checkpoint, prevent_cse=False)
+ def summarize_chunk(query, key, value):
+ attn_weights = jnp.einsum("...qhd,...khd->...qhk", query, key, precision=precision)
+
+ max_score = jnp.max(attn_weights, axis=-1, keepdims=True)
+ max_score = jax.lax.stop_gradient(max_score)
+ exp_weights = jnp.exp(attn_weights - max_score)
+
+ exp_values = jnp.einsum("...vhf,...qhv->...qhf", value, exp_weights, precision=precision)
+ max_score = jnp.einsum("...qhk->...qh", max_score)
+
+ return (exp_values, exp_weights.sum(axis=-1), max_score)
+
+ def chunk_scanner(chunk_idx):
+ # julienne key array
+ key_chunk = jax.lax.dynamic_slice(
+ operand=key,
+ start_indices=[0] * (key.ndim - 3) + [chunk_idx, 0, 0], # [...,k,h,d]
+ slice_sizes=list(key.shape[:-3]) + [key_chunk_size, num_heads, k_features], # [...,k,h,d]
+ )
+
+ # julienne value array
+ value_chunk = jax.lax.dynamic_slice(
+ operand=value,
+ start_indices=[0] * (value.ndim - 3) + [chunk_idx, 0, 0], # [...,v,h,d]
+ slice_sizes=list(value.shape[:-3]) + [key_chunk_size, num_heads, v_features], # [...,v,h,d]
+ )
+
+ return summarize_chunk(query, key_chunk, value_chunk)
+
+ chunk_values, chunk_weights, chunk_max = jax.lax.map(f=chunk_scanner, xs=jnp.arange(0, num_kv, key_chunk_size))
+
+ global_max = jnp.max(chunk_max, axis=0, keepdims=True)
+ max_diffs = jnp.exp(chunk_max - global_max)
+
+ chunk_values *= jnp.expand_dims(max_diffs, axis=-1)
+ chunk_weights *= max_diffs
+
+ all_values = chunk_values.sum(axis=0)
+ all_weights = jnp.expand_dims(chunk_weights, -1).sum(axis=0)
+
+ return all_values / all_weights
+
+
+def jax_memory_efficient_attention(
+ query, key, value, precision=jax.lax.Precision.HIGHEST, query_chunk_size: int = 1024, key_chunk_size: int = 4096
+):
+ r"""
+ Flax Memory-efficient multi-head dot product attention. https://arxiv.org/abs/2112.05682v2
+ https://github.com/AminRezaei0x443/memory-efficient-attention
+
+ Args:
+ query (`jnp.ndarray`): (batch..., query_length, head, query_key_depth_per_head)
+ key (`jnp.ndarray`): (batch..., key_value_length, head, query_key_depth_per_head)
+ value (`jnp.ndarray`): (batch..., key_value_length, head, value_depth_per_head)
+ precision (`jax.lax.Precision`, *optional*, defaults to `jax.lax.Precision.HIGHEST`):
+ numerical precision for computation
+ query_chunk_size (`int`, *optional*, defaults to 1024):
+ chunk size to divide query array value must divide query_length equally without remainder
+ key_chunk_size (`int`, *optional*, defaults to 4096):
+ chunk size to divide key and value array value must divide key_value_length equally without remainder
+
+ Returns:
+ (`jnp.ndarray`) with shape of (batch..., query_length, head, value_depth_per_head)
+ """
+ num_q, num_heads, q_features = query.shape[-3:]
+
+ def chunk_scanner(chunk_idx, _):
+ # julienne query array
+ query_chunk = jax.lax.dynamic_slice(
+ operand=query,
+ start_indices=([0] * (query.ndim - 3)) + [chunk_idx, 0, 0], # [...,q,h,d]
+ slice_sizes=list(query.shape[:-3]) + [min(query_chunk_size, num_q), num_heads, q_features], # [...,q,h,d]
+ )
+
+ return (
+ chunk_idx + query_chunk_size, # unused ignore it
+ _query_chunk_attention(
+ query=query_chunk, key=key, value=value, precision=precision, key_chunk_size=key_chunk_size
+ ),
+ )
+
+ _, res = jax.lax.scan(
+ f=chunk_scanner,
+ init=0,
+ xs=None,
+ length=math.ceil(num_q / query_chunk_size), # start counter # stop counter
+ )
+
+ return jnp.concatenate(res, axis=-3) # fuse the chunked result back
+
+
+class FlaxAttention(nn.Module):
+ r"""
+ A Flax multi-head attention module as described in: https://arxiv.org/abs/1706.03762
+
+ Parameters:
+ query_dim (:obj:`int`):
+ Input hidden states dimension
+ heads (:obj:`int`, *optional*, defaults to 8):
+ Number of heads
+ dim_head (:obj:`int`, *optional*, defaults to 64):
+ Hidden states dimension inside each head
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+ enable memory efficient attention https://arxiv.org/abs/2112.05682
+ split_head_dim (`bool`, *optional*, defaults to `False`):
+ Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+ enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+
+ """
+
+ query_dim: int
+ heads: int = 8
+ dim_head: int = 64
+ dropout: float = 0.0
+ use_memory_efficient_attention: bool = False
+ split_head_dim: bool = False
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ inner_dim = self.dim_head * self.heads
+ self.scale = self.dim_head**-0.5
+
+ # Weights were exported with old names {to_q, to_k, to_v, to_out}
+ self.query = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_q")
+ self.key = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_k")
+ self.value = nn.Dense(inner_dim, use_bias=False, dtype=self.dtype, name="to_v")
+
+ self.proj_attn = nn.Dense(self.query_dim, dtype=self.dtype, name="to_out_0")
+ self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+ def reshape_heads_to_batch_dim(self, tensor):
+ batch_size, seq_len, dim = tensor.shape
+ head_size = self.heads
+ tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size)
+ tensor = jnp.transpose(tensor, (0, 2, 1, 3))
+ tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size)
+ return tensor
+
+ def reshape_batch_dim_to_heads(self, tensor):
+ batch_size, seq_len, dim = tensor.shape
+ head_size = self.heads
+ tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+ tensor = jnp.transpose(tensor, (0, 2, 1, 3))
+ tensor = tensor.reshape(batch_size // head_size, seq_len, dim * head_size)
+ return tensor
+
+ def __call__(self, hidden_states, context=None, deterministic=True):
+ context = hidden_states if context is None else context
+
+ query_proj = self.query(hidden_states)
+ key_proj = self.key(context)
+ value_proj = self.value(context)
+
+ if self.split_head_dim:
+ b = hidden_states.shape[0]
+ query_states = jnp.reshape(query_proj, (b, -1, self.heads, self.dim_head))
+ key_states = jnp.reshape(key_proj, (b, -1, self.heads, self.dim_head))
+ value_states = jnp.reshape(value_proj, (b, -1, self.heads, self.dim_head))
+ else:
+ query_states = self.reshape_heads_to_batch_dim(query_proj)
+ key_states = self.reshape_heads_to_batch_dim(key_proj)
+ value_states = self.reshape_heads_to_batch_dim(value_proj)
+
+ if self.use_memory_efficient_attention:
+ query_states = query_states.transpose(1, 0, 2)
+ key_states = key_states.transpose(1, 0, 2)
+ value_states = value_states.transpose(1, 0, 2)
+
+ # this if statement create a chunk size for each layer of the unet
+ # the chunk size is equal to the query_length dimension of the deepest layer of the unet
+
+ flatten_latent_dim = query_states.shape[-3]
+ if flatten_latent_dim % 64 == 0:
+ query_chunk_size = int(flatten_latent_dim / 64)
+ elif flatten_latent_dim % 16 == 0:
+ query_chunk_size = int(flatten_latent_dim / 16)
+ elif flatten_latent_dim % 4 == 0:
+ query_chunk_size = int(flatten_latent_dim / 4)
+ else:
+ query_chunk_size = int(flatten_latent_dim)
+
+ hidden_states = jax_memory_efficient_attention(
+ query_states, key_states, value_states, query_chunk_size=query_chunk_size, key_chunk_size=4096 * 4
+ )
+
+ hidden_states = hidden_states.transpose(1, 0, 2)
+ else:
+ # compute attentions
+ if self.split_head_dim:
+ attention_scores = jnp.einsum("b t n h, b f n h -> b n f t", key_states, query_states)
+ else:
+ attention_scores = jnp.einsum("b i d, b j d->b i j", query_states, key_states)
+
+ attention_scores = attention_scores * self.scale
+ attention_probs = nn.softmax(attention_scores, axis=-1 if self.split_head_dim else 2)
+
+ # attend to values
+ if self.split_head_dim:
+ hidden_states = jnp.einsum("b n f t, b t n h -> b f n h", attention_probs, value_states)
+ b = hidden_states.shape[0]
+ hidden_states = jnp.reshape(hidden_states, (b, -1, self.heads * self.dim_head))
+ else:
+ hidden_states = jnp.einsum("b i j, b j d -> b i d", attention_probs, value_states)
+ hidden_states = self.reshape_batch_dim_to_heads(hidden_states)
+
+ hidden_states = self.proj_attn(hidden_states)
+ return self.dropout_layer(hidden_states, deterministic=deterministic)
+
+
+class FlaxBasicTransformerBlock(nn.Module):
+ r"""
+ A Flax transformer block layer with `GLU` (Gated Linear Unit) activation function as described in:
+ https://arxiv.org/abs/1706.03762
+
+
+ Parameters:
+ dim (:obj:`int`):
+ Inner hidden states dimension
+ n_heads (:obj:`int`):
+ Number of heads
+ d_head (:obj:`int`):
+ Hidden states dimension inside each head
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ only_cross_attention (`bool`, defaults to `False`):
+ Whether to only apply cross attention.
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+ enable memory efficient attention https://arxiv.org/abs/2112.05682
+ split_head_dim (`bool`, *optional*, defaults to `False`):
+ Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+ enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+ """
+
+ dim: int
+ n_heads: int
+ d_head: int
+ dropout: float = 0.0
+ only_cross_attention: bool = False
+ dtype: jnp.dtype = jnp.float32
+ use_memory_efficient_attention: bool = False
+ split_head_dim: bool = False
+
+ def setup(self):
+ # self attention (or cross_attention if only_cross_attention is True)
+ self.attn1 = FlaxAttention(
+ self.dim,
+ self.n_heads,
+ self.d_head,
+ self.dropout,
+ self.use_memory_efficient_attention,
+ self.split_head_dim,
+ dtype=self.dtype,
+ )
+ # cross attention
+ self.attn2 = FlaxAttention(
+ self.dim,
+ self.n_heads,
+ self.d_head,
+ self.dropout,
+ self.use_memory_efficient_attention,
+ self.split_head_dim,
+ dtype=self.dtype,
+ )
+ self.ff = FlaxFeedForward(dim=self.dim, dropout=self.dropout, dtype=self.dtype)
+ self.norm1 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
+ self.norm2 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
+ self.norm3 = nn.LayerNorm(epsilon=1e-5, dtype=self.dtype)
+ self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+ def __call__(self, hidden_states, context, deterministic=True):
+ # self attention
+ residual = hidden_states
+ if self.only_cross_attention:
+ hidden_states = self.attn1(self.norm1(hidden_states), context, deterministic=deterministic)
+ else:
+ hidden_states = self.attn1(self.norm1(hidden_states), deterministic=deterministic)
+ hidden_states = hidden_states + residual
+
+ # cross attention
+ residual = hidden_states
+ hidden_states = self.attn2(self.norm2(hidden_states), context, deterministic=deterministic)
+ hidden_states = hidden_states + residual
+
+ # feed forward
+ residual = hidden_states
+ hidden_states = self.ff(self.norm3(hidden_states), deterministic=deterministic)
+ hidden_states = hidden_states + residual
+
+ return self.dropout_layer(hidden_states, deterministic=deterministic)
+
+
+class FlaxTransformer2DModel(nn.Module):
+ r"""
+ A Spatial Transformer layer with Gated Linear Unit (GLU) activation function as described in:
+ https://arxiv.org/pdf/1506.02025.pdf
+
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input number of channels
+ n_heads (:obj:`int`):
+ Number of heads
+ d_head (:obj:`int`):
+ Hidden states dimension inside each head
+ depth (:obj:`int`, *optional*, defaults to 1):
+ Number of transformers block
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ use_linear_projection (`bool`, defaults to `False`): tbd
+ only_cross_attention (`bool`, defaults to `False`): tbd
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+ enable memory efficient attention https://arxiv.org/abs/2112.05682
+ split_head_dim (`bool`, *optional*, defaults to `False`):
+ Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+ enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+ """
+
+ in_channels: int
+ n_heads: int
+ d_head: int
+ depth: int = 1
+ dropout: float = 0.0
+ use_linear_projection: bool = False
+ only_cross_attention: bool = False
+ dtype: jnp.dtype = jnp.float32
+ use_memory_efficient_attention: bool = False
+ split_head_dim: bool = False
+
+ def setup(self):
+ self.norm = nn.GroupNorm(num_groups=32, epsilon=1e-5)
+
+ inner_dim = self.n_heads * self.d_head
+ if self.use_linear_projection:
+ self.proj_in = nn.Dense(inner_dim, dtype=self.dtype)
+ else:
+ self.proj_in = nn.Conv(
+ inner_dim,
+ kernel_size=(1, 1),
+ strides=(1, 1),
+ padding="VALID",
+ dtype=self.dtype,
+ )
+
+ self.transformer_blocks = [
+ FlaxBasicTransformerBlock(
+ inner_dim,
+ self.n_heads,
+ self.d_head,
+ dropout=self.dropout,
+ only_cross_attention=self.only_cross_attention,
+ dtype=self.dtype,
+ use_memory_efficient_attention=self.use_memory_efficient_attention,
+ split_head_dim=self.split_head_dim,
+ )
+ for _ in range(self.depth)
+ ]
+
+ if self.use_linear_projection:
+ self.proj_out = nn.Dense(inner_dim, dtype=self.dtype)
+ else:
+ self.proj_out = nn.Conv(
+ inner_dim,
+ kernel_size=(1, 1),
+ strides=(1, 1),
+ padding="VALID",
+ dtype=self.dtype,
+ )
+
+ self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+ def __call__(self, hidden_states, context, deterministic=True):
+ batch, height, width, channels = hidden_states.shape
+ residual = hidden_states
+ hidden_states = self.norm(hidden_states)
+ if self.use_linear_projection:
+ hidden_states = hidden_states.reshape(batch, height * width, channels)
+ hidden_states = self.proj_in(hidden_states)
+ else:
+ hidden_states = self.proj_in(hidden_states)
+ hidden_states = hidden_states.reshape(batch, height * width, channels)
+
+ for transformer_block in self.transformer_blocks:
+ hidden_states = transformer_block(hidden_states, context, deterministic=deterministic)
+
+ if self.use_linear_projection:
+ hidden_states = self.proj_out(hidden_states)
+ hidden_states = hidden_states.reshape(batch, height, width, channels)
+ else:
+ hidden_states = hidden_states.reshape(batch, height, width, channels)
+ hidden_states = self.proj_out(hidden_states)
+
+ hidden_states = hidden_states + residual
+ return self.dropout_layer(hidden_states, deterministic=deterministic)
+
+
+class FlaxFeedForward(nn.Module):
+ r"""
+ Flax module that encapsulates two Linear layers separated by a non-linearity. It is the counterpart of PyTorch's
+ [`FeedForward`] class, with the following simplifications:
+ - The activation function is currently hardcoded to a gated linear unit from:
+ https://arxiv.org/abs/2002.05202
+ - `dim_out` is equal to `dim`.
+ - The number of hidden dimensions is hardcoded to `dim * 4` in [`FlaxGELU`].
+
+ Parameters:
+ dim (:obj:`int`):
+ Inner hidden states dimension
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ dim: int
+ dropout: float = 0.0
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ # The second linear layer needs to be called
+ # net_2 for now to match the index of the Sequential layer
+ self.net_0 = FlaxGEGLU(self.dim, self.dropout, self.dtype)
+ self.net_2 = nn.Dense(self.dim, dtype=self.dtype)
+
+ def __call__(self, hidden_states, deterministic=True):
+ hidden_states = self.net_0(hidden_states, deterministic=deterministic)
+ hidden_states = self.net_2(hidden_states)
+ return hidden_states
+
+
+class FlaxGEGLU(nn.Module):
+ r"""
+ Flax implementation of a Linear layer followed by the variant of the gated linear unit activation function from
+ https://arxiv.org/abs/2002.05202.
+
+ Parameters:
+ dim (:obj:`int`):
+ Input hidden states dimension
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ dim: int
+ dropout: float = 0.0
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ inner_dim = self.dim * 4
+ self.proj = nn.Dense(inner_dim * 2, dtype=self.dtype)
+ self.dropout_layer = nn.Dropout(rate=self.dropout)
+
+ def __call__(self, hidden_states, deterministic=True):
+ hidden_states = self.proj(hidden_states)
+ hidden_linear, hidden_gelu = jnp.split(hidden_states, 2, axis=2)
+ return self.dropout_layer(hidden_linear * nn.gelu(hidden_gelu), deterministic=deterministic)
diff --git a/model/diffusers_c/models/attention_processor.py b/model/diffusers_c/models/attention_processor.py
new file mode 100644
index 0000000000000000000000000000000000000000..5ec8876fc11465effe3ba63fb5d454e55a75d805
--- /dev/null
+++ b/model/diffusers_c/models/attention_processor.py
@@ -0,0 +1,2463 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import inspect
+from importlib import import_module
+from typing import Callable, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..image_processor import IPAdapterMaskProcessor
+from ..utils import USE_PEFT_BACKEND, deprecate, logging
+from ..utils.import_utils import is_xformers_available
+from ..utils.torch_utils import maybe_allow_in_graph
+from .lora import LoRACompatibleLinear, LoRALinearLayer
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+if is_xformers_available():
+ import xformers
+ import xformers.ops
+else:
+ xformers = None
+
+
+@maybe_allow_in_graph
+class Attention(nn.Module):
+ r"""
+ A cross attention layer.
+
+ Parameters:
+ query_dim (`int`):
+ The number of channels in the query.
+ cross_attention_dim (`int`, *optional*):
+ The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`.
+ heads (`int`, *optional*, defaults to 8):
+ The number of heads to use for multi-head attention.
+ dim_head (`int`, *optional*, defaults to 64):
+ The number of channels in each head.
+ dropout (`float`, *optional*, defaults to 0.0):
+ The dropout probability to use.
+ bias (`bool`, *optional*, defaults to False):
+ Set to `True` for the query, key, and value linear layers to contain a bias parameter.
+ upcast_attention (`bool`, *optional*, defaults to False):
+ Set to `True` to upcast the attention computation to `float32`.
+ upcast_softmax (`bool`, *optional*, defaults to False):
+ Set to `True` to upcast the softmax computation to `float32`.
+ cross_attention_norm (`str`, *optional*, defaults to `None`):
+ The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
+ cross_attention_norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use for the group norm in the cross attention.
+ added_kv_proj_dim (`int`, *optional*, defaults to `None`):
+ The number of channels to use for the added key and value projections. If `None`, no projection is used.
+ norm_num_groups (`int`, *optional*, defaults to `None`):
+ The number of groups to use for the group norm in the attention.
+ spatial_norm_dim (`int`, *optional*, defaults to `None`):
+ The number of channels to use for the spatial normalization.
+ out_bias (`bool`, *optional*, defaults to `True`):
+ Set to `True` to use a bias in the output linear layer.
+ scale_qk (`bool`, *optional*, defaults to `True`):
+ Set to `True` to scale the query and key by `1 / sqrt(dim_head)`.
+ only_cross_attention (`bool`, *optional*, defaults to `False`):
+ Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if
+ `added_kv_proj_dim` is not `None`.
+ eps (`float`, *optional*, defaults to 1e-5):
+ An additional value added to the denominator in group normalization that is used for numerical stability.
+ rescale_output_factor (`float`, *optional*, defaults to 1.0):
+ A factor to rescale the output by dividing it with this value.
+ residual_connection (`bool`, *optional*, defaults to `False`):
+ Set to `True` to add the residual connection to the output.
+ _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`):
+ Set to `True` if the attention block is loaded from a deprecated state dict.
+ processor (`AttnProcessor`, *optional*, defaults to `None`):
+ The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and
+ `AttnProcessor` otherwise.
+ """
+
+ def __init__(
+ self,
+ query_dim: int,
+ cross_attention_dim: Optional[int] = None,
+ heads: int = 8,
+ dim_head: int = 64,
+ dropout: float = 0.0,
+ bias: bool = False,
+ upcast_attention: bool = False,
+ upcast_softmax: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ cross_attention_norm_num_groups: int = 32,
+ added_kv_proj_dim: Optional[int] = None,
+ norm_num_groups: Optional[int] = None,
+ spatial_norm_dim: Optional[int] = None,
+ out_bias: bool = True,
+ scale_qk: bool = True,
+ only_cross_attention: bool = False,
+ eps: float = 1e-5,
+ rescale_output_factor: float = 1.0,
+ residual_connection: bool = False,
+ _from_deprecated_attn_block: bool = False,
+ processor: Optional["AttnProcessor"] = None,
+ out_dim: int = None,
+ ):
+ super().__init__()
+ self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+ self.query_dim = query_dim
+ self.use_bias = bias
+ self.is_cross_attention = cross_attention_dim is not None
+ self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+ self.upcast_attention = upcast_attention
+ self.upcast_softmax = upcast_softmax
+ self.rescale_output_factor = rescale_output_factor
+ self.residual_connection = residual_connection
+ self.dropout = dropout
+ self.fused_projections = False
+ self.out_dim = out_dim if out_dim is not None else query_dim
+
+ # we make use of this private variable to know whether this class is loaded
+ # with an deprecated state dict so that we can convert it on the fly
+ self._from_deprecated_attn_block = _from_deprecated_attn_block
+
+ self.scale_qk = scale_qk
+ self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+ self.heads = out_dim // dim_head if out_dim is not None else heads
+ # for slice_size > 0 the attention score computation
+ # is split across the batch axis to save memory
+ # You can set slice_size with `set_attention_slice`
+ self.sliceable_head_dim = heads
+
+ self.added_kv_proj_dim = added_kv_proj_dim
+ self.only_cross_attention = only_cross_attention
+
+ if self.added_kv_proj_dim is None and self.only_cross_attention:
+ raise ValueError(
+ "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
+ )
+
+ if norm_num_groups is not None:
+ self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True)
+ else:
+ self.group_norm = None
+
+ if spatial_norm_dim is not None:
+ self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim)
+ else:
+ self.spatial_norm = None
+
+ if cross_attention_norm is None:
+ self.norm_cross = None
+ elif cross_attention_norm == "layer_norm":
+ self.norm_cross = nn.LayerNorm(self.cross_attention_dim)
+ elif cross_attention_norm == "group_norm":
+ if self.added_kv_proj_dim is not None:
+ # The given `encoder_hidden_states` are initially of shape
+ # (batch_size, seq_len, added_kv_proj_dim) before being projected
+ # to (batch_size, seq_len, cross_attention_dim). The norm is applied
+ # before the projection, so we need to use `added_kv_proj_dim` as
+ # the number of channels for the group norm.
+ norm_cross_num_channels = added_kv_proj_dim
+ else:
+ norm_cross_num_channels = self.cross_attention_dim
+
+ self.norm_cross = nn.GroupNorm(
+ num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True
+ )
+ else:
+ raise ValueError(
+ f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'"
+ )
+
+ if USE_PEFT_BACKEND:
+ linear_cls = nn.Linear
+ else:
+ linear_cls = LoRACompatibleLinear
+
+ self.linear_cls = linear_cls
+ self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias)
+
+ if not self.only_cross_attention:
+ # only relevant for the `AddedKVProcessor` classes
+ self.to_k = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias)
+ self.to_v = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias)
+ else:
+ self.to_k = None
+ self.to_v = None
+
+ if self.added_kv_proj_dim is not None:
+ self.add_k_proj = linear_cls(added_kv_proj_dim, self.inner_dim)
+ self.add_v_proj = linear_cls(added_kv_proj_dim, self.inner_dim)
+
+ self.to_out = nn.ModuleList([])
+ self.to_out.append(linear_cls(self.inner_dim, self.out_dim, bias=out_bias))
+ self.to_out.append(nn.Dropout(dropout))
+
+ # set attention processor
+ # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+ # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+ # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+ if processor is None:
+ processor = (
+ AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
+ )
+ self.set_processor(processor)
+
+ def set_use_memory_efficient_attention_xformers(
+ self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None
+ ) -> None:
+ r"""
+ Set whether to use memory efficient attention from `xformers` or not.
+
+ Args:
+ use_memory_efficient_attention_xformers (`bool`):
+ Whether to use memory efficient attention from `xformers` or not.
+ attention_op (`Callable`, *optional*):
+ The attention operation to use. Defaults to `None` which uses the default attention operation from
+ `xformers`.
+ """
+ is_lora = hasattr(self, "processor") and isinstance(
+ self.processor,
+ LORA_ATTENTION_PROCESSORS,
+ )
+ is_custom_diffusion = hasattr(self, "processor") and isinstance(
+ self.processor,
+ (CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0),
+ )
+ is_added_kv_processor = hasattr(self, "processor") and isinstance(
+ self.processor,
+ (
+ AttnAddedKVProcessor,
+ AttnAddedKVProcessor2_0,
+ SlicedAttnAddedKVProcessor,
+ XFormersAttnAddedKVProcessor,
+ LoRAAttnAddedKVProcessor,
+ ),
+ )
+
+ if use_memory_efficient_attention_xformers:
+ if is_added_kv_processor and (is_lora or is_custom_diffusion):
+ raise NotImplementedError(
+ f"Memory efficient attention is currently not supported for LoRA or custom diffusion for attention processor type {self.processor}"
+ )
+ if not is_xformers_available():
+ raise ModuleNotFoundError(
+ (
+ "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+ " xformers"
+ ),
+ name="xformers",
+ )
+ elif not torch.cuda.is_available():
+ raise ValueError(
+ "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is"
+ " only available for GPU "
+ )
+ else:
+ try:
+ # Make sure we can run the memory efficient attention
+ _ = xformers.ops.memory_efficient_attention(
+ torch.randn((1, 2, 40), device="cuda"),
+ torch.randn((1, 2, 40), device="cuda"),
+ torch.randn((1, 2, 40), device="cuda"),
+ )
+ except Exception as e:
+ raise e
+
+ if is_lora:
+ # TODO (sayakpaul): should we throw a warning if someone wants to use the xformers
+ # variant when using PT 2.0 now that we have LoRAAttnProcessor2_0?
+ processor = LoRAXFormersAttnProcessor(
+ hidden_size=self.processor.hidden_size,
+ cross_attention_dim=self.processor.cross_attention_dim,
+ rank=self.processor.rank,
+ attention_op=attention_op,
+ )
+ processor.load_state_dict(self.processor.state_dict())
+ processor.to(self.processor.to_q_lora.up.weight.device)
+ elif is_custom_diffusion:
+ processor = CustomDiffusionXFormersAttnProcessor(
+ train_kv=self.processor.train_kv,
+ train_q_out=self.processor.train_q_out,
+ hidden_size=self.processor.hidden_size,
+ cross_attention_dim=self.processor.cross_attention_dim,
+ attention_op=attention_op,
+ )
+ processor.load_state_dict(self.processor.state_dict())
+ if hasattr(self.processor, "to_k_custom_diffusion"):
+ processor.to(self.processor.to_k_custom_diffusion.weight.device)
+ elif is_added_kv_processor:
+ # TODO(Patrick, Suraj, William) - currently xformers doesn't work for UnCLIP
+ # which uses this type of cross attention ONLY because the attention mask of format
+ # [0, ..., -10.000, ..., 0, ...,] is not supported
+ # throw warning
+ logger.info(
+ "Memory efficient attention with `xformers` might currently not work correctly if an attention mask is required for the attention operation."
+ )
+ processor = XFormersAttnAddedKVProcessor(attention_op=attention_op)
+ else:
+ processor = XFormersAttnProcessor(attention_op=attention_op)
+ else:
+ if is_lora:
+ attn_processor_class = (
+ LoRAAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else LoRAAttnProcessor
+ )
+ processor = attn_processor_class(
+ hidden_size=self.processor.hidden_size,
+ cross_attention_dim=self.processor.cross_attention_dim,
+ rank=self.processor.rank,
+ )
+ processor.load_state_dict(self.processor.state_dict())
+ processor.to(self.processor.to_q_lora.up.weight.device)
+ elif is_custom_diffusion:
+ attn_processor_class = (
+ CustomDiffusionAttnProcessor2_0
+ if hasattr(F, "scaled_dot_product_attention")
+ else CustomDiffusionAttnProcessor
+ )
+ processor = attn_processor_class(
+ train_kv=self.processor.train_kv,
+ train_q_out=self.processor.train_q_out,
+ hidden_size=self.processor.hidden_size,
+ cross_attention_dim=self.processor.cross_attention_dim,
+ )
+ processor.load_state_dict(self.processor.state_dict())
+ if hasattr(self.processor, "to_k_custom_diffusion"):
+ processor.to(self.processor.to_k_custom_diffusion.weight.device)
+ else:
+ # set attention processor
+ # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+ # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+ # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+ processor = (
+ AttnProcessor2_0()
+ if hasattr(F, "scaled_dot_product_attention") and self.scale_qk
+ else AttnProcessor()
+ )
+
+ self.set_processor(processor)
+
+ def set_attention_slice(self, slice_size: int) -> None:
+ r"""
+ Set the slice size for attention computation.
+
+ Args:
+ slice_size (`int`):
+ The slice size for attention computation.
+ """
+ if slice_size is not None and slice_size > self.sliceable_head_dim:
+ raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.")
+
+ if slice_size is not None and self.added_kv_proj_dim is not None:
+ processor = SlicedAttnAddedKVProcessor(slice_size)
+ elif slice_size is not None:
+ processor = SlicedAttnProcessor(slice_size)
+ elif self.added_kv_proj_dim is not None:
+ processor = AttnAddedKVProcessor()
+ else:
+ # set attention processor
+ # We use the AttnProcessor2_0 by default when torch 2.x is used which uses
+ # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention
+ # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1
+ processor = (
+ AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
+ )
+
+ self.set_processor(processor)
+
+ def set_processor(self, processor: "AttnProcessor") -> None:
+ r"""
+ Set the attention processor to use.
+
+ Args:
+ processor (`AttnProcessor`):
+ The attention processor to use.
+ """
+ # if current processor is in `self._modules` and if passed `processor` is not, we need to
+ # pop `processor` from `self._modules`
+ if (
+ hasattr(self, "processor")
+ and isinstance(self.processor, torch.nn.Module)
+ and not isinstance(processor, torch.nn.Module)
+ ):
+ logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}")
+ self._modules.pop("processor")
+
+ self.processor = processor
+
+ def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor":
+ r"""
+ Get the attention processor in use.
+
+ Args:
+ return_deprecated_lora (`bool`, *optional*, defaults to `False`):
+ Set to `True` to return the deprecated LoRA attention processor.
+
+ Returns:
+ "AttentionProcessor": The attention processor in use.
+ """
+ if not return_deprecated_lora:
+ return self.processor
+
+ # TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible
+ # serialization format for LoRA Attention Processors. It should be deleted once the integration
+ # with PEFT is completed.
+ is_lora_activated = {
+ name: module.lora_layer is not None
+ for name, module in self.named_modules()
+ if hasattr(module, "lora_layer")
+ }
+
+ # 1. if no layer has a LoRA activated we can return the processor as usual
+ if not any(is_lora_activated.values()):
+ return self.processor
+
+ # If doesn't apply LoRA do `add_k_proj` or `add_v_proj`
+ is_lora_activated.pop("add_k_proj", None)
+ is_lora_activated.pop("add_v_proj", None)
+ # 2. else it is not posssible that only some layers have LoRA activated
+ if not all(is_lora_activated.values()):
+ raise ValueError(
+ f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}"
+ )
+
+ # 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor
+ non_lora_processor_cls_name = self.processor.__class__.__name__
+ lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name)
+
+ hidden_size = self.inner_dim
+
+ # now create a LoRA attention processor from the LoRA layers
+ if lora_processor_cls in [LoRAAttnProcessor, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor]:
+ kwargs = {
+ "cross_attention_dim": self.cross_attention_dim,
+ "rank": self.to_q.lora_layer.rank,
+ "network_alpha": self.to_q.lora_layer.network_alpha,
+ "q_rank": self.to_q.lora_layer.rank,
+ "q_hidden_size": self.to_q.lora_layer.out_features,
+ "k_rank": self.to_k.lora_layer.rank,
+ "k_hidden_size": self.to_k.lora_layer.out_features,
+ "v_rank": self.to_v.lora_layer.rank,
+ "v_hidden_size": self.to_v.lora_layer.out_features,
+ "out_rank": self.to_out[0].lora_layer.rank,
+ "out_hidden_size": self.to_out[0].lora_layer.out_features,
+ }
+
+ if hasattr(self.processor, "attention_op"):
+ kwargs["attention_op"] = self.processor.attention_op
+
+ lora_processor = lora_processor_cls(hidden_size, **kwargs)
+ lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())
+ lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())
+ lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())
+ lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())
+ elif lora_processor_cls == LoRAAttnAddedKVProcessor:
+ lora_processor = lora_processor_cls(
+ hidden_size,
+ cross_attention_dim=self.add_k_proj.weight.shape[0],
+ rank=self.to_q.lora_layer.rank,
+ network_alpha=self.to_q.lora_layer.network_alpha,
+ )
+ lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())
+ lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())
+ lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())
+ lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())
+
+ # only save if used
+ if self.add_k_proj.lora_layer is not None:
+ lora_processor.add_k_proj_lora.load_state_dict(self.add_k_proj.lora_layer.state_dict())
+ lora_processor.add_v_proj_lora.load_state_dict(self.add_v_proj.lora_layer.state_dict())
+ else:
+ lora_processor.add_k_proj_lora = None
+ lora_processor.add_v_proj_lora = None
+ else:
+ raise ValueError(f"{lora_processor_cls} does not exist.")
+
+ return lora_processor
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ **cross_attention_kwargs,
+ ) -> torch.Tensor:
+ r"""
+ The forward method of the `Attention` class.
+
+ Args:
+ hidden_states (`torch.Tensor`):
+ The hidden states of the query.
+ encoder_hidden_states (`torch.Tensor`, *optional*):
+ The hidden states of the encoder.
+ attention_mask (`torch.Tensor`, *optional*):
+ The attention mask to use. If `None`, no mask is applied.
+ **cross_attention_kwargs:
+ Additional keyword arguments to pass along to the cross attention.
+
+ Returns:
+ `torch.Tensor`: The output of the attention layer.
+ """
+ # The `Attention` class can call different attention processors / attention functions
+ # here we simply pass along all tensors to the selected processor class
+ # For standard processors that are defined here, `**cross_attention_kwargs` is empty
+
+ attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
+ unused_kwargs = [k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters]
+ if len(unused_kwargs) > 0:
+ logger.warning(
+ f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
+ )
+ cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters}
+
+ return self.processor(
+ self,
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ **cross_attention_kwargs,
+ )
+
+ def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor:
+ r"""
+ Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads`
+ is the number of heads initialized while constructing the `Attention` class.
+
+ Args:
+ tensor (`torch.Tensor`): The tensor to reshape.
+
+ Returns:
+ `torch.Tensor`: The reshaped tensor.
+ """
+ head_size = self.heads
+ batch_size, seq_len, dim = tensor.shape
+ tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim)
+ tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size)
+ return tensor
+
+ def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor:
+ r"""
+ Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is
+ the number of heads initialized while constructing the `Attention` class.
+
+ Args:
+ tensor (`torch.Tensor`): The tensor to reshape.
+ out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is
+ reshaped to `[batch_size * heads, seq_len, dim // heads]`.
+
+ Returns:
+ `torch.Tensor`: The reshaped tensor.
+ """
+ head_size = self.heads
+ if tensor.ndim == 3:
+ batch_size, seq_len, dim = tensor.shape
+ extra_dim = 1
+ else:
+ batch_size, extra_dim, seq_len, dim = tensor.shape
+ tensor = tensor.reshape(batch_size, seq_len * extra_dim, head_size, dim // head_size)
+ tensor = tensor.permute(0, 2, 1, 3)
+
+ if out_dim == 3:
+ tensor = tensor.reshape(batch_size * head_size, seq_len * extra_dim, dim // head_size)
+
+ return tensor
+
+ def get_attention_scores(
+ self, query: torch.Tensor, key: torch.Tensor, attention_mask: torch.Tensor = None
+ ) -> torch.Tensor:
+ r"""
+ Compute the attention scores.
+
+ Args:
+ query (`torch.Tensor`): The query tensor.
+ key (`torch.Tensor`): The key tensor.
+ attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied.
+
+ Returns:
+ `torch.Tensor`: The attention probabilities/scores.
+ """
+ dtype = query.dtype
+ if self.upcast_attention:
+ query = query.float()
+ key = key.float()
+
+ if attention_mask is None:
+ baddbmm_input = torch.empty(
+ query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device
+ )
+ beta = 0
+ else:
+ baddbmm_input = attention_mask
+ beta = 1
+
+ attention_scores = torch.baddbmm(
+ baddbmm_input,
+ query,
+ key.transpose(-1, -2),
+ beta=beta,
+ alpha=self.scale,
+ )
+ del baddbmm_input
+
+ if self.upcast_softmax:
+ attention_scores = attention_scores.float()
+
+ attention_probs = attention_scores.softmax(dim=-1)
+ del attention_scores
+
+ attention_probs = attention_probs.to(dtype)
+
+ return attention_probs
+
+ def prepare_attention_mask(
+ self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3
+ ) -> torch.Tensor:
+ r"""
+ Prepare the attention mask for the attention computation.
+
+ Args:
+ attention_mask (`torch.Tensor`):
+ The attention mask to prepare.
+ target_length (`int`):
+ The target length of the attention mask. This is the length of the attention mask after padding.
+ batch_size (`int`):
+ The batch size, which is used to repeat the attention mask.
+ out_dim (`int`, *optional*, defaults to `3`):
+ The output dimension of the attention mask. Can be either `3` or `4`.
+
+ Returns:
+ `torch.Tensor`: The prepared attention mask.
+ """
+ head_size = self.heads
+ if attention_mask is None:
+ return attention_mask
+
+ current_length: int = attention_mask.shape[-1]
+ if current_length != target_length:
+ if attention_mask.device.type == "mps":
+ # HACK: MPS: Does not support padding by greater than dimension of input tensor.
+ # Instead, we can manually construct the padding tensor.
+ padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length)
+ padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device)
+ attention_mask = torch.cat([attention_mask, padding], dim=2)
+ else:
+ # TODO: for pipelines such as stable-diffusion, padding cross-attn mask:
+ # we want to instead pad by (0, remaining_length), where remaining_length is:
+ # remaining_length: int = target_length - current_length
+ # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding
+ attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+
+ if out_dim == 3:
+ if attention_mask.shape[0] < batch_size * head_size:
+ attention_mask = attention_mask.repeat_interleave(head_size, dim=0)
+ elif out_dim == 4:
+ attention_mask = attention_mask.unsqueeze(1)
+ attention_mask = attention_mask.repeat_interleave(head_size, dim=1)
+
+ return attention_mask
+
+ def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor:
+ r"""
+ Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the
+ `Attention` class.
+
+ Args:
+ encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder.
+
+ Returns:
+ `torch.Tensor`: The normalized encoder hidden states.
+ """
+ assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states"
+
+ if isinstance(self.norm_cross, nn.LayerNorm):
+ encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+ elif isinstance(self.norm_cross, nn.GroupNorm):
+ # Group norm norms along the channels dimension and expects
+ # input to be in the shape of (N, C, *). In this case, we want
+ # to norm along the hidden dimension, so we need to move
+ # (batch_size, sequence_length, hidden_size) ->
+ # (batch_size, hidden_size, sequence_length)
+ encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+ encoder_hidden_states = self.norm_cross(encoder_hidden_states)
+ encoder_hidden_states = encoder_hidden_states.transpose(1, 2)
+ else:
+ assert False
+
+ return encoder_hidden_states
+
+ @torch.no_grad()
+ def fuse_projections(self, fuse=True):
+ device = self.to_q.weight.data.device
+ dtype = self.to_q.weight.data.dtype
+
+ if not self.is_cross_attention:
+ # fetch weight matrices.
+ concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data])
+ in_features = concatenated_weights.shape[1]
+ out_features = concatenated_weights.shape[0]
+
+ # create a new single projection layer and copy over the weights.
+ self.to_qkv = self.linear_cls(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
+ self.to_qkv.weight.copy_(concatenated_weights)
+ if self.use_bias:
+ concatenated_bias = torch.cat([self.to_q.bias.data, self.to_k.bias.data, self.to_v.bias.data])
+ self.to_qkv.bias.copy_(concatenated_bias)
+
+ else:
+ concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data])
+ in_features = concatenated_weights.shape[1]
+ out_features = concatenated_weights.shape[0]
+
+ self.to_kv = self.linear_cls(in_features, out_features, bias=self.use_bias, device=device, dtype=dtype)
+ self.to_kv.weight.copy_(concatenated_weights)
+ if self.use_bias:
+ concatenated_bias = torch.cat([self.to_k.bias.data, self.to_v.bias.data])
+ self.to_kv.bias.copy_(concatenated_bias)
+
+ self.fused_projections = fuse
+
+
+class AttnProcessor:
+ r"""
+ Default processor for performing attention-related computations.
+ """
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ temb: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> torch.Tensor:
+ residual = hidden_states
+
+ args = () if USE_PEFT_BACKEND else (scale,)
+
+ if attn.spatial_norm is not None:
+ hidden_states = attn.spatial_norm(hidden_states, temb)
+
+ input_ndim = hidden_states.ndim
+
+ if input_ndim == 4:
+ batch_size, channel, height, width = hidden_states.shape
+ hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+ if attn.group_norm is not None:
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states, *args)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states, *args)
+ value = attn.to_v(encoder_hidden_states, *args)
+
+ query = attn.head_to_batch_dim(query)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ attention_probs = attn.get_attention_scores(query, key, attention_mask)
+ hidden_states = torch.bmm(attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states, *args)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ if input_ndim == 4:
+ hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+ if attn.residual_connection:
+ hidden_states = hidden_states + residual
+
+ hidden_states = hidden_states / attn.rescale_output_factor
+
+ return hidden_states
+
+
+class CustomDiffusionAttnProcessor(nn.Module):
+ r"""
+ Processor for implementing attention for the Custom Diffusion method.
+
+ Args:
+ train_kv (`bool`, defaults to `True`):
+ Whether to newly train the key and value matrices corresponding to the text features.
+ train_q_out (`bool`, defaults to `True`):
+ Whether to newly train query matrices corresponding to the latent image features.
+ hidden_size (`int`, *optional*, defaults to `None`):
+ The hidden size of the attention layer.
+ cross_attention_dim (`int`, *optional*, defaults to `None`):
+ The number of channels in the `encoder_hidden_states`.
+ out_bias (`bool`, defaults to `True`):
+ Whether to include the bias parameter in `train_q_out`.
+ dropout (`float`, *optional*, defaults to 0.0):
+ The dropout probability to use.
+ """
+
+ def __init__(
+ self,
+ train_kv: bool = True,
+ train_q_out: bool = True,
+ hidden_size: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ out_bias: bool = True,
+ dropout: float = 0.0,
+ ):
+ super().__init__()
+ self.train_kv = train_kv
+ self.train_q_out = train_q_out
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+
+ # `_custom_diffusion` id for easy serialization and loading.
+ if self.train_kv:
+ self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+ self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+ if self.train_q_out:
+ self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)
+ self.to_out_custom_diffusion = nn.ModuleList([])
+ self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))
+ self.to_out_custom_diffusion.append(nn.Dropout(dropout))
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.Tensor:
+ batch_size, sequence_length, _ = hidden_states.shape
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ if self.train_q_out:
+ query = self.to_q_custom_diffusion(hidden_states).to(attn.to_q.weight.dtype)
+ else:
+ query = attn.to_q(hidden_states.to(attn.to_q.weight.dtype))
+
+ if encoder_hidden_states is None:
+ crossattn = False
+ encoder_hidden_states = hidden_states
+ else:
+ crossattn = True
+ if attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ if self.train_kv:
+ key = self.to_k_custom_diffusion(encoder_hidden_states.to(self.to_k_custom_diffusion.weight.dtype))
+ value = self.to_v_custom_diffusion(encoder_hidden_states.to(self.to_v_custom_diffusion.weight.dtype))
+ key = key.to(attn.to_q.weight.dtype)
+ value = value.to(attn.to_q.weight.dtype)
+ else:
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ if crossattn:
+ detach = torch.ones_like(key)
+ detach[:, :1, :] = detach[:, :1, :] * 0.0
+ key = detach * key + (1 - detach) * key.detach()
+ value = detach * value + (1 - detach) * value.detach()
+
+ query = attn.head_to_batch_dim(query)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ attention_probs = attn.get_attention_scores(query, key, attention_mask)
+ hidden_states = torch.bmm(attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ if self.train_q_out:
+ # linear proj
+ hidden_states = self.to_out_custom_diffusion[0](hidden_states)
+ # dropout
+ hidden_states = self.to_out_custom_diffusion[1](hidden_states)
+ else:
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ return hidden_states
+
+
+class AttnAddedKVProcessor:
+ r"""
+ Processor for performing attention-related computations with extra learnable key and value matrices for the text
+ encoder.
+ """
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> torch.Tensor:
+ residual = hidden_states
+
+ args = () if USE_PEFT_BACKEND else (scale,)
+
+ hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+ batch_size, sequence_length, _ = hidden_states.shape
+
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states, *args)
+ query = attn.head_to_batch_dim(query)
+
+ encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states, *args)
+ encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states, *args)
+ encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)
+ encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)
+
+ if not attn.only_cross_attention:
+ key = attn.to_k(hidden_states, *args)
+ value = attn.to_v(hidden_states, *args)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+ key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
+ value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
+ else:
+ key = encoder_hidden_states_key_proj
+ value = encoder_hidden_states_value_proj
+
+ attention_probs = attn.get_attention_scores(query, key, attention_mask)
+ hidden_states = torch.bmm(attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states, *args)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+ hidden_states = hidden_states + residual
+
+ return hidden_states
+
+
+class AttnAddedKVProcessor2_0:
+ r"""
+ Processor for performing scaled dot-product attention (enabled by default if you're using PyTorch 2.0), with extra
+ learnable key and value matrices for the text encoder.
+ """
+
+ def __init__(self):
+ if not hasattr(F, "scaled_dot_product_attention"):
+ raise ImportError(
+ "AttnAddedKVProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+ )
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> torch.Tensor:
+ residual = hidden_states
+
+ args = () if USE_PEFT_BACKEND else (scale,)
+
+ hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+ batch_size, sequence_length, _ = hidden_states.shape
+
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size, out_dim=4)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states, *args)
+ query = attn.head_to_batch_dim(query, out_dim=4)
+
+ encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+ encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+ encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj, out_dim=4)
+ encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj, out_dim=4)
+
+ if not attn.only_cross_attention:
+ key = attn.to_k(hidden_states, *args)
+ value = attn.to_v(hidden_states, *args)
+ key = attn.head_to_batch_dim(key, out_dim=4)
+ value = attn.head_to_batch_dim(value, out_dim=4)
+ key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
+ value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
+ else:
+ key = encoder_hidden_states_key_proj
+ value = encoder_hidden_states_value_proj
+
+ # the output of sdp = (batch, num_heads, seq_len, head_dim)
+ # TODO: add support for attn.scale when we move to Torch 2.1
+ hidden_states = F.scaled_dot_product_attention(
+ query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+ )
+ hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, residual.shape[1])
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states, *args)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+ hidden_states = hidden_states + residual
+
+ return hidden_states
+
+
+class XFormersAttnAddedKVProcessor:
+ r"""
+ Processor for implementing memory efficient attention using xFormers.
+
+ Args:
+ attention_op (`Callable`, *optional*, defaults to `None`):
+ The base
+ [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
+ use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
+ operator.
+ """
+
+ def __init__(self, attention_op: Optional[Callable] = None):
+ self.attention_op = attention_op
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.Tensor:
+ residual = hidden_states
+ hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+ batch_size, sequence_length, _ = hidden_states.shape
+
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states)
+ query = attn.head_to_batch_dim(query)
+
+ encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+ encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+ encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)
+ encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)
+
+ if not attn.only_cross_attention:
+ key = attn.to_k(hidden_states)
+ value = attn.to_v(hidden_states)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+ key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
+ value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
+ else:
+ key = encoder_hidden_states_key_proj
+ value = encoder_hidden_states_value_proj
+
+ hidden_states = xformers.ops.memory_efficient_attention(
+ query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
+ )
+ hidden_states = hidden_states.to(query.dtype)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+ hidden_states = hidden_states + residual
+
+ return hidden_states
+
+
+class XFormersAttnProcessor:
+ r"""
+ Processor for implementing memory efficient attention using xFormers.
+
+ Args:
+ attention_op (`Callable`, *optional*, defaults to `None`):
+ The base
+ [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
+ use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
+ operator.
+ """
+
+ def __init__(self, attention_op: Optional[Callable] = None):
+ self.attention_op = attention_op
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ temb: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ residual = hidden_states
+
+ args = () if USE_PEFT_BACKEND else (scale,)
+
+ if attn.spatial_norm is not None:
+ hidden_states = attn.spatial_norm(hidden_states, temb)
+
+ input_ndim = hidden_states.ndim
+
+ if input_ndim == 4:
+ batch_size, channel, height, width = hidden_states.shape
+ hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+ batch_size, key_tokens, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+
+ attention_mask = attn.prepare_attention_mask(attention_mask, key_tokens, batch_size)
+ if attention_mask is not None:
+ # expand our mask's singleton query_tokens dimension:
+ # [batch*heads, 1, key_tokens] ->
+ # [batch*heads, query_tokens, key_tokens]
+ # so that it can be added as a bias onto the attention scores that xformers computes:
+ # [batch*heads, query_tokens, key_tokens]
+ # we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
+ _, query_tokens, _ = hidden_states.shape
+ attention_mask = attention_mask.expand(-1, query_tokens, -1)
+
+ if attn.group_norm is not None:
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states, *args)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states, *args)
+ value = attn.to_v(encoder_hidden_states, *args)
+
+ query = attn.head_to_batch_dim(query).contiguous()
+ key = attn.head_to_batch_dim(key).contiguous()
+ value = attn.head_to_batch_dim(value).contiguous()
+
+ hidden_states = xformers.ops.memory_efficient_attention(
+ query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
+ )
+ hidden_states = hidden_states.to(query.dtype)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states, *args)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ if input_ndim == 4:
+ hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+ if attn.residual_connection:
+ hidden_states = hidden_states + residual
+
+ hidden_states = hidden_states / attn.rescale_output_factor
+
+ return hidden_states
+
+
+class AttnProcessor2_0:
+ r"""
+ Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+ """
+
+ def __init__(self):
+ if not hasattr(F, "scaled_dot_product_attention"):
+ raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ temb: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ residual = hidden_states
+ if attn.spatial_norm is not None:
+ hidden_states = attn.spatial_norm(hidden_states, temb)
+
+ input_ndim = hidden_states.ndim
+
+ if input_ndim == 4:
+ batch_size, channel, height, width = hidden_states.shape
+ hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+
+ if attention_mask is not None:
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ # scaled_dot_product_attention expects attention_mask shape to be
+ # (batch, heads, source_length, target_length)
+ attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+ if attn.group_norm is not None:
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ args = () if USE_PEFT_BACKEND else (scale,)
+ query = attn.to_q(hidden_states, *args)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states, *args)
+ value = attn.to_v(encoder_hidden_states, *args)
+
+ inner_dim = key.shape[-1]
+ head_dim = inner_dim // attn.heads
+
+ query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ # the output of sdp = (batch, num_heads, seq_len, head_dim)
+ # TODO: add support for attn.scale when we move to Torch 2.1
+ hidden_states = F.scaled_dot_product_attention(
+ query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+ )
+
+ hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+ hidden_states = hidden_states.to(query.dtype)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states, *args)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ if input_ndim == 4:
+ hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+ if attn.residual_connection:
+ hidden_states = hidden_states + residual
+
+ hidden_states = hidden_states / attn.rescale_output_factor
+
+ return hidden_states
+
+
+class FusedAttnProcessor2_0:
+ r"""
+ Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+ It uses fused projection layers. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is currently 🧪 experimental in nature and can change in future.
+
+
+ """
+
+ def __init__(self):
+ if not hasattr(F, "scaled_dot_product_attention"):
+ raise ImportError(
+ "FusedAttnProcessor2_0 requires at least PyTorch 2.0, to use it. Please upgrade PyTorch to > 2.0."
+ )
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ temb: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ residual = hidden_states
+ if attn.spatial_norm is not None:
+ hidden_states = attn.spatial_norm(hidden_states, temb)
+
+ input_ndim = hidden_states.ndim
+
+ if input_ndim == 4:
+ batch_size, channel, height, width = hidden_states.shape
+ hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+
+ if attention_mask is not None:
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ # scaled_dot_product_attention expects attention_mask shape to be
+ # (batch, heads, source_length, target_length)
+ attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+ if attn.group_norm is not None:
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ args = () if USE_PEFT_BACKEND else (scale,)
+ if encoder_hidden_states is None:
+ qkv = attn.to_qkv(hidden_states, *args)
+ split_size = qkv.shape[-1] // 3
+ query, key, value = torch.split(qkv, split_size, dim=-1)
+ else:
+ if attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+ query = attn.to_q(hidden_states, *args)
+
+ kv = attn.to_kv(encoder_hidden_states, *args)
+ split_size = kv.shape[-1] // 2
+ key, value = torch.split(kv, split_size, dim=-1)
+
+ inner_dim = key.shape[-1]
+ head_dim = inner_dim // attn.heads
+
+ query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ # the output of sdp = (batch, num_heads, seq_len, head_dim)
+ # TODO: add support for attn.scale when we move to Torch 2.1
+ hidden_states = F.scaled_dot_product_attention(
+ query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+ )
+
+ hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+ hidden_states = hidden_states.to(query.dtype)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states, *args)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ if input_ndim == 4:
+ hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+ if attn.residual_connection:
+ hidden_states = hidden_states + residual
+
+ hidden_states = hidden_states / attn.rescale_output_factor
+
+ return hidden_states
+
+
+class CustomDiffusionXFormersAttnProcessor(nn.Module):
+ r"""
+ Processor for implementing memory efficient attention using xFormers for the Custom Diffusion method.
+
+ Args:
+ train_kv (`bool`, defaults to `True`):
+ Whether to newly train the key and value matrices corresponding to the text features.
+ train_q_out (`bool`, defaults to `True`):
+ Whether to newly train query matrices corresponding to the latent image features.
+ hidden_size (`int`, *optional*, defaults to `None`):
+ The hidden size of the attention layer.
+ cross_attention_dim (`int`, *optional*, defaults to `None`):
+ The number of channels in the `encoder_hidden_states`.
+ out_bias (`bool`, defaults to `True`):
+ Whether to include the bias parameter in `train_q_out`.
+ dropout (`float`, *optional*, defaults to 0.0):
+ The dropout probability to use.
+ attention_op (`Callable`, *optional*, defaults to `None`):
+ The base
+ [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to use
+ as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best operator.
+ """
+
+ def __init__(
+ self,
+ train_kv: bool = True,
+ train_q_out: bool = False,
+ hidden_size: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ out_bias: bool = True,
+ dropout: float = 0.0,
+ attention_op: Optional[Callable] = None,
+ ):
+ super().__init__()
+ self.train_kv = train_kv
+ self.train_q_out = train_q_out
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+ self.attention_op = attention_op
+
+ # `_custom_diffusion` id for easy serialization and loading.
+ if self.train_kv:
+ self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+ self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+ if self.train_q_out:
+ self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)
+ self.to_out_custom_diffusion = nn.ModuleList([])
+ self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))
+ self.to_out_custom_diffusion.append(nn.Dropout(dropout))
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+ if self.train_q_out:
+ query = self.to_q_custom_diffusion(hidden_states).to(attn.to_q.weight.dtype)
+ else:
+ query = attn.to_q(hidden_states.to(attn.to_q.weight.dtype))
+
+ if encoder_hidden_states is None:
+ crossattn = False
+ encoder_hidden_states = hidden_states
+ else:
+ crossattn = True
+ if attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ if self.train_kv:
+ key = self.to_k_custom_diffusion(encoder_hidden_states.to(self.to_k_custom_diffusion.weight.dtype))
+ value = self.to_v_custom_diffusion(encoder_hidden_states.to(self.to_v_custom_diffusion.weight.dtype))
+ key = key.to(attn.to_q.weight.dtype)
+ value = value.to(attn.to_q.weight.dtype)
+ else:
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ if crossattn:
+ detach = torch.ones_like(key)
+ detach[:, :1, :] = detach[:, :1, :] * 0.0
+ key = detach * key + (1 - detach) * key.detach()
+ value = detach * value + (1 - detach) * value.detach()
+
+ query = attn.head_to_batch_dim(query).contiguous()
+ key = attn.head_to_batch_dim(key).contiguous()
+ value = attn.head_to_batch_dim(value).contiguous()
+
+ hidden_states = xformers.ops.memory_efficient_attention(
+ query, key, value, attn_bias=attention_mask, op=self.attention_op, scale=attn.scale
+ )
+ hidden_states = hidden_states.to(query.dtype)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ if self.train_q_out:
+ # linear proj
+ hidden_states = self.to_out_custom_diffusion[0](hidden_states)
+ # dropout
+ hidden_states = self.to_out_custom_diffusion[1](hidden_states)
+ else:
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ return hidden_states
+
+
+class CustomDiffusionAttnProcessor2_0(nn.Module):
+ r"""
+ Processor for implementing attention for the Custom Diffusion method using PyTorch 2.0’s memory-efficient scaled
+ dot-product attention.
+
+ Args:
+ train_kv (`bool`, defaults to `True`):
+ Whether to newly train the key and value matrices corresponding to the text features.
+ train_q_out (`bool`, defaults to `True`):
+ Whether to newly train query matrices corresponding to the latent image features.
+ hidden_size (`int`, *optional*, defaults to `None`):
+ The hidden size of the attention layer.
+ cross_attention_dim (`int`, *optional*, defaults to `None`):
+ The number of channels in the `encoder_hidden_states`.
+ out_bias (`bool`, defaults to `True`):
+ Whether to include the bias parameter in `train_q_out`.
+ dropout (`float`, *optional*, defaults to 0.0):
+ The dropout probability to use.
+ """
+
+ def __init__(
+ self,
+ train_kv: bool = True,
+ train_q_out: bool = True,
+ hidden_size: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ out_bias: bool = True,
+ dropout: float = 0.0,
+ ):
+ super().__init__()
+ self.train_kv = train_kv
+ self.train_q_out = train_q_out
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+
+ # `_custom_diffusion` id for easy serialization and loading.
+ if self.train_kv:
+ self.to_k_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+ self.to_v_custom_diffusion = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+ if self.train_q_out:
+ self.to_q_custom_diffusion = nn.Linear(hidden_size, hidden_size, bias=False)
+ self.to_out_custom_diffusion = nn.ModuleList([])
+ self.to_out_custom_diffusion.append(nn.Linear(hidden_size, hidden_size, bias=out_bias))
+ self.to_out_custom_diffusion.append(nn.Dropout(dropout))
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ batch_size, sequence_length, _ = hidden_states.shape
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ if self.train_q_out:
+ query = self.to_q_custom_diffusion(hidden_states)
+ else:
+ query = attn.to_q(hidden_states)
+
+ if encoder_hidden_states is None:
+ crossattn = False
+ encoder_hidden_states = hidden_states
+ else:
+ crossattn = True
+ if attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ if self.train_kv:
+ key = self.to_k_custom_diffusion(encoder_hidden_states.to(self.to_k_custom_diffusion.weight.dtype))
+ value = self.to_v_custom_diffusion(encoder_hidden_states.to(self.to_v_custom_diffusion.weight.dtype))
+ key = key.to(attn.to_q.weight.dtype)
+ value = value.to(attn.to_q.weight.dtype)
+
+ else:
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ if crossattn:
+ detach = torch.ones_like(key)
+ detach[:, :1, :] = detach[:, :1, :] * 0.0
+ key = detach * key + (1 - detach) * key.detach()
+ value = detach * value + (1 - detach) * value.detach()
+
+ inner_dim = hidden_states.shape[-1]
+
+ head_dim = inner_dim // attn.heads
+ query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ # the output of sdp = (batch, num_heads, seq_len, head_dim)
+ # TODO: add support for attn.scale when we move to Torch 2.1
+ hidden_states = F.scaled_dot_product_attention(
+ query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+ )
+
+ hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+ hidden_states = hidden_states.to(query.dtype)
+
+ if self.train_q_out:
+ # linear proj
+ hidden_states = self.to_out_custom_diffusion[0](hidden_states)
+ # dropout
+ hidden_states = self.to_out_custom_diffusion[1](hidden_states)
+ else:
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ return hidden_states
+
+
+class SlicedAttnProcessor:
+ r"""
+ Processor for implementing sliced attention.
+
+ Args:
+ slice_size (`int`, *optional*):
+ The number of steps to compute attention. Uses as many slices as `attention_head_dim // slice_size`, and
+ `attention_head_dim` must be a multiple of the `slice_size`.
+ """
+
+ def __init__(self, slice_size: int):
+ self.slice_size = slice_size
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ residual = hidden_states
+
+ input_ndim = hidden_states.ndim
+
+ if input_ndim == 4:
+ batch_size, channel, height, width = hidden_states.shape
+ hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+ if attn.group_norm is not None:
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states)
+ dim = query.shape[-1]
+ query = attn.head_to_batch_dim(query)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ batch_size_attention, query_tokens, _ = query.shape
+ hidden_states = torch.zeros(
+ (batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
+ )
+
+ for i in range(batch_size_attention // self.slice_size):
+ start_idx = i * self.slice_size
+ end_idx = (i + 1) * self.slice_size
+
+ query_slice = query[start_idx:end_idx]
+ key_slice = key[start_idx:end_idx]
+ attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
+
+ attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
+
+ attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
+
+ hidden_states[start_idx:end_idx] = attn_slice
+
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ if input_ndim == 4:
+ hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+ if attn.residual_connection:
+ hidden_states = hidden_states + residual
+
+ hidden_states = hidden_states / attn.rescale_output_factor
+
+ return hidden_states
+
+
+class SlicedAttnAddedKVProcessor:
+ r"""
+ Processor for implementing sliced attention with extra learnable key and value matrices for the text encoder.
+
+ Args:
+ slice_size (`int`, *optional*):
+ The number of steps to compute attention. Uses as many slices as `attention_head_dim // slice_size`, and
+ `attention_head_dim` must be a multiple of the `slice_size`.
+ """
+
+ def __init__(self, slice_size):
+ self.slice_size = slice_size
+
+ def __call__(
+ self,
+ attn: "Attention",
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ temb: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ residual = hidden_states
+
+ if attn.spatial_norm is not None:
+ hidden_states = attn.spatial_norm(hidden_states, temb)
+
+ hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
+
+ batch_size, sequence_length, _ = hidden_states.shape
+
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states)
+ dim = query.shape[-1]
+ query = attn.head_to_batch_dim(query)
+
+ encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+ encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+ encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj)
+ encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj)
+
+ if not attn.only_cross_attention:
+ key = attn.to_k(hidden_states)
+ value = attn.to_v(hidden_states)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+ key = torch.cat([encoder_hidden_states_key_proj, key], dim=1)
+ value = torch.cat([encoder_hidden_states_value_proj, value], dim=1)
+ else:
+ key = encoder_hidden_states_key_proj
+ value = encoder_hidden_states_value_proj
+
+ batch_size_attention, query_tokens, _ = query.shape
+ hidden_states = torch.zeros(
+ (batch_size_attention, query_tokens, dim // attn.heads), device=query.device, dtype=query.dtype
+ )
+
+ for i in range(batch_size_attention // self.slice_size):
+ start_idx = i * self.slice_size
+ end_idx = (i + 1) * self.slice_size
+
+ query_slice = query[start_idx:end_idx]
+ key_slice = key[start_idx:end_idx]
+ attn_mask_slice = attention_mask[start_idx:end_idx] if attention_mask is not None else None
+
+ attn_slice = attn.get_attention_scores(query_slice, key_slice, attn_mask_slice)
+
+ attn_slice = torch.bmm(attn_slice, value[start_idx:end_idx])
+
+ hidden_states[start_idx:end_idx] = attn_slice
+
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
+ hidden_states = hidden_states + residual
+
+ return hidden_states
+
+
+class SpatialNorm(nn.Module):
+ """
+ Spatially conditioned normalization as defined in https://arxiv.org/abs/2209.09002.
+
+ Args:
+ f_channels (`int`):
+ The number of channels for input to group normalization layer, and output of the spatial norm layer.
+ zq_channels (`int`):
+ The number of channels for the quantized vector as described in the paper.
+ """
+
+ def __init__(
+ self,
+ f_channels: int,
+ zq_channels: int,
+ ):
+ super().__init__()
+ self.norm_layer = nn.GroupNorm(num_channels=f_channels, num_groups=32, eps=1e-6, affine=True)
+ self.conv_y = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0)
+ self.conv_b = nn.Conv2d(zq_channels, f_channels, kernel_size=1, stride=1, padding=0)
+
+ def forward(self, f: torch.FloatTensor, zq: torch.FloatTensor) -> torch.FloatTensor:
+ f_size = f.shape[-2:]
+ zq = F.interpolate(zq, size=f_size, mode="nearest")
+ norm_f = self.norm_layer(f)
+ new_f = norm_f * self.conv_y(zq) + self.conv_b(zq)
+ return new_f
+
+
+class LoRAAttnProcessor(nn.Module):
+ def __init__(
+ self,
+ hidden_size: int,
+ cross_attention_dim: Optional[int] = None,
+ rank: int = 4,
+ network_alpha: Optional[int] = None,
+ **kwargs,
+ ):
+ deprecation_message = "Using LoRAAttnProcessor is deprecated. Please use the PEFT backend for all things LoRA. You can install PEFT by running `pip install peft`."
+ deprecate("LoRAAttnProcessor", "0.30.0", deprecation_message, standard_warn=False)
+
+ super().__init__()
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+ self.rank = rank
+
+ q_rank = kwargs.pop("q_rank", None)
+ q_hidden_size = kwargs.pop("q_hidden_size", None)
+ q_rank = q_rank if q_rank is not None else rank
+ q_hidden_size = q_hidden_size if q_hidden_size is not None else hidden_size
+
+ v_rank = kwargs.pop("v_rank", None)
+ v_hidden_size = kwargs.pop("v_hidden_size", None)
+ v_rank = v_rank if v_rank is not None else rank
+ v_hidden_size = v_hidden_size if v_hidden_size is not None else hidden_size
+
+ out_rank = kwargs.pop("out_rank", None)
+ out_hidden_size = kwargs.pop("out_hidden_size", None)
+ out_rank = out_rank if out_rank is not None else rank
+ out_hidden_size = out_hidden_size if out_hidden_size is not None else hidden_size
+
+ self.to_q_lora = LoRALinearLayer(q_hidden_size, q_hidden_size, q_rank, network_alpha)
+ self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
+ self.to_v_lora = LoRALinearLayer(cross_attention_dim or v_hidden_size, v_hidden_size, v_rank, network_alpha)
+ self.to_out_lora = LoRALinearLayer(out_hidden_size, out_hidden_size, out_rank, network_alpha)
+
+ def __call__(self, attn: Attention, hidden_states: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ self_cls_name = self.__class__.__name__
+ deprecate(
+ self_cls_name,
+ "0.26.0",
+ (
+ f"Make sure use {self_cls_name[4:]} instead by setting"
+ "LoRA layers to `self.{to_q,to_k,to_v,to_out[0]}.lora_layer` respectively. This will be done automatically when using"
+ " `LoraLoaderMixin.load_lora_weights`"
+ ),
+ )
+ attn.to_q.lora_layer = self.to_q_lora.to(hidden_states.device)
+ attn.to_k.lora_layer = self.to_k_lora.to(hidden_states.device)
+ attn.to_v.lora_layer = self.to_v_lora.to(hidden_states.device)
+ attn.to_out[0].lora_layer = self.to_out_lora.to(hidden_states.device)
+
+ attn._modules.pop("processor")
+ attn.processor = AttnProcessor()
+ return attn.processor(attn, hidden_states, *args, **kwargs)
+
+
+class LoRAAttnProcessor2_0(nn.Module):
+ def __init__(
+ self,
+ hidden_size: int,
+ cross_attention_dim: Optional[int] = None,
+ rank: int = 4,
+ network_alpha: Optional[int] = None,
+ **kwargs,
+ ):
+ deprecation_message = "Using LoRAAttnProcessor is deprecated. Please use the PEFT backend for all things LoRA. You can install PEFT by running `pip install peft`."
+ deprecate("LoRAAttnProcessor2_0", "0.30.0", deprecation_message, standard_warn=False)
+
+ super().__init__()
+ if not hasattr(F, "scaled_dot_product_attention"):
+ raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+ self.rank = rank
+
+ q_rank = kwargs.pop("q_rank", None)
+ q_hidden_size = kwargs.pop("q_hidden_size", None)
+ q_rank = q_rank if q_rank is not None else rank
+ q_hidden_size = q_hidden_size if q_hidden_size is not None else hidden_size
+
+ v_rank = kwargs.pop("v_rank", None)
+ v_hidden_size = kwargs.pop("v_hidden_size", None)
+ v_rank = v_rank if v_rank is not None else rank
+ v_hidden_size = v_hidden_size if v_hidden_size is not None else hidden_size
+
+ out_rank = kwargs.pop("out_rank", None)
+ out_hidden_size = kwargs.pop("out_hidden_size", None)
+ out_rank = out_rank if out_rank is not None else rank
+ out_hidden_size = out_hidden_size if out_hidden_size is not None else hidden_size
+
+ self.to_q_lora = LoRALinearLayer(q_hidden_size, q_hidden_size, q_rank, network_alpha)
+ self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
+ self.to_v_lora = LoRALinearLayer(cross_attention_dim or v_hidden_size, v_hidden_size, v_rank, network_alpha)
+ self.to_out_lora = LoRALinearLayer(out_hidden_size, out_hidden_size, out_rank, network_alpha)
+
+ def __call__(self, attn: Attention, hidden_states: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ self_cls_name = self.__class__.__name__
+ deprecate(
+ self_cls_name,
+ "0.26.0",
+ (
+ f"Make sure use {self_cls_name[4:]} instead by setting"
+ "LoRA layers to `self.{to_q,to_k,to_v,to_out[0]}.lora_layer` respectively. This will be done automatically when using"
+ " `LoraLoaderMixin.load_lora_weights`"
+ ),
+ )
+ attn.to_q.lora_layer = self.to_q_lora.to(hidden_states.device)
+ attn.to_k.lora_layer = self.to_k_lora.to(hidden_states.device)
+ attn.to_v.lora_layer = self.to_v_lora.to(hidden_states.device)
+ attn.to_out[0].lora_layer = self.to_out_lora.to(hidden_states.device)
+
+ attn._modules.pop("processor")
+ attn.processor = AttnProcessor2_0()
+ return attn.processor(attn, hidden_states, *args, **kwargs)
+
+
+class LoRAXFormersAttnProcessor(nn.Module):
+ r"""
+ Processor for implementing the LoRA attention mechanism with memory efficient attention using xFormers.
+
+ Args:
+ hidden_size (`int`, *optional*):
+ The hidden size of the attention layer.
+ cross_attention_dim (`int`, *optional*):
+ The number of channels in the `encoder_hidden_states`.
+ rank (`int`, defaults to 4):
+ The dimension of the LoRA update matrices.
+ attention_op (`Callable`, *optional*, defaults to `None`):
+ The base
+ [operator](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.AttentionOpBase) to
+ use as the attention operator. It is recommended to set to `None`, and allow xFormers to choose the best
+ operator.
+ network_alpha (`int`, *optional*):
+ Equivalent to `alpha` but it's usage is specific to Kohya (A1111) style LoRAs.
+ kwargs (`dict`):
+ Additional keyword arguments to pass to the `LoRALinearLayer` layers.
+ """
+
+ def __init__(
+ self,
+ hidden_size: int,
+ cross_attention_dim: int,
+ rank: int = 4,
+ attention_op: Optional[Callable] = None,
+ network_alpha: Optional[int] = None,
+ **kwargs,
+ ):
+ super().__init__()
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+ self.rank = rank
+ self.attention_op = attention_op
+
+ q_rank = kwargs.pop("q_rank", None)
+ q_hidden_size = kwargs.pop("q_hidden_size", None)
+ q_rank = q_rank if q_rank is not None else rank
+ q_hidden_size = q_hidden_size if q_hidden_size is not None else hidden_size
+
+ v_rank = kwargs.pop("v_rank", None)
+ v_hidden_size = kwargs.pop("v_hidden_size", None)
+ v_rank = v_rank if v_rank is not None else rank
+ v_hidden_size = v_hidden_size if v_hidden_size is not None else hidden_size
+
+ out_rank = kwargs.pop("out_rank", None)
+ out_hidden_size = kwargs.pop("out_hidden_size", None)
+ out_rank = out_rank if out_rank is not None else rank
+ out_hidden_size = out_hidden_size if out_hidden_size is not None else hidden_size
+
+ self.to_q_lora = LoRALinearLayer(q_hidden_size, q_hidden_size, q_rank, network_alpha)
+ self.to_k_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
+ self.to_v_lora = LoRALinearLayer(cross_attention_dim or v_hidden_size, v_hidden_size, v_rank, network_alpha)
+ self.to_out_lora = LoRALinearLayer(out_hidden_size, out_hidden_size, out_rank, network_alpha)
+
+ def __call__(self, attn: Attention, hidden_states: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ self_cls_name = self.__class__.__name__
+ deprecate(
+ self_cls_name,
+ "0.26.0",
+ (
+ f"Make sure use {self_cls_name[4:]} instead by setting"
+ "LoRA layers to `self.{to_q,to_k,to_v,add_k_proj,add_v_proj,to_out[0]}.lora_layer` respectively. This will be done automatically when using"
+ " `LoraLoaderMixin.load_lora_weights`"
+ ),
+ )
+ attn.to_q.lora_layer = self.to_q_lora.to(hidden_states.device)
+ attn.to_k.lora_layer = self.to_k_lora.to(hidden_states.device)
+ attn.to_v.lora_layer = self.to_v_lora.to(hidden_states.device)
+ attn.to_out[0].lora_layer = self.to_out_lora.to(hidden_states.device)
+
+ attn._modules.pop("processor")
+ attn.processor = XFormersAttnProcessor()
+ return attn.processor(attn, hidden_states, *args, **kwargs)
+
+
+class LoRAAttnAddedKVProcessor(nn.Module):
+ r"""
+ Processor for implementing the LoRA attention mechanism with extra learnable key and value matrices for the text
+ encoder.
+
+ Args:
+ hidden_size (`int`, *optional*):
+ The hidden size of the attention layer.
+ cross_attention_dim (`int`, *optional*, defaults to `None`):
+ The number of channels in the `encoder_hidden_states`.
+ rank (`int`, defaults to 4):
+ The dimension of the LoRA update matrices.
+ network_alpha (`int`, *optional*):
+ Equivalent to `alpha` but it's usage is specific to Kohya (A1111) style LoRAs.
+ kwargs (`dict`):
+ Additional keyword arguments to pass to the `LoRALinearLayer` layers.
+ """
+
+ def __init__(
+ self,
+ hidden_size: int,
+ cross_attention_dim: Optional[int] = None,
+ rank: int = 4,
+ network_alpha: Optional[int] = None,
+ ):
+ super().__init__()
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+ self.rank = rank
+
+ self.to_q_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
+ self.add_k_proj_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
+ self.add_v_proj_lora = LoRALinearLayer(cross_attention_dim or hidden_size, hidden_size, rank, network_alpha)
+ self.to_k_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
+ self.to_v_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
+ self.to_out_lora = LoRALinearLayer(hidden_size, hidden_size, rank, network_alpha)
+
+ def __call__(self, attn: Attention, hidden_states: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ self_cls_name = self.__class__.__name__
+ deprecate(
+ self_cls_name,
+ "0.26.0",
+ (
+ f"Make sure use {self_cls_name[4:]} instead by setting"
+ "LoRA layers to `self.{to_q,to_k,to_v,add_k_proj,add_v_proj,to_out[0]}.lora_layer` respectively. This will be done automatically when using"
+ " `LoraLoaderMixin.load_lora_weights`"
+ ),
+ )
+ attn.to_q.lora_layer = self.to_q_lora.to(hidden_states.device)
+ attn.to_k.lora_layer = self.to_k_lora.to(hidden_states.device)
+ attn.to_v.lora_layer = self.to_v_lora.to(hidden_states.device)
+ attn.to_out[0].lora_layer = self.to_out_lora.to(hidden_states.device)
+
+ attn._modules.pop("processor")
+ attn.processor = AttnAddedKVProcessor()
+ return attn.processor(attn, hidden_states, *args, **kwargs)
+
+
+class IPAdapterAttnProcessor(nn.Module):
+ r"""
+ Attention processor for Multiple IP-Adapater.
+
+ Args:
+ hidden_size (`int`):
+ The hidden size of the attention layer.
+ cross_attention_dim (`int`):
+ The number of channels in the `encoder_hidden_states`.
+ num_tokens (`int`, `Tuple[int]` or `List[int]`, defaults to `(4,)`):
+ The context length of the image features.
+ scale (`float` or List[`float`], defaults to 1.0):
+ the weight scale of image prompt.
+ """
+
+ def __init__(self, hidden_size, cross_attention_dim=None, num_tokens=(4,), scale=1.0):
+ super().__init__()
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+
+ if not isinstance(num_tokens, (tuple, list)):
+ num_tokens = [num_tokens]
+ self.num_tokens = num_tokens
+
+ if not isinstance(scale, list):
+ scale = [scale] * len(num_tokens)
+ if len(scale) != len(num_tokens):
+ raise ValueError("`scale` should be a list of integers with the same length as `num_tokens`.")
+ self.scale = scale
+
+ self.to_k_ip = nn.ModuleList(
+ [nn.Linear(cross_attention_dim, hidden_size, bias=False) for _ in range(len(num_tokens))]
+ )
+ self.to_v_ip = nn.ModuleList(
+ [nn.Linear(cross_attention_dim, hidden_size, bias=False) for _ in range(len(num_tokens))]
+ )
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ temb: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ip_adapter_masks: Optional[torch.FloatTensor] = None,
+ ):
+ residual = hidden_states
+
+ # separate ip_hidden_states from encoder_hidden_states
+ if encoder_hidden_states is not None:
+ if isinstance(encoder_hidden_states, tuple):
+ encoder_hidden_states, ip_hidden_states = encoder_hidden_states
+ else:
+ deprecation_message = (
+ "You have passed a tensor as `encoder_hidden_states`.This is deprecated and will be removed in a future release."
+ " Please make sure to update your script to pass `encoder_hidden_states` as a tuple to supress this warning."
+ )
+ deprecate("encoder_hidden_states not a tuple", "1.0.0", deprecation_message, standard_warn=False)
+ end_pos = encoder_hidden_states.shape[1] - self.num_tokens[0]
+ encoder_hidden_states, ip_hidden_states = (
+ encoder_hidden_states[:, :end_pos, :],
+ [encoder_hidden_states[:, end_pos:, :]],
+ )
+
+ if attn.spatial_norm is not None:
+ hidden_states = attn.spatial_norm(hidden_states, temb)
+
+ input_ndim = hidden_states.ndim
+
+ if input_ndim == 4:
+ batch_size, channel, height, width = hidden_states.shape
+ hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+ if attn.group_norm is not None:
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ query = attn.head_to_batch_dim(query)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ attention_probs = attn.get_attention_scores(query, key, attention_mask)
+ hidden_states = torch.bmm(attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ if ip_adapter_masks is not None:
+ if not isinstance(ip_adapter_masks, torch.Tensor) or ip_adapter_masks.ndim != 4:
+ raise ValueError(
+ " ip_adapter_mask should be a tensor with shape [num_ip_adapter, 1, height, width]."
+ " Please use `IPAdapterMaskProcessor` to preprocess your mask"
+ )
+ if len(ip_adapter_masks) != len(self.scale):
+ raise ValueError(
+ f"Number of ip_adapter_masks ({len(ip_adapter_masks)}) must match number of IP-Adapters ({len(self.scale)})"
+ )
+ else:
+ ip_adapter_masks = [None] * len(self.scale)
+
+ # for ip-adapter
+ for current_ip_hidden_states, scale, to_k_ip, to_v_ip, mask in zip(
+ ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip, ip_adapter_masks
+ ):
+ ip_key = to_k_ip(current_ip_hidden_states)
+ ip_value = to_v_ip(current_ip_hidden_states)
+
+ ip_key = attn.head_to_batch_dim(ip_key)
+ ip_value = attn.head_to_batch_dim(ip_value)
+
+ ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+ current_ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+ current_ip_hidden_states = attn.batch_to_head_dim(current_ip_hidden_states)
+
+ if mask is not None:
+ mask_downsample = IPAdapterMaskProcessor.downsample(
+ mask, batch_size, current_ip_hidden_states.shape[1], current_ip_hidden_states.shape[2]
+ )
+
+ mask_downsample = mask_downsample.to(dtype=query.dtype, device=query.device)
+
+ current_ip_hidden_states = current_ip_hidden_states * mask_downsample
+
+ hidden_states = hidden_states + scale * current_ip_hidden_states
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ if input_ndim == 4:
+ hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+ if attn.residual_connection:
+ hidden_states = hidden_states + residual
+
+ hidden_states = hidden_states / attn.rescale_output_factor
+
+ return hidden_states
+
+
+class IPAdapterAttnProcessor2_0(torch.nn.Module):
+ r"""
+ Attention processor for IP-Adapater for PyTorch 2.0.
+
+ Args:
+ hidden_size (`int`):
+ The hidden size of the attention layer.
+ cross_attention_dim (`int`):
+ The number of channels in the `encoder_hidden_states`.
+ num_tokens (`int`, `Tuple[int]` or `List[int]`, defaults to `(4,)`):
+ The context length of the image features.
+ scale (`float` or `List[float]`, defaults to 1.0):
+ the weight scale of image prompt.
+ """
+
+ def __init__(self, hidden_size, cross_attention_dim=None, num_tokens=(4,), scale=1.0):
+ super().__init__()
+
+ if not hasattr(F, "scaled_dot_product_attention"):
+ raise ImportError(
+ f"{self.__class__.__name__} requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0."
+ )
+
+ self.hidden_size = hidden_size
+ self.cross_attention_dim = cross_attention_dim
+
+ if not isinstance(num_tokens, (tuple, list)):
+ num_tokens = [num_tokens]
+ self.num_tokens = num_tokens
+
+ if not isinstance(scale, list):
+ scale = [scale] * len(num_tokens)
+ if len(scale) != len(num_tokens):
+ raise ValueError("`scale` should be a list of integers with the same length as `num_tokens`.")
+ self.scale = scale
+
+ self.to_k_ip = nn.ModuleList(
+ [nn.Linear(cross_attention_dim, hidden_size, bias=False) for _ in range(len(num_tokens))]
+ )
+ self.to_v_ip = nn.ModuleList(
+ [nn.Linear(cross_attention_dim, hidden_size, bias=False) for _ in range(len(num_tokens))]
+ )
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ temb: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ip_adapter_masks: Optional[torch.FloatTensor] = None,
+ ):
+ residual = hidden_states
+
+ # separate ip_hidden_states from encoder_hidden_states
+ if encoder_hidden_states is not None:
+ if isinstance(encoder_hidden_states, tuple):
+ encoder_hidden_states, ip_hidden_states = encoder_hidden_states
+ else:
+ deprecation_message = (
+ "You have passed a tensor as `encoder_hidden_states`.This is deprecated and will be removed in a future release."
+ " Please make sure to update your script to pass `encoder_hidden_states` as a tuple to supress this warning."
+ )
+ deprecate("encoder_hidden_states not a tuple", "1.0.0", deprecation_message, standard_warn=False)
+ end_pos = encoder_hidden_states.shape[1] - self.num_tokens[0]
+ encoder_hidden_states, ip_hidden_states = (
+ encoder_hidden_states[:, :end_pos, :],
+ [encoder_hidden_states[:, end_pos:, :]],
+ )
+
+ if attn.spatial_norm is not None:
+ hidden_states = attn.spatial_norm(hidden_states, temb)
+
+ input_ndim = hidden_states.ndim
+
+ if input_ndim == 4:
+ batch_size, channel, height, width = hidden_states.shape
+ hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+
+ if attention_mask is not None:
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ # scaled_dot_product_attention expects attention_mask shape to be
+ # (batch, heads, source_length, target_length)
+ attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+ if attn.group_norm is not None:
+ hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+ query = attn.to_q(hidden_states)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ inner_dim = key.shape[-1]
+ head_dim = inner_dim // attn.heads
+
+ query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ # the output of sdp = (batch, num_heads, seq_len, head_dim)
+ # TODO: add support for attn.scale when we move to Torch 2.1
+ hidden_states = F.scaled_dot_product_attention(
+ query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+ )
+
+ hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+ hidden_states = hidden_states.to(query.dtype)
+
+ if ip_adapter_masks is not None:
+ if not isinstance(ip_adapter_masks, torch.Tensor) or ip_adapter_masks.ndim != 4:
+ raise ValueError(
+ " ip_adapter_mask should be a tensor with shape [num_ip_adapter, 1, height, width]."
+ " Please use `IPAdapterMaskProcessor` to preprocess your mask"
+ )
+ if len(ip_adapter_masks) != len(self.scale):
+ raise ValueError(
+ f"Number of ip_adapter_masks ({len(ip_adapter_masks)}) must match number of IP-Adapters ({len(self.scale)})"
+ )
+ else:
+ ip_adapter_masks = [None] * len(self.scale)
+
+ # for ip-adapter
+ for current_ip_hidden_states, scale, to_k_ip, to_v_ip, mask in zip(
+ ip_hidden_states, self.scale, self.to_k_ip, self.to_v_ip, ip_adapter_masks
+ ):
+ ip_key = to_k_ip(current_ip_hidden_states)
+ ip_value = to_v_ip(current_ip_hidden_states)
+
+ ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ # the output of sdp = (batch, num_heads, seq_len, head_dim)
+ # TODO: add support for attn.scale when we move to Torch 2.1
+ current_ip_hidden_states = F.scaled_dot_product_attention(
+ query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+ )
+
+ current_ip_hidden_states = current_ip_hidden_states.transpose(1, 2).reshape(
+ batch_size, -1, attn.heads * head_dim
+ )
+ current_ip_hidden_states = current_ip_hidden_states.to(query.dtype)
+
+ if mask is not None:
+ mask_downsample = IPAdapterMaskProcessor.downsample(
+ mask, batch_size, current_ip_hidden_states.shape[1], current_ip_hidden_states.shape[2]
+ )
+
+ mask_downsample = mask_downsample.to(dtype=query.dtype, device=query.device)
+
+ current_ip_hidden_states = current_ip_hidden_states * mask_downsample
+
+ hidden_states = hidden_states + scale * current_ip_hidden_states
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ if input_ndim == 4:
+ hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+ if attn.residual_connection:
+ hidden_states = hidden_states + residual
+
+ hidden_states = hidden_states / attn.rescale_output_factor
+
+ return hidden_states
+
+
+LORA_ATTENTION_PROCESSORS = (
+ LoRAAttnProcessor,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnAddedKVProcessor,
+)
+
+ADDED_KV_ATTENTION_PROCESSORS = (
+ AttnAddedKVProcessor,
+ SlicedAttnAddedKVProcessor,
+ AttnAddedKVProcessor2_0,
+ XFormersAttnAddedKVProcessor,
+ LoRAAttnAddedKVProcessor,
+)
+
+CROSS_ATTENTION_PROCESSORS = (
+ AttnProcessor,
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ SlicedAttnProcessor,
+ LoRAAttnProcessor,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ IPAdapterAttnProcessor,
+ IPAdapterAttnProcessor2_0,
+)
+
+AttentionProcessor = Union[
+ AttnProcessor,
+ AttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ XFormersAttnProcessor,
+ SlicedAttnProcessor,
+ AttnAddedKVProcessor,
+ SlicedAttnAddedKVProcessor,
+ AttnAddedKVProcessor2_0,
+ XFormersAttnAddedKVProcessor,
+ CustomDiffusionAttnProcessor,
+ CustomDiffusionXFormersAttnProcessor,
+ CustomDiffusionAttnProcessor2_0,
+ # deprecated
+ LoRAAttnProcessor,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnAddedKVProcessor,
+]
diff --git a/model/diffusers_c/models/autoencoders/__init__.py b/model/diffusers_c/models/autoencoders/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..201a40ff17b212e1473d6f8d38f89af60bac6410
--- /dev/null
+++ b/model/diffusers_c/models/autoencoders/__init__.py
@@ -0,0 +1,5 @@
+from .autoencoder_asym_kl import AsymmetricAutoencoderKL
+from .autoencoder_kl import AutoencoderKL
+from .autoencoder_kl_temporal_decoder import AutoencoderKLTemporalDecoder
+from .autoencoder_tiny import AutoencoderTiny
+from .consistency_decoder_vae import ConsistencyDecoderVAE
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/models/autoencoders/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4c570cd0c145f64216bb82004d6398ea3d5996a3
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/models/autoencoders/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..024285c593c6d64f9812d22b3d5dbb16011e4ec4
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_asym_kl.cpython-310.pyc b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_asym_kl.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0ad815e04bd1319e19ad00c799d9277b8409abc5
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_asym_kl.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_asym_kl.cpython-39.pyc b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_asym_kl.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..00266a4f39f27aa9713828056505cdf4e5ea3b3e
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_asym_kl.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl.cpython-310.pyc b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..655e9cef369aa2d8d6b36b751aad6b7b3cd673b8
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl.cpython-39.pyc b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3368e0adef9a31b4361c37b3e2bc1fd9b792b4b7
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl_temporal_decoder.cpython-310.pyc b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl_temporal_decoder.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9c0e596bbcd3f2dcb4ff5d9e2ec69c98533da42b
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl_temporal_decoder.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl_temporal_decoder.cpython-39.pyc b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl_temporal_decoder.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a220db4b3e5fcf120df813a3a3a14cba4e0a9394
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_kl_temporal_decoder.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_tiny.cpython-310.pyc b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_tiny.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2a263cb27902f155ff3228e4e9f6ac8aa85005e8
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_tiny.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_tiny.cpython-39.pyc b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_tiny.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f179fbe20490f6a484faf48a0e3fec8c54f4f828
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/autoencoder_tiny.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/consistency_decoder_vae.cpython-310.pyc b/model/diffusers_c/models/autoencoders/__pycache__/consistency_decoder_vae.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..aee5f25191ad764d56a4029834350f27fa0c2fac
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/consistency_decoder_vae.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/consistency_decoder_vae.cpython-39.pyc b/model/diffusers_c/models/autoencoders/__pycache__/consistency_decoder_vae.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6a9dbf21728ffe37ee318c507cd321ea3f9c6955
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/consistency_decoder_vae.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/vae.cpython-310.pyc b/model/diffusers_c/models/autoencoders/__pycache__/vae.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3bd2d5a4420312fe36183daa703a8602222a12a3
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/vae.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/__pycache__/vae.cpython-39.pyc b/model/diffusers_c/models/autoencoders/__pycache__/vae.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ad2b4164825c617daa5343c43c6e0c3bcc8b2b59
Binary files /dev/null and b/model/diffusers_c/models/autoencoders/__pycache__/vae.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/autoencoders/autoencoder_asym_kl.py b/model/diffusers_c/models/autoencoders/autoencoder_asym_kl.py
new file mode 100644
index 0000000000000000000000000000000000000000..fc2041d2e930ec83264aa4de1273180680a32234
--- /dev/null
+++ b/model/diffusers_c/models/autoencoders/autoencoder_asym_kl.py
@@ -0,0 +1,186 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils.accelerate_utils import apply_forward_hook
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder, MaskConditionDecoder
+
+
+class AsymmetricAutoencoderKL(ModelMixin, ConfigMixin):
+ r"""
+ Designing a Better Asymmetric VQGAN for StableDiffusion https://arxiv.org/abs/2306.04632 . A VAE model with KL loss
+ for encoding images into latents and decoding latent representations into images.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+ out_channels (int, *optional*, defaults to 3): Number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+ Tuple of downsample block types.
+ down_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+ Tuple of down block output channels.
+ layers_per_down_block (`int`, *optional*, defaults to `1`):
+ Number layers for down block.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+ Tuple of upsample block types.
+ up_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+ Tuple of up block output channels.
+ layers_per_up_block (`int`, *optional*, defaults to `1`):
+ Number layers for up block.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
+ sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+ norm_num_groups (`int`, *optional*, defaults to `32`):
+ Number of groups to use for the first normalization layer in ResNet blocks.
+ scaling_factor (`float`, *optional*, defaults to 0.18215):
+ The component-wise standard deviation of the trained latent space computed using the first batch of the
+ training set. This is used to scale the latent space to have unit variance when training the diffusion
+ model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+ diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+ / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+ Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
+ down_block_out_channels: Tuple[int, ...] = (64,),
+ layers_per_down_block: int = 1,
+ up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+ up_block_out_channels: Tuple[int, ...] = (64,),
+ layers_per_up_block: int = 1,
+ act_fn: str = "silu",
+ latent_channels: int = 4,
+ norm_num_groups: int = 32,
+ sample_size: int = 32,
+ scaling_factor: float = 0.18215,
+ ) -> None:
+ super().__init__()
+
+ # pass init params to Encoder
+ self.encoder = Encoder(
+ in_channels=in_channels,
+ out_channels=latent_channels,
+ down_block_types=down_block_types,
+ block_out_channels=down_block_out_channels,
+ layers_per_block=layers_per_down_block,
+ act_fn=act_fn,
+ norm_num_groups=norm_num_groups,
+ double_z=True,
+ )
+
+ # pass init params to Decoder
+ self.decoder = MaskConditionDecoder(
+ in_channels=latent_channels,
+ out_channels=out_channels,
+ up_block_types=up_block_types,
+ block_out_channels=up_block_out_channels,
+ layers_per_block=layers_per_up_block,
+ act_fn=act_fn,
+ norm_num_groups=norm_num_groups,
+ )
+
+ self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
+ self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1)
+
+ self.use_slicing = False
+ self.use_tiling = False
+
+ self.register_to_config(block_out_channels=up_block_out_channels)
+ self.register_to_config(force_upcast=False)
+
+ @apply_forward_hook
+ def encode(
+ self, x: torch.FloatTensor, return_dict: bool = True
+ ) -> Union[AutoencoderKLOutput, Tuple[torch.FloatTensor]]:
+ h = self.encoder(x)
+ moments = self.quant_conv(h)
+ posterior = DiagonalGaussianDistribution(moments)
+
+ if not return_dict:
+ return (posterior,)
+
+ return AutoencoderKLOutput(latent_dist=posterior)
+
+ def _decode(
+ self,
+ z: torch.FloatTensor,
+ image: Optional[torch.FloatTensor] = None,
+ mask: Optional[torch.FloatTensor] = None,
+ return_dict: bool = True,
+ ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+ z = self.post_quant_conv(z)
+ dec = self.decoder(z, image, mask)
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
+
+ @apply_forward_hook
+ def decode(
+ self,
+ z: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ image: Optional[torch.FloatTensor] = None,
+ mask: Optional[torch.FloatTensor] = None,
+ return_dict: bool = True,
+ ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+ decoded = self._decode(z, image, mask).sample
+
+ if not return_dict:
+ return (decoded,)
+
+ return DecoderOutput(sample=decoded)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ mask: Optional[torch.FloatTensor] = None,
+ sample_posterior: bool = False,
+ return_dict: bool = True,
+ generator: Optional[torch.Generator] = None,
+ ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+ r"""
+ Args:
+ sample (`torch.FloatTensor`): Input sample.
+ mask (`torch.FloatTensor`, *optional*, defaults to `None`): Optional inpainting mask.
+ sample_posterior (`bool`, *optional*, defaults to `False`):
+ Whether to sample from the posterior.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+ """
+ x = sample
+ posterior = self.encode(x).latent_dist
+ if sample_posterior:
+ z = posterior.sample(generator=generator)
+ else:
+ z = posterior.mode()
+ dec = self.decode(z, sample, mask).sample
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
diff --git a/model/diffusers_c/models/autoencoders/autoencoder_kl.py b/model/diffusers_c/models/autoencoders/autoencoder_kl.py
new file mode 100644
index 0000000000000000000000000000000000000000..9bbf2023eb992b733eb609bc48a9b05dcb009502
--- /dev/null
+++ b/model/diffusers_c/models/autoencoders/autoencoder_kl.py
@@ -0,0 +1,489 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import FromOriginalVAEMixin
+from ...utils.accelerate_utils import apply_forward_hook
+from ..attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ Attention,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from .vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
+
+
+class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
+ r"""
+ A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+ out_channels (int, *optional*, defaults to 3): Number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+ Tuple of downsample block types.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+ Tuple of upsample block types.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+ Tuple of block output channels.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
+ sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+ scaling_factor (`float`, *optional*, defaults to 0.18215):
+ The component-wise standard deviation of the trained latent space computed using the first batch of the
+ training set. This is used to scale the latent space to have unit variance when training the diffusion
+ model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+ diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+ / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+ Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+ force_upcast (`bool`, *optional*, default to `True`):
+ If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+ can be fine-tuned / trained to a lower range without loosing too much precision in which case
+ `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
+ up_block_types: Tuple[str] = ("UpDecoderBlock2D",),
+ block_out_channels: Tuple[int] = (64,),
+ layers_per_block: int = 1,
+ act_fn: str = "silu",
+ latent_channels: int = 4,
+ norm_num_groups: int = 32,
+ sample_size: int = 32,
+ scaling_factor: float = 0.18215,
+ latents_mean: Optional[Tuple[float]] = None,
+ latents_std: Optional[Tuple[float]] = None,
+ force_upcast: float = True,
+ ):
+ super().__init__()
+
+ # pass init params to Encoder
+ self.encoder = Encoder(
+ in_channels=in_channels,
+ out_channels=latent_channels,
+ down_block_types=down_block_types,
+ block_out_channels=block_out_channels,
+ layers_per_block=layers_per_block,
+ act_fn=act_fn,
+ norm_num_groups=norm_num_groups,
+ double_z=True,
+ )
+
+ # pass init params to Decoder
+ self.decoder = Decoder(
+ in_channels=latent_channels,
+ out_channels=out_channels,
+ up_block_types=up_block_types,
+ block_out_channels=block_out_channels,
+ layers_per_block=layers_per_block,
+ norm_num_groups=norm_num_groups,
+ act_fn=act_fn,
+ )
+
+ self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
+ self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1)
+
+ self.use_slicing = False
+ self.use_tiling = False
+
+ # only relevant if vae tiling is enabled
+ self.tile_sample_min_size = self.config.sample_size
+ sample_size = (
+ self.config.sample_size[0]
+ if isinstance(self.config.sample_size, (list, tuple))
+ else self.config.sample_size
+ )
+ self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
+ self.tile_overlap_factor = 0.25
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if isinstance(module, (Encoder, Decoder)):
+ module.gradient_checkpointing = value
+
+ def enable_tiling(self, use_tiling: bool = True):
+ r"""
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+ processing larger images.
+ """
+ self.use_tiling = use_tiling
+
+ def disable_tiling(self):
+ r"""
+ Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+ decoding in one step.
+ """
+ self.enable_tiling(False)
+
+ def enable_slicing(self):
+ r"""
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+ """
+ self.use_slicing = True
+
+ def disable_slicing(self):
+ r"""
+ Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+ decoding in one step.
+ """
+ self.use_slicing = False
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ @apply_forward_hook
+ def encode(
+ self, x: torch.FloatTensor, return_dict: bool = True
+ ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+ """
+ Encode a batch of images into latents.
+
+ Args:
+ x (`torch.FloatTensor`): Input batch of images.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+ Returns:
+ The latent representations of the encoded images. If `return_dict` is True, a
+ [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+ """
+ if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
+ return self.tiled_encode(x, return_dict=return_dict)
+
+ if self.use_slicing and x.shape[0] > 1:
+ encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
+ h = torch.cat(encoded_slices)
+ else:
+ h = self.encoder(x)
+
+ moments = self.quant_conv(h)
+ posterior = DiagonalGaussianDistribution(moments)
+
+ if not return_dict:
+ return (posterior,)
+
+ return AutoencoderKLOutput(latent_dist=posterior)
+
+ def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+ if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size):
+ return self.tiled_decode(z, return_dict=return_dict)
+
+ z = self.post_quant_conv(z)
+ dec = self.decoder(z)
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
+
+ @apply_forward_hook
+ def decode(
+ self, z: torch.FloatTensor, return_dict: bool = True, generator=None
+ ) -> Union[DecoderOutput, torch.FloatTensor]:
+ """
+ Decode a batch of images.
+
+ Args:
+ z (`torch.FloatTensor`): Input batch of latent vectors.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.vae.DecoderOutput`] or `tuple`:
+ If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+ returned.
+
+ """
+ if self.use_slicing and z.shape[0] > 1:
+ decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)]
+ decoded = torch.cat(decoded_slices)
+ else:
+ decoded = self._decode(z).sample
+
+ if not return_dict:
+ return (decoded,)
+
+ return DecoderOutput(sample=decoded)
+
+ def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+ blend_extent = min(a.shape[2], b.shape[2], blend_extent)
+ for y in range(blend_extent):
+ b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
+ return b
+
+ def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+ blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+ for x in range(blend_extent):
+ b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
+ return b
+
+ def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput:
+ r"""Encode a batch of images using a tiled encoder.
+
+ When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+ steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
+ different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
+ tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
+ output, but they should be much less noticeable.
+
+ Args:
+ x (`torch.FloatTensor`): Input batch of images.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`:
+ If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain
+ `tuple` is returned.
+ """
+ overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
+ blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
+ row_limit = self.tile_latent_min_size - blend_extent
+
+ # Split the image into 512x512 tiles and encode them separately.
+ rows = []
+ for i in range(0, x.shape[2], overlap_size):
+ row = []
+ for j in range(0, x.shape[3], overlap_size):
+ tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
+ tile = self.encoder(tile)
+ tile = self.quant_conv(tile)
+ row.append(tile)
+ rows.append(row)
+ result_rows = []
+ for i, row in enumerate(rows):
+ result_row = []
+ for j, tile in enumerate(row):
+ # blend the above tile and the left tile
+ # to the current tile and add the current tile to the result row
+ if i > 0:
+ tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+ if j > 0:
+ tile = self.blend_h(row[j - 1], tile, blend_extent)
+ result_row.append(tile[:, :, :row_limit, :row_limit])
+ result_rows.append(torch.cat(result_row, dim=3))
+
+ moments = torch.cat(result_rows, dim=2)
+ posterior = DiagonalGaussianDistribution(moments)
+
+ if not return_dict:
+ return (posterior,)
+
+ return AutoencoderKLOutput(latent_dist=posterior)
+
+ def tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+ r"""
+ Decode a batch of images using a tiled decoder.
+
+ Args:
+ z (`torch.FloatTensor`): Input batch of latent vectors.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.vae.DecoderOutput`] or `tuple`:
+ If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+ returned.
+ """
+ overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor))
+ blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor)
+ row_limit = self.tile_sample_min_size - blend_extent
+
+ # Split z into overlapping 64x64 tiles and decode them separately.
+ # The tiles have an overlap to avoid seams between tiles.
+ rows = []
+ for i in range(0, z.shape[2], overlap_size):
+ row = []
+ for j in range(0, z.shape[3], overlap_size):
+ tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size]
+ tile = self.post_quant_conv(tile)
+ decoded = self.decoder(tile)
+ row.append(decoded)
+ rows.append(row)
+ result_rows = []
+ for i, row in enumerate(rows):
+ result_row = []
+ for j, tile in enumerate(row):
+ # blend the above tile and the left tile
+ # to the current tile and add the current tile to the result row
+ if i > 0:
+ tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+ if j > 0:
+ tile = self.blend_h(row[j - 1], tile, blend_extent)
+ result_row.append(tile[:, :, :row_limit, :row_limit])
+ result_rows.append(torch.cat(result_row, dim=3))
+
+ dec = torch.cat(result_rows, dim=2)
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ sample_posterior: bool = False,
+ return_dict: bool = True,
+ generator: Optional[torch.Generator] = None,
+ ) -> Union[DecoderOutput, torch.FloatTensor]:
+ r"""
+ Args:
+ sample (`torch.FloatTensor`): Input sample.
+ sample_posterior (`bool`, *optional*, defaults to `False`):
+ Whether to sample from the posterior.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+ """
+ x = sample
+ posterior = self.encode(x).latent_dist
+ if sample_posterior:
+ z = posterior.sample(generator=generator)
+ else:
+ z = posterior.mode()
+ dec = self.decode(z).sample
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+ def fuse_qkv_projections(self):
+ """
+ Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is 🧪 experimental.
+
+
+ """
+ self.original_attn_processors = None
+
+ for _, attn_processor in self.attn_processors.items():
+ if "Added" in str(attn_processor.__class__.__name__):
+ raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+ self.original_attn_processors = self.attn_processors
+
+ for module in self.modules():
+ if isinstance(module, Attention):
+ module.fuse_projections(fuse=True)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+ def unfuse_qkv_projections(self):
+ """Disables the fused QKV projection if enabled.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ """
+ if self.original_attn_processors is not None:
+ self.set_attn_processor(self.original_attn_processors)
diff --git a/model/diffusers_c/models/autoencoders/autoencoder_kl_temporal_decoder.py b/model/diffusers_c/models/autoencoders/autoencoder_kl_temporal_decoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..b12226fa4bac2a0659d1b24fa19498e505d558ae
--- /dev/null
+++ b/model/diffusers_c/models/autoencoders/autoencoder_kl_temporal_decoder.py
@@ -0,0 +1,399 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import is_torch_version
+from ...utils.accelerate_utils import apply_forward_hook
+from ..attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
+from ..modeling_outputs import AutoencoderKLOutput
+from ..modeling_utils import ModelMixin
+from ..unets.unet_3d_blocks import MidBlockTemporalDecoder, UpBlockTemporalDecoder
+from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder
+
+
+class TemporalDecoder(nn.Module):
+ def __init__(
+ self,
+ in_channels: int = 4,
+ out_channels: int = 3,
+ block_out_channels: Tuple[int] = (128, 256, 512, 512),
+ layers_per_block: int = 2,
+ ):
+ super().__init__()
+ self.layers_per_block = layers_per_block
+
+ self.conv_in = nn.Conv2d(in_channels, block_out_channels[-1], kernel_size=3, stride=1, padding=1)
+ self.mid_block = MidBlockTemporalDecoder(
+ num_layers=self.layers_per_block,
+ in_channels=block_out_channels[-1],
+ out_channels=block_out_channels[-1],
+ attention_head_dim=block_out_channels[-1],
+ )
+
+ # up
+ self.up_blocks = nn.ModuleList([])
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ output_channel = reversed_block_out_channels[0]
+ for i in range(len(block_out_channels)):
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+
+ is_final_block = i == len(block_out_channels) - 1
+ up_block = UpBlockTemporalDecoder(
+ num_layers=self.layers_per_block + 1,
+ in_channels=prev_output_channel,
+ out_channels=output_channel,
+ add_upsample=not is_final_block,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=32, eps=1e-6)
+
+ self.conv_act = nn.SiLU()
+ self.conv_out = torch.nn.Conv2d(
+ in_channels=block_out_channels[0],
+ out_channels=out_channels,
+ kernel_size=3,
+ padding=1,
+ )
+
+ conv_out_kernel_size = (3, 1, 1)
+ padding = [int(k // 2) for k in conv_out_kernel_size]
+ self.time_conv_out = torch.nn.Conv3d(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ kernel_size=conv_out_kernel_size,
+ padding=padding,
+ )
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ image_only_indicator: torch.FloatTensor,
+ num_frames: int = 1,
+ ) -> torch.FloatTensor:
+ r"""The forward method of the `Decoder` class."""
+
+ sample = self.conv_in(sample)
+
+ upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ # middle
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.mid_block),
+ sample,
+ image_only_indicator,
+ use_reentrant=False,
+ )
+ sample = sample.to(upscale_dtype)
+
+ # up
+ for up_block in self.up_blocks:
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(up_block),
+ sample,
+ image_only_indicator,
+ use_reentrant=False,
+ )
+ else:
+ # middle
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.mid_block),
+ sample,
+ image_only_indicator,
+ )
+ sample = sample.to(upscale_dtype)
+
+ # up
+ for up_block in self.up_blocks:
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(up_block),
+ sample,
+ image_only_indicator,
+ )
+ else:
+ # middle
+ sample = self.mid_block(sample, image_only_indicator=image_only_indicator)
+ sample = sample.to(upscale_dtype)
+
+ # up
+ for up_block in self.up_blocks:
+ sample = up_block(sample, image_only_indicator=image_only_indicator)
+
+ # post-process
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ batch_frames, channels, height, width = sample.shape
+ batch_size = batch_frames // num_frames
+ sample = sample[None, :].reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4)
+ sample = self.time_conv_out(sample)
+
+ sample = sample.permute(0, 2, 1, 3, 4).reshape(batch_frames, channels, height, width)
+
+ return sample
+
+
+class AutoencoderKLTemporalDecoder(ModelMixin, ConfigMixin):
+ r"""
+ A VAE model with KL loss for encoding images into latents and decoding latent representations into images.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+ out_channels (int, *optional*, defaults to 3): Number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+ Tuple of downsample block types.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+ Tuple of block output channels.
+ layers_per_block: (`int`, *optional*, defaults to 1): Number of layers per block.
+ latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space.
+ sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+ scaling_factor (`float`, *optional*, defaults to 0.18215):
+ The component-wise standard deviation of the trained latent space computed using the first batch of the
+ training set. This is used to scale the latent space to have unit variance when training the diffusion
+ model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+ diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+ / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+ Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+ force_upcast (`bool`, *optional*, default to `True`):
+ If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+ can be fine-tuned / trained to a lower range without loosing too much precision in which case
+ `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ down_block_types: Tuple[str] = ("DownEncoderBlock2D",),
+ block_out_channels: Tuple[int] = (64,),
+ layers_per_block: int = 1,
+ latent_channels: int = 4,
+ sample_size: int = 32,
+ scaling_factor: float = 0.18215,
+ force_upcast: float = True,
+ ):
+ super().__init__()
+
+ # pass init params to Encoder
+ self.encoder = Encoder(
+ in_channels=in_channels,
+ out_channels=latent_channels,
+ down_block_types=down_block_types,
+ block_out_channels=block_out_channels,
+ layers_per_block=layers_per_block,
+ double_z=True,
+ )
+
+ # pass init params to Decoder
+ self.decoder = TemporalDecoder(
+ in_channels=latent_channels,
+ out_channels=out_channels,
+ block_out_channels=block_out_channels,
+ layers_per_block=layers_per_block,
+ )
+
+ self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
+
+ sample_size = (
+ self.config.sample_size[0]
+ if isinstance(self.config.sample_size, (list, tuple))
+ else self.config.sample_size
+ )
+ self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
+ self.tile_overlap_factor = 0.25
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if isinstance(module, (Encoder, TemporalDecoder)):
+ module.gradient_checkpointing = value
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ @apply_forward_hook
+ def encode(
+ self, x: torch.FloatTensor, return_dict: bool = True
+ ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]:
+ """
+ Encode a batch of images into latents.
+
+ Args:
+ x (`torch.FloatTensor`): Input batch of images.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple.
+
+ Returns:
+ The latent representations of the encoded images. If `return_dict` is True, a
+ [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned.
+ """
+ h = self.encoder(x)
+ moments = self.quant_conv(h)
+ posterior = DiagonalGaussianDistribution(moments)
+
+ if not return_dict:
+ return (posterior,)
+
+ return AutoencoderKLOutput(latent_dist=posterior)
+
+ @apply_forward_hook
+ def decode(
+ self,
+ z: torch.FloatTensor,
+ num_frames: int,
+ return_dict: bool = True,
+ ) -> Union[DecoderOutput, torch.FloatTensor]:
+ """
+ Decode a batch of images.
+
+ Args:
+ z (`torch.FloatTensor`): Input batch of latent vectors.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.vae.DecoderOutput`] or `tuple`:
+ If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is
+ returned.
+
+ """
+ batch_size = z.shape[0] // num_frames
+ image_only_indicator = torch.zeros(batch_size, num_frames, dtype=z.dtype, device=z.device)
+ decoded = self.decoder(z, num_frames=num_frames, image_only_indicator=image_only_indicator)
+
+ if not return_dict:
+ return (decoded,)
+
+ return DecoderOutput(sample=decoded)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ sample_posterior: bool = False,
+ return_dict: bool = True,
+ generator: Optional[torch.Generator] = None,
+ num_frames: int = 1,
+ ) -> Union[DecoderOutput, torch.FloatTensor]:
+ r"""
+ Args:
+ sample (`torch.FloatTensor`): Input sample.
+ sample_posterior (`bool`, *optional*, defaults to `False`):
+ Whether to sample from the posterior.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+ """
+ x = sample
+ posterior = self.encode(x).latent_dist
+ if sample_posterior:
+ z = posterior.sample(generator=generator)
+ else:
+ z = posterior.mode()
+
+ dec = self.decode(z, num_frames=num_frames).sample
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
diff --git a/model/diffusers_c/models/autoencoders/autoencoder_tiny.py b/model/diffusers_c/models/autoencoders/autoencoder_tiny.py
new file mode 100644
index 0000000000000000000000000000000000000000..ef43526cf8a03ea0792904fed7e7bc1ee8d8910c
--- /dev/null
+++ b/model/diffusers_c/models/autoencoders/autoencoder_tiny.py
@@ -0,0 +1,347 @@
+# Copyright 2024 Ollin Boer Bohan and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ...utils.accelerate_utils import apply_forward_hook
+from ..modeling_utils import ModelMixin
+from .vae import DecoderOutput, DecoderTiny, EncoderTiny
+
+
+@dataclass
+class AutoencoderTinyOutput(BaseOutput):
+ """
+ Output of AutoencoderTiny encoding method.
+
+ Args:
+ latents (`torch.Tensor`): Encoded outputs of the `Encoder`.
+
+ """
+
+ latents: torch.Tensor
+
+
+class AutoencoderTiny(ModelMixin, ConfigMixin):
+ r"""
+ A tiny distilled VAE model for encoding images into latents and decoding latent representations into images.
+
+ [`AutoencoderTiny`] is a wrapper around the original implementation of `TAESD`.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for its generic methods implemented for
+ all models (such as downloading or saving).
+
+ Parameters:
+ in_channels (`int`, *optional*, defaults to 3): Number of channels in the input image.
+ out_channels (`int`, *optional*, defaults to 3): Number of channels in the output.
+ encoder_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64, 64, 64, 64)`):
+ Tuple of integers representing the number of output channels for each encoder block. The length of the
+ tuple should be equal to the number of encoder blocks.
+ decoder_block_out_channels (`Tuple[int]`, *optional*, defaults to `(64, 64, 64, 64)`):
+ Tuple of integers representing the number of output channels for each decoder block. The length of the
+ tuple should be equal to the number of decoder blocks.
+ act_fn (`str`, *optional*, defaults to `"relu"`):
+ Activation function to be used throughout the model.
+ latent_channels (`int`, *optional*, defaults to 4):
+ Number of channels in the latent representation. The latent space acts as a compressed representation of
+ the input image.
+ upsampling_scaling_factor (`int`, *optional*, defaults to 2):
+ Scaling factor for upsampling in the decoder. It determines the size of the output image during the
+ upsampling process.
+ num_encoder_blocks (`Tuple[int]`, *optional*, defaults to `(1, 3, 3, 3)`):
+ Tuple of integers representing the number of encoder blocks at each stage of the encoding process. The
+ length of the tuple should be equal to the number of stages in the encoder. Each stage has a different
+ number of encoder blocks.
+ num_decoder_blocks (`Tuple[int]`, *optional*, defaults to `(3, 3, 3, 1)`):
+ Tuple of integers representing the number of decoder blocks at each stage of the decoding process. The
+ length of the tuple should be equal to the number of stages in the decoder. Each stage has a different
+ number of decoder blocks.
+ latent_magnitude (`float`, *optional*, defaults to 3.0):
+ Magnitude of the latent representation. This parameter scales the latent representation values to control
+ the extent of information preservation.
+ latent_shift (float, *optional*, defaults to 0.5):
+ Shift applied to the latent representation. This parameter controls the center of the latent space.
+ scaling_factor (`float`, *optional*, defaults to 1.0):
+ The component-wise standard deviation of the trained latent space computed using the first batch of the
+ training set. This is used to scale the latent space to have unit variance when training the diffusion
+ model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+ diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+ / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+ Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper. For this Autoencoder,
+ however, no such scaling factor was used, hence the value of 1.0 as the default.
+ force_upcast (`bool`, *optional*, default to `False`):
+ If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE
+ can be fine-tuned / trained to a lower range without losing too much precision, in which case
+ `force_upcast` can be set to `False` (see this fp16-friendly
+ [AutoEncoder](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix)).
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ encoder_block_out_channels: Tuple[int, ...] = (64, 64, 64, 64),
+ decoder_block_out_channels: Tuple[int, ...] = (64, 64, 64, 64),
+ act_fn: str = "relu",
+ latent_channels: int = 4,
+ upsampling_scaling_factor: int = 2,
+ num_encoder_blocks: Tuple[int, ...] = (1, 3, 3, 3),
+ num_decoder_blocks: Tuple[int, ...] = (3, 3, 3, 1),
+ latent_magnitude: int = 3,
+ latent_shift: float = 0.5,
+ force_upcast: bool = False,
+ scaling_factor: float = 1.0,
+ ):
+ super().__init__()
+
+ if len(encoder_block_out_channels) != len(num_encoder_blocks):
+ raise ValueError("`encoder_block_out_channels` should have the same length as `num_encoder_blocks`.")
+ if len(decoder_block_out_channels) != len(num_decoder_blocks):
+ raise ValueError("`decoder_block_out_channels` should have the same length as `num_decoder_blocks`.")
+
+ self.encoder = EncoderTiny(
+ in_channels=in_channels,
+ out_channels=latent_channels,
+ num_blocks=num_encoder_blocks,
+ block_out_channels=encoder_block_out_channels,
+ act_fn=act_fn,
+ )
+
+ self.decoder = DecoderTiny(
+ in_channels=latent_channels,
+ out_channels=out_channels,
+ num_blocks=num_decoder_blocks,
+ block_out_channels=decoder_block_out_channels,
+ upsampling_scaling_factor=upsampling_scaling_factor,
+ act_fn=act_fn,
+ )
+
+ self.latent_magnitude = latent_magnitude
+ self.latent_shift = latent_shift
+ self.scaling_factor = scaling_factor
+
+ self.use_slicing = False
+ self.use_tiling = False
+
+ # only relevant if vae tiling is enabled
+ self.spatial_scale_factor = 2**out_channels
+ self.tile_overlap_factor = 0.125
+ self.tile_sample_min_size = 512
+ self.tile_latent_min_size = self.tile_sample_min_size // self.spatial_scale_factor
+
+ self.register_to_config(block_out_channels=decoder_block_out_channels)
+ self.register_to_config(force_upcast=False)
+
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (EncoderTiny, DecoderTiny)):
+ module.gradient_checkpointing = value
+
+ def scale_latents(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ """raw latents -> [0, 1]"""
+ return x.div(2 * self.latent_magnitude).add(self.latent_shift).clamp(0, 1)
+
+ def unscale_latents(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ """[0, 1] -> raw latents"""
+ return x.sub(self.latent_shift).mul(2 * self.latent_magnitude)
+
+ def enable_slicing(self) -> None:
+ r"""
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+ """
+ self.use_slicing = True
+
+ def disable_slicing(self) -> None:
+ r"""
+ Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+ decoding in one step.
+ """
+ self.use_slicing = False
+
+ def enable_tiling(self, use_tiling: bool = True) -> None:
+ r"""
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+ processing larger images.
+ """
+ self.use_tiling = use_tiling
+
+ def disable_tiling(self) -> None:
+ r"""
+ Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+ decoding in one step.
+ """
+ self.enable_tiling(False)
+
+ def _tiled_encode(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ r"""Encode a batch of images using a tiled encoder.
+
+ When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+ steps. This is useful to keep memory use constant regardless of image size. To avoid tiling artifacts, the
+ tiles overlap and are blended together to form a smooth output.
+
+ Args:
+ x (`torch.FloatTensor`): Input batch of images.
+
+ Returns:
+ `torch.FloatTensor`: Encoded batch of images.
+ """
+ # scale of encoder output relative to input
+ sf = self.spatial_scale_factor
+ tile_size = self.tile_sample_min_size
+
+ # number of pixels to blend and to traverse between tile
+ blend_size = int(tile_size * self.tile_overlap_factor)
+ traverse_size = tile_size - blend_size
+
+ # tiles index (up/left)
+ ti = range(0, x.shape[-2], traverse_size)
+ tj = range(0, x.shape[-1], traverse_size)
+
+ # mask for blending
+ blend_masks = torch.stack(
+ torch.meshgrid([torch.arange(tile_size / sf) / (blend_size / sf - 1)] * 2, indexing="ij")
+ )
+ blend_masks = blend_masks.clamp(0, 1).to(x.device)
+
+ # output array
+ out = torch.zeros(x.shape[0], 4, x.shape[-2] // sf, x.shape[-1] // sf, device=x.device)
+ for i in ti:
+ for j in tj:
+ tile_in = x[..., i : i + tile_size, j : j + tile_size]
+ # tile result
+ tile_out = out[..., i // sf : (i + tile_size) // sf, j // sf : (j + tile_size) // sf]
+ tile = self.encoder(tile_in)
+ h, w = tile.shape[-2], tile.shape[-1]
+ # blend tile result into output
+ blend_mask_i = torch.ones_like(blend_masks[0]) if i == 0 else blend_masks[0]
+ blend_mask_j = torch.ones_like(blend_masks[1]) if j == 0 else blend_masks[1]
+ blend_mask = blend_mask_i * blend_mask_j
+ tile, blend_mask = tile[..., :h, :w], blend_mask[..., :h, :w]
+ tile_out.copy_(blend_mask * tile + (1 - blend_mask) * tile_out)
+ return out
+
+ def _tiled_decode(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ r"""Encode a batch of images using a tiled encoder.
+
+ When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+ steps. This is useful to keep memory use constant regardless of image size. To avoid tiling artifacts, the
+ tiles overlap and are blended together to form a smooth output.
+
+ Args:
+ x (`torch.FloatTensor`): Input batch of images.
+
+ Returns:
+ `torch.FloatTensor`: Encoded batch of images.
+ """
+ # scale of decoder output relative to input
+ sf = self.spatial_scale_factor
+ tile_size = self.tile_latent_min_size
+
+ # number of pixels to blend and to traverse between tiles
+ blend_size = int(tile_size * self.tile_overlap_factor)
+ traverse_size = tile_size - blend_size
+
+ # tiles index (up/left)
+ ti = range(0, x.shape[-2], traverse_size)
+ tj = range(0, x.shape[-1], traverse_size)
+
+ # mask for blending
+ blend_masks = torch.stack(
+ torch.meshgrid([torch.arange(tile_size * sf) / (blend_size * sf - 1)] * 2, indexing="ij")
+ )
+ blend_masks = blend_masks.clamp(0, 1).to(x.device)
+
+ # output array
+ out = torch.zeros(x.shape[0], 3, x.shape[-2] * sf, x.shape[-1] * sf, device=x.device)
+ for i in ti:
+ for j in tj:
+ tile_in = x[..., i : i + tile_size, j : j + tile_size]
+ # tile result
+ tile_out = out[..., i * sf : (i + tile_size) * sf, j * sf : (j + tile_size) * sf]
+ tile = self.decoder(tile_in)
+ h, w = tile.shape[-2], tile.shape[-1]
+ # blend tile result into output
+ blend_mask_i = torch.ones_like(blend_masks[0]) if i == 0 else blend_masks[0]
+ blend_mask_j = torch.ones_like(blend_masks[1]) if j == 0 else blend_masks[1]
+ blend_mask = (blend_mask_i * blend_mask_j)[..., :h, :w]
+ tile_out.copy_(blend_mask * tile + (1 - blend_mask) * tile_out)
+ return out
+
+ @apply_forward_hook
+ def encode(
+ self, x: torch.FloatTensor, return_dict: bool = True
+ ) -> Union[AutoencoderTinyOutput, Tuple[torch.FloatTensor]]:
+ if self.use_slicing and x.shape[0] > 1:
+ output = [
+ self._tiled_encode(x_slice) if self.use_tiling else self.encoder(x_slice) for x_slice in x.split(1)
+ ]
+ output = torch.cat(output)
+ else:
+ output = self._tiled_encode(x) if self.use_tiling else self.encoder(x)
+
+ if not return_dict:
+ return (output,)
+
+ return AutoencoderTinyOutput(latents=output)
+
+ @apply_forward_hook
+ def decode(
+ self, x: torch.FloatTensor, generator: Optional[torch.Generator] = None, return_dict: bool = True
+ ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+ if self.use_slicing and x.shape[0] > 1:
+ output = [self._tiled_decode(x_slice) if self.use_tiling else self.decoder(x) for x_slice in x.split(1)]
+ output = torch.cat(output)
+ else:
+ output = self._tiled_decode(x) if self.use_tiling else self.decoder(x)
+
+ if not return_dict:
+ return (output,)
+
+ return DecoderOutput(sample=output)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+ r"""
+ Args:
+ sample (`torch.FloatTensor`): Input sample.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+ """
+ enc = self.encode(sample).latents
+
+ # scale latents to be in [0, 1], then quantize latents to a byte tensor,
+ # as if we were storing the latents in an RGBA uint8 image.
+ scaled_enc = self.scale_latents(enc).mul_(255).round_().byte()
+
+ # unquantize latents back into [0, 1], then unscale latents back to their original range,
+ # as if we were loading the latents from an RGBA uint8 image.
+ unscaled_enc = self.unscale_latents(scaled_enc / 255.0)
+
+ dec = self.decode(unscaled_enc)
+
+ if not return_dict:
+ return (dec,)
+ return DecoderOutput(sample=dec)
diff --git a/model/diffusers_c/models/autoencoders/consistency_decoder_vae.py b/model/diffusers_c/models/autoencoders/consistency_decoder_vae.py
new file mode 100644
index 0000000000000000000000000000000000000000..72c512da983060685375311567363884d652585e
--- /dev/null
+++ b/model/diffusers_c/models/autoencoders/consistency_decoder_vae.py
@@ -0,0 +1,435 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...schedulers import ConsistencyDecoderScheduler
+from ...utils import BaseOutput
+from ...utils.accelerate_utils import apply_forward_hook
+from ...utils.torch_utils import randn_tensor
+from ..attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ..modeling_utils import ModelMixin
+from ..unets.unet_2d import UNet2DModel
+from .vae import DecoderOutput, DiagonalGaussianDistribution, Encoder
+
+
+@dataclass
+class ConsistencyDecoderVAEOutput(BaseOutput):
+ """
+ Output of encoding method.
+
+ Args:
+ latent_dist (`DiagonalGaussianDistribution`):
+ Encoded outputs of `Encoder` represented as the mean and logvar of `DiagonalGaussianDistribution`.
+ `DiagonalGaussianDistribution` allows for sampling latents from the distribution.
+ """
+
+ latent_dist: "DiagonalGaussianDistribution"
+
+
+class ConsistencyDecoderVAE(ModelMixin, ConfigMixin):
+ r"""
+ The consistency decoder used with DALL-E 3.
+
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionPipeline, ConsistencyDecoderVAE
+
+ >>> vae = ConsistencyDecoderVAE.from_pretrained("openai/consistency-decoder", torch_dtype=torch.float16)
+ >>> pipe = StableDiffusionPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", vae=vae, torch_dtype=torch.float16
+ ... ).to("cuda")
+
+ >>> pipe("horse", generator=torch.manual_seed(0)).images
+ ```
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ scaling_factor: float = 0.18215,
+ latent_channels: int = 4,
+ encoder_act_fn: str = "silu",
+ encoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 512),
+ encoder_double_z: bool = True,
+ encoder_down_block_types: Tuple[str, ...] = (
+ "DownEncoderBlock2D",
+ "DownEncoderBlock2D",
+ "DownEncoderBlock2D",
+ "DownEncoderBlock2D",
+ ),
+ encoder_in_channels: int = 3,
+ encoder_layers_per_block: int = 2,
+ encoder_norm_num_groups: int = 32,
+ encoder_out_channels: int = 4,
+ decoder_add_attention: bool = False,
+ decoder_block_out_channels: Tuple[int, ...] = (320, 640, 1024, 1024),
+ decoder_down_block_types: Tuple[str, ...] = (
+ "ResnetDownsampleBlock2D",
+ "ResnetDownsampleBlock2D",
+ "ResnetDownsampleBlock2D",
+ "ResnetDownsampleBlock2D",
+ ),
+ decoder_downsample_padding: int = 1,
+ decoder_in_channels: int = 7,
+ decoder_layers_per_block: int = 3,
+ decoder_norm_eps: float = 1e-05,
+ decoder_norm_num_groups: int = 32,
+ decoder_num_train_timesteps: int = 1024,
+ decoder_out_channels: int = 6,
+ decoder_resnet_time_scale_shift: str = "scale_shift",
+ decoder_time_embedding_type: str = "learned",
+ decoder_up_block_types: Tuple[str, ...] = (
+ "ResnetUpsampleBlock2D",
+ "ResnetUpsampleBlock2D",
+ "ResnetUpsampleBlock2D",
+ "ResnetUpsampleBlock2D",
+ ),
+ ):
+ super().__init__()
+ self.encoder = Encoder(
+ act_fn=encoder_act_fn,
+ block_out_channels=encoder_block_out_channels,
+ double_z=encoder_double_z,
+ down_block_types=encoder_down_block_types,
+ in_channels=encoder_in_channels,
+ layers_per_block=encoder_layers_per_block,
+ norm_num_groups=encoder_norm_num_groups,
+ out_channels=encoder_out_channels,
+ )
+
+ self.decoder_unet = UNet2DModel(
+ add_attention=decoder_add_attention,
+ block_out_channels=decoder_block_out_channels,
+ down_block_types=decoder_down_block_types,
+ downsample_padding=decoder_downsample_padding,
+ in_channels=decoder_in_channels,
+ layers_per_block=decoder_layers_per_block,
+ norm_eps=decoder_norm_eps,
+ norm_num_groups=decoder_norm_num_groups,
+ num_train_timesteps=decoder_num_train_timesteps,
+ out_channels=decoder_out_channels,
+ resnet_time_scale_shift=decoder_resnet_time_scale_shift,
+ time_embedding_type=decoder_time_embedding_type,
+ up_block_types=decoder_up_block_types,
+ )
+ self.decoder_scheduler = ConsistencyDecoderScheduler()
+ self.register_to_config(block_out_channels=encoder_block_out_channels)
+ self.register_to_config(force_upcast=False)
+ self.register_buffer(
+ "means",
+ torch.tensor([0.38862467, 0.02253063, 0.07381133, -0.0171294])[None, :, None, None],
+ persistent=False,
+ )
+ self.register_buffer(
+ "stds", torch.tensor([0.9654121, 1.0440036, 0.76147926, 0.77022034])[None, :, None, None], persistent=False
+ )
+
+ self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1)
+
+ self.use_slicing = False
+ self.use_tiling = False
+
+ # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.enable_tiling
+ def enable_tiling(self, use_tiling: bool = True):
+ r"""
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+ processing larger images.
+ """
+ self.use_tiling = use_tiling
+
+ # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.disable_tiling
+ def disable_tiling(self):
+ r"""
+ Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing
+ decoding in one step.
+ """
+ self.enable_tiling(False)
+
+ # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.enable_slicing
+ def enable_slicing(self):
+ r"""
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+ """
+ self.use_slicing = True
+
+ # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.disable_slicing
+ def disable_slicing(self):
+ r"""
+ Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing
+ decoding in one step.
+ """
+ self.use_slicing = False
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ @apply_forward_hook
+ def encode(
+ self, x: torch.FloatTensor, return_dict: bool = True
+ ) -> Union[ConsistencyDecoderVAEOutput, Tuple[DiagonalGaussianDistribution]]:
+ """
+ Encode a batch of images into latents.
+
+ Args:
+ x (`torch.FloatTensor`): Input batch of images.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether to return a [`~models.consistecy_decoder_vae.ConsistencyDecoderOoutput`] instead of a plain
+ tuple.
+
+ Returns:
+ The latent representations of the encoded images. If `return_dict` is True, a
+ [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] is returned, otherwise a plain `tuple`
+ is returned.
+ """
+ if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size):
+ return self.tiled_encode(x, return_dict=return_dict)
+
+ if self.use_slicing and x.shape[0] > 1:
+ encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)]
+ h = torch.cat(encoded_slices)
+ else:
+ h = self.encoder(x)
+
+ moments = self.quant_conv(h)
+ posterior = DiagonalGaussianDistribution(moments)
+
+ if not return_dict:
+ return (posterior,)
+
+ return ConsistencyDecoderVAEOutput(latent_dist=posterior)
+
+ @apply_forward_hook
+ def decode(
+ self,
+ z: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ num_inference_steps: int = 2,
+ ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+ z = (z * self.config.scaling_factor - self.means) / self.stds
+
+ scale_factor = 2 ** (len(self.config.block_out_channels) - 1)
+ z = F.interpolate(z, mode="nearest", scale_factor=scale_factor)
+
+ batch_size, _, height, width = z.shape
+
+ self.decoder_scheduler.set_timesteps(num_inference_steps, device=self.device)
+
+ x_t = self.decoder_scheduler.init_noise_sigma * randn_tensor(
+ (batch_size, 3, height, width), generator=generator, dtype=z.dtype, device=z.device
+ )
+
+ for t in self.decoder_scheduler.timesteps:
+ model_input = torch.concat([self.decoder_scheduler.scale_model_input(x_t, t), z], dim=1)
+ model_output = self.decoder_unet(model_input, t).sample[:, :3, :, :]
+ prev_sample = self.decoder_scheduler.step(model_output, t, x_t, generator).prev_sample
+ x_t = prev_sample
+
+ x_0 = x_t
+
+ if not return_dict:
+ return (x_0,)
+
+ return DecoderOutput(sample=x_0)
+
+ # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.blend_v
+ def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+ blend_extent = min(a.shape[2], b.shape[2], blend_extent)
+ for y in range(blend_extent):
+ b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent)
+ return b
+
+ # Copied from diffusers.models.autoencoders.autoencoder_kl.AutoencoderKL.blend_h
+ def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor:
+ blend_extent = min(a.shape[3], b.shape[3], blend_extent)
+ for x in range(blend_extent):
+ b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent)
+ return b
+
+ def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> ConsistencyDecoderVAEOutput:
+ r"""Encode a batch of images using a tiled encoder.
+
+ When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several
+ steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is
+ different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the
+ tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the
+ output, but they should be much less noticeable.
+
+ Args:
+ x (`torch.FloatTensor`): Input batch of images.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] instead of a
+ plain tuple.
+
+ Returns:
+ [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] or `tuple`:
+ If return_dict is True, a [`~models.consistency_decoder_vae.ConsistencyDecoderVAEOutput`] is returned,
+ otherwise a plain `tuple` is returned.
+ """
+ overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor))
+ blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor)
+ row_limit = self.tile_latent_min_size - blend_extent
+
+ # Split the image into 512x512 tiles and encode them separately.
+ rows = []
+ for i in range(0, x.shape[2], overlap_size):
+ row = []
+ for j in range(0, x.shape[3], overlap_size):
+ tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size]
+ tile = self.encoder(tile)
+ tile = self.quant_conv(tile)
+ row.append(tile)
+ rows.append(row)
+ result_rows = []
+ for i, row in enumerate(rows):
+ result_row = []
+ for j, tile in enumerate(row):
+ # blend the above tile and the left tile
+ # to the current tile and add the current tile to the result row
+ if i > 0:
+ tile = self.blend_v(rows[i - 1][j], tile, blend_extent)
+ if j > 0:
+ tile = self.blend_h(row[j - 1], tile, blend_extent)
+ result_row.append(tile[:, :, :row_limit, :row_limit])
+ result_rows.append(torch.cat(result_row, dim=3))
+
+ moments = torch.cat(result_rows, dim=2)
+ posterior = DiagonalGaussianDistribution(moments)
+
+ if not return_dict:
+ return (posterior,)
+
+ return ConsistencyDecoderVAEOutput(latent_dist=posterior)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ sample_posterior: bool = False,
+ return_dict: bool = True,
+ generator: Optional[torch.Generator] = None,
+ ) -> Union[DecoderOutput, Tuple[torch.FloatTensor]]:
+ r"""
+ Args:
+ sample (`torch.FloatTensor`): Input sample.
+ sample_posterior (`bool`, *optional*, defaults to `False`):
+ Whether to sample from the posterior.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+ generator (`torch.Generator`, *optional*, defaults to `None`):
+ Generator to use for sampling.
+
+ Returns:
+ [`DecoderOutput`] or `tuple`:
+ If return_dict is True, a [`DecoderOutput`] is returned, otherwise a plain `tuple` is returned.
+ """
+ x = sample
+ posterior = self.encode(x).latent_dist
+ if sample_posterior:
+ z = posterior.sample(generator=generator)
+ else:
+ z = posterior.mode()
+ dec = self.decode(z, generator=generator).sample
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
diff --git a/model/diffusers_c/models/autoencoders/vae.py b/model/diffusers_c/models/autoencoders/vae.py
new file mode 100644
index 0000000000000000000000000000000000000000..cf2e37337c16f19eb8b74769a7d37be1bd2805a0
--- /dev/null
+++ b/model/diffusers_c/models/autoencoders/vae.py
@@ -0,0 +1,983 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from ...utils import BaseOutput, is_torch_version
+from ...utils.torch_utils import randn_tensor
+from ..activations import get_activation
+from ..attention_processor import SpatialNorm
+from ..unets.unet_2d_blocks import (
+ AutoencoderTinyBlock,
+ UNetMidBlock2D,
+ get_down_block,
+ get_up_block,
+)
+
+
+@dataclass
+class DecoderOutput(BaseOutput):
+ r"""
+ Output of decoding method.
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+ The decoded output sample from the last layer of the model.
+ """
+
+ sample: torch.FloatTensor
+
+
+class Encoder(nn.Module):
+ r"""
+ The `Encoder` layer of a variational autoencoder that encodes its input into a latent representation.
+
+ Args:
+ in_channels (`int`, *optional*, defaults to 3):
+ The number of input channels.
+ out_channels (`int`, *optional*, defaults to 3):
+ The number of output channels.
+ down_block_types (`Tuple[str, ...]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+ The types of down blocks to use. See `~diffusers.models.unet_2d_blocks.get_down_block` for available
+ options.
+ block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+ The number of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2):
+ The number of layers per block.
+ norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups for normalization.
+ act_fn (`str`, *optional*, defaults to `"silu"`):
+ The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+ double_z (`bool`, *optional*, defaults to `True`):
+ Whether to double the number of output channels for the last block.
+ """
+
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
+ block_out_channels: Tuple[int, ...] = (64,),
+ layers_per_block: int = 2,
+ norm_num_groups: int = 32,
+ act_fn: str = "silu",
+ double_z: bool = True,
+ mid_block_add_attention=True,
+ ):
+ super().__init__()
+ self.layers_per_block = layers_per_block
+
+ self.conv_in = nn.Conv2d(
+ in_channels,
+ block_out_channels[0],
+ kernel_size=3,
+ stride=1,
+ padding=1,
+ )
+
+ self.mid_block = None
+ self.down_blocks = nn.ModuleList([])
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=self.layers_per_block,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ add_downsample=not is_final_block,
+ resnet_eps=1e-6,
+ downsample_padding=0,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ attention_head_dim=output_channel,
+ temb_channels=None,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ self.mid_block = UNetMidBlock2D(
+ in_channels=block_out_channels[-1],
+ resnet_eps=1e-6,
+ resnet_act_fn=act_fn,
+ output_scale_factor=1,
+ resnet_time_scale_shift="default",
+ attention_head_dim=block_out_channels[-1],
+ resnet_groups=norm_num_groups,
+ temb_channels=None,
+ add_attention=mid_block_add_attention,
+ )
+
+ # out
+ self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[-1], num_groups=norm_num_groups, eps=1e-6)
+ self.conv_act = nn.SiLU()
+
+ conv_out_channels = 2 * out_channels if double_z else out_channels
+ self.conv_out = nn.Conv2d(block_out_channels[-1], conv_out_channels, 3, padding=1)
+
+ self.gradient_checkpointing = False
+
+ def forward(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ r"""The forward method of the `Encoder` class."""
+
+ sample = self.conv_in(sample)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ # down
+ if is_torch_version(">=", "1.11.0"):
+ for down_block in self.down_blocks:
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(down_block), sample, use_reentrant=False
+ )
+ # middle
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.mid_block), sample, use_reentrant=False
+ )
+ else:
+ for down_block in self.down_blocks:
+ sample = torch.utils.checkpoint.checkpoint(create_custom_forward(down_block), sample)
+ # middle
+ sample = torch.utils.checkpoint.checkpoint(create_custom_forward(self.mid_block), sample)
+
+ else:
+ # down
+ for down_block in self.down_blocks:
+ sample = down_block(sample)
+
+ # middle
+ sample = self.mid_block(sample)
+
+ # post-process
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ return sample
+
+
+class Decoder(nn.Module):
+ r"""
+ The `Decoder` layer of a variational autoencoder that decodes its latent representation into an output sample.
+
+ Args:
+ in_channels (`int`, *optional*, defaults to 3):
+ The number of input channels.
+ out_channels (`int`, *optional*, defaults to 3):
+ The number of output channels.
+ up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+ The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
+ block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+ The number of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2):
+ The number of layers per block.
+ norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups for normalization.
+ act_fn (`str`, *optional*, defaults to `"silu"`):
+ The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+ norm_type (`str`, *optional*, defaults to `"group"`):
+ The normalization type to use. Can be either `"group"` or `"spatial"`.
+ """
+
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+ block_out_channels: Tuple[int, ...] = (64,),
+ layers_per_block: int = 2,
+ norm_num_groups: int = 32,
+ act_fn: str = "silu",
+ norm_type: str = "group", # group, spatial
+ mid_block_add_attention=True,
+ ):
+ super().__init__()
+ self.layers_per_block = layers_per_block
+
+ self.conv_in = nn.Conv2d(
+ in_channels,
+ block_out_channels[-1],
+ kernel_size=3,
+ stride=1,
+ padding=1,
+ )
+
+ self.mid_block = None
+ self.up_blocks = nn.ModuleList([])
+
+ temb_channels = in_channels if norm_type == "spatial" else None
+
+ # mid
+ self.mid_block = UNetMidBlock2D(
+ in_channels=block_out_channels[-1],
+ resnet_eps=1e-6,
+ resnet_act_fn=act_fn,
+ output_scale_factor=1,
+ resnet_time_scale_shift="default" if norm_type == "group" else norm_type,
+ attention_head_dim=block_out_channels[-1],
+ resnet_groups=norm_num_groups,
+ temb_channels=temb_channels,
+ add_attention=mid_block_add_attention,
+ )
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+
+ is_final_block = i == len(block_out_channels) - 1
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=self.layers_per_block + 1,
+ in_channels=prev_output_channel,
+ out_channels=output_channel,
+ prev_output_channel=None,
+ add_upsample=not is_final_block,
+ resnet_eps=1e-6,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ attention_head_dim=output_channel,
+ temb_channels=temb_channels,
+ resnet_time_scale_shift=norm_type,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ if norm_type == "spatial":
+ self.conv_norm_out = SpatialNorm(block_out_channels[0], temb_channels)
+ else:
+ self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6)
+ self.conv_act = nn.SiLU()
+ self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ latent_embeds: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ r"""The forward method of the `Decoder` class."""
+
+ sample = self.conv_in(sample)
+
+ upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ # middle
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.mid_block),
+ sample,
+ latent_embeds,
+ use_reentrant=False,
+ )
+ sample = sample.to(upscale_dtype)
+
+ # up
+ for up_block in self.up_blocks:
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(up_block),
+ sample,
+ latent_embeds,
+ use_reentrant=False,
+ )
+ else:
+ # middle
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.mid_block), sample, latent_embeds
+ )
+ sample = sample.to(upscale_dtype)
+
+ # up
+ for up_block in self.up_blocks:
+ sample = torch.utils.checkpoint.checkpoint(create_custom_forward(up_block), sample, latent_embeds)
+ else:
+ # middle
+ sample = self.mid_block(sample, latent_embeds)
+ sample = sample.to(upscale_dtype)
+
+ # up
+ for up_block in self.up_blocks:
+ sample = up_block(sample, latent_embeds)
+
+ # post-process
+ if latent_embeds is None:
+ sample = self.conv_norm_out(sample)
+ else:
+ sample = self.conv_norm_out(sample, latent_embeds)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ return sample
+
+
+class UpSample(nn.Module):
+ r"""
+ The `UpSample` layer of a variational autoencoder that upsamples its input.
+
+ Args:
+ in_channels (`int`, *optional*, defaults to 3):
+ The number of input channels.
+ out_channels (`int`, *optional*, defaults to 3):
+ The number of output channels.
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ ) -> None:
+ super().__init__()
+ self.in_channels = in_channels
+ self.out_channels = out_channels
+ self.deconv = nn.ConvTranspose2d(in_channels, out_channels, kernel_size=4, stride=2, padding=1)
+
+ def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ r"""The forward method of the `UpSample` class."""
+ x = torch.relu(x)
+ x = self.deconv(x)
+ return x
+
+
+class MaskConditionEncoder(nn.Module):
+ """
+ used in AsymmetricAutoencoderKL
+ """
+
+ def __init__(
+ self,
+ in_ch: int,
+ out_ch: int = 192,
+ res_ch: int = 768,
+ stride: int = 16,
+ ) -> None:
+ super().__init__()
+
+ channels = []
+ while stride > 1:
+ stride = stride // 2
+ in_ch_ = out_ch * 2
+ if out_ch > res_ch:
+ out_ch = res_ch
+ if stride == 1:
+ in_ch_ = res_ch
+ channels.append((in_ch_, out_ch))
+ out_ch *= 2
+
+ out_channels = []
+ for _in_ch, _out_ch in channels:
+ out_channels.append(_out_ch)
+ out_channels.append(channels[-1][0])
+
+ layers = []
+ in_ch_ = in_ch
+ for l in range(len(out_channels)):
+ out_ch_ = out_channels[l]
+ if l == 0 or l == 1:
+ layers.append(nn.Conv2d(in_ch_, out_ch_, kernel_size=3, stride=1, padding=1))
+ else:
+ layers.append(nn.Conv2d(in_ch_, out_ch_, kernel_size=4, stride=2, padding=1))
+ in_ch_ = out_ch_
+
+ self.layers = nn.Sequential(*layers)
+
+ def forward(self, x: torch.FloatTensor, mask=None) -> torch.FloatTensor:
+ r"""The forward method of the `MaskConditionEncoder` class."""
+ out = {}
+ for l in range(len(self.layers)):
+ layer = self.layers[l]
+ x = layer(x)
+ out[str(tuple(x.shape))] = x
+ x = torch.relu(x)
+ return out
+
+
+class MaskConditionDecoder(nn.Module):
+ r"""The `MaskConditionDecoder` should be used in combination with [`AsymmetricAutoencoderKL`] to enhance the model's
+ decoder with a conditioner on the mask and masked image.
+
+ Args:
+ in_channels (`int`, *optional*, defaults to 3):
+ The number of input channels.
+ out_channels (`int`, *optional*, defaults to 3):
+ The number of output channels.
+ up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+ The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options.
+ block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`):
+ The number of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2):
+ The number of layers per block.
+ norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups for normalization.
+ act_fn (`str`, *optional*, defaults to `"silu"`):
+ The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+ norm_type (`str`, *optional*, defaults to `"group"`):
+ The normalization type to use. Can be either `"group"` or `"spatial"`.
+ """
+
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+ block_out_channels: Tuple[int, ...] = (64,),
+ layers_per_block: int = 2,
+ norm_num_groups: int = 32,
+ act_fn: str = "silu",
+ norm_type: str = "group", # group, spatial
+ ):
+ super().__init__()
+ self.layers_per_block = layers_per_block
+
+ self.conv_in = nn.Conv2d(
+ in_channels,
+ block_out_channels[-1],
+ kernel_size=3,
+ stride=1,
+ padding=1,
+ )
+
+ self.mid_block = None
+ self.up_blocks = nn.ModuleList([])
+
+ temb_channels = in_channels if norm_type == "spatial" else None
+
+ # mid
+ self.mid_block = UNetMidBlock2D(
+ in_channels=block_out_channels[-1],
+ resnet_eps=1e-6,
+ resnet_act_fn=act_fn,
+ output_scale_factor=1,
+ resnet_time_scale_shift="default" if norm_type == "group" else norm_type,
+ attention_head_dim=block_out_channels[-1],
+ resnet_groups=norm_num_groups,
+ temb_channels=temb_channels,
+ )
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+
+ is_final_block = i == len(block_out_channels) - 1
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=self.layers_per_block + 1,
+ in_channels=prev_output_channel,
+ out_channels=output_channel,
+ prev_output_channel=None,
+ add_upsample=not is_final_block,
+ resnet_eps=1e-6,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ attention_head_dim=output_channel,
+ temb_channels=temb_channels,
+ resnet_time_scale_shift=norm_type,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # condition encoder
+ self.condition_encoder = MaskConditionEncoder(
+ in_ch=out_channels,
+ out_ch=block_out_channels[0],
+ res_ch=block_out_channels[-1],
+ )
+
+ # out
+ if norm_type == "spatial":
+ self.conv_norm_out = SpatialNorm(block_out_channels[0], temb_channels)
+ else:
+ self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-6)
+ self.conv_act = nn.SiLU()
+ self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, 3, padding=1)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ z: torch.FloatTensor,
+ image: Optional[torch.FloatTensor] = None,
+ mask: Optional[torch.FloatTensor] = None,
+ latent_embeds: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ r"""The forward method of the `MaskConditionDecoder` class."""
+ sample = z
+ sample = self.conv_in(sample)
+
+ upscale_dtype = next(iter(self.up_blocks.parameters())).dtype
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ # middle
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.mid_block),
+ sample,
+ latent_embeds,
+ use_reentrant=False,
+ )
+ sample = sample.to(upscale_dtype)
+
+ # condition encoder
+ if image is not None and mask is not None:
+ masked_image = (1 - mask) * image
+ im_x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.condition_encoder),
+ masked_image,
+ mask,
+ use_reentrant=False,
+ )
+
+ # up
+ for up_block in self.up_blocks:
+ if image is not None and mask is not None:
+ sample_ = im_x[str(tuple(sample.shape))]
+ mask_ = nn.functional.interpolate(mask, size=sample.shape[-2:], mode="nearest")
+ sample = sample * mask_ + sample_ * (1 - mask_)
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(up_block),
+ sample,
+ latent_embeds,
+ use_reentrant=False,
+ )
+ if image is not None and mask is not None:
+ sample = sample * mask + im_x[str(tuple(sample.shape))] * (1 - mask)
+ else:
+ # middle
+ sample = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.mid_block), sample, latent_embeds
+ )
+ sample = sample.to(upscale_dtype)
+
+ # condition encoder
+ if image is not None and mask is not None:
+ masked_image = (1 - mask) * image
+ im_x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.condition_encoder),
+ masked_image,
+ mask,
+ )
+
+ # up
+ for up_block in self.up_blocks:
+ if image is not None and mask is not None:
+ sample_ = im_x[str(tuple(sample.shape))]
+ mask_ = nn.functional.interpolate(mask, size=sample.shape[-2:], mode="nearest")
+ sample = sample * mask_ + sample_ * (1 - mask_)
+ sample = torch.utils.checkpoint.checkpoint(create_custom_forward(up_block), sample, latent_embeds)
+ if image is not None and mask is not None:
+ sample = sample * mask + im_x[str(tuple(sample.shape))] * (1 - mask)
+ else:
+ # middle
+ sample = self.mid_block(sample, latent_embeds)
+ sample = sample.to(upscale_dtype)
+
+ # condition encoder
+ if image is not None and mask is not None:
+ masked_image = (1 - mask) * image
+ im_x = self.condition_encoder(masked_image, mask)
+
+ # up
+ for up_block in self.up_blocks:
+ if image is not None and mask is not None:
+ sample_ = im_x[str(tuple(sample.shape))]
+ mask_ = nn.functional.interpolate(mask, size=sample.shape[-2:], mode="nearest")
+ sample = sample * mask_ + sample_ * (1 - mask_)
+ sample = up_block(sample, latent_embeds)
+ if image is not None and mask is not None:
+ sample = sample * mask + im_x[str(tuple(sample.shape))] * (1 - mask)
+
+ # post-process
+ if latent_embeds is None:
+ sample = self.conv_norm_out(sample)
+ else:
+ sample = self.conv_norm_out(sample, latent_embeds)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ return sample
+
+
+class VectorQuantizer(nn.Module):
+ """
+ Improved version over VectorQuantizer, can be used as a drop-in replacement. Mostly avoids costly matrix
+ multiplications and allows for post-hoc remapping of indices.
+ """
+
+ # NOTE: due to a bug the beta term was applied to the wrong term. for
+ # backwards compatibility we use the buggy version by default, but you can
+ # specify legacy=False to fix it.
+ def __init__(
+ self,
+ n_e: int,
+ vq_embed_dim: int,
+ beta: float,
+ remap=None,
+ unknown_index: str = "random",
+ sane_index_shape: bool = False,
+ legacy: bool = True,
+ ):
+ super().__init__()
+ self.n_e = n_e
+ self.vq_embed_dim = vq_embed_dim
+ self.beta = beta
+ self.legacy = legacy
+
+ self.embedding = nn.Embedding(self.n_e, self.vq_embed_dim)
+ self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
+
+ self.remap = remap
+ if self.remap is not None:
+ self.register_buffer("used", torch.tensor(np.load(self.remap)))
+ self.used: torch.Tensor
+ self.re_embed = self.used.shape[0]
+ self.unknown_index = unknown_index # "random" or "extra" or integer
+ if self.unknown_index == "extra":
+ self.unknown_index = self.re_embed
+ self.re_embed = self.re_embed + 1
+ print(
+ f"Remapping {self.n_e} indices to {self.re_embed} indices. "
+ f"Using {self.unknown_index} for unknown indices."
+ )
+ else:
+ self.re_embed = n_e
+
+ self.sane_index_shape = sane_index_shape
+
+ def remap_to_used(self, inds: torch.LongTensor) -> torch.LongTensor:
+ ishape = inds.shape
+ assert len(ishape) > 1
+ inds = inds.reshape(ishape[0], -1)
+ used = self.used.to(inds)
+ match = (inds[:, :, None] == used[None, None, ...]).long()
+ new = match.argmax(-1)
+ unknown = match.sum(2) < 1
+ if self.unknown_index == "random":
+ new[unknown] = torch.randint(0, self.re_embed, size=new[unknown].shape).to(device=new.device)
+ else:
+ new[unknown] = self.unknown_index
+ return new.reshape(ishape)
+
+ def unmap_to_all(self, inds: torch.LongTensor) -> torch.LongTensor:
+ ishape = inds.shape
+ assert len(ishape) > 1
+ inds = inds.reshape(ishape[0], -1)
+ used = self.used.to(inds)
+ if self.re_embed > self.used.shape[0]: # extra token
+ inds[inds >= self.used.shape[0]] = 0 # simply set to zero
+ back = torch.gather(used[None, :][inds.shape[0] * [0], :], 1, inds)
+ return back.reshape(ishape)
+
+ def forward(self, z: torch.FloatTensor) -> Tuple[torch.FloatTensor, torch.FloatTensor, Tuple]:
+ # reshape z -> (batch, height, width, channel) and flatten
+ z = z.permute(0, 2, 3, 1).contiguous()
+ z_flattened = z.view(-1, self.vq_embed_dim)
+
+ # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
+ min_encoding_indices = torch.argmin(torch.cdist(z_flattened, self.embedding.weight), dim=1)
+
+ z_q = self.embedding(min_encoding_indices).view(z.shape)
+ perplexity = None
+ min_encodings = None
+
+ # compute loss for embedding
+ if not self.legacy:
+ loss = self.beta * torch.mean((z_q.detach() - z) ** 2) + torch.mean((z_q - z.detach()) ** 2)
+ else:
+ loss = torch.mean((z_q.detach() - z) ** 2) + self.beta * torch.mean((z_q - z.detach()) ** 2)
+
+ # preserve gradients
+ z_q: torch.FloatTensor = z + (z_q - z).detach()
+
+ # reshape back to match original input shape
+ z_q = z_q.permute(0, 3, 1, 2).contiguous()
+
+ if self.remap is not None:
+ min_encoding_indices = min_encoding_indices.reshape(z.shape[0], -1) # add batch axis
+ min_encoding_indices = self.remap_to_used(min_encoding_indices)
+ min_encoding_indices = min_encoding_indices.reshape(-1, 1) # flatten
+
+ if self.sane_index_shape:
+ min_encoding_indices = min_encoding_indices.reshape(z_q.shape[0], z_q.shape[2], z_q.shape[3])
+
+ return z_q, loss, (perplexity, min_encodings, min_encoding_indices)
+
+ def get_codebook_entry(self, indices: torch.LongTensor, shape: Tuple[int, ...]) -> torch.FloatTensor:
+ # shape specifying (batch, height, width, channel)
+ if self.remap is not None:
+ indices = indices.reshape(shape[0], -1) # add batch axis
+ indices = self.unmap_to_all(indices)
+ indices = indices.reshape(-1) # flatten again
+
+ # get quantized latent vectors
+ z_q: torch.FloatTensor = self.embedding(indices)
+
+ if shape is not None:
+ z_q = z_q.view(shape)
+ # reshape back to match original input shape
+ z_q = z_q.permute(0, 3, 1, 2).contiguous()
+
+ return z_q
+
+
+class DiagonalGaussianDistribution(object):
+ def __init__(self, parameters: torch.Tensor, deterministic: bool = False):
+ self.parameters = parameters
+ self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+ self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+ self.deterministic = deterministic
+ self.std = torch.exp(0.5 * self.logvar)
+ self.var = torch.exp(self.logvar)
+ if self.deterministic:
+ self.var = self.std = torch.zeros_like(
+ self.mean, device=self.parameters.device, dtype=self.parameters.dtype
+ )
+
+ def sample(self, generator: Optional[torch.Generator] = None) -> torch.FloatTensor:
+ # make sure sample is on the same device as the parameters and has same dtype
+ sample = randn_tensor(
+ self.mean.shape,
+ generator=generator,
+ device=self.parameters.device,
+ dtype=self.parameters.dtype,
+ )
+ x = self.mean + self.std * sample
+ return x
+
+ def kl(self, other: "DiagonalGaussianDistribution" = None) -> torch.Tensor:
+ if self.deterministic:
+ return torch.Tensor([0.0])
+ else:
+ if other is None:
+ return 0.5 * torch.sum(
+ torch.pow(self.mean, 2) + self.var - 1.0 - self.logvar,
+ dim=[1, 2, 3],
+ )
+ else:
+ return 0.5 * torch.sum(
+ torch.pow(self.mean - other.mean, 2) / other.var
+ + self.var / other.var
+ - 1.0
+ - self.logvar
+ + other.logvar,
+ dim=[1, 2, 3],
+ )
+
+ def nll(self, sample: torch.Tensor, dims: Tuple[int, ...] = [1, 2, 3]) -> torch.Tensor:
+ if self.deterministic:
+ return torch.Tensor([0.0])
+ logtwopi = np.log(2.0 * np.pi)
+ return 0.5 * torch.sum(
+ logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+ dim=dims,
+ )
+
+ def mode(self) -> torch.Tensor:
+ return self.mean
+
+
+class EncoderTiny(nn.Module):
+ r"""
+ The `EncoderTiny` layer is a simpler version of the `Encoder` layer.
+
+ Args:
+ in_channels (`int`):
+ The number of input channels.
+ out_channels (`int`):
+ The number of output channels.
+ num_blocks (`Tuple[int, ...]`):
+ Each value of the tuple represents a Conv2d layer followed by `value` number of `AutoencoderTinyBlock`'s to
+ use.
+ block_out_channels (`Tuple[int, ...]`):
+ The number of output channels for each block.
+ act_fn (`str`):
+ The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ num_blocks: Tuple[int, ...],
+ block_out_channels: Tuple[int, ...],
+ act_fn: str,
+ ):
+ super().__init__()
+
+ layers = []
+ for i, num_block in enumerate(num_blocks):
+ num_channels = block_out_channels[i]
+
+ if i == 0:
+ layers.append(nn.Conv2d(in_channels, num_channels, kernel_size=3, padding=1))
+ else:
+ layers.append(
+ nn.Conv2d(
+ num_channels,
+ num_channels,
+ kernel_size=3,
+ padding=1,
+ stride=2,
+ bias=False,
+ )
+ )
+
+ for _ in range(num_block):
+ layers.append(AutoencoderTinyBlock(num_channels, num_channels, act_fn))
+
+ layers.append(nn.Conv2d(block_out_channels[-1], out_channels, kernel_size=3, padding=1))
+
+ self.layers = nn.Sequential(*layers)
+ self.gradient_checkpointing = False
+
+ def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ r"""The forward method of the `EncoderTiny` class."""
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x, use_reentrant=False)
+ else:
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x)
+
+ else:
+ # scale image from [-1, 1] to [0, 1] to match TAESD convention
+ x = self.layers(x.add(1).div(2))
+
+ return x
+
+
+class DecoderTiny(nn.Module):
+ r"""
+ The `DecoderTiny` layer is a simpler version of the `Decoder` layer.
+
+ Args:
+ in_channels (`int`):
+ The number of input channels.
+ out_channels (`int`):
+ The number of output channels.
+ num_blocks (`Tuple[int, ...]`):
+ Each value of the tuple represents a Conv2d layer followed by `value` number of `AutoencoderTinyBlock`'s to
+ use.
+ block_out_channels (`Tuple[int, ...]`):
+ The number of output channels for each block.
+ upsampling_scaling_factor (`int`):
+ The scaling factor to use for upsampling.
+ act_fn (`str`):
+ The activation function to use. See `~diffusers.models.activations.get_activation` for available options.
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ num_blocks: Tuple[int, ...],
+ block_out_channels: Tuple[int, ...],
+ upsampling_scaling_factor: int,
+ act_fn: str,
+ ):
+ super().__init__()
+
+ layers = [
+ nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=1),
+ get_activation(act_fn),
+ ]
+
+ for i, num_block in enumerate(num_blocks):
+ is_final_block = i == (len(num_blocks) - 1)
+ num_channels = block_out_channels[i]
+
+ for _ in range(num_block):
+ layers.append(AutoencoderTinyBlock(num_channels, num_channels, act_fn))
+
+ if not is_final_block:
+ layers.append(nn.Upsample(scale_factor=upsampling_scaling_factor))
+
+ conv_out_channel = num_channels if not is_final_block else out_channels
+ layers.append(
+ nn.Conv2d(
+ num_channels,
+ conv_out_channel,
+ kernel_size=3,
+ padding=1,
+ bias=is_final_block,
+ )
+ )
+
+ self.layers = nn.Sequential(*layers)
+ self.gradient_checkpointing = False
+
+ def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ r"""The forward method of the `DecoderTiny` class."""
+ # Clamp.
+ x = torch.tanh(x / 3) * 3
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x, use_reentrant=False)
+ else:
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(self.layers), x)
+
+ else:
+ x = self.layers(x)
+
+ # scale image from [0, 1] to [-1, 1] to match diffusers convention
+ return x.mul(2).sub(1)
diff --git a/model/diffusers_c/models/brushnet.py b/model/diffusers_c/models/brushnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..7b4a745e87f5a77b84f6167e5d12061603a75b33
--- /dev/null
+++ b/model/diffusers_c/models/brushnet.py
@@ -0,0 +1,959 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from .attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from .embeddings import TextImageProjection, TextImageTimeEmbedding, TextTimeEmbedding, TimestepEmbedding, Timesteps
+from .modeling_utils import ModelMixin
+from .unets.unet_2d_blocks import (
+ CrossAttnDownBlock2D,
+ DownBlock2D,
+ get_down_block,
+ get_mid_block,
+ get_up_block,
+)
+from .unets.unet_2d_condition import UNet2DConditionModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class BrushNetOutput(BaseOutput):
+ """
+ The output of [`BrushNetModel`].
+
+ Args:
+ up_block_res_samples (`tuple[torch.Tensor]`):
+ A tuple of upsample activations at different resolutions for each upsampling block. Each tensor should
+ be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+ used to condition the original UNet's upsampling activations.
+ down_block_res_samples (`tuple[torch.Tensor]`):
+ A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
+ be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+ used to condition the original UNet's downsampling activations.
+ mid_down_block_re_sample (`torch.Tensor`):
+ The activation of the midde block (the lowest sample resolution). Each tensor should be of shape
+ `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
+ Output can be used to condition the original UNet's middle block activation.
+ """
+
+ up_block_res_samples: Tuple[torch.Tensor]
+ down_block_res_samples: Tuple[torch.Tensor]
+ mid_block_res_sample: torch.Tensor
+
+
+class BrushNetModel(ModelMixin, ConfigMixin):
+ """
+ A BrushNet model.
+
+ Args:
+ in_channels (`int`, defaults to 4):
+ The number of channels in the input sample.
+ flip_sin_to_cos (`bool`, defaults to `True`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, defaults to 0):
+ The frequency shift to apply to the time embedding.
+ down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+ The tuple of downsample blocks to use.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+ Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+ `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+ The tuple of upsample blocks to use.
+ only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
+ block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, defaults to 2):
+ The number of layers per block.
+ downsample_padding (`int`, defaults to 1):
+ The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, defaults to 1):
+ The scale factor to use for the mid block.
+ act_fn (`str`, defaults to "silu"):
+ The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use for the normalization. If None, normalization and activation layers is skipped
+ in post-processing.
+ norm_eps (`float`, defaults to 1e-5):
+ The epsilon to use for the normalization.
+ cross_attention_dim (`int`, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ encoder_hid_dim (`int`, *optional*, defaults to None):
+ If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+ dimension to `cross_attention_dim`.
+ encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+ If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+ embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+ attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
+ The dimension of the attention heads.
+ use_linear_projection (`bool`, defaults to `False`):
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None,
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+ addition_embed_type (`str`, *optional*, defaults to `None`):
+ Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+ "text". "text" will use the `TextTimeEmbedding` layer.
+ num_class_embeds (`int`, *optional*, defaults to 0):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+ class conditioning with `class_embed_type` equal to `None`.
+ upcast_attention (`bool`, defaults to `False`):
+ resnet_time_scale_shift (`str`, defaults to `"default"`):
+ Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
+ projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
+ The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
+ `class_embed_type="projection"`.
+ brushnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+ The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+ conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
+ The tuple of output channel for each block in the `conditioning_embedding` layer.
+ global_pool_conditions (`bool`, defaults to `False`):
+ TODO(Patrick) - unused parameter.
+ addition_embed_type_num_heads (`int`, defaults to 64):
+ The number of heads to use for the `TextTimeEmbedding` layer.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 4,
+ conditioning_channels: int = 5,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str, ...] = (
+ # "CrossAttnDownBlock2D",
+ # "DownBlock2D",
+ # "DownBlock2D",
+ # "DownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ),
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn", # "UNetMidBlock2D",
+ up_block_types: Tuple[str, ...] = (
+ # "UpBlock2D",
+ # "UpBlock2D",
+ # "UpBlock2D",
+ # "CrossAttnUpBlock2D",
+ "UpBlock2D",
+ "CrossAttnUpBlock2D",
+ "CrossAttnUpBlock2D",
+ "CrossAttnUpBlock2D",
+ ),
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+ layers_per_block: int = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: int = 1280,
+ transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+ encoder_hid_dim: Optional[int] = None,
+ encoder_hid_dim_type: Optional[str] = None,
+ attention_head_dim: Union[int, Tuple[int, ...]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
+ use_linear_projection: bool = False,
+ class_embed_type: Optional[str] = None,
+ addition_embed_type: Optional[str] = None,
+ addition_time_embed_dim: Optional[int] = None,
+ num_class_embeds: Optional[int] = None,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ projection_class_embeddings_input_dim: Optional[int] = None,
+ brushnet_conditioning_channel_order: str = "rgb",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ global_pool_conditions: bool = False,
+ addition_embed_type_num_heads: int = 64,
+ ):
+ super().__init__()
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+ # input
+ conv_in_kernel = 3
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in_condition = nn.Conv2d(
+ in_channels + conditioning_channels,
+ block_out_channels[0],
+ kernel_size=conv_in_kernel,
+ padding=conv_in_padding,
+ )
+
+ # time
+ time_embed_dim = block_out_channels[0] * 4
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+ timestep_input_dim = block_out_channels[0]
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ )
+
+ if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+ encoder_hid_dim_type = "text_proj"
+ self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+ logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+ if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+ )
+
+ if encoder_hid_dim_type == "text_proj":
+ self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+ elif encoder_hid_dim_type == "text_image_proj":
+ # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+ self.encoder_hid_proj = TextImageProjection(
+ text_embed_dim=encoder_hid_dim,
+ image_embed_dim=cross_attention_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+
+ elif encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+ )
+ else:
+ self.encoder_hid_proj = None
+
+ # class embedding
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ elif class_embed_type == "projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+ # 2. it projects from an arbitrary input dimension.
+ #
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+ self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ else:
+ self.class_embedding = None
+
+ if addition_embed_type == "text":
+ if encoder_hid_dim is not None:
+ text_time_embedding_from_dim = encoder_hid_dim
+ else:
+ text_time_embedding_from_dim = cross_attention_dim
+
+ self.add_embedding = TextTimeEmbedding(
+ text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+ )
+ elif addition_embed_type == "text_image":
+ # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+ self.add_embedding = TextImageTimeEmbedding(
+ text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+ )
+ elif addition_embed_type == "text_time":
+ self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+ self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+ elif addition_embed_type is not None:
+ raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+ self.down_blocks = nn.ModuleList([])
+ self.brushnet_down_blocks = nn.ModuleList([])
+
+ if isinstance(only_cross_attention, bool):
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+ if isinstance(attention_head_dim, int):
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ # down
+ output_channel = block_out_channels[0]
+
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[i],
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ downsample_padding=downsample_padding,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ self.down_blocks.append(down_block)
+
+ for _ in range(layers_per_block):
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ if not is_final_block:
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_down_blocks.append(brushnet_block)
+
+ # mid
+ mid_block_channel = block_out_channels[-1]
+
+ brushnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_mid_block = brushnet_block
+
+ self.mid_block = get_mid_block(
+ mid_block_type,
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ in_channels=mid_block_channel,
+ temb_channels=time_embed_dim,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+ only_cross_attention = list(reversed(only_cross_attention))
+
+ output_channel = reversed_block_out_channels[0]
+
+ self.up_blocks = nn.ModuleList([])
+ self.brushnet_up_blocks = nn.ModuleList([])
+
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=layers_per_block + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resolution_idx=i,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=reversed_num_attention_heads[i],
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ for _ in range(layers_per_block + 1):
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_up_blocks.append(brushnet_block)
+
+ if not is_final_block:
+ brushnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ brushnet_block = zero_module(brushnet_block)
+ self.brushnet_up_blocks.append(brushnet_block)
+
+ @classmethod
+ def from_unet(
+ cls,
+ unet: UNet2DConditionModel,
+ brushnet_conditioning_channel_order: str = "rgb",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ load_weights_from_unet: bool = True,
+ conditioning_channels: int = 5,
+ ):
+ r"""
+ Instantiate a [`BrushNetModel`] from [`UNet2DConditionModel`].
+
+ Parameters:
+ unet (`UNet2DConditionModel`):
+ The UNet model weights to copy to the [`BrushNetModel`]. All configuration options are also copied
+ where applicable.
+ """
+ transformer_layers_per_block = (
+ unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
+ )
+ encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
+ encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None
+ addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
+ addition_time_embed_dim = (
+ unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
+ )
+
+ brushnet = cls(
+ in_channels=unet.config.in_channels,
+ conditioning_channels=conditioning_channels,
+ flip_sin_to_cos=unet.config.flip_sin_to_cos,
+ freq_shift=unet.config.freq_shift,
+ # down_block_types=['CrossAttnDownBlock2D','DownBlock2D','DownBlock2D','DownBlock2D'],
+ down_block_types=[
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ],
+ mid_block_type="UNetMidBlock2DCrossAttn",
+ # mid_block_type="UNetMidBlock2DCrossAttn",
+ # up_block_types=['UpBlock2D','UpBlock2D','UpBlock2D','CrossAttnUpBlock2D'],
+ up_block_types=["UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"],
+ only_cross_attention=unet.config.only_cross_attention,
+ block_out_channels=unet.config.block_out_channels,
+ layers_per_block=unet.config.layers_per_block,
+ downsample_padding=unet.config.downsample_padding,
+ mid_block_scale_factor=unet.config.mid_block_scale_factor,
+ act_fn=unet.config.act_fn,
+ norm_num_groups=unet.config.norm_num_groups,
+ norm_eps=unet.config.norm_eps,
+ cross_attention_dim=unet.config.cross_attention_dim,
+ transformer_layers_per_block=transformer_layers_per_block,
+ encoder_hid_dim=encoder_hid_dim,
+ encoder_hid_dim_type=encoder_hid_dim_type,
+ attention_head_dim=unet.config.attention_head_dim,
+ num_attention_heads=unet.config.num_attention_heads,
+ use_linear_projection=unet.config.use_linear_projection,
+ class_embed_type=unet.config.class_embed_type,
+ addition_embed_type=addition_embed_type,
+ addition_time_embed_dim=addition_time_embed_dim,
+ num_class_embeds=unet.config.num_class_embeds,
+ upcast_attention=unet.config.upcast_attention,
+ resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+ projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
+ brushnet_conditioning_channel_order=brushnet_conditioning_channel_order,
+ conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+ )
+
+ if load_weights_from_unet:
+ conv_in_condition_weight = torch.zeros_like(brushnet.conv_in_condition.weight)
+ conv_in_condition_weight[:, :4, ...] = unet.conv_in.weight
+ conv_in_condition_weight[:, 4:8, ...] = unet.conv_in.weight
+ brushnet.conv_in_condition.weight = torch.nn.Parameter(conv_in_condition_weight)
+ brushnet.conv_in_condition.bias = unet.conv_in.bias
+
+ brushnet.time_proj.load_state_dict(unet.time_proj.state_dict())
+ brushnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+ if brushnet.class_embedding:
+ brushnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
+
+ brushnet.down_blocks.load_state_dict(unet.down_blocks.state_dict(), strict=False)
+ brushnet.mid_block.load_state_dict(unet.mid_block.state_dict(), strict=False)
+ brushnet.up_blocks.load_state_dict(unet.up_blocks.state_dict(), strict=False)
+
+ return brushnet
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+ def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+ module.gradient_checkpointing = value
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ brushnet_cond: torch.FloatTensor,
+ conditioning_scale: float = 1.0,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guess_mode: bool = False,
+ return_dict: bool = True,
+ ) -> Union[BrushNetOutput, Tuple[Tuple[torch.FloatTensor, ...], torch.FloatTensor]]:
+ """
+ The [`BrushNetModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor.
+ timestep (`Union[torch.Tensor, float, int]`):
+ The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.Tensor`):
+ The encoder hidden states.
+ brushnet_cond (`torch.FloatTensor`):
+ The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+ conditioning_scale (`float`, defaults to `1.0`):
+ The scale factor for BrushNet outputs.
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+ Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+ timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+ embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ added_cond_kwargs (`dict`):
+ Additional conditions for the Stable Diffusion XL UNet.
+ cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+ A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+ guess_mode (`bool`, defaults to `False`):
+ In this mode, the BrushNet encoder tries its best to recognize the input content of the input even if
+ you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
+ return_dict (`bool`, defaults to `True`):
+ Whether or not to return a [`~models.brushnet.BrushNetOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.brushnet.BrushNetOutput`] **or** `tuple`:
+ If `return_dict` is `True`, a [`~models.brushnet.BrushNetOutput`] is returned, otherwise a tuple is
+ returned where the first element is the sample tensor.
+ """
+ # check channel order
+ channel_order = self.config.brushnet_conditioning_channel_order
+
+ if channel_order == "rgb":
+ # in rgb order by default
+ ...
+ elif channel_order == "bgr":
+ brushnet_cond = torch.flip(brushnet_cond, dims=[1])
+ else:
+ raise ValueError(f"unknown `brushnet_conditioning_channel_order`: {channel_order}")
+
+ # prepare attention_mask
+ if attention_mask is not None:
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+
+ emb = self.time_embedding(t_emb, timestep_cond)
+ aug_emb = None
+
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+ emb = emb + class_emb
+
+ if self.config.addition_embed_type is not None:
+ if self.config.addition_embed_type == "text":
+ aug_emb = self.add_embedding(encoder_hidden_states)
+
+ elif self.config.addition_embed_type == "text_time":
+ if "text_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+ )
+ text_embeds = added_cond_kwargs.get("text_embeds")
+ if "time_ids" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+ )
+ time_ids = added_cond_kwargs.get("time_ids")
+ time_embeds = self.add_time_proj(time_ids.flatten())
+ time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+ add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+ add_embeds = add_embeds.to(emb.dtype)
+ aug_emb = self.add_embedding(add_embeds)
+
+ emb = emb + aug_emb if aug_emb is not None else emb
+
+ # 2. pre-process
+ brushnet_cond = torch.concat([sample, brushnet_cond], 1)
+ sample = self.conv_in_condition(brushnet_cond)
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+ down_block_res_samples += res_samples
+
+ # 4. PaintingNet down blocks
+ brushnet_down_block_res_samples = ()
+ for down_block_res_sample, brushnet_down_block in zip(down_block_res_samples, self.brushnet_down_blocks):
+ down_block_res_sample = brushnet_down_block(down_block_res_sample)
+ brushnet_down_block_res_samples = brushnet_down_block_res_samples + (down_block_res_sample,)
+
+ # 5. mid
+ if self.mid_block is not None:
+ if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = self.mid_block(sample, emb)
+
+ # 6. BrushNet mid blocks
+ brushnet_mid_block_res_sample = self.brushnet_mid_block(sample)
+
+ # 7. up
+ up_block_res_samples = ()
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample, up_res_samples = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ return_res_samples=True,
+ )
+ else:
+ sample, up_res_samples = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ return_res_samples=True,
+ )
+
+ up_block_res_samples += up_res_samples
+
+ # 8. BrushNet up blocks
+ brushnet_up_block_res_samples = ()
+ for up_block_res_sample, brushnet_up_block in zip(up_block_res_samples, self.brushnet_up_blocks):
+ up_block_res_sample = brushnet_up_block(up_block_res_sample)
+ brushnet_up_block_res_samples = brushnet_up_block_res_samples + (up_block_res_sample,)
+
+ # 6. scaling
+ if guess_mode and not self.config.global_pool_conditions:
+ scales = torch.logspace(
+ -1,
+ 0,
+ len(brushnet_down_block_res_samples) + 1 + len(brushnet_up_block_res_samples),
+ device=sample.device,
+ ) # 0.1 to 1.0
+ scales = scales * conditioning_scale
+
+ brushnet_down_block_res_samples = [
+ sample * scale
+ for sample, scale in zip(
+ brushnet_down_block_res_samples, scales[: len(brushnet_down_block_res_samples)]
+ )
+ ]
+ brushnet_mid_block_res_sample = (
+ brushnet_mid_block_res_sample * scales[len(brushnet_down_block_res_samples)]
+ )
+ brushnet_up_block_res_samples = [
+ sample * scale
+ for sample, scale in zip(
+ brushnet_up_block_res_samples, scales[len(brushnet_down_block_res_samples) + 1 :]
+ )
+ ]
+ else:
+ brushnet_down_block_res_samples = [
+ sample * conditioning_scale for sample in brushnet_down_block_res_samples
+ ]
+ brushnet_mid_block_res_sample = brushnet_mid_block_res_sample * conditioning_scale
+ brushnet_up_block_res_samples = [sample * conditioning_scale for sample in brushnet_up_block_res_samples]
+
+ if self.config.global_pool_conditions:
+ brushnet_down_block_res_samples = [
+ torch.mean(sample, dim=(2, 3), keepdim=True) for sample in brushnet_down_block_res_samples
+ ]
+ brushnet_mid_block_res_sample = torch.mean(brushnet_mid_block_res_sample, dim=(2, 3), keepdim=True)
+ brushnet_up_block_res_samples = [
+ torch.mean(sample, dim=(2, 3), keepdim=True) for sample in brushnet_up_block_res_samples
+ ]
+
+ if not return_dict:
+ return (brushnet_down_block_res_samples, brushnet_mid_block_res_sample, brushnet_up_block_res_samples)
+
+ return BrushNetOutput(
+ down_block_res_samples=brushnet_down_block_res_samples,
+ mid_block_res_sample=brushnet_mid_block_res_sample,
+ up_block_res_samples=brushnet_up_block_res_samples,
+ )
+
+
+def zero_module(module):
+ for p in module.parameters():
+ nn.init.zeros_(p)
+ return module
diff --git a/model/diffusers_c/models/controlnet.py b/model/diffusers_c/models/controlnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..130e6430dca6ea5b77e3d8e40ed392672237f974
--- /dev/null
+++ b/model/diffusers_c/models/controlnet.py
@@ -0,0 +1,868 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..loaders import FromOriginalControlNetMixin
+from ..utils import BaseOutput, logging
+from .attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from .embeddings import TextImageProjection, TextImageTimeEmbedding, TextTimeEmbedding, TimestepEmbedding, Timesteps
+from .modeling_utils import ModelMixin
+from .unets.unet_2d_blocks import (
+ CrossAttnDownBlock2D,
+ DownBlock2D,
+ UNetMidBlock2D,
+ UNetMidBlock2DCrossAttn,
+ get_down_block,
+)
+from .unets.unet_2d_condition import UNet2DConditionModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class ControlNetOutput(BaseOutput):
+ """
+ The output of [`ControlNetModel`].
+
+ Args:
+ down_block_res_samples (`tuple[torch.Tensor]`):
+ A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
+ be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+ used to condition the original UNet's downsampling activations.
+ mid_down_block_re_sample (`torch.Tensor`):
+ The activation of the midde block (the lowest sample resolution). Each tensor should be of shape
+ `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
+ Output can be used to condition the original UNet's middle block activation.
+ """
+
+ down_block_res_samples: Tuple[torch.Tensor]
+ mid_block_res_sample: torch.Tensor
+
+
+class ControlNetConditioningEmbedding(nn.Module):
+ """
+ Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
+ [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
+ training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
+ convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
+ (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
+ model) to encode image-space conditions ... into feature maps ..."
+ """
+
+ def __init__(
+ self,
+ conditioning_embedding_channels: int,
+ conditioning_channels: int = 3,
+ block_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
+ ):
+ super().__init__()
+
+ self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+ self.blocks = nn.ModuleList([])
+
+ for i in range(len(block_out_channels) - 1):
+ channel_in = block_out_channels[i]
+ channel_out = block_out_channels[i + 1]
+ self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
+ self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
+
+ self.conv_out = zero_module(
+ nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
+ )
+
+ def forward(self, conditioning):
+ embedding = self.conv_in(conditioning)
+ embedding = F.silu(embedding)
+
+ for block in self.blocks:
+ embedding = block(embedding)
+ embedding = F.silu(embedding)
+
+ embedding = self.conv_out(embedding)
+
+ return embedding
+
+
+class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlNetMixin):
+ """
+ A ControlNet model.
+
+ Args:
+ in_channels (`int`, defaults to 4):
+ The number of channels in the input sample.
+ flip_sin_to_cos (`bool`, defaults to `True`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, defaults to 0):
+ The frequency shift to apply to the time embedding.
+ down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+ The tuple of downsample blocks to use.
+ only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
+ block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, defaults to 2):
+ The number of layers per block.
+ downsample_padding (`int`, defaults to 1):
+ The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, defaults to 1):
+ The scale factor to use for the mid block.
+ act_fn (`str`, defaults to "silu"):
+ The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use for the normalization. If None, normalization and activation layers is skipped
+ in post-processing.
+ norm_eps (`float`, defaults to 1e-5):
+ The epsilon to use for the normalization.
+ cross_attention_dim (`int`, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ encoder_hid_dim (`int`, *optional*, defaults to None):
+ If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+ dimension to `cross_attention_dim`.
+ encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+ If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+ embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+ attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
+ The dimension of the attention heads.
+ use_linear_projection (`bool`, defaults to `False`):
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None,
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+ addition_embed_type (`str`, *optional*, defaults to `None`):
+ Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+ "text". "text" will use the `TextTimeEmbedding` layer.
+ num_class_embeds (`int`, *optional*, defaults to 0):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+ class conditioning with `class_embed_type` equal to `None`.
+ upcast_attention (`bool`, defaults to `False`):
+ resnet_time_scale_shift (`str`, defaults to `"default"`):
+ Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
+ projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
+ The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
+ `class_embed_type="projection"`.
+ controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+ The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+ conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
+ The tuple of output channel for each block in the `conditioning_embedding` layer.
+ global_pool_conditions (`bool`, defaults to `False`):
+ TODO(Patrick) - unused parameter.
+ addition_embed_type_num_heads (`int`, defaults to 64):
+ The number of heads to use for the `TextTimeEmbedding` layer.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 4,
+ conditioning_channels: int = 3,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ),
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+ layers_per_block: int = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: int = 1280,
+ transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+ encoder_hid_dim: Optional[int] = None,
+ encoder_hid_dim_type: Optional[str] = None,
+ attention_head_dim: Union[int, Tuple[int, ...]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
+ use_linear_projection: bool = False,
+ class_embed_type: Optional[str] = None,
+ addition_embed_type: Optional[str] = None,
+ addition_time_embed_dim: Optional[int] = None,
+ num_class_embeds: Optional[int] = None,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ projection_class_embeddings_input_dim: Optional[int] = None,
+ controlnet_conditioning_channel_order: str = "rgb",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ global_pool_conditions: bool = False,
+ addition_embed_type_num_heads: int = 64,
+ ):
+ super().__init__()
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+ # input
+ conv_in_kernel = 3
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in = nn.Conv2d(
+ in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+ )
+
+ # time
+ time_embed_dim = block_out_channels[0] * 4
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+ timestep_input_dim = block_out_channels[0]
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ )
+
+ if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+ encoder_hid_dim_type = "text_proj"
+ self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+ logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+ if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+ )
+
+ if encoder_hid_dim_type == "text_proj":
+ self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+ elif encoder_hid_dim_type == "text_image_proj":
+ # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+ self.encoder_hid_proj = TextImageProjection(
+ text_embed_dim=encoder_hid_dim,
+ image_embed_dim=cross_attention_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+
+ elif encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+ )
+ else:
+ self.encoder_hid_proj = None
+
+ # class embedding
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ elif class_embed_type == "projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+ # 2. it projects from an arbitrary input dimension.
+ #
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+ self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ else:
+ self.class_embedding = None
+
+ if addition_embed_type == "text":
+ if encoder_hid_dim is not None:
+ text_time_embedding_from_dim = encoder_hid_dim
+ else:
+ text_time_embedding_from_dim = cross_attention_dim
+
+ self.add_embedding = TextTimeEmbedding(
+ text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+ )
+ elif addition_embed_type == "text_image":
+ # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+ self.add_embedding = TextImageTimeEmbedding(
+ text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+ )
+ elif addition_embed_type == "text_time":
+ self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+ self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+ elif addition_embed_type is not None:
+ raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+ # control net conditioning embedding
+ self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
+ conditioning_embedding_channels=block_out_channels[0],
+ block_out_channels=conditioning_embedding_out_channels,
+ conditioning_channels=conditioning_channels,
+ )
+
+ self.down_blocks = nn.ModuleList([])
+ self.controlnet_down_blocks = nn.ModuleList([])
+
+ if isinstance(only_cross_attention, bool):
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+ if isinstance(attention_head_dim, int):
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ # down
+ output_channel = block_out_channels[0]
+
+ controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ controlnet_block = zero_module(controlnet_block)
+ self.controlnet_down_blocks.append(controlnet_block)
+
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[i],
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ downsample_padding=downsample_padding,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ self.down_blocks.append(down_block)
+
+ for _ in range(layers_per_block):
+ controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ controlnet_block = zero_module(controlnet_block)
+ self.controlnet_down_blocks.append(controlnet_block)
+
+ if not is_final_block:
+ controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ controlnet_block = zero_module(controlnet_block)
+ self.controlnet_down_blocks.append(controlnet_block)
+
+ # mid
+ mid_block_channel = block_out_channels[-1]
+
+ controlnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1)
+ controlnet_block = zero_module(controlnet_block)
+ self.controlnet_mid_block = controlnet_block
+
+ if mid_block_type == "UNetMidBlock2DCrossAttn":
+ self.mid_block = UNetMidBlock2DCrossAttn(
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ in_channels=mid_block_channel,
+ temb_channels=time_embed_dim,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ )
+ elif mid_block_type == "UNetMidBlock2D":
+ self.mid_block = UNetMidBlock2D(
+ in_channels=block_out_channels[-1],
+ temb_channels=time_embed_dim,
+ num_layers=0,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_groups=norm_num_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ add_attention=False,
+ )
+ else:
+ raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+ @classmethod
+ def from_unet(
+ cls,
+ unet: UNet2DConditionModel,
+ controlnet_conditioning_channel_order: str = "rgb",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ load_weights_from_unet: bool = True,
+ conditioning_channels: int = 3,
+ ):
+ r"""
+ Instantiate a [`ControlNetModel`] from [`UNet2DConditionModel`].
+
+ Parameters:
+ unet (`UNet2DConditionModel`):
+ The UNet model weights to copy to the [`ControlNetModel`]. All configuration options are also copied
+ where applicable.
+ """
+ transformer_layers_per_block = (
+ unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
+ )
+ encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
+ encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None
+ addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
+ addition_time_embed_dim = (
+ unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
+ )
+
+ controlnet = cls(
+ encoder_hid_dim=encoder_hid_dim,
+ encoder_hid_dim_type=encoder_hid_dim_type,
+ addition_embed_type=addition_embed_type,
+ addition_time_embed_dim=addition_time_embed_dim,
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=unet.config.in_channels,
+ flip_sin_to_cos=unet.config.flip_sin_to_cos,
+ freq_shift=unet.config.freq_shift,
+ down_block_types=unet.config.down_block_types,
+ only_cross_attention=unet.config.only_cross_attention,
+ block_out_channels=unet.config.block_out_channels,
+ layers_per_block=unet.config.layers_per_block,
+ downsample_padding=unet.config.downsample_padding,
+ mid_block_scale_factor=unet.config.mid_block_scale_factor,
+ act_fn=unet.config.act_fn,
+ norm_num_groups=unet.config.norm_num_groups,
+ norm_eps=unet.config.norm_eps,
+ cross_attention_dim=unet.config.cross_attention_dim,
+ attention_head_dim=unet.config.attention_head_dim,
+ num_attention_heads=unet.config.num_attention_heads,
+ use_linear_projection=unet.config.use_linear_projection,
+ class_embed_type=unet.config.class_embed_type,
+ num_class_embeds=unet.config.num_class_embeds,
+ upcast_attention=unet.config.upcast_attention,
+ resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+ projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
+ mid_block_type=unet.config.mid_block_type,
+ controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
+ conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+ conditioning_channels=conditioning_channels,
+ )
+
+ if load_weights_from_unet:
+ controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())
+ controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
+ controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+ if controlnet.class_embedding:
+ controlnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
+
+ controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict())
+ controlnet.mid_block.load_state_dict(unet.mid_block.state_dict())
+
+ return controlnet
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+ def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+ module.gradient_checkpointing = value
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ controlnet_cond: torch.FloatTensor,
+ conditioning_scale: float = 1.0,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guess_mode: bool = False,
+ return_dict: bool = True,
+ ) -> Union[ControlNetOutput, Tuple[Tuple[torch.FloatTensor, ...], torch.FloatTensor]]:
+ """
+ The [`ControlNetModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor.
+ timestep (`Union[torch.Tensor, float, int]`):
+ The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.Tensor`):
+ The encoder hidden states.
+ controlnet_cond (`torch.FloatTensor`):
+ The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+ conditioning_scale (`float`, defaults to `1.0`):
+ The scale factor for ControlNet outputs.
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+ Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+ timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+ embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ added_cond_kwargs (`dict`):
+ Additional conditions for the Stable Diffusion XL UNet.
+ cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+ A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+ guess_mode (`bool`, defaults to `False`):
+ In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
+ you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
+ return_dict (`bool`, defaults to `True`):
+ Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
+ If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
+ returned where the first element is the sample tensor.
+ """
+ # check channel order
+ channel_order = self.config.controlnet_conditioning_channel_order
+
+ if channel_order == "rgb":
+ # in rgb order by default
+ ...
+ elif channel_order == "bgr":
+ controlnet_cond = torch.flip(controlnet_cond, dims=[1])
+ else:
+ raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
+
+ # prepare attention_mask
+ if attention_mask is not None:
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+
+ emb = self.time_embedding(t_emb, timestep_cond)
+ aug_emb = None
+
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+ emb = emb + class_emb
+
+ if self.config.addition_embed_type is not None:
+ if self.config.addition_embed_type == "text":
+ aug_emb = self.add_embedding(encoder_hidden_states)
+
+ elif self.config.addition_embed_type == "text_time":
+ if "text_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+ )
+ text_embeds = added_cond_kwargs.get("text_embeds")
+ if "time_ids" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+ )
+ time_ids = added_cond_kwargs.get("time_ids")
+ time_embeds = self.add_time_proj(time_ids.flatten())
+ time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+
+ add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+ add_embeds = add_embeds.to(emb.dtype)
+ aug_emb = self.add_embedding(add_embeds)
+
+ emb = emb + aug_emb if aug_emb is not None else emb
+
+ # 2. pre-process
+ sample = self.conv_in(sample)
+
+ controlnet_cond = self.controlnet_cond_embedding(controlnet_cond)
+ sample = sample + controlnet_cond
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+ down_block_res_samples += res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = self.mid_block(sample, emb)
+
+ # 5. Control net blocks
+
+ controlnet_down_block_res_samples = ()
+
+ for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
+ down_block_res_sample = controlnet_block(down_block_res_sample)
+ controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)
+
+ down_block_res_samples = controlnet_down_block_res_samples
+
+ mid_block_res_sample = self.controlnet_mid_block(sample)
+
+ # 6. scaling
+ if guess_mode and not self.config.global_pool_conditions:
+ scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device) # 0.1 to 1.0
+ scales = scales * conditioning_scale
+ down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
+ mid_block_res_sample = mid_block_res_sample * scales[-1] # last one
+ else:
+ down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
+ mid_block_res_sample = mid_block_res_sample * conditioning_scale
+
+ if self.config.global_pool_conditions:
+ down_block_res_samples = [
+ torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples
+ ]
+ mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True)
+
+ if not return_dict:
+ return (down_block_res_samples, mid_block_res_sample)
+
+ return ControlNetOutput(
+ down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
+ )
+
+
+def zero_module(module):
+ for p in module.parameters():
+ nn.init.zeros_(p)
+ return module
diff --git a/model/diffusers_c/models/controlnet_flax.py b/model/diffusers_c/models/controlnet_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..6f9b201aa1e3ff88630b967e6c0a6b39d3833c67
--- /dev/null
+++ b/model/diffusers_c/models/controlnet_flax.py
@@ -0,0 +1,395 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional, Tuple, Union
+
+import flax
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict
+
+from ..configuration_utils import ConfigMixin, flax_register_to_config
+from ..utils import BaseOutput
+from .embeddings_flax import FlaxTimestepEmbedding, FlaxTimesteps
+from .modeling_flax_utils import FlaxModelMixin
+from .unets.unet_2d_blocks_flax import (
+ FlaxCrossAttnDownBlock2D,
+ FlaxDownBlock2D,
+ FlaxUNetMidBlock2DCrossAttn,
+)
+
+
+@flax.struct.dataclass
+class FlaxControlNetOutput(BaseOutput):
+ """
+ The output of [`FlaxControlNetModel`].
+
+ Args:
+ down_block_res_samples (`jnp.ndarray`):
+ mid_block_res_sample (`jnp.ndarray`):
+ """
+
+ down_block_res_samples: jnp.ndarray
+ mid_block_res_sample: jnp.ndarray
+
+
+class FlaxControlNetConditioningEmbedding(nn.Module):
+ conditioning_embedding_channels: int
+ block_out_channels: Tuple[int, ...] = (16, 32, 96, 256)
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self) -> None:
+ self.conv_in = nn.Conv(
+ self.block_out_channels[0],
+ kernel_size=(3, 3),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ blocks = []
+ for i in range(len(self.block_out_channels) - 1):
+ channel_in = self.block_out_channels[i]
+ channel_out = self.block_out_channels[i + 1]
+ conv1 = nn.Conv(
+ channel_in,
+ kernel_size=(3, 3),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+ blocks.append(conv1)
+ conv2 = nn.Conv(
+ channel_out,
+ kernel_size=(3, 3),
+ strides=(2, 2),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+ blocks.append(conv2)
+ self.blocks = blocks
+
+ self.conv_out = nn.Conv(
+ self.conditioning_embedding_channels,
+ kernel_size=(3, 3),
+ padding=((1, 1), (1, 1)),
+ kernel_init=nn.initializers.zeros_init(),
+ bias_init=nn.initializers.zeros_init(),
+ dtype=self.dtype,
+ )
+
+ def __call__(self, conditioning: jnp.ndarray) -> jnp.ndarray:
+ embedding = self.conv_in(conditioning)
+ embedding = nn.silu(embedding)
+
+ for block in self.blocks:
+ embedding = block(embedding)
+ embedding = nn.silu(embedding)
+
+ embedding = self.conv_out(embedding)
+
+ return embedding
+
+
+@flax_register_to_config
+class FlaxControlNetModel(nn.Module, FlaxModelMixin, ConfigMixin):
+ r"""
+ A ControlNet model.
+
+ This model inherits from [`FlaxModelMixin`]. Check the superclass documentation for it’s generic methods
+ implemented for all models (such as downloading or saving).
+
+ This model is also a Flax Linen [`flax.linen.Module`](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
+ subclass. Use it as a regular Flax Linen module and refer to the Flax documentation for all matters related to its
+ general usage and behavior.
+
+ Inherent JAX features such as the following are supported:
+
+ - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+ - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+ - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+ - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+ Parameters:
+ sample_size (`int`, *optional*):
+ The size of the input sample.
+ in_channels (`int`, *optional*, defaults to 4):
+ The number of channels in the input sample.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("FlaxCrossAttnDownBlock2D", "FlaxCrossAttnDownBlock2D", "FlaxCrossAttnDownBlock2D", "FlaxDownBlock2D")`):
+ The tuple of downsample blocks to use.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2):
+ The number of layers per block.
+ attention_head_dim (`int` or `Tuple[int]`, *optional*, defaults to 8):
+ The dimension of the attention heads.
+ num_attention_heads (`int` or `Tuple[int]`, *optional*):
+ The number of attention heads.
+ cross_attention_dim (`int`, *optional*, defaults to 768):
+ The dimension of the cross attention features.
+ dropout (`float`, *optional*, defaults to 0):
+ Dropout probability for down, up and bottleneck blocks.
+ flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+ controlnet_conditioning_channel_order (`str`, *optional*, defaults to `rgb`):
+ The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+ conditioning_embedding_out_channels (`tuple`, *optional*, defaults to `(16, 32, 96, 256)`):
+ The tuple of output channel for each block in the `conditioning_embedding` layer.
+ """
+
+ sample_size: int = 32
+ in_channels: int = 4
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ )
+ only_cross_attention: Union[bool, Tuple[bool, ...]] = False
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280)
+ layers_per_block: int = 2
+ attention_head_dim: Union[int, Tuple[int, ...]] = 8
+ num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None
+ cross_attention_dim: int = 1280
+ dropout: float = 0.0
+ use_linear_projection: bool = False
+ dtype: jnp.dtype = jnp.float32
+ flip_sin_to_cos: bool = True
+ freq_shift: int = 0
+ controlnet_conditioning_channel_order: str = "rgb"
+ conditioning_embedding_out_channels: Tuple[int, ...] = (16, 32, 96, 256)
+
+ def init_weights(self, rng: jax.Array) -> FrozenDict:
+ # init input tensors
+ sample_shape = (1, self.in_channels, self.sample_size, self.sample_size)
+ sample = jnp.zeros(sample_shape, dtype=jnp.float32)
+ timesteps = jnp.ones((1,), dtype=jnp.int32)
+ encoder_hidden_states = jnp.zeros((1, 1, self.cross_attention_dim), dtype=jnp.float32)
+ controlnet_cond_shape = (1, 3, self.sample_size * 8, self.sample_size * 8)
+ controlnet_cond = jnp.zeros(controlnet_cond_shape, dtype=jnp.float32)
+
+ params_rng, dropout_rng = jax.random.split(rng)
+ rngs = {"params": params_rng, "dropout": dropout_rng}
+
+ return self.init(rngs, sample, timesteps, encoder_hidden_states, controlnet_cond)["params"]
+
+ def setup(self) -> None:
+ block_out_channels = self.block_out_channels
+ time_embed_dim = block_out_channels[0] * 4
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = self.num_attention_heads or self.attention_head_dim
+
+ # input
+ self.conv_in = nn.Conv(
+ block_out_channels[0],
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ # time
+ self.time_proj = FlaxTimesteps(
+ block_out_channels[0], flip_sin_to_cos=self.flip_sin_to_cos, freq_shift=self.config.freq_shift
+ )
+ self.time_embedding = FlaxTimestepEmbedding(time_embed_dim, dtype=self.dtype)
+
+ self.controlnet_cond_embedding = FlaxControlNetConditioningEmbedding(
+ conditioning_embedding_channels=block_out_channels[0],
+ block_out_channels=self.conditioning_embedding_out_channels,
+ )
+
+ only_cross_attention = self.only_cross_attention
+ if isinstance(only_cross_attention, bool):
+ only_cross_attention = (only_cross_attention,) * len(self.down_block_types)
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(self.down_block_types)
+
+ # down
+ down_blocks = []
+ controlnet_down_blocks = []
+
+ output_channel = block_out_channels[0]
+
+ controlnet_block = nn.Conv(
+ output_channel,
+ kernel_size=(1, 1),
+ padding="VALID",
+ kernel_init=nn.initializers.zeros_init(),
+ bias_init=nn.initializers.zeros_init(),
+ dtype=self.dtype,
+ )
+ controlnet_down_blocks.append(controlnet_block)
+
+ for i, down_block_type in enumerate(self.down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ if down_block_type == "CrossAttnDownBlock2D":
+ down_block = FlaxCrossAttnDownBlock2D(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ dropout=self.dropout,
+ num_layers=self.layers_per_block,
+ num_attention_heads=num_attention_heads[i],
+ add_downsample=not is_final_block,
+ use_linear_projection=self.use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ dtype=self.dtype,
+ )
+ else:
+ down_block = FlaxDownBlock2D(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ dropout=self.dropout,
+ num_layers=self.layers_per_block,
+ add_downsample=not is_final_block,
+ dtype=self.dtype,
+ )
+
+ down_blocks.append(down_block)
+
+ for _ in range(self.layers_per_block):
+ controlnet_block = nn.Conv(
+ output_channel,
+ kernel_size=(1, 1),
+ padding="VALID",
+ kernel_init=nn.initializers.zeros_init(),
+ bias_init=nn.initializers.zeros_init(),
+ dtype=self.dtype,
+ )
+ controlnet_down_blocks.append(controlnet_block)
+
+ if not is_final_block:
+ controlnet_block = nn.Conv(
+ output_channel,
+ kernel_size=(1, 1),
+ padding="VALID",
+ kernel_init=nn.initializers.zeros_init(),
+ bias_init=nn.initializers.zeros_init(),
+ dtype=self.dtype,
+ )
+ controlnet_down_blocks.append(controlnet_block)
+
+ self.down_blocks = down_blocks
+ self.controlnet_down_blocks = controlnet_down_blocks
+
+ # mid
+ mid_block_channel = block_out_channels[-1]
+ self.mid_block = FlaxUNetMidBlock2DCrossAttn(
+ in_channels=mid_block_channel,
+ dropout=self.dropout,
+ num_attention_heads=num_attention_heads[-1],
+ use_linear_projection=self.use_linear_projection,
+ dtype=self.dtype,
+ )
+
+ self.controlnet_mid_block = nn.Conv(
+ mid_block_channel,
+ kernel_size=(1, 1),
+ padding="VALID",
+ kernel_init=nn.initializers.zeros_init(),
+ bias_init=nn.initializers.zeros_init(),
+ dtype=self.dtype,
+ )
+
+ def __call__(
+ self,
+ sample: jnp.ndarray,
+ timesteps: Union[jnp.ndarray, float, int],
+ encoder_hidden_states: jnp.ndarray,
+ controlnet_cond: jnp.ndarray,
+ conditioning_scale: float = 1.0,
+ return_dict: bool = True,
+ train: bool = False,
+ ) -> Union[FlaxControlNetOutput, Tuple[Tuple[jnp.ndarray, ...], jnp.ndarray]]:
+ r"""
+ Args:
+ sample (`jnp.ndarray`): (batch, channel, height, width) noisy inputs tensor
+ timestep (`jnp.ndarray` or `float` or `int`): timesteps
+ encoder_hidden_states (`jnp.ndarray`): (batch_size, sequence_length, hidden_size) encoder hidden states
+ controlnet_cond (`jnp.ndarray`): (batch, channel, height, width) the conditional input tensor
+ conditioning_scale (`float`, *optional*, defaults to `1.0`): the scale factor for controlnet outputs
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`models.unets.unet_2d_condition_flax.FlaxUNet2DConditionOutput`] instead of a
+ plain tuple.
+ train (`bool`, *optional*, defaults to `False`):
+ Use deterministic functions and disable dropout when not training.
+
+ Returns:
+ [`~models.unets.unet_2d_condition_flax.FlaxUNet2DConditionOutput`] or `tuple`:
+ [`~models.unets.unet_2d_condition_flax.FlaxUNet2DConditionOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is the sample tensor.
+ """
+ channel_order = self.controlnet_conditioning_channel_order
+ if channel_order == "bgr":
+ controlnet_cond = jnp.flip(controlnet_cond, axis=1)
+
+ # 1. time
+ if not isinstance(timesteps, jnp.ndarray):
+ timesteps = jnp.array([timesteps], dtype=jnp.int32)
+ elif isinstance(timesteps, jnp.ndarray) and len(timesteps.shape) == 0:
+ timesteps = timesteps.astype(dtype=jnp.float32)
+ timesteps = jnp.expand_dims(timesteps, 0)
+
+ t_emb = self.time_proj(timesteps)
+ t_emb = self.time_embedding(t_emb)
+
+ # 2. pre-process
+ sample = jnp.transpose(sample, (0, 2, 3, 1))
+ sample = self.conv_in(sample)
+
+ controlnet_cond = jnp.transpose(controlnet_cond, (0, 2, 3, 1))
+ controlnet_cond = self.controlnet_cond_embedding(controlnet_cond)
+ sample += controlnet_cond
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for down_block in self.down_blocks:
+ if isinstance(down_block, FlaxCrossAttnDownBlock2D):
+ sample, res_samples = down_block(sample, t_emb, encoder_hidden_states, deterministic=not train)
+ else:
+ sample, res_samples = down_block(sample, t_emb, deterministic=not train)
+ down_block_res_samples += res_samples
+
+ # 4. mid
+ sample = self.mid_block(sample, t_emb, encoder_hidden_states, deterministic=not train)
+
+ # 5. contronet blocks
+ controlnet_down_block_res_samples = ()
+ for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
+ down_block_res_sample = controlnet_block(down_block_res_sample)
+ controlnet_down_block_res_samples += (down_block_res_sample,)
+
+ down_block_res_samples = controlnet_down_block_res_samples
+
+ mid_block_res_sample = self.controlnet_mid_block(sample)
+
+ # 6. scaling
+ down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
+ mid_block_res_sample *= conditioning_scale
+
+ if not return_dict:
+ return (down_block_res_samples, mid_block_res_sample)
+
+ return FlaxControlNetOutput(
+ down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
+ )
diff --git a/model/diffusers_c/models/downsampling.py b/model/diffusers_c/models/downsampling.py
new file mode 100644
index 0000000000000000000000000000000000000000..80fb065a6f4c8ed566f0523f4e168fa859b1f83c
--- /dev/null
+++ b/model/diffusers_c/models/downsampling.py
@@ -0,0 +1,338 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..utils import USE_PEFT_BACKEND
+from .lora import LoRACompatibleConv
+from .normalization import RMSNorm
+from .upsampling import upfirdn2d_native
+
+
+class Downsample1D(nn.Module):
+ """A 1D downsampling layer with an optional convolution.
+
+ Parameters:
+ channels (`int`):
+ number of channels in the inputs and outputs.
+ use_conv (`bool`, default `False`):
+ option to use a convolution.
+ out_channels (`int`, optional):
+ number of output channels. Defaults to `channels`.
+ padding (`int`, default `1`):
+ padding for the convolution.
+ name (`str`, default `conv`):
+ name of the downsampling 1D layer.
+ """
+
+ def __init__(
+ self,
+ channels: int,
+ use_conv: bool = False,
+ out_channels: Optional[int] = None,
+ padding: int = 1,
+ name: str = "conv",
+ ):
+ super().__init__()
+ self.channels = channels
+ self.out_channels = out_channels or channels
+ self.use_conv = use_conv
+ self.padding = padding
+ stride = 2
+ self.name = name
+
+ if use_conv:
+ self.conv = nn.Conv1d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
+ else:
+ assert self.channels == self.out_channels
+ self.conv = nn.AvgPool1d(kernel_size=stride, stride=stride)
+
+ def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+ assert inputs.shape[1] == self.channels
+ return self.conv(inputs)
+
+
+class Downsample2D(nn.Module):
+ """A 2D downsampling layer with an optional convolution.
+
+ Parameters:
+ channels (`int`):
+ number of channels in the inputs and outputs.
+ use_conv (`bool`, default `False`):
+ option to use a convolution.
+ out_channels (`int`, optional):
+ number of output channels. Defaults to `channels`.
+ padding (`int`, default `1`):
+ padding for the convolution.
+ name (`str`, default `conv`):
+ name of the downsampling 2D layer.
+ """
+
+ def __init__(
+ self,
+ channels: int,
+ use_conv: bool = False,
+ out_channels: Optional[int] = None,
+ padding: int = 1,
+ name: str = "conv",
+ kernel_size=3,
+ norm_type=None,
+ eps=None,
+ elementwise_affine=None,
+ bias=True,
+ ):
+ super().__init__()
+ self.channels = channels
+ self.out_channels = out_channels or channels
+ self.use_conv = use_conv
+ self.padding = padding
+ stride = 2
+ self.name = name
+ conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+
+ if norm_type == "ln_norm":
+ self.norm = nn.LayerNorm(channels, eps, elementwise_affine)
+ elif norm_type == "rms_norm":
+ self.norm = RMSNorm(channels, eps, elementwise_affine)
+ elif norm_type is None:
+ self.norm = None
+ else:
+ raise ValueError(f"unknown norm_type: {norm_type}")
+
+ if use_conv:
+ conv = conv_cls(
+ self.channels, self.out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=bias
+ )
+ else:
+ assert self.channels == self.out_channels
+ conv = nn.AvgPool2d(kernel_size=stride, stride=stride)
+
+ # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+ if name == "conv":
+ self.Conv2d_0 = conv
+ self.conv = conv
+ elif name == "Conv2d_0":
+ self.conv = conv
+ else:
+ self.conv = conv
+
+ def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
+ assert hidden_states.shape[1] == self.channels
+
+ if self.norm is not None:
+ hidden_states = self.norm(hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+
+ if self.use_conv and self.padding == 0:
+ pad = (0, 1, 0, 1)
+ hidden_states = F.pad(hidden_states, pad, mode="constant", value=0)
+
+ assert hidden_states.shape[1] == self.channels
+
+ if not USE_PEFT_BACKEND:
+ if isinstance(self.conv, LoRACompatibleConv):
+ hidden_states = self.conv(hidden_states, scale)
+ else:
+ hidden_states = self.conv(hidden_states)
+ else:
+ hidden_states = self.conv(hidden_states)
+
+ return hidden_states
+
+
+class FirDownsample2D(nn.Module):
+ """A 2D FIR downsampling layer with an optional convolution.
+
+ Parameters:
+ channels (`int`):
+ number of channels in the inputs and outputs.
+ use_conv (`bool`, default `False`):
+ option to use a convolution.
+ out_channels (`int`, optional):
+ number of output channels. Defaults to `channels`.
+ fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
+ kernel for the FIR filter.
+ """
+
+ def __init__(
+ self,
+ channels: Optional[int] = None,
+ out_channels: Optional[int] = None,
+ use_conv: bool = False,
+ fir_kernel: Tuple[int, int, int, int] = (1, 3, 3, 1),
+ ):
+ super().__init__()
+ out_channels = out_channels if out_channels else channels
+ if use_conv:
+ self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
+ self.fir_kernel = fir_kernel
+ self.use_conv = use_conv
+ self.out_channels = out_channels
+
+ def _downsample_2d(
+ self,
+ hidden_states: torch.FloatTensor,
+ weight: Optional[torch.FloatTensor] = None,
+ kernel: Optional[torch.FloatTensor] = None,
+ factor: int = 2,
+ gain: float = 1,
+ ) -> torch.FloatTensor:
+ """Fused `Conv2d()` followed by `downsample_2d()`.
+ Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
+ efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
+ arbitrary order.
+
+ Args:
+ hidden_states (`torch.FloatTensor`):
+ Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+ weight (`torch.FloatTensor`, *optional*):
+ Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
+ performed by `inChannels = x.shape[0] // numGroups`.
+ kernel (`torch.FloatTensor`, *optional*):
+ FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+ corresponds to average pooling.
+ factor (`int`, *optional*, default to `2`):
+ Integer downsampling factor.
+ gain (`float`, *optional*, default to `1.0`):
+ Scaling factor for signal magnitude.
+
+ Returns:
+ output (`torch.FloatTensor`):
+ Tensor of the shape `[N, C, H // factor, W // factor]` or `[N, H // factor, W // factor, C]`, and same
+ datatype as `x`.
+ """
+
+ assert isinstance(factor, int) and factor >= 1
+ if kernel is None:
+ kernel = [1] * factor
+
+ # setup kernel
+ kernel = torch.tensor(kernel, dtype=torch.float32)
+ if kernel.ndim == 1:
+ kernel = torch.outer(kernel, kernel)
+ kernel /= torch.sum(kernel)
+
+ kernel = kernel * gain
+
+ if self.use_conv:
+ _, _, convH, convW = weight.shape
+ pad_value = (kernel.shape[0] - factor) + (convW - 1)
+ stride_value = [factor, factor]
+ upfirdn_input = upfirdn2d_native(
+ hidden_states,
+ torch.tensor(kernel, device=hidden_states.device),
+ pad=((pad_value + 1) // 2, pad_value // 2),
+ )
+ output = F.conv2d(upfirdn_input, weight, stride=stride_value, padding=0)
+ else:
+ pad_value = kernel.shape[0] - factor
+ output = upfirdn2d_native(
+ hidden_states,
+ torch.tensor(kernel, device=hidden_states.device),
+ down=factor,
+ pad=((pad_value + 1) // 2, pad_value // 2),
+ )
+
+ return output
+
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+ if self.use_conv:
+ downsample_input = self._downsample_2d(hidden_states, weight=self.Conv2d_0.weight, kernel=self.fir_kernel)
+ hidden_states = downsample_input + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
+ else:
+ hidden_states = self._downsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
+
+ return hidden_states
+
+
+# downsample/upsample layer used in k-upscaler, might be able to use FirDownsample2D/DirUpsample2D instead
+class KDownsample2D(nn.Module):
+ r"""A 2D K-downsampling layer.
+
+ Parameters:
+ pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
+ """
+
+ def __init__(self, pad_mode: str = "reflect"):
+ super().__init__()
+ self.pad_mode = pad_mode
+ kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]])
+ self.pad = kernel_1d.shape[1] // 2 - 1
+ self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
+
+ def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+ inputs = F.pad(inputs, (self.pad,) * 4, self.pad_mode)
+ weight = inputs.new_zeros(
+ [
+ inputs.shape[1],
+ inputs.shape[1],
+ self.kernel.shape[0],
+ self.kernel.shape[1],
+ ]
+ )
+ indices = torch.arange(inputs.shape[1], device=inputs.device)
+ kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
+ weight[indices, indices] = kernel
+ return F.conv2d(inputs, weight, stride=2)
+
+
+def downsample_2d(
+ hidden_states: torch.FloatTensor,
+ kernel: Optional[torch.FloatTensor] = None,
+ factor: int = 2,
+ gain: float = 1,
+) -> torch.FloatTensor:
+ r"""Downsample2D a batch of 2D images with the given filter.
+ Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and downsamples each image with the
+ given filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the
+ specified `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its
+ shape is a multiple of the downsampling factor.
+
+ Args:
+ hidden_states (`torch.FloatTensor`)
+ Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+ kernel (`torch.FloatTensor`, *optional*):
+ FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+ corresponds to average pooling.
+ factor (`int`, *optional*, default to `2`):
+ Integer downsampling factor.
+ gain (`float`, *optional*, default to `1.0`):
+ Scaling factor for signal magnitude.
+
+ Returns:
+ output (`torch.FloatTensor`):
+ Tensor of the shape `[N, C, H // factor, W // factor]`
+ """
+
+ assert isinstance(factor, int) and factor >= 1
+ if kernel is None:
+ kernel = [1] * factor
+
+ kernel = torch.tensor(kernel, dtype=torch.float32)
+ if kernel.ndim == 1:
+ kernel = torch.outer(kernel, kernel)
+ kernel /= torch.sum(kernel)
+
+ kernel = kernel * gain
+ pad_value = kernel.shape[0] - factor
+ output = upfirdn2d_native(
+ hidden_states,
+ kernel.to(device=hidden_states.device),
+ down=factor,
+ pad=((pad_value + 1) // 2, pad_value // 2),
+ )
+ return output
diff --git a/model/diffusers_c/models/dual_transformer_2d.py b/model/diffusers_c/models/dual_transformer_2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..b8e40f14d5a84ea740d55bae5ef7ba5c58a0c9a8
--- /dev/null
+++ b/model/diffusers_c/models/dual_transformer_2d.py
@@ -0,0 +1,20 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from ..utils import deprecate
+from .transformers.dual_transformer_2d import DualTransformer2DModel
+
+
+class DualTransformer2DModel(DualTransformer2DModel):
+ deprecation_message = "Importing `DualTransformer2DModel` from `diffusers.models.dual_transformer_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.dual_transformer_2d import DualTransformer2DModel`, instead."
+ deprecate("DualTransformer2DModel", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/embeddings.py b/model/diffusers_c/models/embeddings.py
new file mode 100644
index 0000000000000000000000000000000000000000..49f385d5f4938c81d83794d217f79fabdac7905a
--- /dev/null
+++ b/model/diffusers_c/models/embeddings.py
@@ -0,0 +1,915 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from torch import nn
+
+from ..utils import USE_PEFT_BACKEND, deprecate
+from .activations import get_activation
+from .attention_processor import Attention
+from .lora import LoRACompatibleLinear
+
+
+def get_timestep_embedding(
+ timesteps: torch.Tensor,
+ embedding_dim: int,
+ flip_sin_to_cos: bool = False,
+ downscale_freq_shift: float = 1,
+ scale: float = 1,
+ max_period: int = 10000,
+):
+ """
+ This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
+
+ :param timesteps: a 1-D Tensor of N indices, one per batch element.
+ These may be fractional.
+ :param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the
+ embeddings. :return: an [N x dim] Tensor of positional embeddings.
+ """
+ assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
+
+ half_dim = embedding_dim // 2
+ exponent = -math.log(max_period) * torch.arange(
+ start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
+ )
+ exponent = exponent / (half_dim - downscale_freq_shift)
+
+ emb = torch.exp(exponent)
+ emb = timesteps[:, None].float() * emb[None, :]
+
+ # scale embeddings
+ emb = scale * emb
+
+ # concat sine and cosine embeddings
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
+
+ # flip sine and cosine embeddings
+ if flip_sin_to_cos:
+ emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
+
+ # zero pad
+ if embedding_dim % 2 == 1:
+ emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
+ return emb
+
+
+def get_2d_sincos_pos_embed(
+ embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16
+):
+ """
+ grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or
+ [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+ """
+ if isinstance(grid_size, int):
+ grid_size = (grid_size, grid_size)
+
+ grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size) / interpolation_scale
+ grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size) / interpolation_scale
+ grid = np.meshgrid(grid_w, grid_h) # here w goes first
+ grid = np.stack(grid, axis=0)
+
+ grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
+ pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+ if cls_token and extra_tokens > 0:
+ pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
+ return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+ if embed_dim % 2 != 0:
+ raise ValueError("embed_dim must be divisible by 2")
+
+ # use half of dimensions to encode grid_h
+ emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
+ emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
+
+ emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
+ return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+ """
+ embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)
+ """
+ if embed_dim % 2 != 0:
+ raise ValueError("embed_dim must be divisible by 2")
+
+ omega = np.arange(embed_dim // 2, dtype=np.float64)
+ omega /= embed_dim / 2.0
+ omega = 1.0 / 10000**omega # (D/2,)
+
+ pos = pos.reshape(-1) # (M,)
+ out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product
+
+ emb_sin = np.sin(out) # (M, D/2)
+ emb_cos = np.cos(out) # (M, D/2)
+
+ emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
+ return emb
+
+
+class PatchEmbed(nn.Module):
+ """2D Image to Patch Embedding"""
+
+ def __init__(
+ self,
+ height=224,
+ width=224,
+ patch_size=16,
+ in_channels=3,
+ embed_dim=768,
+ layer_norm=False,
+ flatten=True,
+ bias=True,
+ interpolation_scale=1,
+ ):
+ super().__init__()
+
+ num_patches = (height // patch_size) * (width // patch_size)
+ self.flatten = flatten
+ self.layer_norm = layer_norm
+
+ self.proj = nn.Conv2d(
+ in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
+ )
+ if layer_norm:
+ self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
+ else:
+ self.norm = None
+
+ self.patch_size = patch_size
+ # See:
+ # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L161
+ self.height, self.width = height // patch_size, width // patch_size
+ self.base_size = height // patch_size
+ self.interpolation_scale = interpolation_scale
+ pos_embed = get_2d_sincos_pos_embed(
+ embed_dim, int(num_patches**0.5), base_size=self.base_size, interpolation_scale=self.interpolation_scale
+ )
+ self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+
+ def forward(self, latent):
+ height, width = latent.shape[-2] // self.patch_size, latent.shape[-1] // self.patch_size
+
+ latent = self.proj(latent)
+ if self.flatten:
+ latent = latent.flatten(2).transpose(1, 2) # BCHW -> BNC
+ if self.layer_norm:
+ latent = self.norm(latent)
+
+ # Interpolate positional embeddings if needed.
+ # (For PixArt-Alpha: https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L162C151-L162C160)
+ if self.height != height or self.width != width:
+ pos_embed = get_2d_sincos_pos_embed(
+ embed_dim=self.pos_embed.shape[-1],
+ grid_size=(height, width),
+ base_size=self.base_size,
+ interpolation_scale=self.interpolation_scale,
+ )
+ pos_embed = torch.from_numpy(pos_embed)
+ pos_embed = pos_embed.float().unsqueeze(0).to(latent.device)
+ else:
+ pos_embed = self.pos_embed
+
+ return (latent + pos_embed).to(latent.dtype)
+
+
+class TimestepEmbedding(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ time_embed_dim: int,
+ act_fn: str = "silu",
+ out_dim: int = None,
+ post_act_fn: Optional[str] = None,
+ cond_proj_dim=None,
+ sample_proj_bias=True,
+ ):
+ super().__init__()
+ linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+ self.linear_1 = linear_cls(in_channels, time_embed_dim, sample_proj_bias)
+
+ if cond_proj_dim is not None:
+ self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False)
+ else:
+ self.cond_proj = None
+
+ self.act = get_activation(act_fn)
+
+ if out_dim is not None:
+ time_embed_dim_out = out_dim
+ else:
+ time_embed_dim_out = time_embed_dim
+ self.linear_2 = linear_cls(time_embed_dim, time_embed_dim_out, sample_proj_bias)
+
+ if post_act_fn is None:
+ self.post_act = None
+ else:
+ self.post_act = get_activation(post_act_fn)
+
+ def forward(self, sample, condition=None):
+ if condition is not None:
+ sample = sample + self.cond_proj(condition)
+ sample = self.linear_1(sample)
+
+ if self.act is not None:
+ sample = self.act(sample)
+
+ sample = self.linear_2(sample)
+
+ if self.post_act is not None:
+ sample = self.post_act(sample)
+ return sample
+
+
+class Timesteps(nn.Module):
+ def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float):
+ super().__init__()
+ self.num_channels = num_channels
+ self.flip_sin_to_cos = flip_sin_to_cos
+ self.downscale_freq_shift = downscale_freq_shift
+
+ def forward(self, timesteps):
+ t_emb = get_timestep_embedding(
+ timesteps,
+ self.num_channels,
+ flip_sin_to_cos=self.flip_sin_to_cos,
+ downscale_freq_shift=self.downscale_freq_shift,
+ )
+ return t_emb
+
+
+class GaussianFourierProjection(nn.Module):
+ """Gaussian Fourier embeddings for noise levels."""
+
+ def __init__(
+ self, embedding_size: int = 256, scale: float = 1.0, set_W_to_weight=True, log=True, flip_sin_to_cos=False
+ ):
+ super().__init__()
+ self.weight = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False)
+ self.log = log
+ self.flip_sin_to_cos = flip_sin_to_cos
+
+ if set_W_to_weight:
+ # to delete later
+ self.W = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False)
+
+ self.weight = self.W
+
+ def forward(self, x):
+ if self.log:
+ x = torch.log(x)
+
+ x_proj = x[:, None] * self.weight[None, :] * 2 * np.pi
+
+ if self.flip_sin_to_cos:
+ out = torch.cat([torch.cos(x_proj), torch.sin(x_proj)], dim=-1)
+ else:
+ out = torch.cat([torch.sin(x_proj), torch.cos(x_proj)], dim=-1)
+ return out
+
+
+class SinusoidalPositionalEmbedding(nn.Module):
+ """Apply positional information to a sequence of embeddings.
+
+ Takes in a sequence of embeddings with shape (batch_size, seq_length, embed_dim) and adds positional embeddings to
+ them
+
+ Args:
+ embed_dim: (int): Dimension of the positional embedding.
+ max_seq_length: Maximum sequence length to apply positional embeddings
+
+ """
+
+ def __init__(self, embed_dim: int, max_seq_length: int = 32):
+ super().__init__()
+ position = torch.arange(max_seq_length).unsqueeze(1)
+ div_term = torch.exp(torch.arange(0, embed_dim, 2) * (-math.log(10000.0) / embed_dim))
+ pe = torch.zeros(1, max_seq_length, embed_dim)
+ pe[0, :, 0::2] = torch.sin(position * div_term)
+ pe[0, :, 1::2] = torch.cos(position * div_term)
+ self.register_buffer("pe", pe)
+
+ def forward(self, x):
+ _, seq_length, _ = x.shape
+ x = x + self.pe[:, :seq_length]
+ return x
+
+
+class ImagePositionalEmbeddings(nn.Module):
+ """
+ Converts latent image classes into vector embeddings. Sums the vector embeddings with positional embeddings for the
+ height and width of the latent space.
+
+ For more details, see figure 10 of the dall-e paper: https://arxiv.org/abs/2102.12092
+
+ For VQ-diffusion:
+
+ Output vector embeddings are used as input for the transformer.
+
+ Note that the vector embeddings for the transformer are different than the vector embeddings from the VQVAE.
+
+ Args:
+ num_embed (`int`):
+ Number of embeddings for the latent pixels embeddings.
+ height (`int`):
+ Height of the latent image i.e. the number of height embeddings.
+ width (`int`):
+ Width of the latent image i.e. the number of width embeddings.
+ embed_dim (`int`):
+ Dimension of the produced vector embeddings. Used for the latent pixel, height, and width embeddings.
+ """
+
+ def __init__(
+ self,
+ num_embed: int,
+ height: int,
+ width: int,
+ embed_dim: int,
+ ):
+ super().__init__()
+
+ self.height = height
+ self.width = width
+ self.num_embed = num_embed
+ self.embed_dim = embed_dim
+
+ self.emb = nn.Embedding(self.num_embed, embed_dim)
+ self.height_emb = nn.Embedding(self.height, embed_dim)
+ self.width_emb = nn.Embedding(self.width, embed_dim)
+
+ def forward(self, index):
+ emb = self.emb(index)
+
+ height_emb = self.height_emb(torch.arange(self.height, device=index.device).view(1, self.height))
+
+ # 1 x H x D -> 1 x H x 1 x D
+ height_emb = height_emb.unsqueeze(2)
+
+ width_emb = self.width_emb(torch.arange(self.width, device=index.device).view(1, self.width))
+
+ # 1 x W x D -> 1 x 1 x W x D
+ width_emb = width_emb.unsqueeze(1)
+
+ pos_emb = height_emb + width_emb
+
+ # 1 x H x W x D -> 1 x L xD
+ pos_emb = pos_emb.view(1, self.height * self.width, -1)
+
+ emb = emb + pos_emb[:, : emb.shape[1], :]
+
+ return emb
+
+
+class LabelEmbedding(nn.Module):
+ """
+ Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
+
+ Args:
+ num_classes (`int`): The number of classes.
+ hidden_size (`int`): The size of the vector embeddings.
+ dropout_prob (`float`): The probability of dropping a label.
+ """
+
+ def __init__(self, num_classes, hidden_size, dropout_prob):
+ super().__init__()
+ use_cfg_embedding = dropout_prob > 0
+ self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size)
+ self.num_classes = num_classes
+ self.dropout_prob = dropout_prob
+
+ def token_drop(self, labels, force_drop_ids=None):
+ """
+ Drops labels to enable classifier-free guidance.
+ """
+ if force_drop_ids is None:
+ drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
+ else:
+ drop_ids = torch.tensor(force_drop_ids == 1)
+ labels = torch.where(drop_ids, self.num_classes, labels)
+ return labels
+
+ def forward(self, labels: torch.LongTensor, force_drop_ids=None):
+ use_dropout = self.dropout_prob > 0
+ if (self.training and use_dropout) or (force_drop_ids is not None):
+ labels = self.token_drop(labels, force_drop_ids)
+ embeddings = self.embedding_table(labels)
+ return embeddings
+
+
+class TextImageProjection(nn.Module):
+ def __init__(
+ self,
+ text_embed_dim: int = 1024,
+ image_embed_dim: int = 768,
+ cross_attention_dim: int = 768,
+ num_image_text_embeds: int = 10,
+ ):
+ super().__init__()
+
+ self.num_image_text_embeds = num_image_text_embeds
+ self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim)
+ self.text_proj = nn.Linear(text_embed_dim, cross_attention_dim)
+
+ def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor):
+ batch_size = text_embeds.shape[0]
+
+ # image
+ image_text_embeds = self.image_embeds(image_embeds)
+ image_text_embeds = image_text_embeds.reshape(batch_size, self.num_image_text_embeds, -1)
+
+ # text
+ text_embeds = self.text_proj(text_embeds)
+
+ return torch.cat([image_text_embeds, text_embeds], dim=1)
+
+
+class ImageProjection(nn.Module):
+ def __init__(
+ self,
+ image_embed_dim: int = 768,
+ cross_attention_dim: int = 768,
+ num_image_text_embeds: int = 32,
+ ):
+ super().__init__()
+
+ self.num_image_text_embeds = num_image_text_embeds
+ self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim)
+ self.norm = nn.LayerNorm(cross_attention_dim)
+
+ def forward(self, image_embeds: torch.FloatTensor):
+ batch_size = image_embeds.shape[0]
+
+ # image
+ image_embeds = self.image_embeds(image_embeds)
+ image_embeds = image_embeds.reshape(batch_size, self.num_image_text_embeds, -1)
+ image_embeds = self.norm(image_embeds)
+ return image_embeds
+
+
+class IPAdapterFullImageProjection(nn.Module):
+ def __init__(self, image_embed_dim=1024, cross_attention_dim=1024):
+ super().__init__()
+ from .attention import FeedForward
+
+ self.ff = FeedForward(image_embed_dim, cross_attention_dim, mult=1, activation_fn="gelu")
+ self.norm = nn.LayerNorm(cross_attention_dim)
+
+ def forward(self, image_embeds: torch.FloatTensor):
+ return self.norm(self.ff(image_embeds))
+
+
+class CombinedTimestepLabelEmbeddings(nn.Module):
+ def __init__(self, num_classes, embedding_dim, class_dropout_prob=0.1):
+ super().__init__()
+
+ self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1)
+ self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+ self.class_embedder = LabelEmbedding(num_classes, embedding_dim, class_dropout_prob)
+
+ def forward(self, timestep, class_labels, hidden_dtype=None):
+ timesteps_proj = self.time_proj(timestep)
+ timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (N, D)
+
+ class_labels = self.class_embedder(class_labels) # (N, D)
+
+ conditioning = timesteps_emb + class_labels # (N, D)
+
+ return conditioning
+
+
+class TextTimeEmbedding(nn.Module):
+ def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64):
+ super().__init__()
+ self.norm1 = nn.LayerNorm(encoder_dim)
+ self.pool = AttentionPooling(num_heads, encoder_dim)
+ self.proj = nn.Linear(encoder_dim, time_embed_dim)
+ self.norm2 = nn.LayerNorm(time_embed_dim)
+
+ def forward(self, hidden_states):
+ hidden_states = self.norm1(hidden_states)
+ hidden_states = self.pool(hidden_states)
+ hidden_states = self.proj(hidden_states)
+ hidden_states = self.norm2(hidden_states)
+ return hidden_states
+
+
+class TextImageTimeEmbedding(nn.Module):
+ def __init__(self, text_embed_dim: int = 768, image_embed_dim: int = 768, time_embed_dim: int = 1536):
+ super().__init__()
+ self.text_proj = nn.Linear(text_embed_dim, time_embed_dim)
+ self.text_norm = nn.LayerNorm(time_embed_dim)
+ self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
+
+ def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor):
+ # text
+ time_text_embeds = self.text_proj(text_embeds)
+ time_text_embeds = self.text_norm(time_text_embeds)
+
+ # image
+ time_image_embeds = self.image_proj(image_embeds)
+
+ return time_image_embeds + time_text_embeds
+
+
+class ImageTimeEmbedding(nn.Module):
+ def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536):
+ super().__init__()
+ self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
+ self.image_norm = nn.LayerNorm(time_embed_dim)
+
+ def forward(self, image_embeds: torch.FloatTensor):
+ # image
+ time_image_embeds = self.image_proj(image_embeds)
+ time_image_embeds = self.image_norm(time_image_embeds)
+ return time_image_embeds
+
+
+class ImageHintTimeEmbedding(nn.Module):
+ def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536):
+ super().__init__()
+ self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
+ self.image_norm = nn.LayerNorm(time_embed_dim)
+ self.input_hint_block = nn.Sequential(
+ nn.Conv2d(3, 16, 3, padding=1),
+ nn.SiLU(),
+ nn.Conv2d(16, 16, 3, padding=1),
+ nn.SiLU(),
+ nn.Conv2d(16, 32, 3, padding=1, stride=2),
+ nn.SiLU(),
+ nn.Conv2d(32, 32, 3, padding=1),
+ nn.SiLU(),
+ nn.Conv2d(32, 96, 3, padding=1, stride=2),
+ nn.SiLU(),
+ nn.Conv2d(96, 96, 3, padding=1),
+ nn.SiLU(),
+ nn.Conv2d(96, 256, 3, padding=1, stride=2),
+ nn.SiLU(),
+ nn.Conv2d(256, 4, 3, padding=1),
+ )
+
+ def forward(self, image_embeds: torch.FloatTensor, hint: torch.FloatTensor):
+ # image
+ time_image_embeds = self.image_proj(image_embeds)
+ time_image_embeds = self.image_norm(time_image_embeds)
+ hint = self.input_hint_block(hint)
+ return time_image_embeds, hint
+
+
+class AttentionPooling(nn.Module):
+ # Copied from https://github.com/deep-floyd/IF/blob/2f91391f27dd3c468bf174be5805b4cc92980c0b/deepfloyd_if/model/nn.py#L54
+
+ def __init__(self, num_heads, embed_dim, dtype=None):
+ super().__init__()
+ self.dtype = dtype
+ self.positional_embedding = nn.Parameter(torch.randn(1, embed_dim) / embed_dim**0.5)
+ self.k_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
+ self.q_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
+ self.v_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
+ self.num_heads = num_heads
+ self.dim_per_head = embed_dim // self.num_heads
+
+ def forward(self, x):
+ bs, length, width = x.size()
+
+ def shape(x):
+ # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+ x = x.view(bs, -1, self.num_heads, self.dim_per_head)
+ # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+ x = x.transpose(1, 2)
+ # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+ x = x.reshape(bs * self.num_heads, -1, self.dim_per_head)
+ # (bs*n_heads, length, dim_per_head) --> (bs*n_heads, dim_per_head, length)
+ x = x.transpose(1, 2)
+ return x
+
+ class_token = x.mean(dim=1, keepdim=True) + self.positional_embedding.to(x.dtype)
+ x = torch.cat([class_token, x], dim=1) # (bs, length+1, width)
+
+ # (bs*n_heads, class_token_length, dim_per_head)
+ q = shape(self.q_proj(class_token))
+ # (bs*n_heads, length+class_token_length, dim_per_head)
+ k = shape(self.k_proj(x))
+ v = shape(self.v_proj(x))
+
+ # (bs*n_heads, class_token_length, length+class_token_length):
+ scale = 1 / math.sqrt(math.sqrt(self.dim_per_head))
+ weight = torch.einsum("bct,bcs->bts", q * scale, k * scale) # More stable with f16 than dividing afterwards
+ weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+
+ # (bs*n_heads, dim_per_head, class_token_length)
+ a = torch.einsum("bts,bcs->bct", weight, v)
+
+ # (bs, length+1, width)
+ a = a.reshape(bs, -1, 1).transpose(1, 2)
+
+ return a[:, 0, :] # cls_token
+
+
+def get_fourier_embeds_from_boundingbox(embed_dim, box):
+ """
+ Args:
+ embed_dim: int
+ box: a 3-D tensor [B x N x 4] representing the bounding boxes for GLIGEN pipeline
+ Returns:
+ [B x N x embed_dim] tensor of positional embeddings
+ """
+
+ batch_size, num_boxes = box.shape[:2]
+
+ emb = 100 ** (torch.arange(embed_dim) / embed_dim)
+ emb = emb[None, None, None].to(device=box.device, dtype=box.dtype)
+ emb = emb * box.unsqueeze(-1)
+
+ emb = torch.stack((emb.sin(), emb.cos()), dim=-1)
+ emb = emb.permute(0, 1, 3, 4, 2).reshape(batch_size, num_boxes, embed_dim * 2 * 4)
+
+ return emb
+
+
+class GLIGENTextBoundingboxProjection(nn.Module):
+ def __init__(self, positive_len, out_dim, feature_type="text-only", fourier_freqs=8):
+ super().__init__()
+ self.positive_len = positive_len
+ self.out_dim = out_dim
+
+ self.fourier_embedder_dim = fourier_freqs
+ self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy
+
+ if isinstance(out_dim, tuple):
+ out_dim = out_dim[0]
+
+ if feature_type == "text-only":
+ self.linears = nn.Sequential(
+ nn.Linear(self.positive_len + self.position_dim, 512),
+ nn.SiLU(),
+ nn.Linear(512, 512),
+ nn.SiLU(),
+ nn.Linear(512, out_dim),
+ )
+ self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+
+ elif feature_type == "text-image":
+ self.linears_text = nn.Sequential(
+ nn.Linear(self.positive_len + self.position_dim, 512),
+ nn.SiLU(),
+ nn.Linear(512, 512),
+ nn.SiLU(),
+ nn.Linear(512, out_dim),
+ )
+ self.linears_image = nn.Sequential(
+ nn.Linear(self.positive_len + self.position_dim, 512),
+ nn.SiLU(),
+ nn.Linear(512, 512),
+ nn.SiLU(),
+ nn.Linear(512, out_dim),
+ )
+ self.null_text_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+ self.null_image_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+
+ self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim]))
+
+ def forward(
+ self,
+ boxes,
+ masks,
+ positive_embeddings=None,
+ phrases_masks=None,
+ image_masks=None,
+ phrases_embeddings=None,
+ image_embeddings=None,
+ ):
+ masks = masks.unsqueeze(-1)
+
+ # embedding position (it may includes padding as placeholder)
+ xyxy_embedding = get_fourier_embeds_from_boundingbox(self.fourier_embedder_dim, boxes) # B*N*4 -> B*N*C
+
+ # learnable null embedding
+ xyxy_null = self.null_position_feature.view(1, 1, -1)
+
+ # replace padding with learnable null embedding
+ xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null
+
+ # positionet with text only information
+ if positive_embeddings is not None:
+ # learnable null embedding
+ positive_null = self.null_positive_feature.view(1, 1, -1)
+
+ # replace padding with learnable null embedding
+ positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null
+
+ objs = self.linears(torch.cat([positive_embeddings, xyxy_embedding], dim=-1))
+
+ # positionet with text and image infomation
+ else:
+ phrases_masks = phrases_masks.unsqueeze(-1)
+ image_masks = image_masks.unsqueeze(-1)
+
+ # learnable null embedding
+ text_null = self.null_text_feature.view(1, 1, -1)
+ image_null = self.null_image_feature.view(1, 1, -1)
+
+ # replace padding with learnable null embedding
+ phrases_embeddings = phrases_embeddings * phrases_masks + (1 - phrases_masks) * text_null
+ image_embeddings = image_embeddings * image_masks + (1 - image_masks) * image_null
+
+ objs_text = self.linears_text(torch.cat([phrases_embeddings, xyxy_embedding], dim=-1))
+ objs_image = self.linears_image(torch.cat([image_embeddings, xyxy_embedding], dim=-1))
+ objs = torch.cat([objs_text, objs_image], dim=1)
+
+ return objs
+
+
+class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
+ """
+ For PixArt-Alpha.
+
+ Reference:
+ https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29
+ """
+
+ def __init__(self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False):
+ super().__init__()
+
+ self.outdim = size_emb_dim
+ self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+ self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim)
+
+ self.use_additional_conditions = use_additional_conditions
+ if use_additional_conditions:
+ self.additional_condition_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+ self.resolution_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=size_emb_dim)
+ self.aspect_ratio_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=size_emb_dim)
+
+ def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype):
+ timesteps_proj = self.time_proj(timestep)
+ timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (N, D)
+
+ if self.use_additional_conditions:
+ resolution_emb = self.additional_condition_proj(resolution.flatten()).to(hidden_dtype)
+ resolution_emb = self.resolution_embedder(resolution_emb).reshape(batch_size, -1)
+ aspect_ratio_emb = self.additional_condition_proj(aspect_ratio.flatten()).to(hidden_dtype)
+ aspect_ratio_emb = self.aspect_ratio_embedder(aspect_ratio_emb).reshape(batch_size, -1)
+ conditioning = timesteps_emb + torch.cat([resolution_emb, aspect_ratio_emb], dim=1)
+ else:
+ conditioning = timesteps_emb
+
+ return conditioning
+
+
+class PixArtAlphaTextProjection(nn.Module):
+ """
+ Projects caption embeddings. Also handles dropout for classifier-free guidance.
+
+ Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
+ """
+
+ def __init__(self, in_features, hidden_size, num_tokens=120):
+ super().__init__()
+ self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True)
+ self.act_1 = nn.GELU(approximate="tanh")
+ self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True)
+
+ def forward(self, caption):
+ hidden_states = self.linear_1(caption)
+ hidden_states = self.act_1(hidden_states)
+ hidden_states = self.linear_2(hidden_states)
+ return hidden_states
+
+
+class IPAdapterPlusImageProjection(nn.Module):
+ """Resampler of IP-Adapter Plus.
+
+ Args:
+ ----
+ embed_dims (int): The feature dimension. Defaults to 768.
+ output_dims (int): The number of output channels, that is the same
+ number of the channels in the
+ `unet.config.cross_attention_dim`. Defaults to 1024.
+ hidden_dims (int): The number of hidden channels. Defaults to 1280.
+ depth (int): The number of blocks. Defaults to 8.
+ dim_head (int): The number of head channels. Defaults to 64.
+ heads (int): Parallel attention heads. Defaults to 16.
+ num_queries (int): The number of queries. Defaults to 8.
+ ffn_ratio (float): The expansion ratio of feedforward network hidden
+ layer channels. Defaults to 4.
+ """
+
+ def __init__(
+ self,
+ embed_dims: int = 768,
+ output_dims: int = 1024,
+ hidden_dims: int = 1280,
+ depth: int = 4,
+ dim_head: int = 64,
+ heads: int = 16,
+ num_queries: int = 8,
+ ffn_ratio: float = 4,
+ ) -> None:
+ super().__init__()
+ from .attention import FeedForward # Lazy import to avoid circular import
+
+ self.latents = nn.Parameter(torch.randn(1, num_queries, hidden_dims) / hidden_dims**0.5)
+
+ self.proj_in = nn.Linear(embed_dims, hidden_dims)
+
+ self.proj_out = nn.Linear(hidden_dims, output_dims)
+ self.norm_out = nn.LayerNorm(output_dims)
+
+ self.layers = nn.ModuleList([])
+ for _ in range(depth):
+ self.layers.append(
+ nn.ModuleList(
+ [
+ nn.LayerNorm(hidden_dims),
+ nn.LayerNorm(hidden_dims),
+ Attention(
+ query_dim=hidden_dims,
+ dim_head=dim_head,
+ heads=heads,
+ out_bias=False,
+ ),
+ nn.Sequential(
+ nn.LayerNorm(hidden_dims),
+ FeedForward(hidden_dims, hidden_dims, activation_fn="gelu", mult=ffn_ratio, bias=False),
+ ),
+ ]
+ )
+ )
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ """Forward pass.
+
+ Args:
+ ----
+ x (torch.Tensor): Input Tensor.
+
+ Returns:
+ -------
+ torch.Tensor: Output Tensor.
+ """
+ latents = self.latents.repeat(x.size(0), 1, 1)
+
+ x = self.proj_in(x)
+
+ for ln0, ln1, attn, ff in self.layers:
+ residual = latents
+
+ encoder_hidden_states = ln0(x)
+ latents = ln1(latents)
+ encoder_hidden_states = torch.cat([encoder_hidden_states, latents], dim=-2)
+ latents = attn(latents, encoder_hidden_states) + residual
+ latents = ff(latents) + latents
+
+ latents = self.proj_out(latents)
+ return self.norm_out(latents)
+
+
+class MultiIPAdapterImageProjection(nn.Module):
+ def __init__(self, IPAdapterImageProjectionLayers: Union[List[nn.Module], Tuple[nn.Module]]):
+ super().__init__()
+ self.image_projection_layers = nn.ModuleList(IPAdapterImageProjectionLayers)
+
+ def forward(self, image_embeds: List[torch.FloatTensor]):
+ projected_image_embeds = []
+
+ # currently, we accept `image_embeds` as
+ # 1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim]
+ # 2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim]
+ if not isinstance(image_embeds, list):
+ deprecation_message = (
+ "You have passed a tensor as `image_embeds`.This is deprecated and will be removed in a future release."
+ " Please make sure to update your script to pass `image_embeds` as a list of tensors to supress this warning."
+ )
+ deprecate("image_embeds not a list", "1.0.0", deprecation_message, standard_warn=False)
+ image_embeds = [image_embeds.unsqueeze(1)]
+
+ if len(image_embeds) != len(self.image_projection_layers):
+ raise ValueError(
+ f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}"
+ )
+
+ for image_embed, image_projection_layer in zip(image_embeds, self.image_projection_layers):
+ batch_size, num_images = image_embed.shape[0], image_embed.shape[1]
+ image_embed = image_embed.reshape((batch_size * num_images,) + image_embed.shape[2:])
+ image_embed = image_projection_layer(image_embed)
+ image_embed = image_embed.reshape((batch_size, num_images) + image_embed.shape[1:])
+
+ projected_image_embeds.append(image_embed)
+
+ return projected_image_embeds
diff --git a/model/diffusers_c/models/embeddings_flax.py b/model/diffusers_c/models/embeddings_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..8e343be0d3b71bfcc1f7fb0a1787ad2bc4916dd2
--- /dev/null
+++ b/model/diffusers_c/models/embeddings_flax.py
@@ -0,0 +1,97 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+
+import flax.linen as nn
+import jax.numpy as jnp
+
+
+def get_sinusoidal_embeddings(
+ timesteps: jnp.ndarray,
+ embedding_dim: int,
+ freq_shift: float = 1,
+ min_timescale: float = 1,
+ max_timescale: float = 1.0e4,
+ flip_sin_to_cos: bool = False,
+ scale: float = 1.0,
+) -> jnp.ndarray:
+ """Returns the positional encoding (same as Tensor2Tensor).
+
+ Args:
+ timesteps: a 1-D Tensor of N indices, one per batch element.
+ These may be fractional.
+ embedding_dim: The number of output channels.
+ min_timescale: The smallest time unit (should probably be 0.0).
+ max_timescale: The largest time unit.
+ Returns:
+ a Tensor of timing signals [N, num_channels]
+ """
+ assert timesteps.ndim == 1, "Timesteps should be a 1d-array"
+ assert embedding_dim % 2 == 0, f"Embedding dimension {embedding_dim} should be even"
+ num_timescales = float(embedding_dim // 2)
+ log_timescale_increment = math.log(max_timescale / min_timescale) / (num_timescales - freq_shift)
+ inv_timescales = min_timescale * jnp.exp(jnp.arange(num_timescales, dtype=jnp.float32) * -log_timescale_increment)
+ emb = jnp.expand_dims(timesteps, 1) * jnp.expand_dims(inv_timescales, 0)
+
+ # scale embeddings
+ scaled_time = scale * emb
+
+ if flip_sin_to_cos:
+ signal = jnp.concatenate([jnp.cos(scaled_time), jnp.sin(scaled_time)], axis=1)
+ else:
+ signal = jnp.concatenate([jnp.sin(scaled_time), jnp.cos(scaled_time)], axis=1)
+ signal = jnp.reshape(signal, [jnp.shape(timesteps)[0], embedding_dim])
+ return signal
+
+
+class FlaxTimestepEmbedding(nn.Module):
+ r"""
+ Time step Embedding Module. Learns embeddings for input time steps.
+
+ Args:
+ time_embed_dim (`int`, *optional*, defaults to `32`):
+ Time step embedding dimension
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ time_embed_dim: int = 32
+ dtype: jnp.dtype = jnp.float32
+
+ @nn.compact
+ def __call__(self, temb):
+ temb = nn.Dense(self.time_embed_dim, dtype=self.dtype, name="linear_1")(temb)
+ temb = nn.silu(temb)
+ temb = nn.Dense(self.time_embed_dim, dtype=self.dtype, name="linear_2")(temb)
+ return temb
+
+
+class FlaxTimesteps(nn.Module):
+ r"""
+ Wrapper Module for sinusoidal Time step Embeddings as described in https://arxiv.org/abs/2006.11239
+
+ Args:
+ dim (`int`, *optional*, defaults to `32`):
+ Time step embedding dimension
+ """
+
+ dim: int = 32
+ flip_sin_to_cos: bool = False
+ freq_shift: float = 1
+
+ @nn.compact
+ def __call__(self, timesteps):
+ return get_sinusoidal_embeddings(
+ timesteps, embedding_dim=self.dim, flip_sin_to_cos=self.flip_sin_to_cos, freq_shift=self.freq_shift
+ )
diff --git a/model/diffusers_c/models/lora.py b/model/diffusers_c/models/lora.py
new file mode 100644
index 0000000000000000000000000000000000000000..1bbc96c6f5a7c44731e9e59418d887d0a5cfb40e
--- /dev/null
+++ b/model/diffusers_c/models/lora.py
@@ -0,0 +1,451 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# IMPORTANT: #
+###################################################################
+# ----------------------------------------------------------------#
+# This file is deprecated and will be removed soon #
+# (as soon as PEFT will become a required dependency for LoRA) #
+# ----------------------------------------------------------------#
+###################################################################
+
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..utils import deprecate, logging
+from ..utils.import_utils import is_transformers_available
+
+
+if is_transformers_available():
+ from transformers import CLIPTextModel, CLIPTextModelWithProjection
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def text_encoder_attn_modules(text_encoder):
+ attn_modules = []
+
+ if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+ for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+ name = f"text_model.encoder.layers.{i}.self_attn"
+ mod = layer.self_attn
+ attn_modules.append((name, mod))
+ else:
+ raise ValueError(f"do not know how to get attention modules for: {text_encoder.__class__.__name__}")
+
+ return attn_modules
+
+
+def text_encoder_mlp_modules(text_encoder):
+ mlp_modules = []
+
+ if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
+ for i, layer in enumerate(text_encoder.text_model.encoder.layers):
+ mlp_mod = layer.mlp
+ name = f"text_model.encoder.layers.{i}.mlp"
+ mlp_modules.append((name, mlp_mod))
+ else:
+ raise ValueError(f"do not know how to get mlp modules for: {text_encoder.__class__.__name__}")
+
+ return mlp_modules
+
+
+def adjust_lora_scale_text_encoder(text_encoder, lora_scale: float = 1.0):
+ for _, attn_module in text_encoder_attn_modules(text_encoder):
+ if isinstance(attn_module.q_proj, PatchedLoraProjection):
+ attn_module.q_proj.lora_scale = lora_scale
+ attn_module.k_proj.lora_scale = lora_scale
+ attn_module.v_proj.lora_scale = lora_scale
+ attn_module.out_proj.lora_scale = lora_scale
+
+ for _, mlp_module in text_encoder_mlp_modules(text_encoder):
+ if isinstance(mlp_module.fc1, PatchedLoraProjection):
+ mlp_module.fc1.lora_scale = lora_scale
+ mlp_module.fc2.lora_scale = lora_scale
+
+
+class PatchedLoraProjection(torch.nn.Module):
+ def __init__(self, regular_linear_layer, lora_scale=1, network_alpha=None, rank=4, dtype=None):
+ deprecation_message = "Use of `PatchedLoraProjection` is deprecated. Please switch to PEFT backend by installing PEFT: `pip install peft`."
+ deprecate("PatchedLoraProjection", "1.0.0", deprecation_message)
+
+ super().__init__()
+ from ..models.lora import LoRALinearLayer
+
+ self.regular_linear_layer = regular_linear_layer
+
+ device = self.regular_linear_layer.weight.device
+
+ if dtype is None:
+ dtype = self.regular_linear_layer.weight.dtype
+
+ self.lora_linear_layer = LoRALinearLayer(
+ self.regular_linear_layer.in_features,
+ self.regular_linear_layer.out_features,
+ network_alpha=network_alpha,
+ device=device,
+ dtype=dtype,
+ rank=rank,
+ )
+
+ self.lora_scale = lora_scale
+
+ # overwrite PyTorch's `state_dict` to be sure that only the 'regular_linear_layer' weights are saved
+ # when saving the whole text encoder model and when LoRA is unloaded or fused
+ def state_dict(self, *args, destination=None, prefix="", keep_vars=False):
+ if self.lora_linear_layer is None:
+ return self.regular_linear_layer.state_dict(
+ *args, destination=destination, prefix=prefix, keep_vars=keep_vars
+ )
+
+ return super().state_dict(*args, destination=destination, prefix=prefix, keep_vars=keep_vars)
+
+ def _fuse_lora(self, lora_scale=1.0, safe_fusing=False):
+ if self.lora_linear_layer is None:
+ return
+
+ dtype, device = self.regular_linear_layer.weight.data.dtype, self.regular_linear_layer.weight.data.device
+
+ w_orig = self.regular_linear_layer.weight.data.float()
+ w_up = self.lora_linear_layer.up.weight.data.float()
+ w_down = self.lora_linear_layer.down.weight.data.float()
+
+ if self.lora_linear_layer.network_alpha is not None:
+ w_up = w_up * self.lora_linear_layer.network_alpha / self.lora_linear_layer.rank
+
+ fused_weight = w_orig + (lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
+
+ if safe_fusing and torch.isnan(fused_weight).any().item():
+ raise ValueError(
+ "This LoRA weight seems to be broken. "
+ f"Encountered NaN values when trying to fuse LoRA weights for {self}."
+ "LoRA weights will not be fused."
+ )
+
+ self.regular_linear_layer.weight.data = fused_weight.to(device=device, dtype=dtype)
+
+ # we can drop the lora layer now
+ self.lora_linear_layer = None
+
+ # offload the up and down matrices to CPU to not blow the memory
+ self.w_up = w_up.cpu()
+ self.w_down = w_down.cpu()
+ self.lora_scale = lora_scale
+
+ def _unfuse_lora(self):
+ if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
+ return
+
+ fused_weight = self.regular_linear_layer.weight.data
+ dtype, device = fused_weight.dtype, fused_weight.device
+
+ w_up = self.w_up.to(device=device).float()
+ w_down = self.w_down.to(device).float()
+
+ unfused_weight = fused_weight.float() - (self.lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
+ self.regular_linear_layer.weight.data = unfused_weight.to(device=device, dtype=dtype)
+
+ self.w_up = None
+ self.w_down = None
+
+ def forward(self, input):
+ if self.lora_scale is None:
+ self.lora_scale = 1.0
+ if self.lora_linear_layer is None:
+ return self.regular_linear_layer(input)
+ return self.regular_linear_layer(input) + (self.lora_scale * self.lora_linear_layer(input))
+
+
+class LoRALinearLayer(nn.Module):
+ r"""
+ A linear layer that is used with LoRA.
+
+ Parameters:
+ in_features (`int`):
+ Number of input features.
+ out_features (`int`):
+ Number of output features.
+ rank (`int`, `optional`, defaults to 4):
+ The rank of the LoRA layer.
+ network_alpha (`float`, `optional`, defaults to `None`):
+ The value of the network alpha used for stable learning and preventing underflow. This value has the same
+ meaning as the `--network_alpha` option in the kohya-ss trainer script. See
+ https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+ device (`torch.device`, `optional`, defaults to `None`):
+ The device to use for the layer's weights.
+ dtype (`torch.dtype`, `optional`, defaults to `None`):
+ The dtype to use for the layer's weights.
+ """
+
+ def __init__(
+ self,
+ in_features: int,
+ out_features: int,
+ rank: int = 4,
+ network_alpha: Optional[float] = None,
+ device: Optional[Union[torch.device, str]] = None,
+ dtype: Optional[torch.dtype] = None,
+ ):
+ super().__init__()
+
+ self.down = nn.Linear(in_features, rank, bias=False, device=device, dtype=dtype)
+ self.up = nn.Linear(rank, out_features, bias=False, device=device, dtype=dtype)
+ # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
+ # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+ self.network_alpha = network_alpha
+ self.rank = rank
+ self.out_features = out_features
+ self.in_features = in_features
+
+ nn.init.normal_(self.down.weight, std=1 / rank)
+ nn.init.zeros_(self.up.weight)
+
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+ orig_dtype = hidden_states.dtype
+ dtype = self.down.weight.dtype
+
+ down_hidden_states = self.down(hidden_states.to(dtype))
+ up_hidden_states = self.up(down_hidden_states)
+
+ if self.network_alpha is not None:
+ up_hidden_states *= self.network_alpha / self.rank
+
+ return up_hidden_states.to(orig_dtype)
+
+
+class LoRAConv2dLayer(nn.Module):
+ r"""
+ A convolutional layer that is used with LoRA.
+
+ Parameters:
+ in_features (`int`):
+ Number of input features.
+ out_features (`int`):
+ Number of output features.
+ rank (`int`, `optional`, defaults to 4):
+ The rank of the LoRA layer.
+ kernel_size (`int` or `tuple` of two `int`, `optional`, defaults to 1):
+ The kernel size of the convolution.
+ stride (`int` or `tuple` of two `int`, `optional`, defaults to 1):
+ The stride of the convolution.
+ padding (`int` or `tuple` of two `int` or `str`, `optional`, defaults to 0):
+ The padding of the convolution.
+ network_alpha (`float`, `optional`, defaults to `None`):
+ The value of the network alpha used for stable learning and preventing underflow. This value has the same
+ meaning as the `--network_alpha` option in the kohya-ss trainer script. See
+ https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+ """
+
+ def __init__(
+ self,
+ in_features: int,
+ out_features: int,
+ rank: int = 4,
+ kernel_size: Union[int, Tuple[int, int]] = (1, 1),
+ stride: Union[int, Tuple[int, int]] = (1, 1),
+ padding: Union[int, Tuple[int, int], str] = 0,
+ network_alpha: Optional[float] = None,
+ ):
+ super().__init__()
+
+ self.down = nn.Conv2d(in_features, rank, kernel_size=kernel_size, stride=stride, padding=padding, bias=False)
+ # according to the official kohya_ss trainer kernel_size are always fixed for the up layer
+ # # see: https://github.com/bmaltais/kohya_ss/blob/2accb1305979ba62f5077a23aabac23b4c37e935/networks/lora_diffusers.py#L129
+ self.up = nn.Conv2d(rank, out_features, kernel_size=(1, 1), stride=(1, 1), bias=False)
+
+ # This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
+ # See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
+ self.network_alpha = network_alpha
+ self.rank = rank
+
+ nn.init.normal_(self.down.weight, std=1 / rank)
+ nn.init.zeros_(self.up.weight)
+
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+ orig_dtype = hidden_states.dtype
+ dtype = self.down.weight.dtype
+
+ down_hidden_states = self.down(hidden_states.to(dtype))
+ up_hidden_states = self.up(down_hidden_states)
+
+ if self.network_alpha is not None:
+ up_hidden_states *= self.network_alpha / self.rank
+
+ return up_hidden_states.to(orig_dtype)
+
+
+class LoRACompatibleConv(nn.Conv2d):
+ """
+ A convolutional layer that can be used with LoRA.
+ """
+
+ def __init__(self, *args, lora_layer: Optional[LoRAConv2dLayer] = None, **kwargs):
+ deprecation_message = "Use of `LoRACompatibleConv` is deprecated. Please switch to PEFT backend by installing PEFT: `pip install peft`."
+ deprecate("LoRACompatibleConv", "1.0.0", deprecation_message)
+
+ super().__init__(*args, **kwargs)
+ self.lora_layer = lora_layer
+
+ def set_lora_layer(self, lora_layer: Optional[LoRAConv2dLayer]):
+ deprecation_message = "Use of `set_lora_layer()` is deprecated. Please switch to PEFT backend by installing PEFT: `pip install peft`."
+ deprecate("set_lora_layer", "1.0.0", deprecation_message)
+
+ self.lora_layer = lora_layer
+
+ def _fuse_lora(self, lora_scale: float = 1.0, safe_fusing: bool = False):
+ if self.lora_layer is None:
+ return
+
+ dtype, device = self.weight.data.dtype, self.weight.data.device
+
+ w_orig = self.weight.data.float()
+ w_up = self.lora_layer.up.weight.data.float()
+ w_down = self.lora_layer.down.weight.data.float()
+
+ if self.lora_layer.network_alpha is not None:
+ w_up = w_up * self.lora_layer.network_alpha / self.lora_layer.rank
+
+ fusion = torch.mm(w_up.flatten(start_dim=1), w_down.flatten(start_dim=1))
+ fusion = fusion.reshape((w_orig.shape))
+ fused_weight = w_orig + (lora_scale * fusion)
+
+ if safe_fusing and torch.isnan(fused_weight).any().item():
+ raise ValueError(
+ "This LoRA weight seems to be broken. "
+ f"Encountered NaN values when trying to fuse LoRA weights for {self}."
+ "LoRA weights will not be fused."
+ )
+
+ self.weight.data = fused_weight.to(device=device, dtype=dtype)
+
+ # we can drop the lora layer now
+ self.lora_layer = None
+
+ # offload the up and down matrices to CPU to not blow the memory
+ self.w_up = w_up.cpu()
+ self.w_down = w_down.cpu()
+ self._lora_scale = lora_scale
+
+ def _unfuse_lora(self):
+ if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
+ return
+
+ fused_weight = self.weight.data
+ dtype, device = fused_weight.data.dtype, fused_weight.data.device
+
+ self.w_up = self.w_up.to(device=device).float()
+ self.w_down = self.w_down.to(device).float()
+
+ fusion = torch.mm(self.w_up.flatten(start_dim=1), self.w_down.flatten(start_dim=1))
+ fusion = fusion.reshape((fused_weight.shape))
+ unfused_weight = fused_weight.float() - (self._lora_scale * fusion)
+ self.weight.data = unfused_weight.to(device=device, dtype=dtype)
+
+ self.w_up = None
+ self.w_down = None
+
+ def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+ if self.padding_mode != "zeros":
+ hidden_states = F.pad(hidden_states, self._reversed_padding_repeated_twice, mode=self.padding_mode)
+ padding = (0, 0)
+ else:
+ padding = self.padding
+
+ original_outputs = F.conv2d(
+ hidden_states, self.weight, self.bias, self.stride, padding, self.dilation, self.groups
+ )
+
+ if self.lora_layer is None:
+ return original_outputs
+ else:
+ return original_outputs + (scale * self.lora_layer(hidden_states))
+
+
+class LoRACompatibleLinear(nn.Linear):
+ """
+ A Linear layer that can be used with LoRA.
+ """
+
+ def __init__(self, *args, lora_layer: Optional[LoRALinearLayer] = None, **kwargs):
+ deprecation_message = "Use of `LoRACompatibleLinear` is deprecated. Please switch to PEFT backend by installing PEFT: `pip install peft`."
+ deprecate("LoRACompatibleLinear", "1.0.0", deprecation_message)
+
+ super().__init__(*args, **kwargs)
+ self.lora_layer = lora_layer
+
+ def set_lora_layer(self, lora_layer: Optional[LoRALinearLayer]):
+ deprecation_message = "Use of `set_lora_layer()` is deprecated. Please switch to PEFT backend by installing PEFT: `pip install peft`."
+ deprecate("set_lora_layer", "1.0.0", deprecation_message)
+ self.lora_layer = lora_layer
+
+ def _fuse_lora(self, lora_scale: float = 1.0, safe_fusing: bool = False):
+ if self.lora_layer is None:
+ return
+
+ dtype, device = self.weight.data.dtype, self.weight.data.device
+
+ w_orig = self.weight.data.float()
+ w_up = self.lora_layer.up.weight.data.float()
+ w_down = self.lora_layer.down.weight.data.float()
+
+ if self.lora_layer.network_alpha is not None:
+ w_up = w_up * self.lora_layer.network_alpha / self.lora_layer.rank
+
+ fused_weight = w_orig + (lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
+
+ if safe_fusing and torch.isnan(fused_weight).any().item():
+ raise ValueError(
+ "This LoRA weight seems to be broken. "
+ f"Encountered NaN values when trying to fuse LoRA weights for {self}."
+ "LoRA weights will not be fused."
+ )
+
+ self.weight.data = fused_weight.to(device=device, dtype=dtype)
+
+ # we can drop the lora layer now
+ self.lora_layer = None
+
+ # offload the up and down matrices to CPU to not blow the memory
+ self.w_up = w_up.cpu()
+ self.w_down = w_down.cpu()
+ self._lora_scale = lora_scale
+
+ def _unfuse_lora(self):
+ if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
+ return
+
+ fused_weight = self.weight.data
+ dtype, device = fused_weight.dtype, fused_weight.device
+
+ w_up = self.w_up.to(device=device).float()
+ w_down = self.w_down.to(device).float()
+
+ unfused_weight = fused_weight.float() - (self._lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
+ self.weight.data = unfused_weight.to(device=device, dtype=dtype)
+
+ self.w_up = None
+ self.w_down = None
+
+ def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+ if self.lora_layer is None:
+ out = super().forward(hidden_states)
+ return out
+ else:
+ out = super().forward(hidden_states) + (scale * self.lora_layer(hidden_states))
+ return out
diff --git a/model/diffusers_c/models/modeling_flax_pytorch_utils.py b/model/diffusers_c/models/modeling_flax_pytorch_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..247219915891af19a5e787c5c1c9f41fec696b93
--- /dev/null
+++ b/model/diffusers_c/models/modeling_flax_pytorch_utils.py
@@ -0,0 +1,134 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch - Flax general utilities."""
+
+import re
+
+import jax.numpy as jnp
+from flax.traverse_util import flatten_dict, unflatten_dict
+from jax.random import PRNGKey
+
+from ..utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+def rename_key(key):
+ regex = r"\w+[.]\d+"
+ pats = re.findall(regex, key)
+ for pat in pats:
+ key = key.replace(pat, "_".join(pat.split(".")))
+ return key
+
+
+#####################
+# PyTorch => Flax #
+#####################
+
+
+# Adapted from https://github.com/huggingface/transformers/blob/c603c80f46881ae18b2ca50770ef65fa4033eacd/src/transformers/modeling_flax_pytorch_utils.py#L69
+# and https://github.com/patil-suraj/stable-diffusion-jax/blob/main/stable_diffusion_jax/convert_diffusers_to_jax.py
+def rename_key_and_reshape_tensor(pt_tuple_key, pt_tensor, random_flax_state_dict):
+ """Rename PT weight names to corresponding Flax weight names and reshape tensor if necessary"""
+ # conv norm or layer norm
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)
+
+ # rename attention layers
+ if len(pt_tuple_key) > 1:
+ for rename_from, rename_to in (
+ ("to_out_0", "proj_attn"),
+ ("to_k", "key"),
+ ("to_v", "value"),
+ ("to_q", "query"),
+ ):
+ if pt_tuple_key[-2] == rename_from:
+ weight_name = pt_tuple_key[-1]
+ weight_name = "kernel" if weight_name == "weight" else weight_name
+ renamed_pt_tuple_key = pt_tuple_key[:-2] + (rename_to, weight_name)
+ if renamed_pt_tuple_key in random_flax_state_dict:
+ assert random_flax_state_dict[renamed_pt_tuple_key].shape == pt_tensor.T.shape
+ return renamed_pt_tuple_key, pt_tensor.T
+
+ if (
+ any("norm" in str_ for str_ in pt_tuple_key)
+ and (pt_tuple_key[-1] == "bias")
+ and (pt_tuple_key[:-1] + ("bias",) not in random_flax_state_dict)
+ and (pt_tuple_key[:-1] + ("scale",) in random_flax_state_dict)
+ ):
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)
+ return renamed_pt_tuple_key, pt_tensor
+ elif pt_tuple_key[-1] in ["weight", "gamma"] and pt_tuple_key[:-1] + ("scale",) in random_flax_state_dict:
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("scale",)
+ return renamed_pt_tuple_key, pt_tensor
+
+ # embedding
+ if pt_tuple_key[-1] == "weight" and pt_tuple_key[:-1] + ("embedding",) in random_flax_state_dict:
+ pt_tuple_key = pt_tuple_key[:-1] + ("embedding",)
+ return renamed_pt_tuple_key, pt_tensor
+
+ # conv layer
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("kernel",)
+ if pt_tuple_key[-1] == "weight" and pt_tensor.ndim == 4:
+ pt_tensor = pt_tensor.transpose(2, 3, 1, 0)
+ return renamed_pt_tuple_key, pt_tensor
+
+ # linear layer
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("kernel",)
+ if pt_tuple_key[-1] == "weight":
+ pt_tensor = pt_tensor.T
+ return renamed_pt_tuple_key, pt_tensor
+
+ # old PyTorch layer norm weight
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("weight",)
+ if pt_tuple_key[-1] == "gamma":
+ return renamed_pt_tuple_key, pt_tensor
+
+ # old PyTorch layer norm bias
+ renamed_pt_tuple_key = pt_tuple_key[:-1] + ("bias",)
+ if pt_tuple_key[-1] == "beta":
+ return renamed_pt_tuple_key, pt_tensor
+
+ return pt_tuple_key, pt_tensor
+
+
+def convert_pytorch_state_dict_to_flax(pt_state_dict, flax_model, init_key=42):
+ # Step 1: Convert pytorch tensor to numpy
+ pt_state_dict = {k: v.numpy() for k, v in pt_state_dict.items()}
+
+ # Step 2: Since the model is stateless, get random Flax params
+ random_flax_params = flax_model.init_weights(PRNGKey(init_key))
+
+ random_flax_state_dict = flatten_dict(random_flax_params)
+ flax_state_dict = {}
+
+ # Need to change some parameters name to match Flax names
+ for pt_key, pt_tensor in pt_state_dict.items():
+ renamed_pt_key = rename_key(pt_key)
+ pt_tuple_key = tuple(renamed_pt_key.split("."))
+
+ # Correctly rename weight parameters
+ flax_key, flax_tensor = rename_key_and_reshape_tensor(pt_tuple_key, pt_tensor, random_flax_state_dict)
+
+ if flax_key in random_flax_state_dict:
+ if flax_tensor.shape != random_flax_state_dict[flax_key].shape:
+ raise ValueError(
+ f"PyTorch checkpoint seems to be incorrect. Weight {pt_key} was expected to be of shape "
+ f"{random_flax_state_dict[flax_key].shape}, but is {flax_tensor.shape}."
+ )
+
+ # also add unexpected weight so that warning is thrown
+ flax_state_dict[flax_key] = jnp.asarray(flax_tensor)
+
+ return unflatten_dict(flax_state_dict)
diff --git a/model/diffusers_c/models/modeling_flax_utils.py b/model/diffusers_c/models/modeling_flax_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..34177b66bb10939a5372e4f58db3fee2d4dd8126
--- /dev/null
+++ b/model/diffusers_c/models/modeling_flax_utils.py
@@ -0,0 +1,565 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from pickle import UnpicklingError
+from typing import Any, Dict, Union
+
+import jax
+import jax.numpy as jnp
+import msgpack.exceptions
+from flax.core.frozen_dict import FrozenDict, unfreeze
+from flax.serialization import from_bytes, to_bytes
+from flax.traverse_util import flatten_dict, unflatten_dict
+from huggingface_hub import create_repo, hf_hub_download
+from huggingface_hub.utils import (
+ EntryNotFoundError,
+ RepositoryNotFoundError,
+ RevisionNotFoundError,
+ validate_hf_hub_args,
+)
+from requests import HTTPError
+
+from .. import __version__, is_torch_available
+from ..utils import (
+ CONFIG_NAME,
+ FLAX_WEIGHTS_NAME,
+ HUGGINGFACE_CO_RESOLVE_ENDPOINT,
+ WEIGHTS_NAME,
+ PushToHubMixin,
+ logging,
+)
+from .modeling_flax_pytorch_utils import convert_pytorch_state_dict_to_flax
+
+
+logger = logging.get_logger(__name__)
+
+
+class FlaxModelMixin(PushToHubMixin):
+ r"""
+ Base class for all Flax models.
+
+ [`FlaxModelMixin`] takes care of storing the model configuration and provides methods for loading, downloading and
+ saving models.
+
+ - **config_name** ([`str`]) -- Filename to save a model to when calling [`~FlaxModelMixin.save_pretrained`].
+ """
+
+ config_name = CONFIG_NAME
+ _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
+ _flax_internal_args = ["name", "parent", "dtype"]
+
+ @classmethod
+ def _from_config(cls, config, **kwargs):
+ """
+ All context managers that the model should be initialized under go here.
+ """
+ return cls(config, **kwargs)
+
+ def _cast_floating_to(self, params: Union[Dict, FrozenDict], dtype: jnp.dtype, mask: Any = None) -> Any:
+ """
+ Helper method to cast floating-point values of given parameter `PyTree` to given `dtype`.
+ """
+
+ # taken from https://github.com/deepmind/jmp/blob/3a8318abc3292be38582794dbf7b094e6583b192/jmp/_src/policy.py#L27
+ def conditional_cast(param):
+ if isinstance(param, jnp.ndarray) and jnp.issubdtype(param.dtype, jnp.floating):
+ param = param.astype(dtype)
+ return param
+
+ if mask is None:
+ return jax.tree_map(conditional_cast, params)
+
+ flat_params = flatten_dict(params)
+ flat_mask, _ = jax.tree_flatten(mask)
+
+ for masked, key in zip(flat_mask, flat_params.keys()):
+ if masked:
+ param = flat_params[key]
+ flat_params[key] = conditional_cast(param)
+
+ return unflatten_dict(flat_params)
+
+ def to_bf16(self, params: Union[Dict, FrozenDict], mask: Any = None):
+ r"""
+ Cast the floating-point `params` to `jax.numpy.bfloat16`. This returns a new `params` tree and does not cast
+ the `params` in place.
+
+ This method can be used on a TPU to explicitly convert the model parameters to bfloat16 precision to do full
+ half-precision training or to save weights in bfloat16 for inference in order to save memory and improve speed.
+
+ Arguments:
+ params (`Union[Dict, FrozenDict]`):
+ A `PyTree` of model parameters.
+ mask (`Union[Dict, FrozenDict]`):
+ A `PyTree` with same structure as the `params` tree. The leaves should be booleans. It should be `True`
+ for params you want to cast, and `False` for those you want to skip.
+
+ Examples:
+
+ ```python
+ >>> from diffusers import FlaxUNet2DConditionModel
+
+ >>> # load model
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> # By default, the model parameters will be in fp32 precision, to cast these to bfloat16 precision
+ >>> params = model.to_bf16(params)
+ >>> # If you don't want to cast certain parameters (for example layer norm bias and scale)
+ >>> # then pass the mask as follows
+ >>> from flax import traverse_util
+
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> flat_params = traverse_util.flatten_dict(params)
+ >>> mask = {
+ ... path: (path[-2] != ("LayerNorm", "bias") and path[-2:] != ("LayerNorm", "scale"))
+ ... for path in flat_params
+ ... }
+ >>> mask = traverse_util.unflatten_dict(mask)
+ >>> params = model.to_bf16(params, mask)
+ ```"""
+ return self._cast_floating_to(params, jnp.bfloat16, mask)
+
+ def to_fp32(self, params: Union[Dict, FrozenDict], mask: Any = None):
+ r"""
+ Cast the floating-point `params` to `jax.numpy.float32`. This method can be used to explicitly convert the
+ model parameters to fp32 precision. This returns a new `params` tree and does not cast the `params` in place.
+
+ Arguments:
+ params (`Union[Dict, FrozenDict]`):
+ A `PyTree` of model parameters.
+ mask (`Union[Dict, FrozenDict]`):
+ A `PyTree` with same structure as the `params` tree. The leaves should be booleans. It should be `True`
+ for params you want to cast, and `False` for those you want to skip.
+
+ Examples:
+
+ ```python
+ >>> from diffusers import FlaxUNet2DConditionModel
+
+ >>> # Download model and configuration from huggingface.co
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> # By default, the model params will be in fp32, to illustrate the use of this method,
+ >>> # we'll first cast to fp16 and back to fp32
+ >>> params = model.to_f16(params)
+ >>> # now cast back to fp32
+ >>> params = model.to_fp32(params)
+ ```"""
+ return self._cast_floating_to(params, jnp.float32, mask)
+
+ def to_fp16(self, params: Union[Dict, FrozenDict], mask: Any = None):
+ r"""
+ Cast the floating-point `params` to `jax.numpy.float16`. This returns a new `params` tree and does not cast the
+ `params` in place.
+
+ This method can be used on a GPU to explicitly convert the model parameters to float16 precision to do full
+ half-precision training or to save weights in float16 for inference in order to save memory and improve speed.
+
+ Arguments:
+ params (`Union[Dict, FrozenDict]`):
+ A `PyTree` of model parameters.
+ mask (`Union[Dict, FrozenDict]`):
+ A `PyTree` with same structure as the `params` tree. The leaves should be booleans. It should be `True`
+ for params you want to cast, and `False` for those you want to skip.
+
+ Examples:
+
+ ```python
+ >>> from diffusers import FlaxUNet2DConditionModel
+
+ >>> # load model
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> # By default, the model params will be in fp32, to cast these to float16
+ >>> params = model.to_fp16(params)
+ >>> # If you want don't want to cast certain parameters (for example layer norm bias and scale)
+ >>> # then pass the mask as follows
+ >>> from flax import traverse_util
+
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> flat_params = traverse_util.flatten_dict(params)
+ >>> mask = {
+ ... path: (path[-2] != ("LayerNorm", "bias") and path[-2:] != ("LayerNorm", "scale"))
+ ... for path in flat_params
+ ... }
+ >>> mask = traverse_util.unflatten_dict(mask)
+ >>> params = model.to_fp16(params, mask)
+ ```"""
+ return self._cast_floating_to(params, jnp.float16, mask)
+
+ def init_weights(self, rng: jax.Array) -> Dict:
+ raise NotImplementedError(f"init_weights method has to be implemented for {self}")
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(
+ cls,
+ pretrained_model_name_or_path: Union[str, os.PathLike],
+ dtype: jnp.dtype = jnp.float32,
+ *model_args,
+ **kwargs,
+ ):
+ r"""
+ Instantiate a pretrained Flax model from a pretrained model configuration.
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`):
+ Can be either:
+
+ - A string, the *model id* (for example `runwayml/stable-diffusion-v1-5`) of a pretrained model
+ hosted on the Hub.
+ - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+ using [`~FlaxModelMixin.save_pretrained`].
+ dtype (`jax.numpy.dtype`, *optional*, defaults to `jax.numpy.float32`):
+ The data type of the computation. Can be one of `jax.numpy.float32`, `jax.numpy.float16` (on GPUs) and
+ `jax.numpy.bfloat16` (on TPUs).
+
+ This can be used to enable mixed-precision training or half-precision inference on GPUs or TPUs. If
+ specified, all the computation will be performed with the given `dtype`.
+
+
+
+ This only specifies the dtype of the *computation* and does not influence the dtype of model
+ parameters.
+
+ If you wish to change the dtype of the model parameters, see [`~FlaxModelMixin.to_fp16`] and
+ [`~FlaxModelMixin.to_bf16`].
+
+
+
+ model_args (sequence of positional arguments, *optional*):
+ All remaining positional arguments are passed to the underlying model's `__init__` method.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ local_files_only(`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ from_pt (`bool`, *optional*, defaults to `False`):
+ Load the model weights from a PyTorch checkpoint save file.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to update the configuration object (after it is loaded) and initiate the model (for
+ example, `output_attentions=True`). Behaves differently depending on whether a `config` is provided or
+ automatically loaded:
+
+ - If a configuration is provided with `config`, `kwargs` are directly passed to the underlying
+ model's `__init__` method (we assume all relevant updates to the configuration have already been
+ done).
+ - If a configuration is not provided, `kwargs` are first passed to the configuration class
+ initialization function [`~ConfigMixin.from_config`]. Each key of the `kwargs` that corresponds
+ to a configuration attribute is used to override said attribute with the supplied `kwargs` value.
+ Remaining keys that do not correspond to any configuration attribute are passed to the underlying
+ model's `__init__` function.
+
+ Examples:
+
+ ```python
+ >>> from diffusers import FlaxUNet2DConditionModel
+
+ >>> # Download model and configuration from huggingface.co and cache.
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> # Model was saved using *save_pretrained('./test/saved_model/')* (for example purposes, not runnable).
+ >>> model, params = FlaxUNet2DConditionModel.from_pretrained("./test/saved_model/")
+ ```
+
+ If you get the error message below, you need to finetune the weights for your downstream task:
+
+ ```bash
+ Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+ - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+ ```
+ """
+ config = kwargs.pop("config", None)
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ from_pt = kwargs.pop("from_pt", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ local_files_only = kwargs.pop("local_files_only", False)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ subfolder = kwargs.pop("subfolder", None)
+
+ user_agent = {
+ "diffusers": __version__,
+ "file_type": "model",
+ "framework": "flax",
+ }
+
+ # Load config if we don't provide one
+ if config is None:
+ config, unused_kwargs = cls.load_config(
+ pretrained_model_name_or_path,
+ cache_dir=cache_dir,
+ return_unused_kwargs=True,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ **kwargs,
+ )
+
+ model, model_kwargs = cls.from_config(config, dtype=dtype, return_unused_kwargs=True, **unused_kwargs)
+
+ # Load model
+ pretrained_path_with_subfolder = (
+ pretrained_model_name_or_path
+ if subfolder is None
+ else os.path.join(pretrained_model_name_or_path, subfolder)
+ )
+ if os.path.isdir(pretrained_path_with_subfolder):
+ if from_pt:
+ if not os.path.isfile(os.path.join(pretrained_path_with_subfolder, WEIGHTS_NAME)):
+ raise EnvironmentError(
+ f"Error no file named {WEIGHTS_NAME} found in directory {pretrained_path_with_subfolder} "
+ )
+ model_file = os.path.join(pretrained_path_with_subfolder, WEIGHTS_NAME)
+ elif os.path.isfile(os.path.join(pretrained_path_with_subfolder, FLAX_WEIGHTS_NAME)):
+ # Load from a Flax checkpoint
+ model_file = os.path.join(pretrained_path_with_subfolder, FLAX_WEIGHTS_NAME)
+ # Check if pytorch weights exist instead
+ elif os.path.isfile(os.path.join(pretrained_path_with_subfolder, WEIGHTS_NAME)):
+ raise EnvironmentError(
+ f"{WEIGHTS_NAME} file found in directory {pretrained_path_with_subfolder}. Please load the model"
+ " using `from_pt=True`."
+ )
+ else:
+ raise EnvironmentError(
+ f"Error no file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME} found in directory "
+ f"{pretrained_path_with_subfolder}."
+ )
+ else:
+ try:
+ model_file = hf_hub_download(
+ pretrained_model_name_or_path,
+ filename=FLAX_WEIGHTS_NAME if not from_pt else WEIGHTS_NAME,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ proxies=proxies,
+ resume_download=resume_download,
+ local_files_only=local_files_only,
+ token=token,
+ user_agent=user_agent,
+ subfolder=subfolder,
+ revision=revision,
+ )
+
+ except RepositoryNotFoundError:
+ raise EnvironmentError(
+ f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier "
+ "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a "
+ "token having permission to this repo with `token` or log in with `huggingface-cli "
+ "login`."
+ )
+ except RevisionNotFoundError:
+ raise EnvironmentError(
+ f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for "
+ "this model name. Check the model page at "
+ f"'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions."
+ )
+ except EntryNotFoundError:
+ raise EnvironmentError(
+ f"{pretrained_model_name_or_path} does not appear to have a file named {FLAX_WEIGHTS_NAME}."
+ )
+ except HTTPError as err:
+ raise EnvironmentError(
+ f"There was a specific connection error when trying to load {pretrained_model_name_or_path}:\n"
+ f"{err}"
+ )
+ except ValueError:
+ raise EnvironmentError(
+ f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it"
+ f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a"
+ f" directory containing a file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME}.\nCheckout your"
+ " internet connection or see how to run the library in offline mode at"
+ " 'https://huggingface.co/docs/transformers/installation#offline-mode'."
+ )
+ except EnvironmentError:
+ raise EnvironmentError(
+ f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it from "
+ "'https://huggingface.co/models', make sure you don't have a local directory with the same name. "
+ f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory "
+ f"containing a file named {FLAX_WEIGHTS_NAME} or {WEIGHTS_NAME}."
+ )
+
+ if from_pt:
+ if is_torch_available():
+ from .modeling_utils import load_state_dict
+ else:
+ raise EnvironmentError(
+ "Can't load the model in PyTorch format because PyTorch is not installed. "
+ "Please, install PyTorch or use native Flax weights."
+ )
+
+ # Step 1: Get the pytorch file
+ pytorch_model_file = load_state_dict(model_file)
+
+ # Step 2: Convert the weights
+ state = convert_pytorch_state_dict_to_flax(pytorch_model_file, model)
+ else:
+ try:
+ with open(model_file, "rb") as state_f:
+ state = from_bytes(cls, state_f.read())
+ except (UnpicklingError, msgpack.exceptions.ExtraData) as e:
+ try:
+ with open(model_file) as f:
+ if f.read().startswith("version"):
+ raise OSError(
+ "You seem to have cloned a repository without having git-lfs installed. Please"
+ " install git-lfs and run `git lfs install` followed by `git lfs pull` in the"
+ " folder you cloned."
+ )
+ else:
+ raise ValueError from e
+ except (UnicodeDecodeError, ValueError):
+ raise EnvironmentError(f"Unable to convert {model_file} to Flax deserializable object. ")
+ # make sure all arrays are stored as jnp.ndarray
+ # NOTE: This is to prevent a bug this will be fixed in Flax >= v0.3.4:
+ # https://github.com/google/flax/issues/1261
+ state = jax.tree_util.tree_map(lambda x: jax.device_put(x, jax.local_devices(backend="cpu")[0]), state)
+
+ # flatten dicts
+ state = flatten_dict(state)
+
+ params_shape_tree = jax.eval_shape(model.init_weights, rng=jax.random.PRNGKey(0))
+ required_params = set(flatten_dict(unfreeze(params_shape_tree)).keys())
+
+ shape_state = flatten_dict(unfreeze(params_shape_tree))
+
+ missing_keys = required_params - set(state.keys())
+ unexpected_keys = set(state.keys()) - required_params
+
+ if missing_keys:
+ logger.warning(
+ f"The checkpoint {pretrained_model_name_or_path} is missing required keys: {missing_keys}. "
+ "Make sure to call model.init_weights to initialize the missing weights."
+ )
+ cls._missing_keys = missing_keys
+
+ for key in state.keys():
+ if key in shape_state and state[key].shape != shape_state[key].shape:
+ raise ValueError(
+ f"Trying to load the pretrained weight for {key} failed: checkpoint has shape "
+ f"{state[key].shape} which is incompatible with the model shape {shape_state[key].shape}. "
+ )
+
+ # remove unexpected keys to not be saved again
+ for unexpected_key in unexpected_keys:
+ del state[unexpected_key]
+
+ if len(unexpected_keys) > 0:
+ logger.warning(
+ f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
+ f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
+ f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task or"
+ " with another architecture."
+ )
+ else:
+ logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
+
+ if len(missing_keys) > 0:
+ logger.warning(
+ f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
+ f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
+ " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
+ )
+ else:
+ logger.info(
+ f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
+ f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the checkpoint"
+ f" was trained on, you can already use {model.__class__.__name__} for predictions without further"
+ " training."
+ )
+
+ return model, unflatten_dict(state)
+
+ def save_pretrained(
+ self,
+ save_directory: Union[str, os.PathLike],
+ params: Union[Dict, FrozenDict],
+ is_main_process: bool = True,
+ push_to_hub: bool = False,
+ **kwargs,
+ ):
+ """
+ Save a model and its configuration file to a directory so that it can be reloaded using the
+ [`~FlaxModelMixin.from_pretrained`] class method.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to save a model and its configuration file to. Will be created if it doesn't exist.
+ params (`Union[Dict, FrozenDict]`):
+ A `PyTree` of model parameters.
+ is_main_process (`bool`, *optional*, defaults to `True`):
+ Whether the process calling this is the main process or not. Useful during distributed training and you
+ need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+ process to avoid race conditions.
+ push_to_hub (`bool`, *optional*, defaults to `False`):
+ Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the
+ repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+ namespace).
+ kwargs (`Dict[str, Any]`, *optional*):
+ Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+ """
+ if os.path.isfile(save_directory):
+ logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+ return
+
+ os.makedirs(save_directory, exist_ok=True)
+
+ if push_to_hub:
+ commit_message = kwargs.pop("commit_message", None)
+ private = kwargs.pop("private", False)
+ create_pr = kwargs.pop("create_pr", False)
+ token = kwargs.pop("token", None)
+ repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
+ repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id
+
+ model_to_save = self
+
+ # Attach architecture to the config
+ # Save the config
+ if is_main_process:
+ model_to_save.save_config(save_directory)
+
+ # save model
+ output_model_file = os.path.join(save_directory, FLAX_WEIGHTS_NAME)
+ with open(output_model_file, "wb") as f:
+ model_bytes = to_bytes(params)
+ f.write(model_bytes)
+
+ logger.info(f"Model weights saved in {output_model_file}")
+
+ if push_to_hub:
+ self._upload_folder(
+ save_directory,
+ repo_id,
+ token=token,
+ commit_message=commit_message,
+ create_pr=create_pr,
+ )
diff --git a/model/diffusers_c/models/modeling_outputs.py b/model/diffusers_c/models/modeling_outputs.py
new file mode 100644
index 0000000000000000000000000000000000000000..8dfee5fec1819b93316266c81402f36e8225321a
--- /dev/null
+++ b/model/diffusers_c/models/modeling_outputs.py
@@ -0,0 +1,17 @@
+from dataclasses import dataclass
+
+from ..utils import BaseOutput
+
+
+@dataclass
+class AutoencoderKLOutput(BaseOutput):
+ """
+ Output of AutoencoderKL encoding method.
+
+ Args:
+ latent_dist (`DiagonalGaussianDistribution`):
+ Encoded outputs of `Encoder` represented as the mean and logvar of `DiagonalGaussianDistribution`.
+ `DiagonalGaussianDistribution` allows for sampling latents from the distribution.
+ """
+
+ latent_dist: "DiagonalGaussianDistribution" # noqa: F821
diff --git a/model/diffusers_c/models/modeling_pytorch_flax_utils.py b/model/diffusers_c/models/modeling_pytorch_flax_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..eed2c79f6d509d4fa28601d21c65b8fe5fed62ae
--- /dev/null
+++ b/model/diffusers_c/models/modeling_pytorch_flax_utils.py
@@ -0,0 +1,160 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch - Flax general utilities."""
+
+from pickle import UnpicklingError
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.serialization import from_bytes
+from flax.traverse_util import flatten_dict
+
+from ..utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+#####################
+# Flax => PyTorch #
+#####################
+
+
+# from https://github.com/huggingface/transformers/blob/main/src/transformers/modeling_flax_pytorch_utils.py#L224-L352
+def load_flax_checkpoint_in_pytorch_model(pt_model, model_file):
+ try:
+ with open(model_file, "rb") as flax_state_f:
+ flax_state = from_bytes(None, flax_state_f.read())
+ except UnpicklingError as e:
+ try:
+ with open(model_file) as f:
+ if f.read().startswith("version"):
+ raise OSError(
+ "You seem to have cloned a repository without having git-lfs installed. Please"
+ " install git-lfs and run `git lfs install` followed by `git lfs pull` in the"
+ " folder you cloned."
+ )
+ else:
+ raise ValueError from e
+ except (UnicodeDecodeError, ValueError):
+ raise EnvironmentError(f"Unable to convert {model_file} to Flax deserializable object. ")
+
+ return load_flax_weights_in_pytorch_model(pt_model, flax_state)
+
+
+def load_flax_weights_in_pytorch_model(pt_model, flax_state):
+ """Load flax checkpoints in a PyTorch model"""
+
+ try:
+ import torch # noqa: F401
+ except ImportError:
+ logger.error(
+ "Loading Flax weights in PyTorch requires both PyTorch and Flax to be installed. Please see"
+ " https://pytorch.org/ and https://flax.readthedocs.io/en/latest/installation.html for installation"
+ " instructions."
+ )
+ raise
+
+ # check if we have bf16 weights
+ is_type_bf16 = flatten_dict(jax.tree_util.tree_map(lambda x: x.dtype == jnp.bfloat16, flax_state)).values()
+ if any(is_type_bf16):
+ # convert all weights to fp32 if they are bf16 since torch.from_numpy can-not handle bf16
+
+ # and bf16 is not fully supported in PT yet.
+ logger.warning(
+ "Found ``bfloat16`` weights in Flax model. Casting all ``bfloat16`` weights to ``float32`` "
+ "before loading those in PyTorch model."
+ )
+ flax_state = jax.tree_util.tree_map(
+ lambda params: params.astype(np.float32) if params.dtype == jnp.bfloat16 else params, flax_state
+ )
+
+ pt_model.base_model_prefix = ""
+
+ flax_state_dict = flatten_dict(flax_state, sep=".")
+ pt_model_dict = pt_model.state_dict()
+
+ # keep track of unexpected & missing keys
+ unexpected_keys = []
+ missing_keys = set(pt_model_dict.keys())
+
+ for flax_key_tuple, flax_tensor in flax_state_dict.items():
+ flax_key_tuple_array = flax_key_tuple.split(".")
+
+ if flax_key_tuple_array[-1] == "kernel" and flax_tensor.ndim == 4:
+ flax_key_tuple_array = flax_key_tuple_array[:-1] + ["weight"]
+ flax_tensor = jnp.transpose(flax_tensor, (3, 2, 0, 1))
+ elif flax_key_tuple_array[-1] == "kernel":
+ flax_key_tuple_array = flax_key_tuple_array[:-1] + ["weight"]
+ flax_tensor = flax_tensor.T
+ elif flax_key_tuple_array[-1] == "scale":
+ flax_key_tuple_array = flax_key_tuple_array[:-1] + ["weight"]
+
+ if "time_embedding" not in flax_key_tuple_array:
+ for i, flax_key_tuple_string in enumerate(flax_key_tuple_array):
+ flax_key_tuple_array[i] = (
+ flax_key_tuple_string.replace("_0", ".0")
+ .replace("_1", ".1")
+ .replace("_2", ".2")
+ .replace("_3", ".3")
+ .replace("_4", ".4")
+ .replace("_5", ".5")
+ .replace("_6", ".6")
+ .replace("_7", ".7")
+ .replace("_8", ".8")
+ .replace("_9", ".9")
+ )
+
+ flax_key = ".".join(flax_key_tuple_array)
+
+ if flax_key in pt_model_dict:
+ if flax_tensor.shape != pt_model_dict[flax_key].shape:
+ raise ValueError(
+ f"Flax checkpoint seems to be incorrect. Weight {flax_key_tuple} was expected "
+ f"to be of shape {pt_model_dict[flax_key].shape}, but is {flax_tensor.shape}."
+ )
+ else:
+ # add weight to pytorch dict
+ flax_tensor = np.asarray(flax_tensor) if not isinstance(flax_tensor, np.ndarray) else flax_tensor
+ pt_model_dict[flax_key] = torch.from_numpy(flax_tensor)
+ # remove from missing keys
+ missing_keys.remove(flax_key)
+ else:
+ # weight is not expected by PyTorch model
+ unexpected_keys.append(flax_key)
+
+ pt_model.load_state_dict(pt_model_dict)
+
+ # re-transform missing_keys to list
+ missing_keys = list(missing_keys)
+
+ if len(unexpected_keys) > 0:
+ logger.warning(
+ "Some weights of the Flax model were not used when initializing the PyTorch model"
+ f" {pt_model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are initializing"
+ f" {pt_model.__class__.__name__} from a Flax model trained on another task or with another architecture"
+ " (e.g. initializing a BertForSequenceClassification model from a FlaxBertForPreTraining model).\n- This"
+ f" IS NOT expected if you are initializing {pt_model.__class__.__name__} from a Flax model that you expect"
+ " to be exactly identical (e.g. initializing a BertForSequenceClassification model from a"
+ " FlaxBertForSequenceClassification model)."
+ )
+ if len(missing_keys) > 0:
+ logger.warning(
+ f"Some weights of {pt_model.__class__.__name__} were not initialized from the Flax model and are newly"
+ f" initialized: {missing_keys}\nYou should probably TRAIN this model on a down-stream task to be able to"
+ " use it for predictions and inference."
+ )
+
+ return pt_model
diff --git a/model/diffusers_c/models/modeling_utils.py b/model/diffusers_c/models/modeling_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..19ecef4a130aff97235d866980b251801e955ae6
--- /dev/null
+++ b/model/diffusers_c/models/modeling_utils.py
@@ -0,0 +1,1020 @@
+# Copyright 2024 The HuggingFace Inc. team.
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import itertools
+import os
+import re
+from collections import OrderedDict
+from functools import partial
+from typing import Any, Callable, List, Optional, Tuple, Union
+
+import safetensors
+import torch
+from huggingface_hub import create_repo
+from huggingface_hub.utils import validate_hf_hub_args
+from torch import Tensor, nn
+
+from .. import __version__
+from ..utils import (
+ CONFIG_NAME,
+ FLAX_WEIGHTS_NAME,
+ SAFETENSORS_FILE_EXTENSION,
+ SAFETENSORS_WEIGHTS_NAME,
+ WEIGHTS_NAME,
+ _add_variant,
+ _get_model_file,
+ deprecate,
+ is_accelerate_available,
+ is_torch_version,
+ logging,
+)
+from ..utils.hub_utils import PushToHubMixin, load_or_create_model_card, populate_model_card
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_torch_version(">=", "1.9.0"):
+ _LOW_CPU_MEM_USAGE_DEFAULT = True
+else:
+ _LOW_CPU_MEM_USAGE_DEFAULT = False
+
+
+if is_accelerate_available():
+ import accelerate
+ from accelerate.utils import set_module_tensor_to_device
+ from accelerate.utils.versions import is_torch_version
+
+
+def get_parameter_device(parameter: torch.nn.Module) -> torch.device:
+ try:
+ parameters_and_buffers = itertools.chain(parameter.parameters(), parameter.buffers())
+ return next(parameters_and_buffers).device
+ except StopIteration:
+ # For torch.nn.DataParallel compatibility in PyTorch 1.5
+
+ def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]:
+ tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
+ return tuples
+
+ gen = parameter._named_members(get_members_fn=find_tensor_attributes)
+ first_tuple = next(gen)
+ return first_tuple[1].device
+
+
+def get_parameter_dtype(parameter: torch.nn.Module) -> torch.dtype:
+ try:
+ params = tuple(parameter.parameters())
+ if len(params) > 0:
+ return params[0].dtype
+
+ buffers = tuple(parameter.buffers())
+ if len(buffers) > 0:
+ return buffers[0].dtype
+
+ except StopIteration:
+ # For torch.nn.DataParallel compatibility in PyTorch 1.5
+
+ def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]:
+ tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)]
+ return tuples
+
+ gen = parameter._named_members(get_members_fn=find_tensor_attributes)
+ first_tuple = next(gen)
+ return first_tuple[1].dtype
+
+
+def load_state_dict(checkpoint_file: Union[str, os.PathLike], variant: Optional[str] = None):
+ """
+ Reads a checkpoint file, returning properly formatted errors if they arise.
+ """
+ try:
+ file_extension = os.path.basename(checkpoint_file).split(".")[-1]
+ if file_extension == SAFETENSORS_FILE_EXTENSION:
+ return safetensors.torch.load_file(checkpoint_file, device="cpu")
+ else:
+ return torch.load(checkpoint_file, map_location="cpu")
+ except Exception as e:
+ try:
+ with open(checkpoint_file) as f:
+ if f.read().startswith("version"):
+ raise OSError(
+ "You seem to have cloned a repository without having git-lfs installed. Please install "
+ "git-lfs and run `git lfs install` followed by `git lfs pull` in the folder "
+ "you cloned."
+ )
+ else:
+ raise ValueError(
+ f"Unable to locate the file {checkpoint_file} which is necessary to load this pretrained "
+ "model. Make sure you have saved the model properly."
+ ) from e
+ except (UnicodeDecodeError, ValueError):
+ raise OSError(
+ f"Unable to load weights from checkpoint file for '{checkpoint_file}' " f"at '{checkpoint_file}'. "
+ )
+
+
+def load_model_dict_into_meta(
+ model,
+ state_dict: OrderedDict,
+ device: Optional[Union[str, torch.device]] = None,
+ dtype: Optional[Union[str, torch.dtype]] = None,
+ model_name_or_path: Optional[str] = None,
+) -> List[str]:
+ device = device or torch.device("cpu")
+ dtype = dtype or torch.float32
+
+ accepts_dtype = "dtype" in set(inspect.signature(set_module_tensor_to_device).parameters.keys())
+
+ unexpected_keys = []
+ empty_state_dict = model.state_dict()
+ for param_name, param in state_dict.items():
+ if param_name not in empty_state_dict:
+ unexpected_keys.append(param_name)
+ continue
+
+ if empty_state_dict[param_name].shape != param.shape:
+ model_name_or_path_str = f"{model_name_or_path} " if model_name_or_path is not None else ""
+ raise ValueError(
+ f"Cannot load {model_name_or_path_str}because {param_name} expected shape {empty_state_dict[param_name]}, but got {param.shape}. If you want to instead overwrite randomly initialized weights, please make sure to pass both `low_cpu_mem_usage=False` and `ignore_mismatched_sizes=True`. For more information, see also: https://github.com/huggingface/diffusers/issues/1619#issuecomment-1345604389 as an example."
+ )
+
+ if accepts_dtype:
+ set_module_tensor_to_device(model, param_name, device, value=param, dtype=dtype)
+ else:
+ set_module_tensor_to_device(model, param_name, device, value=param)
+ return unexpected_keys
+
+
+def _load_state_dict_into_model(model_to_load, state_dict: OrderedDict) -> List[str]:
+ # Convert old format to new format if needed from a PyTorch state_dict
+ # copy state_dict so _load_from_state_dict can modify it
+ state_dict = state_dict.copy()
+ error_msgs = []
+
+ # PyTorch's `_load_from_state_dict` does not copy parameters in a module's descendants
+ # so we need to apply the function recursively.
+ def load(module: torch.nn.Module, prefix: str = ""):
+ args = (state_dict, prefix, {}, True, [], [], error_msgs)
+ module._load_from_state_dict(*args)
+
+ for name, child in module._modules.items():
+ if child is not None:
+ load(child, prefix + name + ".")
+
+ load(model_to_load)
+
+ return error_msgs
+
+
+class ModelMixin(torch.nn.Module, PushToHubMixin):
+ r"""
+ Base class for all models.
+
+ [`ModelMixin`] takes care of storing the model configuration and provides methods for loading, downloading and
+ saving models.
+
+ - **config_name** ([`str`]) -- Filename to save a model to when calling [`~models.ModelMixin.save_pretrained`].
+ """
+
+ config_name = CONFIG_NAME
+ _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
+ _supports_gradient_checkpointing = False
+ _keys_to_ignore_on_load_unexpected = None
+
+ def __init__(self):
+ super().__init__()
+
+ def __getattr__(self, name: str) -> Any:
+ """The only reason we overwrite `getattr` here is to gracefully deprecate accessing
+ config attributes directly. See https://github.com/huggingface/diffusers/pull/3129 We need to overwrite
+ __getattr__ here in addition so that we don't trigger `torch.nn.Module`'s __getattr__':
+ https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
+ """
+
+ is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)
+ is_attribute = name in self.__dict__
+
+ if is_in_config and not is_attribute:
+ deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'unet.config.{name}'."
+ deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False, stacklevel=3)
+ return self._internal_dict[name]
+
+ # call PyTorch's https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
+ return super().__getattr__(name)
+
+ @property
+ def is_gradient_checkpointing(self) -> bool:
+ """
+ Whether gradient checkpointing is activated for this model or not.
+ """
+ return any(hasattr(m, "gradient_checkpointing") and m.gradient_checkpointing for m in self.modules())
+
+ def enable_gradient_checkpointing(self) -> None:
+ """
+ Activates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
+ *checkpoint activations* in other frameworks).
+ """
+ if not self._supports_gradient_checkpointing:
+ raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
+ self.apply(partial(self._set_gradient_checkpointing, value=True))
+
+ def disable_gradient_checkpointing(self) -> None:
+ """
+ Deactivates gradient checkpointing for the current model (may be referred to as *activation checkpointing* or
+ *checkpoint activations* in other frameworks).
+ """
+ if self._supports_gradient_checkpointing:
+ self.apply(partial(self._set_gradient_checkpointing, value=False))
+
+ def set_use_memory_efficient_attention_xformers(
+ self, valid: bool, attention_op: Optional[Callable] = None
+ ) -> None:
+ # Recursively walk through all the children.
+ # Any children which exposes the set_use_memory_efficient_attention_xformers method
+ # gets the message
+ def fn_recursive_set_mem_eff(module: torch.nn.Module):
+ if hasattr(module, "set_use_memory_efficient_attention_xformers"):
+ module.set_use_memory_efficient_attention_xformers(valid, attention_op)
+
+ for child in module.children():
+ fn_recursive_set_mem_eff(child)
+
+ for module in self.children():
+ if isinstance(module, torch.nn.Module):
+ fn_recursive_set_mem_eff(module)
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None) -> None:
+ r"""
+ Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
+
+ When this option is enabled, you should observe lower GPU memory usage and a potential speed up during
+ inference. Speed up during training is not guaranteed.
+
+
+
+ ⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes
+ precedent.
+
+
+
+ Parameters:
+ attention_op (`Callable`, *optional*):
+ Override the default `None` operator for use as `op` argument to the
+ [`memory_efficient_attention()`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention)
+ function of xFormers.
+
+ Examples:
+
+ ```py
+ >>> import torch
+ >>> from diffusers import UNet2DConditionModel
+ >>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp
+
+ >>> model = UNet2DConditionModel.from_pretrained(
+ ... "stabilityai/stable-diffusion-2-1", subfolder="unet", torch_dtype=torch.float16
+ ... )
+ >>> model = model.to("cuda")
+ >>> model.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
+ ```
+ """
+ self.set_use_memory_efficient_attention_xformers(True, attention_op)
+
+ def disable_xformers_memory_efficient_attention(self) -> None:
+ r"""
+ Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
+ """
+ self.set_use_memory_efficient_attention_xformers(False)
+
+ def save_pretrained(
+ self,
+ save_directory: Union[str, os.PathLike],
+ is_main_process: bool = True,
+ save_function: Optional[Callable] = None,
+ safe_serialization: bool = True,
+ variant: Optional[str] = None,
+ push_to_hub: bool = False,
+ **kwargs,
+ ):
+ """
+ Save a model and its configuration file to a directory so that it can be reloaded using the
+ [`~models.ModelMixin.from_pretrained`] class method.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to save a model and its configuration file to. Will be created if it doesn't exist.
+ is_main_process (`bool`, *optional*, defaults to `True`):
+ Whether the process calling this is the main process or not. Useful during distributed training and you
+ need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+ process to avoid race conditions.
+ save_function (`Callable`):
+ The function to use to save the state dictionary. Useful during distributed training when you need to
+ replace `torch.save` with another method. Can be configured with the environment variable
+ `DIFFUSERS_SAVE_MODE`.
+ safe_serialization (`bool`, *optional*, defaults to `True`):
+ Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+ variant (`str`, *optional*):
+ If specified, weights are saved in the format `pytorch_model..bin`.
+ push_to_hub (`bool`, *optional*, defaults to `False`):
+ Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the
+ repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+ namespace).
+ kwargs (`Dict[str, Any]`, *optional*):
+ Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+ """
+ if os.path.isfile(save_directory):
+ logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+ return
+
+ os.makedirs(save_directory, exist_ok=True)
+
+ if push_to_hub:
+ commit_message = kwargs.pop("commit_message", None)
+ private = kwargs.pop("private", False)
+ create_pr = kwargs.pop("create_pr", False)
+ token = kwargs.pop("token", None)
+ repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
+ repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id
+
+ # Only save the model itself if we are using distributed training
+ model_to_save = self
+
+ # Attach architecture to the config
+ # Save the config
+ if is_main_process:
+ model_to_save.save_config(save_directory)
+
+ # Save the model
+ state_dict = model_to_save.state_dict()
+
+ weights_name = SAFETENSORS_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
+ weights_name = _add_variant(weights_name, variant)
+
+ # Save the model
+ if safe_serialization:
+ safetensors.torch.save_file(
+ state_dict, os.path.join(save_directory, weights_name), metadata={"format": "pt"}
+ )
+ else:
+ torch.save(state_dict, os.path.join(save_directory, weights_name))
+
+ logger.info(f"Model weights saved in {os.path.join(save_directory, weights_name)}")
+
+ if push_to_hub:
+ # Create a new empty model card and eventually tag it
+ model_card = load_or_create_model_card(repo_id, token=token)
+ model_card = populate_model_card(model_card)
+ model_card.save(os.path.join(save_directory, "README.md"))
+
+ self._upload_folder(
+ save_directory,
+ repo_id,
+ token=token,
+ commit_message=commit_message,
+ create_pr=create_pr,
+ )
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
+ r"""
+ Instantiate a pretrained PyTorch model from a pretrained model configuration.
+
+ The model is set in evaluation mode - `model.eval()` - by default, and dropout modules are deactivated. To
+ train the model, set it back in training mode with `model.train()`.
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+ the Hub.
+ - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+ with [`~ModelMixin.save_pretrained`].
+
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
+ dtype is automatically derived from the model's weights.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info (`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only(`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ from_flax (`bool`, *optional*, defaults to `False`):
+ Load the model weights from a Flax checkpoint save file.
+ subfolder (`str`, *optional*, defaults to `""`):
+ The subfolder location of a model file within a larger model repository on the Hub or locally.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+ device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+ A map that specifies where each submodule should go. It doesn't need to be defined for each
+ parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
+ same device.
+
+ Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
+ more information about each option see [designing a device
+ map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+ max_memory (`Dict`, *optional*):
+ A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
+ each GPU and the available CPU RAM if unset.
+ offload_folder (`str` or `os.PathLike`, *optional*):
+ The path to offload weights if `device_map` contains the value `"disk"`.
+ offload_state_dict (`bool`, *optional*):
+ If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
+ the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
+ when there is some disk offload.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ variant (`str`, *optional*):
+ Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
+ loading `from_flax`.
+ use_safetensors (`bool`, *optional*, defaults to `None`):
+ If set to `None`, the `safetensors` weights are downloaded if they're available **and** if the
+ `safetensors` library is installed. If set to `True`, the model is forcibly loaded from `safetensors`
+ weights. If set to `False`, `safetensors` weights are not loaded.
+
+
+
+ To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
+ `huggingface-cli login`. You can also activate the special
+ ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a
+ firewalled environment.
+
+
+
+ Example:
+
+ ```py
+ from diffusers import UNet2DConditionModel
+
+ unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")
+ ```
+
+ If you get the error message below, you need to finetune the weights for your downstream task:
+
+ ```bash
+ Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+ - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+ ```
+ """
+ cache_dir = kwargs.pop("cache_dir", None)
+ ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
+ force_download = kwargs.pop("force_download", False)
+ from_flax = kwargs.pop("from_flax", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ output_loading_info = kwargs.pop("output_loading_info", False)
+ local_files_only = kwargs.pop("local_files_only", None)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ torch_dtype = kwargs.pop("torch_dtype", None)
+ subfolder = kwargs.pop("subfolder", None)
+ device_map = kwargs.pop("device_map", None)
+ max_memory = kwargs.pop("max_memory", None)
+ offload_folder = kwargs.pop("offload_folder", None)
+ offload_state_dict = kwargs.pop("offload_state_dict", False)
+ low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+ variant = kwargs.pop("variant", None)
+ use_safetensors = kwargs.pop("use_safetensors", None)
+
+ allow_pickle = False
+ if use_safetensors is None:
+ use_safetensors = True
+ allow_pickle = True
+
+ if low_cpu_mem_usage and not is_accelerate_available():
+ low_cpu_mem_usage = False
+ logger.warning(
+ "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+ " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+ " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+ " install accelerate\n```\n."
+ )
+
+ if device_map is not None and not is_accelerate_available():
+ raise NotImplementedError(
+ "Loading and dispatching requires `accelerate`. Please make sure to install accelerate or set"
+ " `device_map=None`. You can install accelerate with `pip install accelerate`."
+ )
+
+ # Check if we can handle device_map and dispatching the weights
+ if device_map is not None and not is_torch_version(">=", "1.9.0"):
+ raise NotImplementedError(
+ "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
+ " `device_map=None`."
+ )
+
+ if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+ raise NotImplementedError(
+ "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+ " `low_cpu_mem_usage=False`."
+ )
+
+ if low_cpu_mem_usage is False and device_map is not None:
+ raise ValueError(
+ f"You cannot set `low_cpu_mem_usage` to `False` while using device_map={device_map} for loading and"
+ " dispatching. Please make sure to set `low_cpu_mem_usage=True`."
+ )
+
+ # Load config if we don't provide a configuration
+ config_path = pretrained_model_name_or_path
+
+ user_agent = {
+ "diffusers": __version__,
+ "file_type": "model",
+ "framework": "pytorch",
+ }
+
+ # load config
+ config, unused_kwargs, commit_hash = cls.load_config(
+ config_path,
+ cache_dir=cache_dir,
+ return_unused_kwargs=True,
+ return_commit_hash=True,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ device_map=device_map,
+ max_memory=max_memory,
+ offload_folder=offload_folder,
+ offload_state_dict=offload_state_dict,
+ user_agent=user_agent,
+ **kwargs,
+ )
+
+ # load model
+ model_file = None
+ if from_flax:
+ model_file = _get_model_file(
+ pretrained_model_name_or_path,
+ weights_name=FLAX_WEIGHTS_NAME,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ commit_hash=commit_hash,
+ )
+ model = cls.from_config(config, **unused_kwargs)
+
+ # Convert the weights
+ from .modeling_pytorch_flax_utils import load_flax_checkpoint_in_pytorch_model
+
+ model = load_flax_checkpoint_in_pytorch_model(model, model_file)
+ else:
+ if use_safetensors:
+ try:
+ model_file = _get_model_file(
+ pretrained_model_name_or_path,
+ weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ commit_hash=commit_hash,
+ )
+ except IOError as e:
+ if not allow_pickle:
+ raise e
+ pass
+ if model_file is None:
+ model_file = _get_model_file(
+ pretrained_model_name_or_path,
+ weights_name=_add_variant(WEIGHTS_NAME, variant),
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ subfolder=subfolder,
+ user_agent=user_agent,
+ commit_hash=commit_hash,
+ )
+
+ if low_cpu_mem_usage:
+ # Instantiate model with empty weights
+ with accelerate.init_empty_weights():
+ model = cls.from_config(config, **unused_kwargs)
+
+ # if device_map is None, load the state dict and move the params from meta device to the cpu
+ if device_map is None:
+ param_device = "cpu"
+ state_dict = load_state_dict(model_file, variant=variant)
+ model._convert_deprecated_attention_blocks(state_dict)
+ # move the params from meta device to cpu
+ missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
+ if len(missing_keys) > 0:
+ raise ValueError(
+ f"Cannot load {cls} from {pretrained_model_name_or_path} because the following keys are"
+ f" missing: \n {', '.join(missing_keys)}. \n Please make sure to pass"
+ " `low_cpu_mem_usage=False` and `device_map=None` if you want to randomly initialize"
+ " those weights or else make sure your checkpoint file is correct."
+ )
+
+ unexpected_keys = load_model_dict_into_meta(
+ model,
+ state_dict,
+ device=param_device,
+ dtype=torch_dtype,
+ model_name_or_path=pretrained_model_name_or_path,
+ )
+
+ if cls._keys_to_ignore_on_load_unexpected is not None:
+ for pat in cls._keys_to_ignore_on_load_unexpected:
+ unexpected_keys = [k for k in unexpected_keys if re.search(pat, k) is None]
+
+ if len(unexpected_keys) > 0:
+ logger.warn(
+ f"Some weights of the model checkpoint were not used when initializing {cls.__name__}: \n {[', '.join(unexpected_keys)]}"
+ )
+
+ else: # else let accelerate handle loading and dispatching.
+ # Load weights and dispatch according to the device_map
+ # by default the device_map is None and the weights are loaded on the CPU
+ try:
+ accelerate.load_checkpoint_and_dispatch(
+ model,
+ model_file,
+ device_map,
+ max_memory=max_memory,
+ offload_folder=offload_folder,
+ offload_state_dict=offload_state_dict,
+ dtype=torch_dtype,
+ )
+ except AttributeError as e:
+ # When using accelerate loading, we do not have the ability to load the state
+ # dict and rename the weight names manually. Additionally, accelerate skips
+ # torch loading conventions and directly writes into `module.{_buffers, _parameters}`
+ # (which look like they should be private variables?), so we can't use the standard hooks
+ # to rename parameters on load. We need to mimic the original weight names so the correct
+ # attributes are available. After we have loaded the weights, we convert the deprecated
+ # names to the new non-deprecated names. Then we _greatly encourage_ the user to convert
+ # the weights so we don't have to do this again.
+
+ if "'Attention' object has no attribute" in str(e):
+ logger.warn(
+ f"Taking `{str(e)}` while using `accelerate.load_checkpoint_and_dispatch` to mean {pretrained_model_name_or_path}"
+ " was saved with deprecated attention block weight names. We will load it with the deprecated attention block"
+ " names and convert them on the fly to the new attention block format. Please re-save the model after this conversion,"
+ " so we don't have to do the on the fly renaming in the future. If the model is from a hub checkpoint,"
+ " please also re-upload it or open a PR on the original repository."
+ )
+ model._temp_convert_self_to_deprecated_attention_blocks()
+ accelerate.load_checkpoint_and_dispatch(
+ model,
+ model_file,
+ device_map,
+ max_memory=max_memory,
+ offload_folder=offload_folder,
+ offload_state_dict=offload_state_dict,
+ dtype=torch_dtype,
+ )
+ model._undo_temp_convert_self_to_deprecated_attention_blocks()
+ else:
+ raise e
+
+ loading_info = {
+ "missing_keys": [],
+ "unexpected_keys": [],
+ "mismatched_keys": [],
+ "error_msgs": [],
+ }
+ else:
+ model = cls.from_config(config, **unused_kwargs)
+
+ state_dict = load_state_dict(model_file, variant=variant)
+ model._convert_deprecated_attention_blocks(state_dict)
+
+ model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(
+ model,
+ state_dict,
+ model_file,
+ pretrained_model_name_or_path,
+ ignore_mismatched_sizes=ignore_mismatched_sizes,
+ )
+
+ loading_info = {
+ "missing_keys": missing_keys,
+ "unexpected_keys": unexpected_keys,
+ "mismatched_keys": mismatched_keys,
+ "error_msgs": error_msgs,
+ }
+
+ if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
+ raise ValueError(
+ f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}."
+ )
+ elif torch_dtype is not None:
+ model = model.to(torch_dtype)
+
+ model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+
+ # Set model in evaluation mode to deactivate DropOut modules by default
+ model.eval()
+ if output_loading_info:
+ return model, loading_info
+
+ return model
+
+ @classmethod
+ def _load_pretrained_model(
+ cls,
+ model,
+ state_dict: OrderedDict,
+ resolved_archive_file,
+ pretrained_model_name_or_path: Union[str, os.PathLike],
+ ignore_mismatched_sizes: bool = False,
+ ):
+ # Retrieve missing & unexpected_keys
+ model_state_dict = model.state_dict()
+ loaded_keys = list(state_dict.keys())
+
+ expected_keys = list(model_state_dict.keys())
+
+ original_loaded_keys = loaded_keys
+
+ missing_keys = list(set(expected_keys) - set(loaded_keys))
+ unexpected_keys = list(set(loaded_keys) - set(expected_keys))
+
+ # Make sure we are able to load base models as well as derived models (with heads)
+ model_to_load = model
+
+ def _find_mismatched_keys(
+ state_dict,
+ model_state_dict,
+ loaded_keys,
+ ignore_mismatched_sizes,
+ ):
+ mismatched_keys = []
+ if ignore_mismatched_sizes:
+ for checkpoint_key in loaded_keys:
+ model_key = checkpoint_key
+
+ if (
+ model_key in model_state_dict
+ and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
+ ):
+ mismatched_keys.append(
+ (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
+ )
+ del state_dict[checkpoint_key]
+ return mismatched_keys
+
+ if state_dict is not None:
+ # Whole checkpoint
+ mismatched_keys = _find_mismatched_keys(
+ state_dict,
+ model_state_dict,
+ original_loaded_keys,
+ ignore_mismatched_sizes,
+ )
+ error_msgs = _load_state_dict_into_model(model_to_load, state_dict)
+
+ if len(error_msgs) > 0:
+ error_msg = "\n\t".join(error_msgs)
+ if "size mismatch" in error_msg:
+ error_msg += (
+ "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
+ )
+ raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")
+
+ if len(unexpected_keys) > 0:
+ logger.warning(
+ f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
+ f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
+ f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task"
+ " or with another architecture (e.g. initializing a BertForSequenceClassification model from a"
+ " BertForPreTraining model).\n- This IS NOT expected if you are initializing"
+ f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly"
+ " identical (initializing a BertForSequenceClassification model from a"
+ " BertForSequenceClassification model)."
+ )
+ else:
+ logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
+ if len(missing_keys) > 0:
+ logger.warning(
+ f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
+ f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
+ " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
+ )
+ elif len(mismatched_keys) == 0:
+ logger.info(
+ f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
+ f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the"
+ f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions"
+ " without further training."
+ )
+ if len(mismatched_keys) > 0:
+ mismatched_warning = "\n".join(
+ [
+ f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
+ for key, shape1, shape2 in mismatched_keys
+ ]
+ )
+ logger.warning(
+ f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
+ f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"
+ f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be"
+ " able to use it for predictions and inference."
+ )
+
+ return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs
+
+ @property
+ def device(self) -> torch.device:
+ """
+ `torch.device`: The device on which the module is (assuming that all the module parameters are on the same
+ device).
+ """
+ return get_parameter_device(self)
+
+ @property
+ def dtype(self) -> torch.dtype:
+ """
+ `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype).
+ """
+ return get_parameter_dtype(self)
+
+ def num_parameters(self, only_trainable: bool = False, exclude_embeddings: bool = False) -> int:
+ """
+ Get number of (trainable or non-embedding) parameters in the module.
+
+ Args:
+ only_trainable (`bool`, *optional*, defaults to `False`):
+ Whether or not to return only the number of trainable parameters.
+ exclude_embeddings (`bool`, *optional*, defaults to `False`):
+ Whether or not to return only the number of non-embedding parameters.
+
+ Returns:
+ `int`: The number of parameters.
+
+ Example:
+
+ ```py
+ from diffusers import UNet2DConditionModel
+
+ model_id = "runwayml/stable-diffusion-v1-5"
+ unet = UNet2DConditionModel.from_pretrained(model_id, subfolder="unet")
+ unet.num_parameters(only_trainable=True)
+ 859520964
+ ```
+ """
+
+ if exclude_embeddings:
+ embedding_param_names = [
+ f"{name}.weight"
+ for name, module_type in self.named_modules()
+ if isinstance(module_type, torch.nn.Embedding)
+ ]
+ non_embedding_parameters = [
+ parameter for name, parameter in self.named_parameters() if name not in embedding_param_names
+ ]
+ return sum(p.numel() for p in non_embedding_parameters if p.requires_grad or not only_trainable)
+ else:
+ return sum(p.numel() for p in self.parameters() if p.requires_grad or not only_trainable)
+
+ def _convert_deprecated_attention_blocks(self, state_dict: OrderedDict) -> None:
+ deprecated_attention_block_paths = []
+
+ def recursive_find_attn_block(name, module):
+ if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block:
+ deprecated_attention_block_paths.append(name)
+
+ for sub_name, sub_module in module.named_children():
+ sub_name = sub_name if name == "" else f"{name}.{sub_name}"
+ recursive_find_attn_block(sub_name, sub_module)
+
+ recursive_find_attn_block("", self)
+
+ # NOTE: we have to check if the deprecated parameters are in the state dict
+ # because it is possible we are loading from a state dict that was already
+ # converted
+
+ for path in deprecated_attention_block_paths:
+ # group_norm path stays the same
+
+ # query -> to_q
+ if f"{path}.query.weight" in state_dict:
+ state_dict[f"{path}.to_q.weight"] = state_dict.pop(f"{path}.query.weight")
+ if f"{path}.query.bias" in state_dict:
+ state_dict[f"{path}.to_q.bias"] = state_dict.pop(f"{path}.query.bias")
+
+ # key -> to_k
+ if f"{path}.key.weight" in state_dict:
+ state_dict[f"{path}.to_k.weight"] = state_dict.pop(f"{path}.key.weight")
+ if f"{path}.key.bias" in state_dict:
+ state_dict[f"{path}.to_k.bias"] = state_dict.pop(f"{path}.key.bias")
+
+ # value -> to_v
+ if f"{path}.value.weight" in state_dict:
+ state_dict[f"{path}.to_v.weight"] = state_dict.pop(f"{path}.value.weight")
+ if f"{path}.value.bias" in state_dict:
+ state_dict[f"{path}.to_v.bias"] = state_dict.pop(f"{path}.value.bias")
+
+ # proj_attn -> to_out.0
+ if f"{path}.proj_attn.weight" in state_dict:
+ state_dict[f"{path}.to_out.0.weight"] = state_dict.pop(f"{path}.proj_attn.weight")
+ if f"{path}.proj_attn.bias" in state_dict:
+ state_dict[f"{path}.to_out.0.bias"] = state_dict.pop(f"{path}.proj_attn.bias")
+
+ def _temp_convert_self_to_deprecated_attention_blocks(self) -> None:
+ deprecated_attention_block_modules = []
+
+ def recursive_find_attn_block(module):
+ if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block:
+ deprecated_attention_block_modules.append(module)
+
+ for sub_module in module.children():
+ recursive_find_attn_block(sub_module)
+
+ recursive_find_attn_block(self)
+
+ for module in deprecated_attention_block_modules:
+ module.query = module.to_q
+ module.key = module.to_k
+ module.value = module.to_v
+ module.proj_attn = module.to_out[0]
+
+ # We don't _have_ to delete the old attributes, but it's helpful to ensure
+ # that _all_ the weights are loaded into the new attributes and we're not
+ # making an incorrect assumption that this model should be converted when
+ # it really shouldn't be.
+ del module.to_q
+ del module.to_k
+ del module.to_v
+ del module.to_out
+
+ def _undo_temp_convert_self_to_deprecated_attention_blocks(self) -> None:
+ deprecated_attention_block_modules = []
+
+ def recursive_find_attn_block(module) -> None:
+ if hasattr(module, "_from_deprecated_attn_block") and module._from_deprecated_attn_block:
+ deprecated_attention_block_modules.append(module)
+
+ for sub_module in module.children():
+ recursive_find_attn_block(sub_module)
+
+ recursive_find_attn_block(self)
+
+ for module in deprecated_attention_block_modules:
+ module.to_q = module.query
+ module.to_k = module.key
+ module.to_v = module.value
+ module.to_out = nn.ModuleList([module.proj_attn, nn.Dropout(module.dropout)])
+
+ del module.query
+ del module.key
+ del module.value
+ del module.proj_attn
diff --git a/model/diffusers_c/models/normalization.py b/model/diffusers_c/models/normalization.py
new file mode 100644
index 0000000000000000000000000000000000000000..0c38dc8a933830f00c4353c4b40ca8da7e1e8385
--- /dev/null
+++ b/model/diffusers_c/models/normalization.py
@@ -0,0 +1,253 @@
+# Copyright 2024 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numbers
+from typing import Dict, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..utils import is_torch_version
+from .activations import get_activation
+from .embeddings import CombinedTimestepLabelEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings
+
+
+class AdaLayerNorm(nn.Module):
+ r"""
+ Norm layer modified to incorporate timestep embeddings.
+
+ Parameters:
+ embedding_dim (`int`): The size of each embedding vector.
+ num_embeddings (`int`): The size of the embeddings dictionary.
+ """
+
+ def __init__(self, embedding_dim: int, num_embeddings: int):
+ super().__init__()
+ self.emb = nn.Embedding(num_embeddings, embedding_dim)
+ self.silu = nn.SiLU()
+ self.linear = nn.Linear(embedding_dim, embedding_dim * 2)
+ self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False)
+
+ def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor:
+ emb = self.linear(self.silu(self.emb(timestep)))
+ scale, shift = torch.chunk(emb, 2)
+ x = self.norm(x) * (1 + scale) + shift
+ return x
+
+
+class AdaLayerNormZero(nn.Module):
+ r"""
+ Norm layer adaptive layer norm zero (adaLN-Zero).
+
+ Parameters:
+ embedding_dim (`int`): The size of each embedding vector.
+ num_embeddings (`int`): The size of the embeddings dictionary.
+ """
+
+ def __init__(self, embedding_dim: int, num_embeddings: int):
+ super().__init__()
+
+ self.emb = CombinedTimestepLabelEmbeddings(num_embeddings, embedding_dim)
+
+ self.silu = nn.SiLU()
+ self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
+ self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6)
+
+ def forward(
+ self,
+ x: torch.Tensor,
+ timestep: torch.Tensor,
+ class_labels: torch.LongTensor,
+ hidden_dtype: Optional[torch.dtype] = None,
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+ emb = self.linear(self.silu(self.emb(timestep, class_labels, hidden_dtype=hidden_dtype)))
+ shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1)
+ x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None]
+ return x, gate_msa, shift_mlp, scale_mlp, gate_mlp
+
+
+class AdaLayerNormSingle(nn.Module):
+ r"""
+ Norm layer adaptive layer norm single (adaLN-single).
+
+ As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3).
+
+ Parameters:
+ embedding_dim (`int`): The size of each embedding vector.
+ use_additional_conditions (`bool`): To use additional conditions for normalization or not.
+ """
+
+ def __init__(self, embedding_dim: int, use_additional_conditions: bool = False):
+ super().__init__()
+
+ self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings(
+ embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions
+ )
+
+ self.silu = nn.SiLU()
+ self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
+
+ def forward(
+ self,
+ timestep: torch.Tensor,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ batch_size: Optional[int] = None,
+ hidden_dtype: Optional[torch.dtype] = None,
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+ # No modulation happening here.
+ embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype)
+ return self.linear(self.silu(embedded_timestep)), embedded_timestep
+
+
+class AdaGroupNorm(nn.Module):
+ r"""
+ GroupNorm layer modified to incorporate timestep embeddings.
+
+ Parameters:
+ embedding_dim (`int`): The size of each embedding vector.
+ num_embeddings (`int`): The size of the embeddings dictionary.
+ num_groups (`int`): The number of groups to separate the channels into.
+ act_fn (`str`, *optional*, defaults to `None`): The activation function to use.
+ eps (`float`, *optional*, defaults to `1e-5`): The epsilon value to use for numerical stability.
+ """
+
+ def __init__(
+ self, embedding_dim: int, out_dim: int, num_groups: int, act_fn: Optional[str] = None, eps: float = 1e-5
+ ):
+ super().__init__()
+ self.num_groups = num_groups
+ self.eps = eps
+
+ if act_fn is None:
+ self.act = None
+ else:
+ self.act = get_activation(act_fn)
+
+ self.linear = nn.Linear(embedding_dim, out_dim * 2)
+
+ def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor:
+ if self.act:
+ emb = self.act(emb)
+ emb = self.linear(emb)
+ emb = emb[:, :, None, None]
+ scale, shift = emb.chunk(2, dim=1)
+
+ x = F.group_norm(x, self.num_groups, eps=self.eps)
+ x = x * (1 + scale) + shift
+ return x
+
+
+class AdaLayerNormContinuous(nn.Module):
+ def __init__(
+ self,
+ embedding_dim: int,
+ conditioning_embedding_dim: int,
+ # NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters
+ # because the output is immediately scaled and shifted by the projected conditioning embeddings.
+ # Note that AdaLayerNorm does not let the norm layer have scale and shift parameters.
+ # However, this is how it was implemented in the original code, and it's rather likely you should
+ # set `elementwise_affine` to False.
+ elementwise_affine=True,
+ eps=1e-5,
+ bias=True,
+ norm_type="layer_norm",
+ ):
+ super().__init__()
+ self.silu = nn.SiLU()
+ self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias)
+ if norm_type == "layer_norm":
+ self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias)
+ elif norm_type == "rms_norm":
+ self.norm = RMSNorm(embedding_dim, eps, elementwise_affine)
+ else:
+ raise ValueError(f"unknown norm_type {norm_type}")
+
+ def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
+ emb = self.linear(self.silu(conditioning_embedding))
+ scale, shift = torch.chunk(emb, 2, dim=1)
+ x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+ return x
+
+
+if is_torch_version(">=", "2.1.0"):
+ LayerNorm = nn.LayerNorm
+else:
+ # Has optional bias parameter compared to torch layer norm
+ # TODO: replace with torch layernorm once min required torch version >= 2.1
+ class LayerNorm(nn.Module):
+ def __init__(self, dim, eps: float = 1e-5, elementwise_affine: bool = True, bias: bool = True):
+ super().__init__()
+
+ self.eps = eps
+
+ if isinstance(dim, numbers.Integral):
+ dim = (dim,)
+
+ self.dim = torch.Size(dim)
+
+ if elementwise_affine:
+ self.weight = nn.Parameter(torch.ones(dim))
+ self.bias = nn.Parameter(torch.zeros(dim)) if bias else None
+ else:
+ self.weight = None
+ self.bias = None
+
+ def forward(self, input):
+ return F.layer_norm(input, self.dim, self.weight, self.bias, self.eps)
+
+
+class RMSNorm(nn.Module):
+ def __init__(self, dim, eps: float, elementwise_affine: bool = True):
+ super().__init__()
+
+ self.eps = eps
+
+ if isinstance(dim, numbers.Integral):
+ dim = (dim,)
+
+ self.dim = torch.Size(dim)
+
+ if elementwise_affine:
+ self.weight = nn.Parameter(torch.ones(dim))
+ else:
+ self.weight = None
+
+ def forward(self, hidden_states):
+ input_dtype = hidden_states.dtype
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+ hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
+
+ if self.weight is not None:
+ # convert into half-precision if necessary
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
+ hidden_states = hidden_states.to(self.weight.dtype)
+ hidden_states = hidden_states * self.weight
+ else:
+ hidden_states = hidden_states.to(input_dtype)
+
+ return hidden_states
+
+
+class GlobalResponseNorm(nn.Module):
+ # Taken from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105
+ def __init__(self, dim):
+ super().__init__()
+ self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim))
+ self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim))
+
+ def forward(self, x):
+ gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True)
+ nx = gx / (gx.mean(dim=-1, keepdim=True) + 1e-6)
+ return self.gamma * (x * nx) + self.beta + x
diff --git a/model/diffusers_c/models/prior_transformer.py b/model/diffusers_c/models/prior_transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..328835a9538136564cd811ac0992d5ba7b7e5b83
--- /dev/null
+++ b/model/diffusers_c/models/prior_transformer.py
@@ -0,0 +1,12 @@
+from ..utils import deprecate
+from .transformers.prior_transformer import PriorTransformer, PriorTransformerOutput
+
+
+class PriorTransformerOutput(PriorTransformerOutput):
+ deprecation_message = "Importing `PriorTransformerOutput` from `diffusers.models.prior_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.prior_transformer import PriorTransformerOutput`, instead."
+ deprecate("PriorTransformerOutput", "0.29", deprecation_message)
+
+
+class PriorTransformer(PriorTransformer):
+ deprecation_message = "Importing `PriorTransformer` from `diffusers.models.prior_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.prior_transformer import PriorTransformer`, instead."
+ deprecate("PriorTransformer", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/resnet.py b/model/diffusers_c/models/resnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..84cb31f430a021baf3144e6cdb65caafc20d021b
--- /dev/null
+++ b/model/diffusers_c/models/resnet.py
@@ -0,0 +1,829 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+# `TemporalConvLayer` Copyright 2024 Alibaba DAMO-VILAB, The ModelScope Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import partial
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..utils import USE_PEFT_BACKEND
+from .activations import get_activation
+from .attention_processor import SpatialNorm
+from .downsampling import ( # noqa
+ Downsample1D,
+ Downsample2D,
+ FirDownsample2D,
+ KDownsample2D,
+ downsample_2d,
+)
+from .lora import LoRACompatibleConv, LoRACompatibleLinear
+from .normalization import AdaGroupNorm
+from .upsampling import ( # noqa
+ FirUpsample2D,
+ KUpsample2D,
+ Upsample1D,
+ Upsample2D,
+ upfirdn2d_native,
+ upsample_2d,
+)
+
+
+class ResnetBlockCondNorm2D(nn.Module):
+ r"""
+ A Resnet block that use normalization layer that incorporate conditioning information.
+
+ Parameters:
+ in_channels (`int`): The number of channels in the input.
+ out_channels (`int`, *optional*, default to be `None`):
+ The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+ dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+ temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+ groups (`int`, *optional*, default to `32`): The number of groups to use for the first normalization layer.
+ groups_out (`int`, *optional*, default to None):
+ The number of groups to use for the second normalization layer. if set to None, same as `groups`.
+ eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+ non_linearity (`str`, *optional*, default to `"swish"`): the activation function to use.
+ time_embedding_norm (`str`, *optional*, default to `"ada_group"` ):
+ The normalization layer for time embedding `temb`. Currently only support "ada_group" or "spatial".
+ kernel (`torch.FloatTensor`, optional, default to None): FIR filter, see
+ [`~models.resnet.FirUpsample2D`] and [`~models.resnet.FirDownsample2D`].
+ output_scale_factor (`float`, *optional*, default to be `1.0`): the scale factor to use for the output.
+ use_in_shortcut (`bool`, *optional*, default to `True`):
+ If `True`, add a 1x1 nn.conv2d layer for skip-connection.
+ up (`bool`, *optional*, default to `False`): If `True`, add an upsample layer.
+ down (`bool`, *optional*, default to `False`): If `True`, add a downsample layer.
+ conv_shortcut_bias (`bool`, *optional*, default to `True`): If `True`, adds a learnable bias to the
+ `conv_shortcut` output.
+ conv_2d_out_channels (`int`, *optional*, default to `None`): the number of channels in the output.
+ If None, same as `out_channels`.
+ """
+
+ def __init__(
+ self,
+ *,
+ in_channels: int,
+ out_channels: Optional[int] = None,
+ conv_shortcut: bool = False,
+ dropout: float = 0.0,
+ temb_channels: int = 512,
+ groups: int = 32,
+ groups_out: Optional[int] = None,
+ eps: float = 1e-6,
+ non_linearity: str = "swish",
+ time_embedding_norm: str = "ada_group", # ada_group, spatial
+ output_scale_factor: float = 1.0,
+ use_in_shortcut: Optional[bool] = None,
+ up: bool = False,
+ down: bool = False,
+ conv_shortcut_bias: bool = True,
+ conv_2d_out_channels: Optional[int] = None,
+ ):
+ super().__init__()
+ self.in_channels = in_channels
+ out_channels = in_channels if out_channels is None else out_channels
+ self.out_channels = out_channels
+ self.use_conv_shortcut = conv_shortcut
+ self.up = up
+ self.down = down
+ self.output_scale_factor = output_scale_factor
+ self.time_embedding_norm = time_embedding_norm
+
+ conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+
+ if groups_out is None:
+ groups_out = groups
+
+ if self.time_embedding_norm == "ada_group": # ada_group
+ self.norm1 = AdaGroupNorm(temb_channels, in_channels, groups, eps=eps)
+ elif self.time_embedding_norm == "spatial":
+ self.norm1 = SpatialNorm(in_channels, temb_channels)
+ else:
+ raise ValueError(f" unsupported time_embedding_norm: {self.time_embedding_norm}")
+
+ self.conv1 = conv_cls(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+ if self.time_embedding_norm == "ada_group": # ada_group
+ self.norm2 = AdaGroupNorm(temb_channels, out_channels, groups_out, eps=eps)
+ elif self.time_embedding_norm == "spatial": # spatial
+ self.norm2 = SpatialNorm(out_channels, temb_channels)
+ else:
+ raise ValueError(f" unsupported time_embedding_norm: {self.time_embedding_norm}")
+
+ self.dropout = torch.nn.Dropout(dropout)
+
+ conv_2d_out_channels = conv_2d_out_channels or out_channels
+ self.conv2 = conv_cls(out_channels, conv_2d_out_channels, kernel_size=3, stride=1, padding=1)
+
+ self.nonlinearity = get_activation(non_linearity)
+
+ self.upsample = self.downsample = None
+ if self.up:
+ self.upsample = Upsample2D(in_channels, use_conv=False)
+ elif self.down:
+ self.downsample = Downsample2D(in_channels, use_conv=False, padding=1, name="op")
+
+ self.use_in_shortcut = self.in_channels != conv_2d_out_channels if use_in_shortcut is None else use_in_shortcut
+
+ self.conv_shortcut = None
+ if self.use_in_shortcut:
+ self.conv_shortcut = conv_cls(
+ in_channels,
+ conv_2d_out_channels,
+ kernel_size=1,
+ stride=1,
+ padding=0,
+ bias=conv_shortcut_bias,
+ )
+
+ def forward(
+ self,
+ input_tensor: torch.FloatTensor,
+ temb: torch.FloatTensor,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ hidden_states = input_tensor
+
+ hidden_states = self.norm1(hidden_states, temb)
+
+ hidden_states = self.nonlinearity(hidden_states)
+
+ if self.upsample is not None:
+ # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+ if hidden_states.shape[0] >= 64:
+ input_tensor = input_tensor.contiguous()
+ hidden_states = hidden_states.contiguous()
+ input_tensor = self.upsample(input_tensor, scale=scale)
+ hidden_states = self.upsample(hidden_states, scale=scale)
+
+ elif self.downsample is not None:
+ input_tensor = self.downsample(input_tensor, scale=scale)
+ hidden_states = self.downsample(hidden_states, scale=scale)
+
+ hidden_states = self.conv1(hidden_states, scale) if not USE_PEFT_BACKEND else self.conv1(hidden_states)
+
+ hidden_states = self.norm2(hidden_states, temb)
+
+ hidden_states = self.nonlinearity(hidden_states)
+
+ hidden_states = self.dropout(hidden_states)
+ hidden_states = self.conv2(hidden_states, scale) if not USE_PEFT_BACKEND else self.conv2(hidden_states)
+
+ if self.conv_shortcut is not None:
+ input_tensor = (
+ self.conv_shortcut(input_tensor, scale) if not USE_PEFT_BACKEND else self.conv_shortcut(input_tensor)
+ )
+
+ output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+ return output_tensor
+
+
+class ResnetBlock2D(nn.Module):
+ r"""
+ A Resnet block.
+
+ Parameters:
+ in_channels (`int`): The number of channels in the input.
+ out_channels (`int`, *optional*, default to be `None`):
+ The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+ dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+ temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+ groups (`int`, *optional*, default to `32`): The number of groups to use for the first normalization layer.
+ groups_out (`int`, *optional*, default to None):
+ The number of groups to use for the second normalization layer. if set to None, same as `groups`.
+ eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+ non_linearity (`str`, *optional*, default to `"swish"`): the activation function to use.
+ time_embedding_norm (`str`, *optional*, default to `"default"` ): Time scale shift config.
+ By default, apply timestep embedding conditioning with a simple shift mechanism. Choose "scale_shift"
+ for a stronger conditioning with scale and shift.
+ kernel (`torch.FloatTensor`, optional, default to None): FIR filter, see
+ [`~models.resnet.FirUpsample2D`] and [`~models.resnet.FirDownsample2D`].
+ output_scale_factor (`float`, *optional*, default to be `1.0`): the scale factor to use for the output.
+ use_in_shortcut (`bool`, *optional*, default to `True`):
+ If `True`, add a 1x1 nn.conv2d layer for skip-connection.
+ up (`bool`, *optional*, default to `False`): If `True`, add an upsample layer.
+ down (`bool`, *optional*, default to `False`): If `True`, add a downsample layer.
+ conv_shortcut_bias (`bool`, *optional*, default to `True`): If `True`, adds a learnable bias to the
+ `conv_shortcut` output.
+ conv_2d_out_channels (`int`, *optional*, default to `None`): the number of channels in the output.
+ If None, same as `out_channels`.
+ """
+
+ def __init__(
+ self,
+ *,
+ in_channels: int,
+ out_channels: Optional[int] = None,
+ conv_shortcut: bool = False,
+ dropout: float = 0.0,
+ temb_channels: int = 512,
+ groups: int = 32,
+ groups_out: Optional[int] = None,
+ pre_norm: bool = True,
+ eps: float = 1e-6,
+ non_linearity: str = "swish",
+ skip_time_act: bool = False,
+ time_embedding_norm: str = "default", # default, scale_shift,
+ kernel: Optional[torch.FloatTensor] = None,
+ output_scale_factor: float = 1.0,
+ use_in_shortcut: Optional[bool] = None,
+ up: bool = False,
+ down: bool = False,
+ conv_shortcut_bias: bool = True,
+ conv_2d_out_channels: Optional[int] = None,
+ ):
+ super().__init__()
+ if time_embedding_norm == "ada_group":
+ raise ValueError(
+ "This class cannot be used with `time_embedding_norm==ada_group`, please use `ResnetBlockCondNorm2D` instead",
+ )
+ if time_embedding_norm == "spatial":
+ raise ValueError(
+ "This class cannot be used with `time_embedding_norm==spatial`, please use `ResnetBlockCondNorm2D` instead",
+ )
+
+ self.pre_norm = True
+ self.in_channels = in_channels
+ out_channels = in_channels if out_channels is None else out_channels
+ self.out_channels = out_channels
+ self.use_conv_shortcut = conv_shortcut
+ self.up = up
+ self.down = down
+ self.output_scale_factor = output_scale_factor
+ self.time_embedding_norm = time_embedding_norm
+ self.skip_time_act = skip_time_act
+
+ linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+ conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+
+ if groups_out is None:
+ groups_out = groups
+
+ self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+
+ self.conv1 = conv_cls(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+ if temb_channels is not None:
+ if self.time_embedding_norm == "default":
+ self.time_emb_proj = linear_cls(temb_channels, out_channels)
+ elif self.time_embedding_norm == "scale_shift":
+ self.time_emb_proj = linear_cls(temb_channels, 2 * out_channels)
+ else:
+ raise ValueError(f"unknown time_embedding_norm : {self.time_embedding_norm} ")
+ else:
+ self.time_emb_proj = None
+
+ self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+
+ self.dropout = torch.nn.Dropout(dropout)
+ conv_2d_out_channels = conv_2d_out_channels or out_channels
+ self.conv2 = conv_cls(out_channels, conv_2d_out_channels, kernel_size=3, stride=1, padding=1)
+
+ self.nonlinearity = get_activation(non_linearity)
+
+ self.upsample = self.downsample = None
+ if self.up:
+ if kernel == "fir":
+ fir_kernel = (1, 3, 3, 1)
+ self.upsample = lambda x: upsample_2d(x, kernel=fir_kernel)
+ elif kernel == "sde_vp":
+ self.upsample = partial(F.interpolate, scale_factor=2.0, mode="nearest")
+ else:
+ self.upsample = Upsample2D(in_channels, use_conv=False)
+ elif self.down:
+ if kernel == "fir":
+ fir_kernel = (1, 3, 3, 1)
+ self.downsample = lambda x: downsample_2d(x, kernel=fir_kernel)
+ elif kernel == "sde_vp":
+ self.downsample = partial(F.avg_pool2d, kernel_size=2, stride=2)
+ else:
+ self.downsample = Downsample2D(in_channels, use_conv=False, padding=1, name="op")
+
+ self.use_in_shortcut = self.in_channels != conv_2d_out_channels if use_in_shortcut is None else use_in_shortcut
+
+ self.conv_shortcut = None
+ if self.use_in_shortcut:
+ self.conv_shortcut = conv_cls(
+ in_channels,
+ conv_2d_out_channels,
+ kernel_size=1,
+ stride=1,
+ padding=0,
+ bias=conv_shortcut_bias,
+ )
+
+ def forward(
+ self,
+ input_tensor: torch.FloatTensor,
+ temb: torch.FloatTensor,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ hidden_states = input_tensor
+
+ hidden_states = self.norm1(hidden_states)
+ hidden_states = self.nonlinearity(hidden_states)
+
+ if self.upsample is not None:
+ # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+ if hidden_states.shape[0] >= 64:
+ input_tensor = input_tensor.contiguous()
+ hidden_states = hidden_states.contiguous()
+ input_tensor = (
+ self.upsample(input_tensor, scale=scale)
+ if isinstance(self.upsample, Upsample2D)
+ else self.upsample(input_tensor)
+ )
+ hidden_states = (
+ self.upsample(hidden_states, scale=scale)
+ if isinstance(self.upsample, Upsample2D)
+ else self.upsample(hidden_states)
+ )
+ elif self.downsample is not None:
+ input_tensor = (
+ self.downsample(input_tensor, scale=scale)
+ if isinstance(self.downsample, Downsample2D)
+ else self.downsample(input_tensor)
+ )
+ hidden_states = (
+ self.downsample(hidden_states, scale=scale)
+ if isinstance(self.downsample, Downsample2D)
+ else self.downsample(hidden_states)
+ )
+
+ hidden_states = self.conv1(hidden_states, scale) if not USE_PEFT_BACKEND else self.conv1(hidden_states)
+
+ if self.time_emb_proj is not None:
+ if not self.skip_time_act:
+ temb = self.nonlinearity(temb)
+ temb = (
+ self.time_emb_proj(temb, scale)[:, :, None, None]
+ if not USE_PEFT_BACKEND
+ else self.time_emb_proj(temb)[:, :, None, None]
+ )
+
+ if self.time_embedding_norm == "default":
+ if temb is not None:
+ hidden_states = hidden_states + temb
+ hidden_states = self.norm2(hidden_states)
+ elif self.time_embedding_norm == "scale_shift":
+ if temb is None:
+ raise ValueError(
+ f" `temb` should not be None when `time_embedding_norm` is {self.time_embedding_norm}"
+ )
+ time_scale, time_shift = torch.chunk(temb, 2, dim=1)
+ hidden_states = self.norm2(hidden_states)
+ hidden_states = hidden_states * (1 + time_scale) + time_shift
+ else:
+ hidden_states = self.norm2(hidden_states)
+
+ hidden_states = self.nonlinearity(hidden_states)
+
+ hidden_states = self.dropout(hidden_states)
+ hidden_states = self.conv2(hidden_states, scale) if not USE_PEFT_BACKEND else self.conv2(hidden_states)
+
+ if self.conv_shortcut is not None:
+ input_tensor = (
+ self.conv_shortcut(input_tensor, scale) if not USE_PEFT_BACKEND else self.conv_shortcut(input_tensor)
+ )
+
+ output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+ return output_tensor
+
+
+# unet_rl.py
+def rearrange_dims(tensor: torch.Tensor) -> torch.Tensor:
+ if len(tensor.shape) == 2:
+ return tensor[:, :, None]
+ if len(tensor.shape) == 3:
+ return tensor[:, :, None, :]
+ elif len(tensor.shape) == 4:
+ return tensor[:, :, 0, :]
+ else:
+ raise ValueError(f"`len(tensor)`: {len(tensor)} has to be 2, 3 or 4.")
+
+
+class Conv1dBlock(nn.Module):
+ """
+ Conv1d --> GroupNorm --> Mish
+
+ Parameters:
+ inp_channels (`int`): Number of input channels.
+ out_channels (`int`): Number of output channels.
+ kernel_size (`int` or `tuple`): Size of the convolving kernel.
+ n_groups (`int`, default `8`): Number of groups to separate the channels into.
+ activation (`str`, defaults to `mish`): Name of the activation function.
+ """
+
+ def __init__(
+ self,
+ inp_channels: int,
+ out_channels: int,
+ kernel_size: Union[int, Tuple[int, int]],
+ n_groups: int = 8,
+ activation: str = "mish",
+ ):
+ super().__init__()
+
+ self.conv1d = nn.Conv1d(inp_channels, out_channels, kernel_size, padding=kernel_size // 2)
+ self.group_norm = nn.GroupNorm(n_groups, out_channels)
+ self.mish = get_activation(activation)
+
+ def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+ intermediate_repr = self.conv1d(inputs)
+ intermediate_repr = rearrange_dims(intermediate_repr)
+ intermediate_repr = self.group_norm(intermediate_repr)
+ intermediate_repr = rearrange_dims(intermediate_repr)
+ output = self.mish(intermediate_repr)
+ return output
+
+
+# unet_rl.py
+class ResidualTemporalBlock1D(nn.Module):
+ """
+ Residual 1D block with temporal convolutions.
+
+ Parameters:
+ inp_channels (`int`): Number of input channels.
+ out_channels (`int`): Number of output channels.
+ embed_dim (`int`): Embedding dimension.
+ kernel_size (`int` or `tuple`): Size of the convolving kernel.
+ activation (`str`, defaults `mish`): It is possible to choose the right activation function.
+ """
+
+ def __init__(
+ self,
+ inp_channels: int,
+ out_channels: int,
+ embed_dim: int,
+ kernel_size: Union[int, Tuple[int, int]] = 5,
+ activation: str = "mish",
+ ):
+ super().__init__()
+ self.conv_in = Conv1dBlock(inp_channels, out_channels, kernel_size)
+ self.conv_out = Conv1dBlock(out_channels, out_channels, kernel_size)
+
+ self.time_emb_act = get_activation(activation)
+ self.time_emb = nn.Linear(embed_dim, out_channels)
+
+ self.residual_conv = (
+ nn.Conv1d(inp_channels, out_channels, 1) if inp_channels != out_channels else nn.Identity()
+ )
+
+ def forward(self, inputs: torch.Tensor, t: torch.Tensor) -> torch.Tensor:
+ """
+ Args:
+ inputs : [ batch_size x inp_channels x horizon ]
+ t : [ batch_size x embed_dim ]
+
+ returns:
+ out : [ batch_size x out_channels x horizon ]
+ """
+ t = self.time_emb_act(t)
+ t = self.time_emb(t)
+ out = self.conv_in(inputs) + rearrange_dims(t)
+ out = self.conv_out(out)
+ return out + self.residual_conv(inputs)
+
+
+class TemporalConvLayer(nn.Module):
+ """
+ Temporal convolutional layer that can be used for video (sequence of images) input Code mostly copied from:
+ https://github.com/modelscope/modelscope/blob/1509fdb973e5871f37148a4b5e5964cafd43e64d/modelscope/models/multi_modal/video_synthesis/unet_sd.py#L1016
+
+ Parameters:
+ in_dim (`int`): Number of input channels.
+ out_dim (`int`): Number of output channels.
+ dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
+ """
+
+ def __init__(
+ self,
+ in_dim: int,
+ out_dim: Optional[int] = None,
+ dropout: float = 0.0,
+ norm_num_groups: int = 32,
+ ):
+ super().__init__()
+ out_dim = out_dim or in_dim
+ self.in_dim = in_dim
+ self.out_dim = out_dim
+
+ # conv layers
+ self.conv1 = nn.Sequential(
+ nn.GroupNorm(norm_num_groups, in_dim),
+ nn.SiLU(),
+ nn.Conv3d(in_dim, out_dim, (3, 1, 1), padding=(1, 0, 0)),
+ )
+ self.conv2 = nn.Sequential(
+ nn.GroupNorm(norm_num_groups, out_dim),
+ nn.SiLU(),
+ nn.Dropout(dropout),
+ nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+ )
+ self.conv3 = nn.Sequential(
+ nn.GroupNorm(norm_num_groups, out_dim),
+ nn.SiLU(),
+ nn.Dropout(dropout),
+ nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+ )
+ self.conv4 = nn.Sequential(
+ nn.GroupNorm(norm_num_groups, out_dim),
+ nn.SiLU(),
+ nn.Dropout(dropout),
+ nn.Conv3d(out_dim, in_dim, (3, 1, 1), padding=(1, 0, 0)),
+ )
+
+ # zero out the last layer params,so the conv block is identity
+ nn.init.zeros_(self.conv4[-1].weight)
+ nn.init.zeros_(self.conv4[-1].bias)
+
+ def forward(self, hidden_states: torch.Tensor, num_frames: int = 1) -> torch.Tensor:
+ hidden_states = (
+ hidden_states[None, :].reshape((-1, num_frames) + hidden_states.shape[1:]).permute(0, 2, 1, 3, 4)
+ )
+
+ identity = hidden_states
+ hidden_states = self.conv1(hidden_states)
+ hidden_states = self.conv2(hidden_states)
+ hidden_states = self.conv3(hidden_states)
+ hidden_states = self.conv4(hidden_states)
+
+ hidden_states = identity + hidden_states
+
+ hidden_states = hidden_states.permute(0, 2, 1, 3, 4).reshape(
+ (hidden_states.shape[0] * hidden_states.shape[2], -1) + hidden_states.shape[3:]
+ )
+ return hidden_states
+
+
+class TemporalResnetBlock(nn.Module):
+ r"""
+ A Resnet block.
+
+ Parameters:
+ in_channels (`int`): The number of channels in the input.
+ out_channels (`int`, *optional*, default to be `None`):
+ The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+ temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+ eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: Optional[int] = None,
+ temb_channels: int = 512,
+ eps: float = 1e-6,
+ ):
+ super().__init__()
+ self.in_channels = in_channels
+ out_channels = in_channels if out_channels is None else out_channels
+ self.out_channels = out_channels
+
+ kernel_size = (3, 1, 1)
+ padding = [k // 2 for k in kernel_size]
+
+ self.norm1 = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=eps, affine=True)
+ self.conv1 = nn.Conv3d(
+ in_channels,
+ out_channels,
+ kernel_size=kernel_size,
+ stride=1,
+ padding=padding,
+ )
+
+ if temb_channels is not None:
+ self.time_emb_proj = nn.Linear(temb_channels, out_channels)
+ else:
+ self.time_emb_proj = None
+
+ self.norm2 = torch.nn.GroupNorm(num_groups=32, num_channels=out_channels, eps=eps, affine=True)
+
+ self.dropout = torch.nn.Dropout(0.0)
+ self.conv2 = nn.Conv3d(
+ out_channels,
+ out_channels,
+ kernel_size=kernel_size,
+ stride=1,
+ padding=padding,
+ )
+
+ self.nonlinearity = get_activation("silu")
+
+ self.use_in_shortcut = self.in_channels != out_channels
+
+ self.conv_shortcut = None
+ if self.use_in_shortcut:
+ self.conv_shortcut = nn.Conv3d(
+ in_channels,
+ out_channels,
+ kernel_size=1,
+ stride=1,
+ padding=0,
+ )
+
+ def forward(self, input_tensor: torch.FloatTensor, temb: torch.FloatTensor) -> torch.FloatTensor:
+ hidden_states = input_tensor
+
+ hidden_states = self.norm1(hidden_states)
+ hidden_states = self.nonlinearity(hidden_states)
+ hidden_states = self.conv1(hidden_states)
+
+ if self.time_emb_proj is not None:
+ temb = self.nonlinearity(temb)
+ temb = self.time_emb_proj(temb)[:, :, :, None, None]
+ temb = temb.permute(0, 2, 1, 3, 4)
+ hidden_states = hidden_states + temb
+
+ hidden_states = self.norm2(hidden_states)
+ hidden_states = self.nonlinearity(hidden_states)
+ hidden_states = self.dropout(hidden_states)
+ hidden_states = self.conv2(hidden_states)
+
+ if self.conv_shortcut is not None:
+ input_tensor = self.conv_shortcut(input_tensor)
+
+ output_tensor = input_tensor + hidden_states
+
+ return output_tensor
+
+
+# VideoResBlock
+class SpatioTemporalResBlock(nn.Module):
+ r"""
+ A SpatioTemporal Resnet block.
+
+ Parameters:
+ in_channels (`int`): The number of channels in the input.
+ out_channels (`int`, *optional*, default to be `None`):
+ The number of output channels for the first conv2d layer. If None, same as `in_channels`.
+ temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
+ eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the spatial resenet.
+ temporal_eps (`float`, *optional*, defaults to `eps`): The epsilon to use for the temporal resnet.
+ merge_factor (`float`, *optional*, defaults to `0.5`): The merge factor to use for the temporal mixing.
+ merge_strategy (`str`, *optional*, defaults to `learned_with_images`):
+ The merge strategy to use for the temporal mixing.
+ switch_spatial_to_temporal_mix (`bool`, *optional*, defaults to `False`):
+ If `True`, switch the spatial and temporal mixing.
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: Optional[int] = None,
+ temb_channels: int = 512,
+ eps: float = 1e-6,
+ temporal_eps: Optional[float] = None,
+ merge_factor: float = 0.5,
+ merge_strategy="learned_with_images",
+ switch_spatial_to_temporal_mix: bool = False,
+ ):
+ super().__init__()
+
+ self.spatial_res_block = ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=eps,
+ )
+
+ self.temporal_res_block = TemporalResnetBlock(
+ in_channels=out_channels if out_channels is not None else in_channels,
+ out_channels=out_channels if out_channels is not None else in_channels,
+ temb_channels=temb_channels,
+ eps=temporal_eps if temporal_eps is not None else eps,
+ )
+
+ self.time_mixer = AlphaBlender(
+ alpha=merge_factor,
+ merge_strategy=merge_strategy,
+ switch_spatial_to_temporal_mix=switch_spatial_to_temporal_mix,
+ )
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ image_only_indicator: Optional[torch.Tensor] = None,
+ ):
+ num_frames = image_only_indicator.shape[-1]
+ hidden_states = self.spatial_res_block(hidden_states, temb)
+
+ batch_frames, channels, height, width = hidden_states.shape
+ batch_size = batch_frames // num_frames
+
+ hidden_states_mix = (
+ hidden_states[None, :].reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4)
+ )
+ hidden_states = (
+ hidden_states[None, :].reshape(batch_size, num_frames, channels, height, width).permute(0, 2, 1, 3, 4)
+ )
+
+ if temb is not None:
+ temb = temb.reshape(batch_size, num_frames, -1)
+
+ hidden_states = self.temporal_res_block(hidden_states, temb)
+ hidden_states = self.time_mixer(
+ x_spatial=hidden_states_mix,
+ x_temporal=hidden_states,
+ image_only_indicator=image_only_indicator,
+ )
+
+ hidden_states = hidden_states.permute(0, 2, 1, 3, 4).reshape(batch_frames, channels, height, width)
+ return hidden_states
+
+
+class AlphaBlender(nn.Module):
+ r"""
+ A module to blend spatial and temporal features.
+
+ Parameters:
+ alpha (`float`): The initial value of the blending factor.
+ merge_strategy (`str`, *optional*, defaults to `learned_with_images`):
+ The merge strategy to use for the temporal mixing.
+ switch_spatial_to_temporal_mix (`bool`, *optional*, defaults to `False`):
+ If `True`, switch the spatial and temporal mixing.
+ """
+
+ strategies = ["learned", "fixed", "learned_with_images"]
+
+ def __init__(
+ self,
+ alpha: float,
+ merge_strategy: str = "learned_with_images",
+ switch_spatial_to_temporal_mix: bool = False,
+ ):
+ super().__init__()
+ self.merge_strategy = merge_strategy
+ self.switch_spatial_to_temporal_mix = switch_spatial_to_temporal_mix # For TemporalVAE
+
+ if merge_strategy not in self.strategies:
+ raise ValueError(f"merge_strategy needs to be in {self.strategies}")
+
+ if self.merge_strategy == "fixed":
+ self.register_buffer("mix_factor", torch.Tensor([alpha]))
+ elif self.merge_strategy == "learned" or self.merge_strategy == "learned_with_images":
+ self.register_parameter("mix_factor", torch.nn.Parameter(torch.Tensor([alpha])))
+ else:
+ raise ValueError(f"Unknown merge strategy {self.merge_strategy}")
+
+ def get_alpha(self, image_only_indicator: torch.Tensor, ndims: int) -> torch.Tensor:
+ if self.merge_strategy == "fixed":
+ alpha = self.mix_factor
+
+ elif self.merge_strategy == "learned":
+ alpha = torch.sigmoid(self.mix_factor)
+
+ elif self.merge_strategy == "learned_with_images":
+ if image_only_indicator is None:
+ raise ValueError("Please provide image_only_indicator to use learned_with_images merge strategy")
+
+ alpha = torch.where(
+ image_only_indicator.bool(),
+ torch.ones(1, 1, device=image_only_indicator.device),
+ torch.sigmoid(self.mix_factor)[..., None],
+ )
+
+ # (batch, channel, frames, height, width)
+ if ndims == 5:
+ alpha = alpha[:, None, :, None, None]
+ # (batch*frames, height*width, channels)
+ elif ndims == 3:
+ alpha = alpha.reshape(-1)[:, None, None]
+ else:
+ raise ValueError(f"Unexpected ndims {ndims}. Dimensions should be 3 or 5")
+
+ else:
+ raise NotImplementedError
+
+ return alpha
+
+ def forward(
+ self,
+ x_spatial: torch.Tensor,
+ x_temporal: torch.Tensor,
+ image_only_indicator: Optional[torch.Tensor] = None,
+ ) -> torch.Tensor:
+ alpha = self.get_alpha(image_only_indicator, x_spatial.ndim)
+ alpha = alpha.to(x_spatial.dtype)
+
+ if self.switch_spatial_to_temporal_mix:
+ alpha = 1.0 - alpha
+
+ x = alpha * x_spatial + (1.0 - alpha) * x_temporal
+ return x
diff --git a/model/diffusers_c/models/resnet_flax.py b/model/diffusers_c/models/resnet_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..f8bb4788d9a5620cbe19607a5f80d00b1b12ddca
--- /dev/null
+++ b/model/diffusers_c/models/resnet_flax.py
@@ -0,0 +1,124 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+
+
+class FlaxUpsample2D(nn.Module):
+ out_channels: int
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ self.conv = nn.Conv(
+ self.out_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ def __call__(self, hidden_states):
+ batch, height, width, channels = hidden_states.shape
+ hidden_states = jax.image.resize(
+ hidden_states,
+ shape=(batch, height * 2, width * 2, channels),
+ method="nearest",
+ )
+ hidden_states = self.conv(hidden_states)
+ return hidden_states
+
+
+class FlaxDownsample2D(nn.Module):
+ out_channels: int
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ self.conv = nn.Conv(
+ self.out_channels,
+ kernel_size=(3, 3),
+ strides=(2, 2),
+ padding=((1, 1), (1, 1)), # padding="VALID",
+ dtype=self.dtype,
+ )
+
+ def __call__(self, hidden_states):
+ # pad = ((0, 0), (0, 1), (0, 1), (0, 0)) # pad height and width dim
+ # hidden_states = jnp.pad(hidden_states, pad_width=pad)
+ hidden_states = self.conv(hidden_states)
+ return hidden_states
+
+
+class FlaxResnetBlock2D(nn.Module):
+ in_channels: int
+ out_channels: int = None
+ dropout_prob: float = 0.0
+ use_nin_shortcut: bool = None
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ out_channels = self.in_channels if self.out_channels is None else self.out_channels
+
+ self.norm1 = nn.GroupNorm(num_groups=32, epsilon=1e-5)
+ self.conv1 = nn.Conv(
+ out_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ self.time_emb_proj = nn.Dense(out_channels, dtype=self.dtype)
+
+ self.norm2 = nn.GroupNorm(num_groups=32, epsilon=1e-5)
+ self.dropout = nn.Dropout(self.dropout_prob)
+ self.conv2 = nn.Conv(
+ out_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ use_nin_shortcut = self.in_channels != out_channels if self.use_nin_shortcut is None else self.use_nin_shortcut
+
+ self.conv_shortcut = None
+ if use_nin_shortcut:
+ self.conv_shortcut = nn.Conv(
+ out_channels,
+ kernel_size=(1, 1),
+ strides=(1, 1),
+ padding="VALID",
+ dtype=self.dtype,
+ )
+
+ def __call__(self, hidden_states, temb, deterministic=True):
+ residual = hidden_states
+ hidden_states = self.norm1(hidden_states)
+ hidden_states = nn.swish(hidden_states)
+ hidden_states = self.conv1(hidden_states)
+
+ temb = self.time_emb_proj(nn.swish(temb))
+ temb = jnp.expand_dims(jnp.expand_dims(temb, 1), 1)
+ hidden_states = hidden_states + temb
+
+ hidden_states = self.norm2(hidden_states)
+ hidden_states = nn.swish(hidden_states)
+ hidden_states = self.dropout(hidden_states, deterministic)
+ hidden_states = self.conv2(hidden_states)
+
+ if self.conv_shortcut is not None:
+ residual = self.conv_shortcut(residual)
+
+ return hidden_states + residual
diff --git a/model/diffusers_c/models/t5_film_transformer.py b/model/diffusers_c/models/t5_film_transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..6aa5ff7449de1ad0933a17b63b514ea3361916c7
--- /dev/null
+++ b/model/diffusers_c/models/t5_film_transformer.py
@@ -0,0 +1,70 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from ..utils import deprecate
+from .transformers.t5_film_transformer import (
+ DecoderLayer,
+ NewGELUActivation,
+ T5DenseGatedActDense,
+ T5FilmDecoder,
+ T5FiLMLayer,
+ T5LayerCrossAttention,
+ T5LayerFFCond,
+ T5LayerNorm,
+ T5LayerSelfAttentionCond,
+)
+
+
+class T5FilmDecoder(T5FilmDecoder):
+ deprecation_message = "Importing `T5FilmDecoder` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import T5FilmDecoder`, instead."
+ deprecate("T5FilmDecoder", "0.29", deprecation_message)
+
+
+class DecoderLayer(DecoderLayer):
+ deprecation_message = "Importing `DecoderLayer` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import DecoderLayer`, instead."
+ deprecate("DecoderLayer", "0.29", deprecation_message)
+
+
+class T5LayerSelfAttentionCond(T5LayerSelfAttentionCond):
+ deprecation_message = "Importing `T5LayerSelfAttentionCond` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import T5LayerSelfAttentionCond`, instead."
+ deprecate("T5LayerSelfAttentionCond", "0.29", deprecation_message)
+
+
+class T5LayerCrossAttention(T5LayerCrossAttention):
+ deprecation_message = "Importing `T5LayerCrossAttention` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import T5LayerCrossAttention`, instead."
+ deprecate("T5LayerCrossAttention", "0.29", deprecation_message)
+
+
+class T5LayerFFCond(T5LayerFFCond):
+ deprecation_message = "Importing `T5LayerFFCond` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import T5LayerFFCond`, instead."
+ deprecate("T5LayerFFCond", "0.29", deprecation_message)
+
+
+class T5DenseGatedActDense(T5DenseGatedActDense):
+ deprecation_message = "Importing `T5DenseGatedActDense` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import T5DenseGatedActDense`, instead."
+ deprecate("T5DenseGatedActDense", "0.29", deprecation_message)
+
+
+class T5LayerNorm(T5LayerNorm):
+ deprecation_message = "Importing `T5LayerNorm` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import T5LayerNorm`, instead."
+ deprecate("T5LayerNorm", "0.29", deprecation_message)
+
+
+class NewGELUActivation(NewGELUActivation):
+ deprecation_message = "Importing `T5LayerNorm` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import NewGELUActivation`, instead."
+ deprecate("NewGELUActivation", "0.29", deprecation_message)
+
+
+class T5FiLMLayer(T5FiLMLayer):
+ deprecation_message = "Importing `T5FiLMLayer` from `diffusers.models.t5_film_transformer` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.t5_film_transformer import T5FiLMLayer`, instead."
+ deprecate("T5FiLMLayer", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/transformer_2d.py b/model/diffusers_c/models/transformer_2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..5d8ef1347af1ba0d8feff652ccdcb211b94df5f6
--- /dev/null
+++ b/model/diffusers_c/models/transformer_2d.py
@@ -0,0 +1,25 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from ..utils import deprecate
+from .transformers.transformer_2d import Transformer2DModel, Transformer2DModelOutput
+
+
+class Transformer2DModelOutput(Transformer2DModelOutput):
+ deprecation_message = "Importing `Transformer2DModelOutput` from `diffusers.models.transformer_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.transformer_2d import Transformer2DModelOutput`, instead."
+ deprecate("Transformer2DModelOutput", "0.29", deprecation_message)
+
+
+class Transformer2DModel(Transformer2DModel):
+ deprecation_message = "Importing `Transformer2DModel` from `diffusers.models.transformer_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.transformer_2d import Transformer2DModel`, instead."
+ deprecate("Transformer2DModel", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/transformer_temporal.py b/model/diffusers_c/models/transformer_temporal.py
new file mode 100644
index 0000000000000000000000000000000000000000..83c7a8e67b5623a27931a3a7b8c3fc2cfc92f90c
--- /dev/null
+++ b/model/diffusers_c/models/transformer_temporal.py
@@ -0,0 +1,34 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from ..utils import deprecate
+from .transformers.transformer_temporal import (
+ TransformerSpatioTemporalModel,
+ TransformerTemporalModel,
+ TransformerTemporalModelOutput,
+)
+
+
+class TransformerTemporalModelOutput(TransformerTemporalModelOutput):
+ deprecation_message = "Importing `TransformerTemporalModelOutput` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.tranformer_temporal import TransformerTemporalModelOutput`, instead."
+ deprecate("TransformerTemporalModelOutput", "0.29", deprecation_message)
+
+
+class TransformerTemporalModel(TransformerTemporalModel):
+ deprecation_message = "Importing `TransformerTemporalModel` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.tranformer_temporal import TransformerTemporalModel`, instead."
+ deprecate("TransformerTemporalModel", "0.29", deprecation_message)
+
+
+class TransformerSpatioTemporalModel(TransformerSpatioTemporalModel):
+ deprecation_message = "Importing `TransformerSpatioTemporalModel` from `diffusers.models.transformer_temporal` is deprecated and this will be removed in a future version. Please use `from diffusers.models.transformers.tranformer_temporal import TransformerSpatioTemporalModel`, instead."
+ deprecate("TransformerTemporalModelOutput", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/transformers/__init__.py b/model/diffusers_c/models/transformers/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..dc78a72b2fb84b4ffb7663a6e2908883b4110924
--- /dev/null
+++ b/model/diffusers_c/models/transformers/__init__.py
@@ -0,0 +1,9 @@
+from ...utils import is_torch_available
+
+
+if is_torch_available():
+ from .dual_transformer_2d import DualTransformer2DModel
+ from .prior_transformer import PriorTransformer
+ from .t5_film_transformer import T5FilmDecoder
+ from .transformer_2d import Transformer2DModel
+ from .transformer_temporal import TransformerTemporalModel
diff --git a/model/diffusers_c/models/transformers/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/models/transformers/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f3e5966c6b7b1d77951c9b396028e86d420b8f1c
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/models/transformers/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0021a12869b50b53a46d104c2cfc9cae71f2d0cf
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/dual_transformer_2d.cpython-310.pyc b/model/diffusers_c/models/transformers/__pycache__/dual_transformer_2d.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5bab864a4802253510c04cec857fd8511d424943
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/dual_transformer_2d.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/dual_transformer_2d.cpython-39.pyc b/model/diffusers_c/models/transformers/__pycache__/dual_transformer_2d.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ab987ff6482742f6c5d6d234c8708a6a406b8215
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/dual_transformer_2d.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/prior_transformer.cpython-310.pyc b/model/diffusers_c/models/transformers/__pycache__/prior_transformer.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5b4c9dbb8b8130c6ce7166c41f9a80114d634a36
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/prior_transformer.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/prior_transformer.cpython-39.pyc b/model/diffusers_c/models/transformers/__pycache__/prior_transformer.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4f0bc51831d1c5010f5abc828cfadf26fe0d63d7
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/prior_transformer.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/t5_film_transformer.cpython-310.pyc b/model/diffusers_c/models/transformers/__pycache__/t5_film_transformer.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..efab429d32ab5b7df9e804e84ee54c06a6e0a660
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/t5_film_transformer.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/t5_film_transformer.cpython-39.pyc b/model/diffusers_c/models/transformers/__pycache__/t5_film_transformer.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2d644dc4128d1d61fd5147d6ade39b1cfd6942f1
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/t5_film_transformer.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/transformer_2d.cpython-310.pyc b/model/diffusers_c/models/transformers/__pycache__/transformer_2d.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f5520a17da3a79ab7e084abe361dbd896c1aca5e
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/transformer_2d.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/transformer_2d.cpython-39.pyc b/model/diffusers_c/models/transformers/__pycache__/transformer_2d.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6f7e617b9464d4a0859190591cd4369761cfeadf
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/transformer_2d.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/transformer_temporal.cpython-310.pyc b/model/diffusers_c/models/transformers/__pycache__/transformer_temporal.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..48a316d4a16b56b1d6d829c5b0d4506154274fa8
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/transformer_temporal.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/transformers/__pycache__/transformer_temporal.cpython-39.pyc b/model/diffusers_c/models/transformers/__pycache__/transformer_temporal.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1bf4b3b25ffbbc3a081dcdc85c2fbaa70768336d
Binary files /dev/null and b/model/diffusers_c/models/transformers/__pycache__/transformer_temporal.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/transformers/dual_transformer_2d.py b/model/diffusers_c/models/transformers/dual_transformer_2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..96849bd28bb1c866a6c1d9160b53e957ef3e2c35
--- /dev/null
+++ b/model/diffusers_c/models/transformers/dual_transformer_2d.py
@@ -0,0 +1,155 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional
+
+from torch import nn
+
+from .transformer_2d import Transformer2DModel, Transformer2DModelOutput
+
+
+class DualTransformer2DModel(nn.Module):
+ """
+ Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference.
+
+ Parameters:
+ num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+ in_channels (`int`, *optional*):
+ Pass if the input is continuous. The number of channels in the input and output.
+ num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+ dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use.
+ cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
+ sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
+ Note that this is fixed at training time as it is used for learning a number of position embeddings. See
+ `ImagePositionalEmbeddings`.
+ num_vector_embeds (`int`, *optional*):
+ Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
+ Includes the class for the masked latent pixel.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+ num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
+ The number of diffusion steps used during training. Note that this is fixed at training time as it is used
+ to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
+ up to but not more than steps than `num_embeds_ada_norm`.
+ attention_bias (`bool`, *optional*):
+ Configure if the TransformerBlocks' attention should contain a bias parameter.
+ """
+
+ def __init__(
+ self,
+ num_attention_heads: int = 16,
+ attention_head_dim: int = 88,
+ in_channels: Optional[int] = None,
+ num_layers: int = 1,
+ dropout: float = 0.0,
+ norm_num_groups: int = 32,
+ cross_attention_dim: Optional[int] = None,
+ attention_bias: bool = False,
+ sample_size: Optional[int] = None,
+ num_vector_embeds: Optional[int] = None,
+ activation_fn: str = "geglu",
+ num_embeds_ada_norm: Optional[int] = None,
+ ):
+ super().__init__()
+ self.transformers = nn.ModuleList(
+ [
+ Transformer2DModel(
+ num_attention_heads=num_attention_heads,
+ attention_head_dim=attention_head_dim,
+ in_channels=in_channels,
+ num_layers=num_layers,
+ dropout=dropout,
+ norm_num_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ attention_bias=attention_bias,
+ sample_size=sample_size,
+ num_vector_embeds=num_vector_embeds,
+ activation_fn=activation_fn,
+ num_embeds_ada_norm=num_embeds_ada_norm,
+ )
+ for _ in range(2)
+ ]
+ )
+
+ # Variables that can be set by a pipeline:
+
+ # The ratio of transformer1 to transformer2's output states to be combined during inference
+ self.mix_ratio = 0.5
+
+ # The shape of `encoder_hidden_states` is expected to be
+ # `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)`
+ self.condition_lengths = [77, 257]
+
+ # Which transformer to use to encode which condition.
+ # E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])`
+ self.transformer_index_for_condition = [1, 0]
+
+ def forward(
+ self,
+ hidden_states,
+ encoder_hidden_states,
+ timestep=None,
+ attention_mask=None,
+ cross_attention_kwargs=None,
+ return_dict: bool = True,
+ ):
+ """
+ Args:
+ hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
+ When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
+ hidden_states.
+ encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+ Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+ self-attention.
+ timestep ( `torch.long`, *optional*):
+ Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
+ attention_mask (`torch.FloatTensor`, *optional*):
+ Optional attention mask to be applied in Attention.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`:
+ [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When
+ returning a tuple, the first element is the sample tensor.
+ """
+ input_states = hidden_states
+
+ encoded_states = []
+ tokens_start = 0
+ # attention_mask is not used yet
+ for i in range(2):
+ # for each of the two transformers, pass the corresponding condition tokens
+ condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.condition_lengths[i]]
+ transformer_index = self.transformer_index_for_condition[i]
+ encoded_state = self.transformers[transformer_index](
+ input_states,
+ encoder_hidden_states=condition_state,
+ timestep=timestep,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+ encoded_states.append(encoded_state - input_states)
+ tokens_start += self.condition_lengths[i]
+
+ output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio)
+ output_states = output_states + input_states
+
+ if not return_dict:
+ return (output_states,)
+
+ return Transformer2DModelOutput(sample=output_states)
diff --git a/model/diffusers_c/models/transformers/prior_transformer.py b/model/diffusers_c/models/transformers/prior_transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..990eabe2c37a1c00ed2ec0c6b249b648f3a9bc34
--- /dev/null
+++ b/model/diffusers_c/models/transformers/prior_transformer.py
@@ -0,0 +1,380 @@
+from dataclasses import dataclass
+from typing import Dict, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
+from ...utils import BaseOutput
+from ..attention import BasicTransformerBlock
+from ..attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+
+
+@dataclass
+class PriorTransformerOutput(BaseOutput):
+ """
+ The output of [`PriorTransformer`].
+
+ Args:
+ predicted_image_embedding (`torch.FloatTensor` of shape `(batch_size, embedding_dim)`):
+ The predicted CLIP image embedding conditioned on the CLIP text embedding input.
+ """
+
+ predicted_image_embedding: torch.FloatTensor
+
+
+class PriorTransformer(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin):
+ """
+ A Prior Transformer model.
+
+ Parameters:
+ num_attention_heads (`int`, *optional*, defaults to 32): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head.
+ num_layers (`int`, *optional*, defaults to 20): The number of layers of Transformer blocks to use.
+ embedding_dim (`int`, *optional*, defaults to 768): The dimension of the model input `hidden_states`
+ num_embeddings (`int`, *optional*, defaults to 77):
+ The number of embeddings of the model input `hidden_states`
+ additional_embeddings (`int`, *optional*, defaults to 4): The number of additional tokens appended to the
+ projected `hidden_states`. The actual length of the used `hidden_states` is `num_embeddings +
+ additional_embeddings`.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ time_embed_act_fn (`str`, *optional*, defaults to 'silu'):
+ The activation function to use to create timestep embeddings.
+ norm_in_type (`str`, *optional*, defaults to None): The normalization layer to apply on hidden states before
+ passing to Transformer blocks. Set it to `None` if normalization is not needed.
+ embedding_proj_norm_type (`str`, *optional*, defaults to None):
+ The normalization layer to apply on the input `proj_embedding`. Set it to `None` if normalization is not
+ needed.
+ encoder_hid_proj_type (`str`, *optional*, defaults to `linear`):
+ The projection layer to apply on the input `encoder_hidden_states`. Set it to `None` if
+ `encoder_hidden_states` is `None`.
+ added_emb_type (`str`, *optional*, defaults to `prd`): Additional embeddings to condition the model.
+ Choose from `prd` or `None`. if choose `prd`, it will prepend a token indicating the (quantized) dot
+ product between the text embedding and image embedding as proposed in the unclip paper
+ https://arxiv.org/abs/2204.06125 If it is `None`, no additional embeddings will be prepended.
+ time_embed_dim (`int, *optional*, defaults to None): The dimension of timestep embeddings.
+ If None, will be set to `num_attention_heads * attention_head_dim`
+ embedding_proj_dim (`int`, *optional*, default to None):
+ The dimension of `proj_embedding`. If None, will be set to `embedding_dim`.
+ clip_embed_dim (`int`, *optional*, default to None):
+ The dimension of the output. If None, will be set to `embedding_dim`.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ num_attention_heads: int = 32,
+ attention_head_dim: int = 64,
+ num_layers: int = 20,
+ embedding_dim: int = 768,
+ num_embeddings=77,
+ additional_embeddings=4,
+ dropout: float = 0.0,
+ time_embed_act_fn: str = "silu",
+ norm_in_type: Optional[str] = None, # layer
+ embedding_proj_norm_type: Optional[str] = None, # layer
+ encoder_hid_proj_type: Optional[str] = "linear", # linear
+ added_emb_type: Optional[str] = "prd", # prd
+ time_embed_dim: Optional[int] = None,
+ embedding_proj_dim: Optional[int] = None,
+ clip_embed_dim: Optional[int] = None,
+ ):
+ super().__init__()
+ self.num_attention_heads = num_attention_heads
+ self.attention_head_dim = attention_head_dim
+ inner_dim = num_attention_heads * attention_head_dim
+ self.additional_embeddings = additional_embeddings
+
+ time_embed_dim = time_embed_dim or inner_dim
+ embedding_proj_dim = embedding_proj_dim or embedding_dim
+ clip_embed_dim = clip_embed_dim or embedding_dim
+
+ self.time_proj = Timesteps(inner_dim, True, 0)
+ self.time_embedding = TimestepEmbedding(inner_dim, time_embed_dim, out_dim=inner_dim, act_fn=time_embed_act_fn)
+
+ self.proj_in = nn.Linear(embedding_dim, inner_dim)
+
+ if embedding_proj_norm_type is None:
+ self.embedding_proj_norm = None
+ elif embedding_proj_norm_type == "layer":
+ self.embedding_proj_norm = nn.LayerNorm(embedding_proj_dim)
+ else:
+ raise ValueError(f"unsupported embedding_proj_norm_type: {embedding_proj_norm_type}")
+
+ self.embedding_proj = nn.Linear(embedding_proj_dim, inner_dim)
+
+ if encoder_hid_proj_type is None:
+ self.encoder_hidden_states_proj = None
+ elif encoder_hid_proj_type == "linear":
+ self.encoder_hidden_states_proj = nn.Linear(embedding_dim, inner_dim)
+ else:
+ raise ValueError(f"unsupported encoder_hid_proj_type: {encoder_hid_proj_type}")
+
+ self.positional_embedding = nn.Parameter(torch.zeros(1, num_embeddings + additional_embeddings, inner_dim))
+
+ if added_emb_type == "prd":
+ self.prd_embedding = nn.Parameter(torch.zeros(1, 1, inner_dim))
+ elif added_emb_type is None:
+ self.prd_embedding = None
+ else:
+ raise ValueError(
+ f"`added_emb_type`: {added_emb_type} is not supported. Make sure to choose one of `'prd'` or `None`."
+ )
+
+ self.transformer_blocks = nn.ModuleList(
+ [
+ BasicTransformerBlock(
+ inner_dim,
+ num_attention_heads,
+ attention_head_dim,
+ dropout=dropout,
+ activation_fn="gelu",
+ attention_bias=True,
+ )
+ for d in range(num_layers)
+ ]
+ )
+
+ if norm_in_type == "layer":
+ self.norm_in = nn.LayerNorm(inner_dim)
+ elif norm_in_type is None:
+ self.norm_in = None
+ else:
+ raise ValueError(f"Unsupported norm_in_type: {norm_in_type}.")
+
+ self.norm_out = nn.LayerNorm(inner_dim)
+
+ self.proj_to_clip_embeddings = nn.Linear(inner_dim, clip_embed_dim)
+
+ causal_attention_mask = torch.full(
+ [num_embeddings + additional_embeddings, num_embeddings + additional_embeddings], -10000.0
+ )
+ causal_attention_mask.triu_(1)
+ causal_attention_mask = causal_attention_mask[None, ...]
+ self.register_buffer("causal_attention_mask", causal_attention_mask, persistent=False)
+
+ self.clip_mean = nn.Parameter(torch.zeros(1, clip_embed_dim))
+ self.clip_std = nn.Parameter(torch.zeros(1, clip_embed_dim))
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ def forward(
+ self,
+ hidden_states,
+ timestep: Union[torch.Tensor, float, int],
+ proj_embedding: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.BoolTensor] = None,
+ return_dict: bool = True,
+ ):
+ """
+ The [`PriorTransformer`] forward method.
+
+ Args:
+ hidden_states (`torch.FloatTensor` of shape `(batch_size, embedding_dim)`):
+ The currently predicted image embeddings.
+ timestep (`torch.LongTensor`):
+ Current denoising step.
+ proj_embedding (`torch.FloatTensor` of shape `(batch_size, embedding_dim)`):
+ Projected embedding vector the denoising process is conditioned on.
+ encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, num_embeddings, embedding_dim)`):
+ Hidden states of the text embeddings the denoising process is conditioned on.
+ attention_mask (`torch.BoolTensor` of shape `(batch_size, num_embeddings)`):
+ Text mask for the text embeddings.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.prior_transformer.PriorTransformerOutput`] instead of a plain
+ tuple.
+
+ Returns:
+ [`~models.prior_transformer.PriorTransformerOutput`] or `tuple`:
+ If return_dict is True, a [`~models.prior_transformer.PriorTransformerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+ """
+ batch_size = hidden_states.shape[0]
+
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ timesteps = torch.tensor([timesteps], dtype=torch.long, device=hidden_states.device)
+ elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(hidden_states.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps * torch.ones(batch_size, dtype=timesteps.dtype, device=timesteps.device)
+
+ timesteps_projected = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might be fp16, so we need to cast here.
+ timesteps_projected = timesteps_projected.to(dtype=self.dtype)
+ time_embeddings = self.time_embedding(timesteps_projected)
+
+ if self.embedding_proj_norm is not None:
+ proj_embedding = self.embedding_proj_norm(proj_embedding)
+
+ proj_embeddings = self.embedding_proj(proj_embedding)
+ if self.encoder_hidden_states_proj is not None and encoder_hidden_states is not None:
+ encoder_hidden_states = self.encoder_hidden_states_proj(encoder_hidden_states)
+ elif self.encoder_hidden_states_proj is not None and encoder_hidden_states is None:
+ raise ValueError("`encoder_hidden_states_proj` requires `encoder_hidden_states` to be set")
+
+ hidden_states = self.proj_in(hidden_states)
+
+ positional_embeddings = self.positional_embedding.to(hidden_states.dtype)
+
+ additional_embeds = []
+ additional_embeddings_len = 0
+
+ if encoder_hidden_states is not None:
+ additional_embeds.append(encoder_hidden_states)
+ additional_embeddings_len += encoder_hidden_states.shape[1]
+
+ if len(proj_embeddings.shape) == 2:
+ proj_embeddings = proj_embeddings[:, None, :]
+
+ if len(hidden_states.shape) == 2:
+ hidden_states = hidden_states[:, None, :]
+
+ additional_embeds = additional_embeds + [
+ proj_embeddings,
+ time_embeddings[:, None, :],
+ hidden_states,
+ ]
+
+ if self.prd_embedding is not None:
+ prd_embedding = self.prd_embedding.to(hidden_states.dtype).expand(batch_size, -1, -1)
+ additional_embeds.append(prd_embedding)
+
+ hidden_states = torch.cat(
+ additional_embeds,
+ dim=1,
+ )
+
+ # Allow positional_embedding to not include the `addtional_embeddings` and instead pad it with zeros for these additional tokens
+ additional_embeddings_len = additional_embeddings_len + proj_embeddings.shape[1] + 1
+ if positional_embeddings.shape[1] < hidden_states.shape[1]:
+ positional_embeddings = F.pad(
+ positional_embeddings,
+ (
+ 0,
+ 0,
+ additional_embeddings_len,
+ self.prd_embedding.shape[1] if self.prd_embedding is not None else 0,
+ ),
+ value=0.0,
+ )
+
+ hidden_states = hidden_states + positional_embeddings
+
+ if attention_mask is not None:
+ attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+ attention_mask = F.pad(attention_mask, (0, self.additional_embeddings), value=0.0)
+ attention_mask = (attention_mask[:, None, :] + self.causal_attention_mask).to(hidden_states.dtype)
+ attention_mask = attention_mask.repeat_interleave(self.config.num_attention_heads, dim=0)
+
+ if self.norm_in is not None:
+ hidden_states = self.norm_in(hidden_states)
+
+ for block in self.transformer_blocks:
+ hidden_states = block(hidden_states, attention_mask=attention_mask)
+
+ hidden_states = self.norm_out(hidden_states)
+
+ if self.prd_embedding is not None:
+ hidden_states = hidden_states[:, -1]
+ else:
+ hidden_states = hidden_states[:, additional_embeddings_len:]
+
+ predicted_image_embedding = self.proj_to_clip_embeddings(hidden_states)
+
+ if not return_dict:
+ return (predicted_image_embedding,)
+
+ return PriorTransformerOutput(predicted_image_embedding=predicted_image_embedding)
+
+ def post_process_latents(self, prior_latents):
+ prior_latents = (prior_latents * self.clip_std) + self.clip_mean
+ return prior_latents
diff --git a/model/diffusers_c/models/transformers/t5_film_transformer.py b/model/diffusers_c/models/transformers/t5_film_transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..bff98db021cfb23259e709e352e14918ed24ffb1
--- /dev/null
+++ b/model/diffusers_c/models/transformers/t5_film_transformer.py
@@ -0,0 +1,438 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ..attention_processor import Attention
+from ..embeddings import get_timestep_embedding
+from ..modeling_utils import ModelMixin
+
+
+class T5FilmDecoder(ModelMixin, ConfigMixin):
+ r"""
+ T5 style decoder with FiLM conditioning.
+
+ Args:
+ input_dims (`int`, *optional*, defaults to `128`):
+ The number of input dimensions.
+ targets_length (`int`, *optional*, defaults to `256`):
+ The length of the targets.
+ d_model (`int`, *optional*, defaults to `768`):
+ Size of the input hidden states.
+ num_layers (`int`, *optional*, defaults to `12`):
+ The number of `DecoderLayer`'s to use.
+ num_heads (`int`, *optional*, defaults to `12`):
+ The number of attention heads to use.
+ d_kv (`int`, *optional*, defaults to `64`):
+ Size of the key-value projection vectors.
+ d_ff (`int`, *optional*, defaults to `2048`):
+ The number of dimensions in the intermediate feed-forward layer of `DecoderLayer`'s.
+ dropout_rate (`float`, *optional*, defaults to `0.1`):
+ Dropout probability.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ input_dims: int = 128,
+ targets_length: int = 256,
+ max_decoder_noise_time: float = 2000.0,
+ d_model: int = 768,
+ num_layers: int = 12,
+ num_heads: int = 12,
+ d_kv: int = 64,
+ d_ff: int = 2048,
+ dropout_rate: float = 0.1,
+ ):
+ super().__init__()
+
+ self.conditioning_emb = nn.Sequential(
+ nn.Linear(d_model, d_model * 4, bias=False),
+ nn.SiLU(),
+ nn.Linear(d_model * 4, d_model * 4, bias=False),
+ nn.SiLU(),
+ )
+
+ self.position_encoding = nn.Embedding(targets_length, d_model)
+ self.position_encoding.weight.requires_grad = False
+
+ self.continuous_inputs_projection = nn.Linear(input_dims, d_model, bias=False)
+
+ self.dropout = nn.Dropout(p=dropout_rate)
+
+ self.decoders = nn.ModuleList()
+ for lyr_num in range(num_layers):
+ # FiLM conditional T5 decoder
+ lyr = DecoderLayer(d_model=d_model, d_kv=d_kv, num_heads=num_heads, d_ff=d_ff, dropout_rate=dropout_rate)
+ self.decoders.append(lyr)
+
+ self.decoder_norm = T5LayerNorm(d_model)
+
+ self.post_dropout = nn.Dropout(p=dropout_rate)
+ self.spec_out = nn.Linear(d_model, input_dims, bias=False)
+
+ def encoder_decoder_mask(self, query_input: torch.FloatTensor, key_input: torch.FloatTensor) -> torch.FloatTensor:
+ mask = torch.mul(query_input.unsqueeze(-1), key_input.unsqueeze(-2))
+ return mask.unsqueeze(-3)
+
+ def forward(self, encodings_and_masks, decoder_input_tokens, decoder_noise_time):
+ batch, _, _ = decoder_input_tokens.shape
+ assert decoder_noise_time.shape == (batch,)
+
+ # decoder_noise_time is in [0, 1), so rescale to expected timing range.
+ time_steps = get_timestep_embedding(
+ decoder_noise_time * self.config.max_decoder_noise_time,
+ embedding_dim=self.config.d_model,
+ max_period=self.config.max_decoder_noise_time,
+ ).to(dtype=self.dtype)
+
+ conditioning_emb = self.conditioning_emb(time_steps).unsqueeze(1)
+
+ assert conditioning_emb.shape == (batch, 1, self.config.d_model * 4)
+
+ seq_length = decoder_input_tokens.shape[1]
+
+ # If we want to use relative positions for audio context, we can just offset
+ # this sequence by the length of encodings_and_masks.
+ decoder_positions = torch.broadcast_to(
+ torch.arange(seq_length, device=decoder_input_tokens.device),
+ (batch, seq_length),
+ )
+
+ position_encodings = self.position_encoding(decoder_positions)
+
+ inputs = self.continuous_inputs_projection(decoder_input_tokens)
+ inputs += position_encodings
+ y = self.dropout(inputs)
+
+ # decoder: No padding present.
+ decoder_mask = torch.ones(
+ decoder_input_tokens.shape[:2], device=decoder_input_tokens.device, dtype=inputs.dtype
+ )
+
+ # Translate encoding masks to encoder-decoder masks.
+ encodings_and_encdec_masks = [(x, self.encoder_decoder_mask(decoder_mask, y)) for x, y in encodings_and_masks]
+
+ # cross attend style: concat encodings
+ encoded = torch.cat([x[0] for x in encodings_and_encdec_masks], dim=1)
+ encoder_decoder_mask = torch.cat([x[1] for x in encodings_and_encdec_masks], dim=-1)
+
+ for lyr in self.decoders:
+ y = lyr(
+ y,
+ conditioning_emb=conditioning_emb,
+ encoder_hidden_states=encoded,
+ encoder_attention_mask=encoder_decoder_mask,
+ )[0]
+
+ y = self.decoder_norm(y)
+ y = self.post_dropout(y)
+
+ spec_out = self.spec_out(y)
+ return spec_out
+
+
+class DecoderLayer(nn.Module):
+ r"""
+ T5 decoder layer.
+
+ Args:
+ d_model (`int`):
+ Size of the input hidden states.
+ d_kv (`int`):
+ Size of the key-value projection vectors.
+ num_heads (`int`):
+ Number of attention heads.
+ d_ff (`int`):
+ Size of the intermediate feed-forward layer.
+ dropout_rate (`float`):
+ Dropout probability.
+ layer_norm_epsilon (`float`, *optional*, defaults to `1e-6`):
+ A small value used for numerical stability to avoid dividing by zero.
+ """
+
+ def __init__(
+ self, d_model: int, d_kv: int, num_heads: int, d_ff: int, dropout_rate: float, layer_norm_epsilon: float = 1e-6
+ ):
+ super().__init__()
+ self.layer = nn.ModuleList()
+
+ # cond self attention: layer 0
+ self.layer.append(
+ T5LayerSelfAttentionCond(d_model=d_model, d_kv=d_kv, num_heads=num_heads, dropout_rate=dropout_rate)
+ )
+
+ # cross attention: layer 1
+ self.layer.append(
+ T5LayerCrossAttention(
+ d_model=d_model,
+ d_kv=d_kv,
+ num_heads=num_heads,
+ dropout_rate=dropout_rate,
+ layer_norm_epsilon=layer_norm_epsilon,
+ )
+ )
+
+ # Film Cond MLP + dropout: last layer
+ self.layer.append(
+ T5LayerFFCond(d_model=d_model, d_ff=d_ff, dropout_rate=dropout_rate, layer_norm_epsilon=layer_norm_epsilon)
+ )
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ conditioning_emb: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.Tensor] = None,
+ encoder_attention_mask: Optional[torch.Tensor] = None,
+ encoder_decoder_position_bias=None,
+ ) -> Tuple[torch.FloatTensor]:
+ hidden_states = self.layer[0](
+ hidden_states,
+ conditioning_emb=conditioning_emb,
+ attention_mask=attention_mask,
+ )
+
+ if encoder_hidden_states is not None:
+ encoder_extended_attention_mask = torch.where(encoder_attention_mask > 0, 0, -1e10).to(
+ encoder_hidden_states.dtype
+ )
+
+ hidden_states = self.layer[1](
+ hidden_states,
+ key_value_states=encoder_hidden_states,
+ attention_mask=encoder_extended_attention_mask,
+ )
+
+ # Apply Film Conditional Feed Forward layer
+ hidden_states = self.layer[-1](hidden_states, conditioning_emb)
+
+ return (hidden_states,)
+
+
+class T5LayerSelfAttentionCond(nn.Module):
+ r"""
+ T5 style self-attention layer with conditioning.
+
+ Args:
+ d_model (`int`):
+ Size of the input hidden states.
+ d_kv (`int`):
+ Size of the key-value projection vectors.
+ num_heads (`int`):
+ Number of attention heads.
+ dropout_rate (`float`):
+ Dropout probability.
+ """
+
+ def __init__(self, d_model: int, d_kv: int, num_heads: int, dropout_rate: float):
+ super().__init__()
+ self.layer_norm = T5LayerNorm(d_model)
+ self.FiLMLayer = T5FiLMLayer(in_features=d_model * 4, out_features=d_model)
+ self.attention = Attention(query_dim=d_model, heads=num_heads, dim_head=d_kv, out_bias=False, scale_qk=False)
+ self.dropout = nn.Dropout(dropout_rate)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ conditioning_emb: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ # pre_self_attention_layer_norm
+ normed_hidden_states = self.layer_norm(hidden_states)
+
+ if conditioning_emb is not None:
+ normed_hidden_states = self.FiLMLayer(normed_hidden_states, conditioning_emb)
+
+ # Self-attention block
+ attention_output = self.attention(normed_hidden_states)
+
+ hidden_states = hidden_states + self.dropout(attention_output)
+
+ return hidden_states
+
+
+class T5LayerCrossAttention(nn.Module):
+ r"""
+ T5 style cross-attention layer.
+
+ Args:
+ d_model (`int`):
+ Size of the input hidden states.
+ d_kv (`int`):
+ Size of the key-value projection vectors.
+ num_heads (`int`):
+ Number of attention heads.
+ dropout_rate (`float`):
+ Dropout probability.
+ layer_norm_epsilon (`float`):
+ A small value used for numerical stability to avoid dividing by zero.
+ """
+
+ def __init__(self, d_model: int, d_kv: int, num_heads: int, dropout_rate: float, layer_norm_epsilon: float):
+ super().__init__()
+ self.attention = Attention(query_dim=d_model, heads=num_heads, dim_head=d_kv, out_bias=False, scale_qk=False)
+ self.layer_norm = T5LayerNorm(d_model, eps=layer_norm_epsilon)
+ self.dropout = nn.Dropout(dropout_rate)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ key_value_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ normed_hidden_states = self.layer_norm(hidden_states)
+ attention_output = self.attention(
+ normed_hidden_states,
+ encoder_hidden_states=key_value_states,
+ attention_mask=attention_mask.squeeze(1),
+ )
+ layer_output = hidden_states + self.dropout(attention_output)
+ return layer_output
+
+
+class T5LayerFFCond(nn.Module):
+ r"""
+ T5 style feed-forward conditional layer.
+
+ Args:
+ d_model (`int`):
+ Size of the input hidden states.
+ d_ff (`int`):
+ Size of the intermediate feed-forward layer.
+ dropout_rate (`float`):
+ Dropout probability.
+ layer_norm_epsilon (`float`):
+ A small value used for numerical stability to avoid dividing by zero.
+ """
+
+ def __init__(self, d_model: int, d_ff: int, dropout_rate: float, layer_norm_epsilon: float):
+ super().__init__()
+ self.DenseReluDense = T5DenseGatedActDense(d_model=d_model, d_ff=d_ff, dropout_rate=dropout_rate)
+ self.film = T5FiLMLayer(in_features=d_model * 4, out_features=d_model)
+ self.layer_norm = T5LayerNorm(d_model, eps=layer_norm_epsilon)
+ self.dropout = nn.Dropout(dropout_rate)
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, conditioning_emb: Optional[torch.FloatTensor] = None
+ ) -> torch.FloatTensor:
+ forwarded_states = self.layer_norm(hidden_states)
+ if conditioning_emb is not None:
+ forwarded_states = self.film(forwarded_states, conditioning_emb)
+
+ forwarded_states = self.DenseReluDense(forwarded_states)
+ hidden_states = hidden_states + self.dropout(forwarded_states)
+ return hidden_states
+
+
+class T5DenseGatedActDense(nn.Module):
+ r"""
+ T5 style feed-forward layer with gated activations and dropout.
+
+ Args:
+ d_model (`int`):
+ Size of the input hidden states.
+ d_ff (`int`):
+ Size of the intermediate feed-forward layer.
+ dropout_rate (`float`):
+ Dropout probability.
+ """
+
+ def __init__(self, d_model: int, d_ff: int, dropout_rate: float):
+ super().__init__()
+ self.wi_0 = nn.Linear(d_model, d_ff, bias=False)
+ self.wi_1 = nn.Linear(d_model, d_ff, bias=False)
+ self.wo = nn.Linear(d_ff, d_model, bias=False)
+ self.dropout = nn.Dropout(dropout_rate)
+ self.act = NewGELUActivation()
+
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+ hidden_gelu = self.act(self.wi_0(hidden_states))
+ hidden_linear = self.wi_1(hidden_states)
+ hidden_states = hidden_gelu * hidden_linear
+ hidden_states = self.dropout(hidden_states)
+
+ hidden_states = self.wo(hidden_states)
+ return hidden_states
+
+
+class T5LayerNorm(nn.Module):
+ r"""
+ T5 style layer normalization module.
+
+ Args:
+ hidden_size (`int`):
+ Size of the input hidden states.
+ eps (`float`, `optional`, defaults to `1e-6`):
+ A small value used for numerical stability to avoid dividing by zero.
+ """
+
+ def __init__(self, hidden_size: int, eps: float = 1e-6):
+ """
+ Construct a layernorm module in the T5 style. No bias and no subtraction of mean.
+ """
+ super().__init__()
+ self.weight = nn.Parameter(torch.ones(hidden_size))
+ self.variance_epsilon = eps
+
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+ # T5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean
+ # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus variance is calculated
+ # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for
+ # half-precision inputs is done in fp32
+
+ variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+ # convert into half-precision if necessary
+ if self.weight.dtype in [torch.float16, torch.bfloat16]:
+ hidden_states = hidden_states.to(self.weight.dtype)
+
+ return self.weight * hidden_states
+
+
+class NewGELUActivation(nn.Module):
+ """
+ Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT). Also see
+ the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
+ """
+
+ def forward(self, input: torch.Tensor) -> torch.Tensor:
+ return 0.5 * input * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (input + 0.044715 * torch.pow(input, 3.0))))
+
+
+class T5FiLMLayer(nn.Module):
+ """
+ T5 style FiLM Layer.
+
+ Args:
+ in_features (`int`):
+ Number of input features.
+ out_features (`int`):
+ Number of output features.
+ """
+
+ def __init__(self, in_features: int, out_features: int):
+ super().__init__()
+ self.scale_bias = nn.Linear(in_features, out_features * 2, bias=False)
+
+ def forward(self, x: torch.FloatTensor, conditioning_emb: torch.FloatTensor) -> torch.FloatTensor:
+ emb = self.scale_bias(conditioning_emb)
+ scale, shift = torch.chunk(emb, 2, -1)
+ x = x * (1 + scale) + shift
+ return x
diff --git a/model/diffusers_c/models/transformers/transformer_2d.py b/model/diffusers_c/models/transformers/transformer_2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..8b391eeebfd9a3876b3984e98e1d1e0927d73271
--- /dev/null
+++ b/model/diffusers_c/models/transformers/transformer_2d.py
@@ -0,0 +1,470 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from ..attention import BasicTransformerBlock
+from ..embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection
+from ..lora import LoRACompatibleConv, LoRACompatibleLinear
+from ..modeling_utils import ModelMixin
+from ..normalization import AdaLayerNormSingle
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+ """
+ The output of [`Transformer2DModel`].
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+ The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+ distributions for the unnoised latent pixels.
+ """
+
+ sample: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+ """
+ A 2D Transformer model for image-like data.
+
+ Parameters:
+ num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+ in_channels (`int`, *optional*):
+ The number of channels in the input and output (specify if the input is **continuous**).
+ num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+ sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+ This is fixed during training since it is used to learn a number of position embeddings.
+ num_vector_embeds (`int`, *optional*):
+ The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+ Includes the class for the masked latent pixel.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+ num_embeds_ada_norm ( `int`, *optional*):
+ The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+ `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+ added to the hidden states.
+
+ During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+ attention_bias (`bool`, *optional*):
+ Configure if the `TransformerBlocks` attention should contain a bias parameter.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ num_attention_heads: int = 16,
+ attention_head_dim: int = 88,
+ in_channels: Optional[int] = None,
+ out_channels: Optional[int] = None,
+ num_layers: int = 1,
+ dropout: float = 0.0,
+ norm_num_groups: int = 32,
+ cross_attention_dim: Optional[int] = None,
+ attention_bias: bool = False,
+ sample_size: Optional[int] = None,
+ num_vector_embeds: Optional[int] = None,
+ patch_size: Optional[int] = None,
+ activation_fn: str = "geglu",
+ num_embeds_ada_norm: Optional[int] = None,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ double_self_attention: bool = False,
+ upcast_attention: bool = False,
+ norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+ norm_elementwise_affine: bool = True,
+ norm_eps: float = 1e-5,
+ attention_type: str = "default",
+ caption_channels: int = None,
+ interpolation_scale: float = None,
+ ):
+ super().__init__()
+ if patch_size is not None:
+ if norm_type not in ["ada_norm", "ada_norm_zero", "ada_norm_single"]:
+ raise NotImplementedError(
+ f"Forward pass is not implemented when `patch_size` is not None and `norm_type` is '{norm_type}'."
+ )
+ elif norm_type in ["ada_norm", "ada_norm_zero"] and num_embeds_ada_norm is None:
+ raise ValueError(
+ f"When using a `patch_size` and this `norm_type` ({norm_type}), `num_embeds_ada_norm` cannot be None."
+ )
+
+ self.use_linear_projection = use_linear_projection
+ self.num_attention_heads = num_attention_heads
+ self.attention_head_dim = attention_head_dim
+ inner_dim = num_attention_heads * attention_head_dim
+
+ conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+ linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+ # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+ # Define whether input is continuous or discrete depending on configuration
+ self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+ self.is_input_vectorized = num_vector_embeds is not None
+ self.is_input_patches = in_channels is not None and patch_size is not None
+
+ if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+ deprecation_message = (
+ f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+ " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+ " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+ " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+ " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+ )
+ deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
+ norm_type = "ada_norm"
+
+ if self.is_input_continuous and self.is_input_vectorized:
+ raise ValueError(
+ f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+ " sure that either `in_channels` or `num_vector_embeds` is None."
+ )
+ elif self.is_input_vectorized and self.is_input_patches:
+ raise ValueError(
+ f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+ " sure that either `num_vector_embeds` or `num_patches` is None."
+ )
+ elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+ raise ValueError(
+ f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+ f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+ )
+
+ # 2. Define input layers
+ if self.is_input_continuous:
+ self.in_channels = in_channels
+
+ self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+ if use_linear_projection:
+ self.proj_in = linear_cls(in_channels, inner_dim)
+ else:
+ self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+ elif self.is_input_vectorized:
+ assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+ assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
+
+ self.height = sample_size
+ self.width = sample_size
+ self.num_vector_embeds = num_vector_embeds
+ self.num_latent_pixels = self.height * self.width
+
+ self.latent_image_embedding = ImagePositionalEmbeddings(
+ num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
+ )
+ elif self.is_input_patches:
+ assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
+
+ self.height = sample_size
+ self.width = sample_size
+
+ self.patch_size = patch_size
+ interpolation_scale = (
+ interpolation_scale if interpolation_scale is not None else max(self.config.sample_size // 64, 1)
+ )
+ self.pos_embed = PatchEmbed(
+ height=sample_size,
+ width=sample_size,
+ patch_size=patch_size,
+ in_channels=in_channels,
+ embed_dim=inner_dim,
+ interpolation_scale=interpolation_scale,
+ )
+
+ # 3. Define transformers blocks
+ self.transformer_blocks = nn.ModuleList(
+ [
+ BasicTransformerBlock(
+ inner_dim,
+ num_attention_heads,
+ attention_head_dim,
+ dropout=dropout,
+ cross_attention_dim=cross_attention_dim,
+ activation_fn=activation_fn,
+ num_embeds_ada_norm=num_embeds_ada_norm,
+ attention_bias=attention_bias,
+ only_cross_attention=only_cross_attention,
+ double_self_attention=double_self_attention,
+ upcast_attention=upcast_attention,
+ norm_type=norm_type,
+ norm_elementwise_affine=norm_elementwise_affine,
+ norm_eps=norm_eps,
+ attention_type=attention_type,
+ )
+ for d in range(num_layers)
+ ]
+ )
+
+ # 4. Define output layers
+ self.out_channels = in_channels if out_channels is None else out_channels
+ if self.is_input_continuous:
+ # TODO: should use out_channels for continuous projections
+ if use_linear_projection:
+ self.proj_out = linear_cls(inner_dim, in_channels)
+ else:
+ self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+ elif self.is_input_vectorized:
+ self.norm_out = nn.LayerNorm(inner_dim)
+ self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
+ elif self.is_input_patches and norm_type != "ada_norm_single":
+ self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+ self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
+ self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+ elif self.is_input_patches and norm_type == "ada_norm_single":
+ self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+ self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+ self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+
+ # 5. PixArt-Alpha blocks.
+ self.adaln_single = None
+ self.use_additional_conditions = False
+ if norm_type == "ada_norm_single":
+ self.use_additional_conditions = self.config.sample_size == 128
+ # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+ # additional conditions until we find better name
+ self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
+
+ self.caption_projection = None
+ if caption_channels is not None:
+ self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+ self.gradient_checkpointing = False
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if hasattr(module, "gradient_checkpointing"):
+ module.gradient_checkpointing = value
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ encoder_hidden_states: Optional[torch.Tensor] = None,
+ timestep: Optional[torch.LongTensor] = None,
+ added_cond_kwargs: Dict[str, torch.Tensor] = None,
+ class_labels: Optional[torch.LongTensor] = None,
+ cross_attention_kwargs: Dict[str, Any] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ encoder_attention_mask: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ ):
+ """
+ The [`Transformer2DModel`] forward method.
+
+ Args:
+ hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+ Input `hidden_states`.
+ encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+ Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+ self-attention.
+ timestep ( `torch.LongTensor`, *optional*):
+ Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+ class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+ Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+ `AdaLayerZeroNorm`.
+ cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ attention_mask ( `torch.Tensor`, *optional*):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ encoder_attention_mask ( `torch.Tensor`, *optional*):
+ Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+ * Mask `(batch, sequence_length)` True = keep, False = discard.
+ * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+ If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+ above. This bias will be added to the cross-attention scores.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+ tuple.
+
+ Returns:
+ If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+ `tuple` where the first element is the sample tensor.
+ """
+ # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+ # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+ # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+ # expects mask of shape:
+ # [batch, key_tokens]
+ # adds singleton query_tokens dimension:
+ # [batch, 1, key_tokens]
+ # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+ # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+ # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+ if attention_mask is not None and attention_mask.ndim == 2:
+ # assume that mask is expressed as:
+ # (1 = keep, 0 = discard)
+ # convert mask into a bias that can be added to attention scores:
+ # (keep = +0, discard = -10000.0)
+ attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # convert encoder_attention_mask to a bias the same way we do for attention_mask
+ if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+ encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+ encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+ # Retrieve lora scale.
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+ # 1. Input
+ if self.is_input_continuous:
+ batch, _, height, width = hidden_states.shape
+ residual = hidden_states
+
+ hidden_states = self.norm(hidden_states)
+ if not self.use_linear_projection:
+ hidden_states = (
+ self.proj_in(hidden_states, scale=lora_scale)
+ if not USE_PEFT_BACKEND
+ else self.proj_in(hidden_states)
+ )
+ inner_dim = hidden_states.shape[1]
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+ else:
+ inner_dim = hidden_states.shape[1]
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+ hidden_states = (
+ self.proj_in(hidden_states, scale=lora_scale)
+ if not USE_PEFT_BACKEND
+ else self.proj_in(hidden_states)
+ )
+
+ elif self.is_input_vectorized:
+ hidden_states = self.latent_image_embedding(hidden_states)
+ elif self.is_input_patches:
+ height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
+ hidden_states = self.pos_embed(hidden_states)
+
+ if self.adaln_single is not None:
+ if self.use_additional_conditions and added_cond_kwargs is None:
+ raise ValueError(
+ "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
+ )
+ batch_size = hidden_states.shape[0]
+ timestep, embedded_timestep = self.adaln_single(
+ timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+ )
+
+ # 2. Blocks
+ if self.caption_projection is not None:
+ batch_size = hidden_states.shape[0]
+ encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+ encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+ for block in self.transformer_blocks:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block),
+ hidden_states,
+ attention_mask,
+ encoder_hidden_states,
+ encoder_attention_mask,
+ timestep,
+ cross_attention_kwargs,
+ class_labels,
+ **ckpt_kwargs,
+ )
+ else:
+ hidden_states = block(
+ hidden_states,
+ attention_mask=attention_mask,
+ encoder_hidden_states=encoder_hidden_states,
+ encoder_attention_mask=encoder_attention_mask,
+ timestep=timestep,
+ cross_attention_kwargs=cross_attention_kwargs,
+ class_labels=class_labels,
+ )
+
+ # 3. Output
+ if self.is_input_continuous:
+ if not self.use_linear_projection:
+ hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+ hidden_states = (
+ self.proj_out(hidden_states, scale=lora_scale)
+ if not USE_PEFT_BACKEND
+ else self.proj_out(hidden_states)
+ )
+ else:
+ hidden_states = (
+ self.proj_out(hidden_states, scale=lora_scale)
+ if not USE_PEFT_BACKEND
+ else self.proj_out(hidden_states)
+ )
+ hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+ output = hidden_states + residual
+ elif self.is_input_vectorized:
+ hidden_states = self.norm_out(hidden_states)
+ logits = self.out(hidden_states)
+ # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
+ logits = logits.permute(0, 2, 1)
+
+ # log(p(x_0))
+ output = F.log_softmax(logits.double(), dim=1).float()
+
+ if self.is_input_patches:
+ if self.config.norm_type != "ada_norm_single":
+ conditioning = self.transformer_blocks[0].norm1.emb(
+ timestep, class_labels, hidden_dtype=hidden_states.dtype
+ )
+ shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+ hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+ hidden_states = self.proj_out_2(hidden_states)
+ elif self.config.norm_type == "ada_norm_single":
+ shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+ hidden_states = self.norm_out(hidden_states)
+ # Modulation
+ hidden_states = hidden_states * (1 + scale) + shift
+ hidden_states = self.proj_out(hidden_states)
+ hidden_states = hidden_states.squeeze(1)
+
+ # unpatchify
+ if self.adaln_single is None:
+ height = width = int(hidden_states.shape[1] ** 0.5)
+ hidden_states = hidden_states.reshape(
+ shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+ )
+ hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+ output = hidden_states.reshape(
+ shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+ )
+
+ if not return_dict:
+ return (output,)
+
+ return Transformer2DModelOutput(sample=output)
diff --git a/model/diffusers_c/models/transformers/transformer_temporal.py b/model/diffusers_c/models/transformers/transformer_temporal.py
new file mode 100644
index 0000000000000000000000000000000000000000..9c61eaee2d2146bf80efdfdba58a81619369bdd7
--- /dev/null
+++ b/model/diffusers_c/models/transformers/transformer_temporal.py
@@ -0,0 +1,379 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ..attention import BasicTransformerBlock, TemporalBasicTransformerBlock
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from ..resnet import AlphaBlender
+
+
+@dataclass
+class TransformerTemporalModelOutput(BaseOutput):
+ """
+ The output of [`TransformerTemporalModel`].
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size x num_frames, num_channels, height, width)`):
+ The hidden states output conditioned on `encoder_hidden_states` input.
+ """
+
+ sample: torch.FloatTensor
+
+
+class TransformerTemporalModel(ModelMixin, ConfigMixin):
+ """
+ A Transformer model for video-like data.
+
+ Parameters:
+ num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+ in_channels (`int`, *optional*):
+ The number of channels in the input and output (specify if the input is **continuous**).
+ num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+ attention_bias (`bool`, *optional*):
+ Configure if the `TransformerBlock` attention should contain a bias parameter.
+ sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+ This is fixed during training since it is used to learn a number of position embeddings.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`):
+ Activation function to use in feed-forward. See `diffusers.models.activations.get_activation` for supported
+ activation functions.
+ norm_elementwise_affine (`bool`, *optional*):
+ Configure if the `TransformerBlock` should use learnable elementwise affine parameters for normalization.
+ double_self_attention (`bool`, *optional*):
+ Configure if each `TransformerBlock` should contain two self-attention layers.
+ positional_embeddings: (`str`, *optional*):
+ The type of positional embeddings to apply to the sequence input before passing use.
+ num_positional_embeddings: (`int`, *optional*):
+ The maximum length of the sequence over which to apply positional embeddings.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ num_attention_heads: int = 16,
+ attention_head_dim: int = 88,
+ in_channels: Optional[int] = None,
+ out_channels: Optional[int] = None,
+ num_layers: int = 1,
+ dropout: float = 0.0,
+ norm_num_groups: int = 32,
+ cross_attention_dim: Optional[int] = None,
+ attention_bias: bool = False,
+ sample_size: Optional[int] = None,
+ activation_fn: str = "geglu",
+ norm_elementwise_affine: bool = True,
+ double_self_attention: bool = True,
+ positional_embeddings: Optional[str] = None,
+ num_positional_embeddings: Optional[int] = None,
+ ):
+ super().__init__()
+ self.num_attention_heads = num_attention_heads
+ self.attention_head_dim = attention_head_dim
+ inner_dim = num_attention_heads * attention_head_dim
+
+ self.in_channels = in_channels
+
+ self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+ self.proj_in = nn.Linear(in_channels, inner_dim)
+
+ # 3. Define transformers blocks
+ self.transformer_blocks = nn.ModuleList(
+ [
+ BasicTransformerBlock(
+ inner_dim,
+ num_attention_heads,
+ attention_head_dim,
+ dropout=dropout,
+ cross_attention_dim=cross_attention_dim,
+ activation_fn=activation_fn,
+ attention_bias=attention_bias,
+ double_self_attention=double_self_attention,
+ norm_elementwise_affine=norm_elementwise_affine,
+ positional_embeddings=positional_embeddings,
+ num_positional_embeddings=num_positional_embeddings,
+ )
+ for d in range(num_layers)
+ ]
+ )
+
+ self.proj_out = nn.Linear(inner_dim, in_channels)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.LongTensor] = None,
+ timestep: Optional[torch.LongTensor] = None,
+ class_labels: torch.LongTensor = None,
+ num_frames: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ return_dict: bool = True,
+ ) -> TransformerTemporalModelOutput:
+ """
+ The [`TransformerTemporal`] forward method.
+
+ Args:
+ hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+ Input hidden_states.
+ encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+ Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+ self-attention.
+ timestep ( `torch.LongTensor`, *optional*):
+ Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+ class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+ Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+ `AdaLayerZeroNorm`.
+ num_frames (`int`, *optional*, defaults to 1):
+ The number of frames to be processed per batch. This is used to reshape the hidden states.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+ tuple.
+
+ Returns:
+ [`~models.transformer_temporal.TransformerTemporalModelOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.transformer_temporal.TransformerTemporalModelOutput`] is
+ returned, otherwise a `tuple` where the first element is the sample tensor.
+ """
+ # 1. Input
+ batch_frames, channel, height, width = hidden_states.shape
+ batch_size = batch_frames // num_frames
+
+ residual = hidden_states
+
+ hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, channel, height, width)
+ hidden_states = hidden_states.permute(0, 2, 1, 3, 4)
+
+ hidden_states = self.norm(hidden_states)
+ hidden_states = hidden_states.permute(0, 3, 4, 2, 1).reshape(batch_size * height * width, num_frames, channel)
+
+ hidden_states = self.proj_in(hidden_states)
+
+ # 2. Blocks
+ for block in self.transformer_blocks:
+ hidden_states = block(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ timestep=timestep,
+ cross_attention_kwargs=cross_attention_kwargs,
+ class_labels=class_labels,
+ )
+
+ # 3. Output
+ hidden_states = self.proj_out(hidden_states)
+ hidden_states = (
+ hidden_states[None, None, :]
+ .reshape(batch_size, height, width, num_frames, channel)
+ .permute(0, 3, 4, 1, 2)
+ .contiguous()
+ )
+ hidden_states = hidden_states.reshape(batch_frames, channel, height, width)
+
+ output = hidden_states + residual
+
+ if not return_dict:
+ return (output,)
+
+ return TransformerTemporalModelOutput(sample=output)
+
+
+class TransformerSpatioTemporalModel(nn.Module):
+ """
+ A Transformer model for video-like data.
+
+ Parameters:
+ num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+ in_channels (`int`, *optional*):
+ The number of channels in the input and output (specify if the input is **continuous**).
+ out_channels (`int`, *optional*):
+ The number of channels in the output (specify if the input is **continuous**).
+ num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+ cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+ """
+
+ def __init__(
+ self,
+ num_attention_heads: int = 16,
+ attention_head_dim: int = 88,
+ in_channels: int = 320,
+ out_channels: Optional[int] = None,
+ num_layers: int = 1,
+ cross_attention_dim: Optional[int] = None,
+ ):
+ super().__init__()
+ self.num_attention_heads = num_attention_heads
+ self.attention_head_dim = attention_head_dim
+
+ inner_dim = num_attention_heads * attention_head_dim
+ self.inner_dim = inner_dim
+
+ # 2. Define input layers
+ self.in_channels = in_channels
+ self.norm = torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6)
+ self.proj_in = nn.Linear(in_channels, inner_dim)
+
+ # 3. Define transformers blocks
+ self.transformer_blocks = nn.ModuleList(
+ [
+ BasicTransformerBlock(
+ inner_dim,
+ num_attention_heads,
+ attention_head_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+ for d in range(num_layers)
+ ]
+ )
+
+ time_mix_inner_dim = inner_dim
+ self.temporal_transformer_blocks = nn.ModuleList(
+ [
+ TemporalBasicTransformerBlock(
+ inner_dim,
+ time_mix_inner_dim,
+ num_attention_heads,
+ attention_head_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+ for _ in range(num_layers)
+ ]
+ )
+
+ time_embed_dim = in_channels * 4
+ self.time_pos_embed = TimestepEmbedding(in_channels, time_embed_dim, out_dim=in_channels)
+ self.time_proj = Timesteps(in_channels, True, 0)
+ self.time_mixer = AlphaBlender(alpha=0.5, merge_strategy="learned_with_images")
+
+ # 4. Define output layers
+ self.out_channels = in_channels if out_channels is None else out_channels
+ # TODO: should use out_channels for continuous projections
+ self.proj_out = nn.Linear(inner_dim, in_channels)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ encoder_hidden_states: Optional[torch.Tensor] = None,
+ image_only_indicator: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ ):
+ """
+ Args:
+ hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
+ Input hidden_states.
+ num_frames (`int`):
+ The number of frames to be processed per batch. This is used to reshape the hidden states.
+ encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+ Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+ self-attention.
+ image_only_indicator (`torch.LongTensor` of shape `(batch size, num_frames)`, *optional*):
+ A tensor indicating whether the input contains only images. 1 indicates that the input contains only
+ images, 0 indicates that the input contains video frames.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.transformer_temporal.TransformerTemporalModelOutput`] instead of a plain
+ tuple.
+
+ Returns:
+ [`~models.transformer_temporal.TransformerTemporalModelOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.transformer_temporal.TransformerTemporalModelOutput`] is
+ returned, otherwise a `tuple` where the first element is the sample tensor.
+ """
+ # 1. Input
+ batch_frames, _, height, width = hidden_states.shape
+ num_frames = image_only_indicator.shape[-1]
+ batch_size = batch_frames // num_frames
+
+ time_context = encoder_hidden_states
+ time_context_first_timestep = time_context[None, :].reshape(
+ batch_size, num_frames, -1, time_context.shape[-1]
+ )[:, 0]
+ time_context = time_context_first_timestep[None, :].broadcast_to(
+ height * width, batch_size, 1, time_context.shape[-1]
+ )
+ time_context = time_context.reshape(height * width * batch_size, 1, time_context.shape[-1])
+
+ residual = hidden_states
+
+ hidden_states = self.norm(hidden_states)
+ inner_dim = hidden_states.shape[1]
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch_frames, height * width, inner_dim)
+ hidden_states = self.proj_in(hidden_states)
+
+ num_frames_emb = torch.arange(num_frames, device=hidden_states.device)
+ num_frames_emb = num_frames_emb.repeat(batch_size, 1)
+ num_frames_emb = num_frames_emb.reshape(-1)
+ t_emb = self.time_proj(num_frames_emb)
+
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=hidden_states.dtype)
+
+ emb = self.time_pos_embed(t_emb)
+ emb = emb[:, None, :]
+
+ # 2. Blocks
+ for block, temporal_block in zip(self.transformer_blocks, self.temporal_transformer_blocks):
+ if self.training and self.gradient_checkpointing:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ block,
+ hidden_states,
+ None,
+ encoder_hidden_states,
+ None,
+ use_reentrant=False,
+ )
+ else:
+ hidden_states = block(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ )
+
+ hidden_states_mix = hidden_states
+ hidden_states_mix = hidden_states_mix + emb
+
+ hidden_states_mix = temporal_block(
+ hidden_states_mix,
+ num_frames=num_frames,
+ encoder_hidden_states=time_context,
+ )
+ hidden_states = self.time_mixer(
+ x_spatial=hidden_states,
+ x_temporal=hidden_states_mix,
+ image_only_indicator=image_only_indicator,
+ )
+
+ # 3. Output
+ hidden_states = self.proj_out(hidden_states)
+ hidden_states = hidden_states.reshape(batch_frames, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+ output = hidden_states + residual
+
+ if not return_dict:
+ return (output,)
+
+ return TransformerTemporalModelOutput(sample=output)
diff --git a/model/diffusers_c/models/unet_1d.py b/model/diffusers_c/models/unet_1d.py
new file mode 100644
index 0000000000000000000000000000000000000000..e857c90cae4014f2e4908bc464375af1ffd027e9
--- /dev/null
+++ b/model/diffusers_c/models/unet_1d.py
@@ -0,0 +1,26 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..utils import deprecate
+from .unets.unet_1d import UNet1DModel, UNet1DOutput
+
+
+class UNet1DOutput(UNet1DOutput):
+ deprecation_message = "Importing `UNet1DOutput` from `diffusers.models.unet_1d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d import UNet1DOutput`, instead."
+ deprecate("UNet1DOutput", "0.29", deprecation_message)
+
+
+class UNet1DModel(UNet1DModel):
+ deprecation_message = "Importing `UNet1DModel` from `diffusers.models.unet_1d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d import UNet1DModel`, instead."
+ deprecate("UNet1DModel", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/unet_1d_blocks.py b/model/diffusers_c/models/unet_1d_blocks.py
new file mode 100644
index 0000000000000000000000000000000000000000..6b0f09457d17ff9a71f9cecc61e92b4a56724dba
--- /dev/null
+++ b/model/diffusers_c/models/unet_1d_blocks.py
@@ -0,0 +1,203 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from ..utils import deprecate
+from .unets.unet_1d_blocks import (
+ AttnDownBlock1D,
+ AttnUpBlock1D,
+ DownBlock1D,
+ DownBlock1DNoSkip,
+ DownResnetBlock1D,
+ Downsample1d,
+ MidResTemporalBlock1D,
+ OutConv1DBlock,
+ OutValueFunctionBlock,
+ ResConvBlock,
+ SelfAttention1d,
+ UNetMidBlock1D,
+ UpBlock1D,
+ UpBlock1DNoSkip,
+ UpResnetBlock1D,
+ Upsample1d,
+ ValueFunctionMidBlock1D,
+)
+
+
+class DownResnetBlock1D(DownResnetBlock1D):
+ deprecation_message = "Importing `DownResnetBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import DownResnetBlock1D`, instead."
+ deprecate("DownResnetBlock1D", "0.29", deprecation_message)
+
+
+class UpResnetBlock1D(UpResnetBlock1D):
+ deprecation_message = "Importing `UpResnetBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import UpResnetBlock1D`, instead."
+ deprecate("UpResnetBlock1D", "0.29", deprecation_message)
+
+
+class ValueFunctionMidBlock1D(ValueFunctionMidBlock1D):
+ deprecation_message = "Importing `ValueFunctionMidBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import ValueFunctionMidBlock1D`, instead."
+ deprecate("ValueFunctionMidBlock1D", "0.29", deprecation_message)
+
+
+class OutConv1DBlock(OutConv1DBlock):
+ deprecation_message = "Importing `OutConv1DBlock` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import OutConv1DBlock`, instead."
+ deprecate("OutConv1DBlock", "0.29", deprecation_message)
+
+
+class OutValueFunctionBlock(OutValueFunctionBlock):
+ deprecation_message = "Importing `OutValueFunctionBlock` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import OutValueFunctionBlock`, instead."
+ deprecate("OutValueFunctionBlock", "0.29", deprecation_message)
+
+
+class Downsample1d(Downsample1d):
+ deprecation_message = "Importing `Downsample1d` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import Downsample1d`, instead."
+ deprecate("Downsample1d", "0.29", deprecation_message)
+
+
+class Upsample1d(Upsample1d):
+ deprecation_message = "Importing `Upsample1d` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import Upsample1d`, instead."
+ deprecate("Upsample1d", "0.29", deprecation_message)
+
+
+class SelfAttention1d(SelfAttention1d):
+ deprecation_message = "Importing `SelfAttention1d` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import SelfAttention1d`, instead."
+ deprecate("SelfAttention1d", "0.29", deprecation_message)
+
+
+class ResConvBlock(ResConvBlock):
+ deprecation_message = "Importing `ResConvBlock` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import ResConvBlock`, instead."
+ deprecate("ResConvBlock", "0.29", deprecation_message)
+
+
+class UNetMidBlock1D(UNetMidBlock1D):
+ deprecation_message = "Importing `UNetMidBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import UNetMidBlock1D`, instead."
+ deprecate("UNetMidBlock1D", "0.29", deprecation_message)
+
+
+class AttnDownBlock1D(AttnDownBlock1D):
+ deprecation_message = "Importing `AttnDownBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import AttnDownBlock1D`, instead."
+ deprecate("AttnDownBlock1D", "0.29", deprecation_message)
+
+
+class DownBlock1D(DownBlock1D):
+ deprecation_message = "Importing `DownBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import DownBlock1D`, instead."
+ deprecate("DownBlock1D", "0.29", deprecation_message)
+
+
+class DownBlock1DNoSkip(DownBlock1DNoSkip):
+ deprecation_message = "Importing `DownBlock1DNoSkip` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import DownBlock1DNoSkip`, instead."
+ deprecate("DownBlock1DNoSkip", "0.29", deprecation_message)
+
+
+class AttnUpBlock1D(AttnUpBlock1D):
+ deprecation_message = "Importing `AttnUpBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import AttnUpBlock1D`, instead."
+ deprecate("AttnUpBlock1D", "0.29", deprecation_message)
+
+
+class UpBlock1D(UpBlock1D):
+ deprecation_message = "Importing `UpBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import UpBlock1D`, instead."
+ deprecate("UpBlock1D", "0.29", deprecation_message)
+
+
+class UpBlock1DNoSkip(UpBlock1DNoSkip):
+ deprecation_message = "Importing `UpBlock1DNoSkip` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import UpBlock1DNoSkip`, instead."
+ deprecate("UpBlock1DNoSkip", "0.29", deprecation_message)
+
+
+class MidResTemporalBlock1D(MidResTemporalBlock1D):
+ deprecation_message = "Importing `MidResTemporalBlock1D` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import MidResTemporalBlock1D`, instead."
+ deprecate("MidResTemporalBlock1D", "0.29", deprecation_message)
+
+
+def get_down_block(
+ down_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ add_downsample: bool,
+):
+ deprecation_message = "Importing `get_down_block` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import get_down_block`, instead."
+ deprecate("get_down_block", "0.29", deprecation_message)
+
+ from .unets.unet_1d_blocks import get_down_block
+
+ return get_down_block(
+ down_block_type=down_block_type,
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ )
+
+
+def get_up_block(
+ up_block_type: str, num_layers: int, in_channels: int, out_channels: int, temb_channels: int, add_upsample: bool
+):
+ deprecation_message = "Importing `get_up_block` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import get_up_block`, instead."
+ deprecate("get_up_block", "0.29", deprecation_message)
+
+ from .unets.unet_1d_blocks import get_up_block
+
+ return get_up_block(
+ up_block_type=up_block_type,
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ )
+
+
+def get_mid_block(
+ mid_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ mid_channels: int,
+ out_channels: int,
+ embed_dim: int,
+ add_downsample: bool,
+):
+ deprecation_message = "Importing `get_mid_block` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import get_mid_block`, instead."
+ deprecate("get_mid_block", "0.29", deprecation_message)
+
+ from .unets.unet_1d_blocks import get_mid_block
+
+ return get_mid_block(
+ mid_block_type=mid_block_type,
+ num_layers=num_layers,
+ in_channels=in_channels,
+ mid_channels=mid_channels,
+ out_channels=out_channels,
+ embed_dim=embed_dim,
+ add_downsample=add_downsample,
+ )
+
+
+def get_out_block(
+ *, out_block_type: str, num_groups_out: int, embed_dim: int, out_channels: int, act_fn: str, fc_dim: int
+):
+ deprecation_message = "Importing `get_out_block` from `diffusers.models.unet_1d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_1d_blocks import get_out_block`, instead."
+ deprecate("get_out_block", "0.29", deprecation_message)
+
+ from .unets.unet_1d_blocks import get_out_block
+
+ return get_out_block(
+ out_block_type=out_block_type,
+ num_groups_out=num_groups_out,
+ embed_dim=embed_dim,
+ out_channels=out_channels,
+ act_fn=act_fn,
+ fc_dim=fc_dim,
+ )
diff --git a/model/diffusers_c/models/unet_2d.py b/model/diffusers_c/models/unet_2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..21f1fea68d6c476c659f4c7669fa955684ac6443
--- /dev/null
+++ b/model/diffusers_c/models/unet_2d.py
@@ -0,0 +1,27 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from ..utils import deprecate
+from .unets.unet_2d import UNet2DModel, UNet2DOutput
+
+
+class UNet2DOutput(UNet2DOutput):
+ deprecation_message = "Importing `UNet2DOutput` from `diffusers.models.unet_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d import UNet2DOutput`, instead."
+ deprecate("UNet2DOutput", "0.29", deprecation_message)
+
+
+class UNet2DModel(UNet2DModel):
+ deprecation_message = "Importing `UNet2DModel` from `diffusers.models.unet_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d import UNet2DModel`, instead."
+ deprecate("UNet2DModel", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/unet_2d_blocks.py b/model/diffusers_c/models/unet_2d_blocks.py
new file mode 100644
index 0000000000000000000000000000000000000000..d7ee5d69e7421f46cba36ac588b0e633040d0ff8
--- /dev/null
+++ b/model/diffusers_c/models/unet_2d_blocks.py
@@ -0,0 +1,392 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+from ..utils import deprecate
+from .unets.unet_2d_blocks import (
+ AttnDownBlock2D,
+ AttnDownEncoderBlock2D,
+ AttnSkipDownBlock2D,
+ AttnSkipUpBlock2D,
+ AttnUpBlock2D,
+ AttnUpDecoderBlock2D,
+ AutoencoderTinyBlock,
+ CrossAttnDownBlock2D,
+ CrossAttnUpBlock2D,
+ DownBlock2D,
+ KAttentionBlock,
+ KCrossAttnDownBlock2D,
+ KCrossAttnUpBlock2D,
+ KDownBlock2D,
+ KUpBlock2D,
+ ResnetDownsampleBlock2D,
+ ResnetUpsampleBlock2D,
+ SimpleCrossAttnDownBlock2D,
+ SimpleCrossAttnUpBlock2D,
+ SkipDownBlock2D,
+ SkipUpBlock2D,
+ UNetMidBlock2D,
+ UNetMidBlock2DCrossAttn,
+ UNetMidBlock2DSimpleCrossAttn,
+ UpBlock2D,
+ UpDecoderBlock2D,
+)
+
+
+def get_down_block(
+ down_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ add_downsample: bool,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ resnet_groups: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ downsample_padding: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: float = 1.0,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = None,
+ downsample_type: Optional[str] = None,
+ dropout: float = 0.0,
+):
+ deprecation_message = "Importing `get_down_block` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import get_down_block`, instead."
+ deprecate("get_down_block", "0.29", deprecation_message)
+
+ from .unets.unet_2d_blocks import get_down_block
+
+ return get_down_block(
+ down_block_type=down_block_type,
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ transformer_layers_per_block=transformer_layers_per_block,
+ num_attention_heads=num_attention_heads,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ downsample_padding=downsample_padding,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ resnet_out_scale_factor=resnet_out_scale_factor,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim,
+ downsample_type=downsample_type,
+ dropout=dropout,
+ )
+
+
+def get_mid_block(
+ mid_block_type: str,
+ temb_channels: int,
+ in_channels: int,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ resnet_groups: int,
+ output_scale_factor: float = 1.0,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ mid_block_only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = 1,
+ dropout: float = 0.0,
+):
+ if mid_block_type == "UNetMidBlock2DCrossAttn":
+ return UNetMidBlock2DCrossAttn(
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ resnet_groups=resnet_groups,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+ return UNetMidBlock2DSimpleCrossAttn(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ only_cross_attention=mid_block_only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+ elif mid_block_type == "UNetMidBlock2D":
+ return UNetMidBlock2D(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ num_layers=0,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ add_attention=False,
+ )
+ elif mid_block_type == "MidBlock2D":
+ return MidBlock2D(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resnet_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ )
+ elif mid_block_type is None:
+ return None
+ else:
+ raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+
+def get_up_block(
+ up_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ add_upsample: bool,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ resolution_idx: Optional[int] = None,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ resnet_groups: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: float = 1.0,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = None,
+ upsample_type: Optional[str] = None,
+ dropout: float = 0.0,
+):
+ deprecation_message = "Importing `get_up_block` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import get_up_block`, instead."
+ deprecate("get_up_block", "0.29", deprecation_message)
+
+ from .unets.unet_2d_blocks import get_up_block
+
+ return get_up_block(
+ up_block_type=up_block_type,
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resolution_idx=resolution_idx,
+ transformer_layers_per_block=transformer_layers_per_block,
+ num_attention_heads=num_attention_heads,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ resnet_out_scale_factor=resnet_out_scale_factor,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim,
+ upsample_type=upsample_type,
+ dropout=dropout,
+ )
+
+
+class AutoencoderTinyBlock(AutoencoderTinyBlock):
+ deprecation_message = "Importing `AutoencoderTinyBlock` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import AutoencoderTinyBlock`, instead."
+ deprecate("AutoencoderTinyBlock", "0.29", deprecation_message)
+
+
+class UNetMidBlock2D(UNetMidBlock2D):
+ deprecation_message = "Importing `UNetMidBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import UNetMidBlock2D`, instead."
+ deprecate("UNetMidBlock2D", "0.29", deprecation_message)
+
+
+class UNetMidBlock2DCrossAttn(UNetMidBlock2DCrossAttn):
+ deprecation_message = "Importing `UNetMidBlock2DCrossAttn` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import UNetMidBlock2DCrossAttn`, instead."
+ deprecate("UNetMidBlock2DCrossAttn", "0.29", deprecation_message)
+
+
+class UNetMidBlock2DSimpleCrossAttn(UNetMidBlock2DSimpleCrossAttn):
+ deprecation_message = "Importing `UNetMidBlock2DSimpleCrossAttn` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import UNetMidBlock2DSimpleCrossAttn`, instead."
+ deprecate("UNetMidBlock2DSimpleCrossAttn", "0.29", deprecation_message)
+
+
+class MidBlock2D(MidBlock2D):
+ deprecation_message = "Importing `MidBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import MidBlock2D`, instead."
+ deprecate("MidBlock2D", "0.29", deprecation_message)
+
+
+class AttnDownBlock2D(AttnDownBlock2D):
+ deprecation_message = "Importing `AttnDownBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import AttnDownBlock2D`, instead."
+ deprecate("AttnDownBlock2D", "0.29", deprecation_message)
+
+
+class CrossAttnDownBlock2D(CrossAttnDownBlock2D):
+ deprecation_message = "Importing `AttnDownBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import CrossAttnDownBlock2D`, instead."
+ deprecate("CrossAttnDownBlock2D", "0.29", deprecation_message)
+
+
+class DownBlock2D(DownBlock2D):
+ deprecation_message = "Importing `DownBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import DownBlock2D`, instead."
+ deprecate("DownBlock2D", "0.29", deprecation_message)
+
+
+class AttnDownEncoderBlock2D(AttnDownEncoderBlock2D):
+ deprecation_message = "Importing `AttnDownEncoderBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import AttnDownEncoderBlock2D`, instead."
+ deprecate("AttnDownEncoderBlock2D", "0.29", deprecation_message)
+
+
+class AttnSkipDownBlock2D(AttnSkipDownBlock2D):
+ deprecation_message = "Importing `AttnSkipDownBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import AttnSkipDownBlock2D`, instead."
+ deprecate("AttnSkipDownBlock2D", "0.29", deprecation_message)
+
+
+class SkipDownBlock2D(SkipDownBlock2D):
+ deprecation_message = "Importing `SkipDownBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import SkipDownBlock2D`, instead."
+ deprecate("SkipDownBlock2D", "0.29", deprecation_message)
+
+
+class ResnetDownsampleBlock2D(ResnetDownsampleBlock2D):
+ deprecation_message = "Importing `ResnetDownsampleBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import ResnetDownsampleBlock2D`, instead."
+ deprecate("ResnetDownsampleBlock2D", "0.29", deprecation_message)
+
+
+class SimpleCrossAttnDownBlock2D(SimpleCrossAttnDownBlock2D):
+ deprecation_message = "Importing `SimpleCrossAttnDownBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import SimpleCrossAttnDownBlock2D`, instead."
+ deprecate("SimpleCrossAttnDownBlock2D", "0.29", deprecation_message)
+
+
+class KDownBlock2D(KDownBlock2D):
+ deprecation_message = "Importing `KDownBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import KDownBlock2D`, instead."
+ deprecate("KDownBlock2D", "0.29", deprecation_message)
+
+
+class KCrossAttnDownBlock2D(KCrossAttnDownBlock2D):
+ deprecation_message = "Importing `KCrossAttnDownBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import KCrossAttnDownBlock2D`, instead."
+ deprecate("KCrossAttnDownBlock2D", "0.29", deprecation_message)
+
+
+class AttnUpBlock2D(AttnUpBlock2D):
+ deprecation_message = "Importing `AttnUpBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import AttnUpBlock2D`, instead."
+ deprecate("AttnUpBlock2D", "0.29", deprecation_message)
+
+
+class CrossAttnUpBlock2D(CrossAttnUpBlock2D):
+ deprecation_message = "Importing `CrossAttnUpBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import CrossAttnUpBlock2D`, instead."
+ deprecate("CrossAttnUpBlock2D", "0.29", deprecation_message)
+
+
+class UpBlock2D(UpBlock2D):
+ deprecation_message = "Importing `UpBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import UpBlock2D`, instead."
+ deprecate("UpBlock2D", "0.29", deprecation_message)
+
+
+class UpDecoderBlock2D(UpDecoderBlock2D):
+ deprecation_message = "Importing `UpDecoderBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import UpDecoderBlock2D`, instead."
+ deprecate("UpDecoderBlock2D", "0.29", deprecation_message)
+
+
+class AttnUpDecoderBlock2D(AttnUpDecoderBlock2D):
+ deprecation_message = "Importing `AttnUpDecoderBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import AttnUpDecoderBlock2D`, instead."
+ deprecate("AttnUpDecoderBlock2D", "0.29", deprecation_message)
+
+
+class AttnSkipUpBlock2D(AttnSkipUpBlock2D):
+ deprecation_message = "Importing `AttnSkipUpBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import AttnSkipUpBlock2D`, instead."
+ deprecate("AttnSkipUpBlock2D", "0.29", deprecation_message)
+
+
+class SkipUpBlock2D(SkipUpBlock2D):
+ deprecation_message = "Importing `SkipUpBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import SkipUpBlock2D`, instead."
+ deprecate("SkipUpBlock2D", "0.29", deprecation_message)
+
+
+class ResnetUpsampleBlock2D(ResnetUpsampleBlock2D):
+ deprecation_message = "Importing `ResnetUpsampleBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import ResnetUpsampleBlock2D`, instead."
+ deprecate("ResnetUpsampleBlock2D", "0.29", deprecation_message)
+
+
+class SimpleCrossAttnUpBlock2D(SimpleCrossAttnUpBlock2D):
+ deprecation_message = "Importing `SimpleCrossAttnUpBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import SimpleCrossAttnUpBlock2D`, instead."
+ deprecate("SimpleCrossAttnUpBlock2D", "0.29", deprecation_message)
+
+
+class KUpBlock2D(KUpBlock2D):
+ deprecation_message = "Importing `KUpBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import KUpBlock2D`, instead."
+ deprecate("KUpBlock2D", "0.29", deprecation_message)
+
+
+class KCrossAttnUpBlock2D(KCrossAttnUpBlock2D):
+ deprecation_message = "Importing `KCrossAttnUpBlock2D` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import KCrossAttnUpBlock2D`, instead."
+ deprecate("KCrossAttnUpBlock2D", "0.29", deprecation_message)
+
+
+# can potentially later be renamed to `No-feed-forward` attention
+class KAttentionBlock(KAttentionBlock):
+ deprecation_message = "Importing `KAttentionBlock` from `diffusers.models.unet_2d_blocks` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_blocks import KAttentionBlock`, instead."
+ deprecate("KAttentionBlock", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/unet_2d_condition.py b/model/diffusers_c/models/unet_2d_condition.py
new file mode 100644
index 0000000000000000000000000000000000000000..85a3e7b0919753f06be4a7406df47004a57f441c
--- /dev/null
+++ b/model/diffusers_c/models/unet_2d_condition.py
@@ -0,0 +1,25 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from ..utils import deprecate
+from .unets.unet_2d_condition import UNet2DConditionModel, UNet2DConditionOutput
+
+
+class UNet2DConditionOutput(UNet2DConditionOutput):
+ deprecation_message = "Importing `UNet2DConditionOutput` from `diffusers.models.unet_2d_condition` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_condition import UNet2DConditionOutput`, instead."
+ deprecate("UNet2DConditionOutput", "0.29", deprecation_message)
+
+
+class UNet2DConditionModel(UNet2DConditionModel):
+ deprecation_message = "Importing `UNet2DConditionModel` from `diffusers.models.unet_2d_condition` is deprecated and this will be removed in a future version. Please use `from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel`, instead."
+ deprecate("UNet2DConditionModel", "0.29", deprecation_message)
diff --git a/model/diffusers_c/models/unets/__init__.py b/model/diffusers_c/models/unets/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..9ef04fb62606ed30699be8ac6cb38b011a692070
--- /dev/null
+++ b/model/diffusers_c/models/unets/__init__.py
@@ -0,0 +1,18 @@
+from ...utils import is_flax_available, is_torch_available
+
+
+if is_torch_available():
+ from .unet_1d import UNet1DModel
+ from .unet_2d import UNet2DModel
+ from .unet_2d_condition import UNet2DConditionModel
+ from .unet_3d_condition import UNet3DConditionModel
+ from .unet_i2vgen_xl import I2VGenXLUNet
+ from .unet_kandinsky3 import Kandinsky3UNet
+ from .unet_motion_model import MotionAdapter, UNetMotionModel
+ from .unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
+ from .unet_stable_cascade import StableCascadeUNet
+ from .uvit_2d import UVit2DModel
+
+
+if is_flax_available():
+ from .unet_2d_condition_flax import FlaxUNet2DConditionModel
diff --git a/model/diffusers_c/models/unets/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2b74a1b08880fd7cd19ca05e252ef50e33c2ba1e
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3e8c8393950948f15d9a7dd08aa18f67f6916434
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_1d.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_1d.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d9089d82de24663b605dd37728edcab98038de08
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_1d.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_1d.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_1d.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c60d787c6f034cd6926acbc8e4ba72f235d3f157
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_1d.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_1d_blocks.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_1d_blocks.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..26cf7b1b3a2bd1554c328a99eabc25d133fafdc3
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_1d_blocks.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_1d_blocks.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_1d_blocks.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ae78b30d88fd3358af9430f9dca130581395550f
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_1d_blocks.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_2d.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_2d.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3293c7993728295695a4ca30f5c1e2b1e3846458
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_2d.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_2d.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_2d.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..426d6113b60300dbf7089eb2b000192686b75210
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_2d.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f1a3d891c6085aac0010ae19523496e7e24edace
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c43e755b0dbbb3664af06e6c427c1319f62ea91f
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks_flax.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks_flax.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8f81f3e0e2a81b05fc66e98a4afa6a7750b694ee
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_2d_blocks_flax.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_2d_condition.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_2d_condition.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b068e0a876983ba69fc466b9f67f4da0a095af4f
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_2d_condition.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_2d_condition.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_2d_condition.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..adc3e3de6d5246e6b0f9746be8d0927766d6a726
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_2d_condition.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_2d_condition_flax.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_2d_condition_flax.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d7dc5bad16e80ddc01c92513da5e303ff29eea82
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_2d_condition_flax.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_3d_blocks.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_3d_blocks.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..724726a2d635e50ef6a57abdb2da314456bf1b8b
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_3d_blocks.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_3d_blocks.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_3d_blocks.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e05313a0b84340485ce2380c3b36dfbae58b56e9
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_3d_blocks.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_3d_condition.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_3d_condition.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..72ed5dcbc25da15d6ddf8a7c01b861ad10bf568d
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_3d_condition.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_3d_condition.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_3d_condition.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..25bdd6d2cc5764f57041ee37d4321effa6688356
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_3d_condition.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_i2vgen_xl.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_i2vgen_xl.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..bca5e5ae2b09aac10e06311d2c56ce8951b1caa5
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_i2vgen_xl.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_i2vgen_xl.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_i2vgen_xl.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..583dd150da61d3f2057b7091e4d9e003be78d888
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_i2vgen_xl.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_kandinsky3.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_kandinsky3.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..23838bfe2423a99238d94e10daffec9be5b88597
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_kandinsky3.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_kandinsky3.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_kandinsky3.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..223e11d4fa63fb4a05e2d8c35183dd0ebc399aba
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_kandinsky3.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_motion_model.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_motion_model.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f9d5b88275ec621f3531be1ce6cd72642046e9d0
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_motion_model.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_motion_model.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_motion_model.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8fa01dc4f5e1dfb9d9cb9bd5b61dcc3ef42dd7ca
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_motion_model.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_spatio_temporal_condition.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_spatio_temporal_condition.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b47b7a177abb37dfc19c9c4ee92f35a79220909a
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_spatio_temporal_condition.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_spatio_temporal_condition.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_spatio_temporal_condition.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2ee0d3ede852cd8171c606d10252f0d5aba8c6fa
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_spatio_temporal_condition.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_stable_cascade.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/unet_stable_cascade.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..793627e7946ba9917b0787510b2db31293293031
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_stable_cascade.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/unet_stable_cascade.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/unet_stable_cascade.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..485c30f545f3d083d3eb54a04c89db4fe3217c8f
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/unet_stable_cascade.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/uvit_2d.cpython-310.pyc b/model/diffusers_c/models/unets/__pycache__/uvit_2d.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ae24d486bc93d06c2335779debe3a42f85288f72
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/uvit_2d.cpython-310.pyc differ
diff --git a/model/diffusers_c/models/unets/__pycache__/uvit_2d.cpython-39.pyc b/model/diffusers_c/models/unets/__pycache__/uvit_2d.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0f05effb9492fcca36fd8c14d07b952a6880f8fc
Binary files /dev/null and b/model/diffusers_c/models/unets/__pycache__/uvit_2d.cpython-39.pyc differ
diff --git a/model/diffusers_c/models/unets/unet_1d.py b/model/diffusers_c/models/unets/unet_1d.py
new file mode 100644
index 0000000000000000000000000000000000000000..59d70f67c959e1c87d13c4cfef30353dafaf5c49
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_1d.py
@@ -0,0 +1,255 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ..embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from .unet_1d_blocks import get_down_block, get_mid_block, get_out_block, get_up_block
+
+
+@dataclass
+class UNet1DOutput(BaseOutput):
+ """
+ The output of [`UNet1DModel`].
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, sample_size)`):
+ The hidden states output from the last layer of the model.
+ """
+
+ sample: torch.FloatTensor
+
+
+class UNet1DModel(ModelMixin, ConfigMixin):
+ r"""
+ A 1D UNet model that takes a noisy sample and a timestep and returns a sample shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int`, *optional*): Default length of sample. Should be adaptable at runtime.
+ in_channels (`int`, *optional*, defaults to 2): Number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 2): Number of channels in the output.
+ extra_in_channels (`int`, *optional*, defaults to 0):
+ Number of additional channels to be added to the input of the first down block. Useful for cases where the
+ input data has more channels than what the model was initially designed for.
+ time_embedding_type (`str`, *optional*, defaults to `"fourier"`): Type of time embedding to use.
+ freq_shift (`float`, *optional*, defaults to 0.0): Frequency shift for Fourier time embedding.
+ flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+ Whether to flip sin to cos for Fourier time embedding.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("DownBlock1DNoSkip", "DownBlock1D", "AttnDownBlock1D")`):
+ Tuple of downsample block types.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("AttnUpBlock1D", "UpBlock1D", "UpBlock1DNoSkip")`):
+ Tuple of upsample block types.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(32, 32, 64)`):
+ Tuple of block output channels.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock1D"`): Block type for middle of UNet.
+ out_block_type (`str`, *optional*, defaults to `None`): Optional output processing block of UNet.
+ act_fn (`str`, *optional*, defaults to `None`): Optional activation function in UNet blocks.
+ norm_num_groups (`int`, *optional*, defaults to 8): The number of groups for normalization.
+ layers_per_block (`int`, *optional*, defaults to 1): The number of layers per block.
+ downsample_each_block (`int`, *optional*, defaults to `False`):
+ Experimental feature for using a UNet without upsampling.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: int = 65536,
+ sample_rate: Optional[int] = None,
+ in_channels: int = 2,
+ out_channels: int = 2,
+ extra_in_channels: int = 0,
+ time_embedding_type: str = "fourier",
+ flip_sin_to_cos: bool = True,
+ use_timestep_embedding: bool = False,
+ freq_shift: float = 0.0,
+ down_block_types: Tuple[str] = ("DownBlock1DNoSkip", "DownBlock1D", "AttnDownBlock1D"),
+ up_block_types: Tuple[str] = ("AttnUpBlock1D", "UpBlock1D", "UpBlock1DNoSkip"),
+ mid_block_type: Tuple[str] = "UNetMidBlock1D",
+ out_block_type: str = None,
+ block_out_channels: Tuple[int] = (32, 32, 64),
+ act_fn: str = None,
+ norm_num_groups: int = 8,
+ layers_per_block: int = 1,
+ downsample_each_block: bool = False,
+ ):
+ super().__init__()
+ self.sample_size = sample_size
+
+ # time
+ if time_embedding_type == "fourier":
+ self.time_proj = GaussianFourierProjection(
+ embedding_size=8, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+ )
+ timestep_input_dim = 2 * block_out_channels[0]
+ elif time_embedding_type == "positional":
+ self.time_proj = Timesteps(
+ block_out_channels[0], flip_sin_to_cos=flip_sin_to_cos, downscale_freq_shift=freq_shift
+ )
+ timestep_input_dim = block_out_channels[0]
+
+ if use_timestep_embedding:
+ time_embed_dim = block_out_channels[0] * 4
+ self.time_mlp = TimestepEmbedding(
+ in_channels=timestep_input_dim,
+ time_embed_dim=time_embed_dim,
+ act_fn=act_fn,
+ out_dim=block_out_channels[0],
+ )
+
+ self.down_blocks = nn.ModuleList([])
+ self.mid_block = None
+ self.up_blocks = nn.ModuleList([])
+ self.out_block = None
+
+ # down
+ output_channel = in_channels
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+
+ if i == 0:
+ input_channel += extra_in_channels
+
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=block_out_channels[0],
+ add_downsample=not is_final_block or downsample_each_block,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ self.mid_block = get_mid_block(
+ mid_block_type,
+ in_channels=block_out_channels[-1],
+ mid_channels=block_out_channels[-1],
+ out_channels=block_out_channels[-1],
+ embed_dim=block_out_channels[0],
+ num_layers=layers_per_block,
+ add_downsample=downsample_each_block,
+ )
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ output_channel = reversed_block_out_channels[0]
+ if out_block_type is None:
+ final_upsample_channels = out_channels
+ else:
+ final_upsample_channels = block_out_channels[0]
+
+ for i, up_block_type in enumerate(up_block_types):
+ prev_output_channel = output_channel
+ output_channel = (
+ reversed_block_out_channels[i + 1] if i < len(up_block_types) - 1 else final_upsample_channels
+ )
+
+ is_final_block = i == len(block_out_channels) - 1
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=layers_per_block,
+ in_channels=prev_output_channel,
+ out_channels=output_channel,
+ temb_channels=block_out_channels[0],
+ add_upsample=not is_final_block,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ num_groups_out = norm_num_groups if norm_num_groups is not None else min(block_out_channels[0] // 4, 32)
+ self.out_block = get_out_block(
+ out_block_type=out_block_type,
+ num_groups_out=num_groups_out,
+ embed_dim=block_out_channels[0],
+ out_channels=out_channels,
+ act_fn=act_fn,
+ fc_dim=block_out_channels[-1] // 4,
+ )
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ return_dict: bool = True,
+ ) -> Union[UNet1DOutput, Tuple]:
+ r"""
+ The [`UNet1DModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch_size, num_channels, sample_size)`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unet_1d.UNet1DOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.unet_1d.UNet1DOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unet_1d.UNet1DOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is the sample tensor.
+ """
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
+ elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ timestep_embed = self.time_proj(timesteps)
+ if self.config.use_timestep_embedding:
+ timestep_embed = self.time_mlp(timestep_embed)
+ else:
+ timestep_embed = timestep_embed[..., None]
+ timestep_embed = timestep_embed.repeat([1, 1, sample.shape[2]]).to(sample.dtype)
+ timestep_embed = timestep_embed.broadcast_to((sample.shape[:1] + timestep_embed.shape[1:]))
+
+ # 2. down
+ down_block_res_samples = ()
+ for downsample_block in self.down_blocks:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=timestep_embed)
+ down_block_res_samples += res_samples
+
+ # 3. mid
+ if self.mid_block:
+ sample = self.mid_block(sample, timestep_embed)
+
+ # 4. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ res_samples = down_block_res_samples[-1:]
+ down_block_res_samples = down_block_res_samples[:-1]
+ sample = upsample_block(sample, res_hidden_states_tuple=res_samples, temb=timestep_embed)
+
+ # 5. post-process
+ if self.out_block:
+ sample = self.out_block(sample, timestep_embed)
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet1DOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/unet_1d_blocks.py b/model/diffusers_c/models/unets/unet_1d_blocks.py
new file mode 100644
index 0000000000000000000000000000000000000000..e3163cd1d5b373f328097dfc24b083bd5d542318
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_1d_blocks.py
@@ -0,0 +1,702 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..activations import get_activation
+from ..resnet import Downsample1D, ResidualTemporalBlock1D, Upsample1D, rearrange_dims
+
+
+class DownResnetBlock1D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: Optional[int] = None,
+ num_layers: int = 1,
+ conv_shortcut: bool = False,
+ temb_channels: int = 32,
+ groups: int = 32,
+ groups_out: Optional[int] = None,
+ non_linearity: Optional[str] = None,
+ time_embedding_norm: str = "default",
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ ):
+ super().__init__()
+ self.in_channels = in_channels
+ out_channels = in_channels if out_channels is None else out_channels
+ self.out_channels = out_channels
+ self.use_conv_shortcut = conv_shortcut
+ self.time_embedding_norm = time_embedding_norm
+ self.add_downsample = add_downsample
+ self.output_scale_factor = output_scale_factor
+
+ if groups_out is None:
+ groups_out = groups
+
+ # there will always be at least one resnet
+ resnets = [ResidualTemporalBlock1D(in_channels, out_channels, embed_dim=temb_channels)]
+
+ for _ in range(num_layers):
+ resnets.append(ResidualTemporalBlock1D(out_channels, out_channels, embed_dim=temb_channels))
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if non_linearity is None:
+ self.nonlinearity = None
+ else:
+ self.nonlinearity = get_activation(non_linearity)
+
+ self.downsample = None
+ if add_downsample:
+ self.downsample = Downsample1D(out_channels, use_conv=True, padding=1)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ output_states = ()
+
+ hidden_states = self.resnets[0](hidden_states, temb)
+ for resnet in self.resnets[1:]:
+ hidden_states = resnet(hidden_states, temb)
+
+ output_states += (hidden_states,)
+
+ if self.nonlinearity is not None:
+ hidden_states = self.nonlinearity(hidden_states)
+
+ if self.downsample is not None:
+ hidden_states = self.downsample(hidden_states)
+
+ return hidden_states, output_states
+
+
+class UpResnetBlock1D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: Optional[int] = None,
+ num_layers: int = 1,
+ temb_channels: int = 32,
+ groups: int = 32,
+ groups_out: Optional[int] = None,
+ non_linearity: Optional[str] = None,
+ time_embedding_norm: str = "default",
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ self.in_channels = in_channels
+ out_channels = in_channels if out_channels is None else out_channels
+ self.out_channels = out_channels
+ self.time_embedding_norm = time_embedding_norm
+ self.add_upsample = add_upsample
+ self.output_scale_factor = output_scale_factor
+
+ if groups_out is None:
+ groups_out = groups
+
+ # there will always be at least one resnet
+ resnets = [ResidualTemporalBlock1D(2 * in_channels, out_channels, embed_dim=temb_channels)]
+
+ for _ in range(num_layers):
+ resnets.append(ResidualTemporalBlock1D(out_channels, out_channels, embed_dim=temb_channels))
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if non_linearity is None:
+ self.nonlinearity = None
+ else:
+ self.nonlinearity = get_activation(non_linearity)
+
+ self.upsample = None
+ if add_upsample:
+ self.upsample = Upsample1D(out_channels, use_conv_transpose=True)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Optional[Tuple[torch.FloatTensor, ...]] = None,
+ temb: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ if res_hidden_states_tuple is not None:
+ res_hidden_states = res_hidden_states_tuple[-1]
+ hidden_states = torch.cat((hidden_states, res_hidden_states), dim=1)
+
+ hidden_states = self.resnets[0](hidden_states, temb)
+ for resnet in self.resnets[1:]:
+ hidden_states = resnet(hidden_states, temb)
+
+ if self.nonlinearity is not None:
+ hidden_states = self.nonlinearity(hidden_states)
+
+ if self.upsample is not None:
+ hidden_states = self.upsample(hidden_states)
+
+ return hidden_states
+
+
+class ValueFunctionMidBlock1D(nn.Module):
+ def __init__(self, in_channels: int, out_channels: int, embed_dim: int):
+ super().__init__()
+ self.in_channels = in_channels
+ self.out_channels = out_channels
+ self.embed_dim = embed_dim
+
+ self.res1 = ResidualTemporalBlock1D(in_channels, in_channels // 2, embed_dim=embed_dim)
+ self.down1 = Downsample1D(out_channels // 2, use_conv=True)
+ self.res2 = ResidualTemporalBlock1D(in_channels // 2, in_channels // 4, embed_dim=embed_dim)
+ self.down2 = Downsample1D(out_channels // 4, use_conv=True)
+
+ def forward(self, x: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ x = self.res1(x, temb)
+ x = self.down1(x)
+ x = self.res2(x, temb)
+ x = self.down2(x)
+ return x
+
+
+class MidResTemporalBlock1D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ embed_dim: int,
+ num_layers: int = 1,
+ add_downsample: bool = False,
+ add_upsample: bool = False,
+ non_linearity: Optional[str] = None,
+ ):
+ super().__init__()
+ self.in_channels = in_channels
+ self.out_channels = out_channels
+ self.add_downsample = add_downsample
+
+ # there will always be at least one resnet
+ resnets = [ResidualTemporalBlock1D(in_channels, out_channels, embed_dim=embed_dim)]
+
+ for _ in range(num_layers):
+ resnets.append(ResidualTemporalBlock1D(out_channels, out_channels, embed_dim=embed_dim))
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if non_linearity is None:
+ self.nonlinearity = None
+ else:
+ self.nonlinearity = get_activation(non_linearity)
+
+ self.upsample = None
+ if add_upsample:
+ self.upsample = Downsample1D(out_channels, use_conv=True)
+
+ self.downsample = None
+ if add_downsample:
+ self.downsample = Downsample1D(out_channels, use_conv=True)
+
+ if self.upsample and self.downsample:
+ raise ValueError("Block cannot downsample and upsample")
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: torch.FloatTensor) -> torch.FloatTensor:
+ hidden_states = self.resnets[0](hidden_states, temb)
+ for resnet in self.resnets[1:]:
+ hidden_states = resnet(hidden_states, temb)
+
+ if self.upsample:
+ hidden_states = self.upsample(hidden_states)
+ if self.downsample:
+ self.downsample = self.downsample(hidden_states)
+
+ return hidden_states
+
+
+class OutConv1DBlock(nn.Module):
+ def __init__(self, num_groups_out: int, out_channels: int, embed_dim: int, act_fn: str):
+ super().__init__()
+ self.final_conv1d_1 = nn.Conv1d(embed_dim, embed_dim, 5, padding=2)
+ self.final_conv1d_gn = nn.GroupNorm(num_groups_out, embed_dim)
+ self.final_conv1d_act = get_activation(act_fn)
+ self.final_conv1d_2 = nn.Conv1d(embed_dim, out_channels, 1)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = self.final_conv1d_1(hidden_states)
+ hidden_states = rearrange_dims(hidden_states)
+ hidden_states = self.final_conv1d_gn(hidden_states)
+ hidden_states = rearrange_dims(hidden_states)
+ hidden_states = self.final_conv1d_act(hidden_states)
+ hidden_states = self.final_conv1d_2(hidden_states)
+ return hidden_states
+
+
+class OutValueFunctionBlock(nn.Module):
+ def __init__(self, fc_dim: int, embed_dim: int, act_fn: str = "mish"):
+ super().__init__()
+ self.final_block = nn.ModuleList(
+ [
+ nn.Linear(fc_dim + embed_dim, fc_dim // 2),
+ get_activation(act_fn),
+ nn.Linear(fc_dim // 2, 1),
+ ]
+ )
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: torch.FloatTensor) -> torch.FloatTensor:
+ hidden_states = hidden_states.view(hidden_states.shape[0], -1)
+ hidden_states = torch.cat((hidden_states, temb), dim=-1)
+ for layer in self.final_block:
+ hidden_states = layer(hidden_states)
+
+ return hidden_states
+
+
+_kernels = {
+ "linear": [1 / 8, 3 / 8, 3 / 8, 1 / 8],
+ "cubic": [-0.01171875, -0.03515625, 0.11328125, 0.43359375, 0.43359375, 0.11328125, -0.03515625, -0.01171875],
+ "lanczos3": [
+ 0.003689131001010537,
+ 0.015056144446134567,
+ -0.03399861603975296,
+ -0.066637322306633,
+ 0.13550527393817902,
+ 0.44638532400131226,
+ 0.44638532400131226,
+ 0.13550527393817902,
+ -0.066637322306633,
+ -0.03399861603975296,
+ 0.015056144446134567,
+ 0.003689131001010537,
+ ],
+}
+
+
+class Downsample1d(nn.Module):
+ def __init__(self, kernel: str = "linear", pad_mode: str = "reflect"):
+ super().__init__()
+ self.pad_mode = pad_mode
+ kernel_1d = torch.tensor(_kernels[kernel])
+ self.pad = kernel_1d.shape[0] // 2 - 1
+ self.register_buffer("kernel", kernel_1d)
+
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+ hidden_states = F.pad(hidden_states, (self.pad,) * 2, self.pad_mode)
+ weight = hidden_states.new_zeros([hidden_states.shape[1], hidden_states.shape[1], self.kernel.shape[0]])
+ indices = torch.arange(hidden_states.shape[1], device=hidden_states.device)
+ kernel = self.kernel.to(weight)[None, :].expand(hidden_states.shape[1], -1)
+ weight[indices, indices] = kernel
+ return F.conv1d(hidden_states, weight, stride=2)
+
+
+class Upsample1d(nn.Module):
+ def __init__(self, kernel: str = "linear", pad_mode: str = "reflect"):
+ super().__init__()
+ self.pad_mode = pad_mode
+ kernel_1d = torch.tensor(_kernels[kernel]) * 2
+ self.pad = kernel_1d.shape[0] // 2 - 1
+ self.register_buffer("kernel", kernel_1d)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = F.pad(hidden_states, ((self.pad + 1) // 2,) * 2, self.pad_mode)
+ weight = hidden_states.new_zeros([hidden_states.shape[1], hidden_states.shape[1], self.kernel.shape[0]])
+ indices = torch.arange(hidden_states.shape[1], device=hidden_states.device)
+ kernel = self.kernel.to(weight)[None, :].expand(hidden_states.shape[1], -1)
+ weight[indices, indices] = kernel
+ return F.conv_transpose1d(hidden_states, weight, stride=2, padding=self.pad * 2 + 1)
+
+
+class SelfAttention1d(nn.Module):
+ def __init__(self, in_channels: int, n_head: int = 1, dropout_rate: float = 0.0):
+ super().__init__()
+ self.channels = in_channels
+ self.group_norm = nn.GroupNorm(1, num_channels=in_channels)
+ self.num_heads = n_head
+
+ self.query = nn.Linear(self.channels, self.channels)
+ self.key = nn.Linear(self.channels, self.channels)
+ self.value = nn.Linear(self.channels, self.channels)
+
+ self.proj_attn = nn.Linear(self.channels, self.channels, bias=True)
+
+ self.dropout = nn.Dropout(dropout_rate, inplace=True)
+
+ def transpose_for_scores(self, projection: torch.Tensor) -> torch.Tensor:
+ new_projection_shape = projection.size()[:-1] + (self.num_heads, -1)
+ # move heads to 2nd position (B, T, H * D) -> (B, T, H, D) -> (B, H, T, D)
+ new_projection = projection.view(new_projection_shape).permute(0, 2, 1, 3)
+ return new_projection
+
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+ residual = hidden_states
+ batch, channel_dim, seq = hidden_states.shape
+
+ hidden_states = self.group_norm(hidden_states)
+ hidden_states = hidden_states.transpose(1, 2)
+
+ query_proj = self.query(hidden_states)
+ key_proj = self.key(hidden_states)
+ value_proj = self.value(hidden_states)
+
+ query_states = self.transpose_for_scores(query_proj)
+ key_states = self.transpose_for_scores(key_proj)
+ value_states = self.transpose_for_scores(value_proj)
+
+ scale = 1 / math.sqrt(math.sqrt(key_states.shape[-1]))
+
+ attention_scores = torch.matmul(query_states * scale, key_states.transpose(-1, -2) * scale)
+ attention_probs = torch.softmax(attention_scores, dim=-1)
+
+ # compute attention output
+ hidden_states = torch.matmul(attention_probs, value_states)
+
+ hidden_states = hidden_states.permute(0, 2, 1, 3).contiguous()
+ new_hidden_states_shape = hidden_states.size()[:-2] + (self.channels,)
+ hidden_states = hidden_states.view(new_hidden_states_shape)
+
+ # compute next hidden_states
+ hidden_states = self.proj_attn(hidden_states)
+ hidden_states = hidden_states.transpose(1, 2)
+ hidden_states = self.dropout(hidden_states)
+
+ output = hidden_states + residual
+
+ return output
+
+
+class ResConvBlock(nn.Module):
+ def __init__(self, in_channels: int, mid_channels: int, out_channels: int, is_last: bool = False):
+ super().__init__()
+ self.is_last = is_last
+ self.has_conv_skip = in_channels != out_channels
+
+ if self.has_conv_skip:
+ self.conv_skip = nn.Conv1d(in_channels, out_channels, 1, bias=False)
+
+ self.conv_1 = nn.Conv1d(in_channels, mid_channels, 5, padding=2)
+ self.group_norm_1 = nn.GroupNorm(1, mid_channels)
+ self.gelu_1 = nn.GELU()
+ self.conv_2 = nn.Conv1d(mid_channels, out_channels, 5, padding=2)
+
+ if not self.is_last:
+ self.group_norm_2 = nn.GroupNorm(1, out_channels)
+ self.gelu_2 = nn.GELU()
+
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+ residual = self.conv_skip(hidden_states) if self.has_conv_skip else hidden_states
+
+ hidden_states = self.conv_1(hidden_states)
+ hidden_states = self.group_norm_1(hidden_states)
+ hidden_states = self.gelu_1(hidden_states)
+ hidden_states = self.conv_2(hidden_states)
+
+ if not self.is_last:
+ hidden_states = self.group_norm_2(hidden_states)
+ hidden_states = self.gelu_2(hidden_states)
+
+ output = hidden_states + residual
+ return output
+
+
+class UNetMidBlock1D(nn.Module):
+ def __init__(self, mid_channels: int, in_channels: int, out_channels: Optional[int] = None):
+ super().__init__()
+
+ out_channels = in_channels if out_channels is None else out_channels
+
+ # there is always at least one resnet
+ self.down = Downsample1d("cubic")
+ resnets = [
+ ResConvBlock(in_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, out_channels),
+ ]
+ attentions = [
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(out_channels, out_channels // 32),
+ ]
+ self.up = Upsample1d(kernel="cubic")
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = self.down(hidden_states)
+ for attn, resnet in zip(self.attentions, self.resnets):
+ hidden_states = resnet(hidden_states)
+ hidden_states = attn(hidden_states)
+
+ hidden_states = self.up(hidden_states)
+
+ return hidden_states
+
+
+class AttnDownBlock1D(nn.Module):
+ def __init__(self, out_channels: int, in_channels: int, mid_channels: Optional[int] = None):
+ super().__init__()
+ mid_channels = out_channels if mid_channels is None else mid_channels
+
+ self.down = Downsample1d("cubic")
+ resnets = [
+ ResConvBlock(in_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, out_channels),
+ ]
+ attentions = [
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(out_channels, out_channels // 32),
+ ]
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = self.down(hidden_states)
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states)
+ hidden_states = attn(hidden_states)
+
+ return hidden_states, (hidden_states,)
+
+
+class DownBlock1D(nn.Module):
+ def __init__(self, out_channels: int, in_channels: int, mid_channels: Optional[int] = None):
+ super().__init__()
+ mid_channels = out_channels if mid_channels is None else mid_channels
+
+ self.down = Downsample1d("cubic")
+ resnets = [
+ ResConvBlock(in_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, out_channels),
+ ]
+
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = self.down(hidden_states)
+
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states)
+
+ return hidden_states, (hidden_states,)
+
+
+class DownBlock1DNoSkip(nn.Module):
+ def __init__(self, out_channels: int, in_channels: int, mid_channels: Optional[int] = None):
+ super().__init__()
+ mid_channels = out_channels if mid_channels is None else mid_channels
+
+ resnets = [
+ ResConvBlock(in_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, out_channels),
+ ]
+
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = torch.cat([hidden_states, temb], dim=1)
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states)
+
+ return hidden_states, (hidden_states,)
+
+
+class AttnUpBlock1D(nn.Module):
+ def __init__(self, in_channels: int, out_channels: int, mid_channels: Optional[int] = None):
+ super().__init__()
+ mid_channels = out_channels if mid_channels is None else mid_channels
+
+ resnets = [
+ ResConvBlock(2 * in_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, out_channels),
+ ]
+ attentions = [
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(mid_channels, mid_channels // 32),
+ SelfAttention1d(out_channels, out_channels // 32),
+ ]
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+ self.up = Upsample1d(kernel="cubic")
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ res_hidden_states = res_hidden_states_tuple[-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states)
+ hidden_states = attn(hidden_states)
+
+ hidden_states = self.up(hidden_states)
+
+ return hidden_states
+
+
+class UpBlock1D(nn.Module):
+ def __init__(self, in_channels: int, out_channels: int, mid_channels: Optional[int] = None):
+ super().__init__()
+ mid_channels = in_channels if mid_channels is None else mid_channels
+
+ resnets = [
+ ResConvBlock(2 * in_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, out_channels),
+ ]
+
+ self.resnets = nn.ModuleList(resnets)
+ self.up = Upsample1d(kernel="cubic")
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ res_hidden_states = res_hidden_states_tuple[-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states)
+
+ hidden_states = self.up(hidden_states)
+
+ return hidden_states
+
+
+class UpBlock1DNoSkip(nn.Module):
+ def __init__(self, in_channels: int, out_channels: int, mid_channels: Optional[int] = None):
+ super().__init__()
+ mid_channels = in_channels if mid_channels is None else mid_channels
+
+ resnets = [
+ ResConvBlock(2 * in_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, mid_channels),
+ ResConvBlock(mid_channels, mid_channels, out_channels, is_last=True),
+ ]
+
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ res_hidden_states = res_hidden_states_tuple[-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states)
+
+ return hidden_states
+
+
+DownBlockType = Union[DownResnetBlock1D, DownBlock1D, AttnDownBlock1D, DownBlock1DNoSkip]
+MidBlockType = Union[MidResTemporalBlock1D, ValueFunctionMidBlock1D, UNetMidBlock1D]
+OutBlockType = Union[OutConv1DBlock, OutValueFunctionBlock]
+UpBlockType = Union[UpResnetBlock1D, UpBlock1D, AttnUpBlock1D, UpBlock1DNoSkip]
+
+
+def get_down_block(
+ down_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ add_downsample: bool,
+) -> DownBlockType:
+ if down_block_type == "DownResnetBlock1D":
+ return DownResnetBlock1D(
+ in_channels=in_channels,
+ num_layers=num_layers,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ )
+ elif down_block_type == "DownBlock1D":
+ return DownBlock1D(out_channels=out_channels, in_channels=in_channels)
+ elif down_block_type == "AttnDownBlock1D":
+ return AttnDownBlock1D(out_channels=out_channels, in_channels=in_channels)
+ elif down_block_type == "DownBlock1DNoSkip":
+ return DownBlock1DNoSkip(out_channels=out_channels, in_channels=in_channels)
+ raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+ up_block_type: str, num_layers: int, in_channels: int, out_channels: int, temb_channels: int, add_upsample: bool
+) -> UpBlockType:
+ if up_block_type == "UpResnetBlock1D":
+ return UpResnetBlock1D(
+ in_channels=in_channels,
+ num_layers=num_layers,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ )
+ elif up_block_type == "UpBlock1D":
+ return UpBlock1D(in_channels=in_channels, out_channels=out_channels)
+ elif up_block_type == "AttnUpBlock1D":
+ return AttnUpBlock1D(in_channels=in_channels, out_channels=out_channels)
+ elif up_block_type == "UpBlock1DNoSkip":
+ return UpBlock1DNoSkip(in_channels=in_channels, out_channels=out_channels)
+ raise ValueError(f"{up_block_type} does not exist.")
+
+
+def get_mid_block(
+ mid_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ mid_channels: int,
+ out_channels: int,
+ embed_dim: int,
+ add_downsample: bool,
+) -> MidBlockType:
+ if mid_block_type == "MidResTemporalBlock1D":
+ return MidResTemporalBlock1D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ embed_dim=embed_dim,
+ add_downsample=add_downsample,
+ )
+ elif mid_block_type == "ValueFunctionMidBlock1D":
+ return ValueFunctionMidBlock1D(in_channels=in_channels, out_channels=out_channels, embed_dim=embed_dim)
+ elif mid_block_type == "UNetMidBlock1D":
+ return UNetMidBlock1D(in_channels=in_channels, mid_channels=mid_channels, out_channels=out_channels)
+ raise ValueError(f"{mid_block_type} does not exist.")
+
+
+def get_out_block(
+ *, out_block_type: str, num_groups_out: int, embed_dim: int, out_channels: int, act_fn: str, fc_dim: int
+) -> Optional[OutBlockType]:
+ if out_block_type == "OutConv1DBlock":
+ return OutConv1DBlock(num_groups_out, out_channels, embed_dim, act_fn)
+ elif out_block_type == "ValueFunction":
+ return OutValueFunctionBlock(fc_dim, embed_dim, act_fn)
+ return None
diff --git a/model/diffusers_c/models/unets/unet_2d.py b/model/diffusers_c/models/unets/unet_2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..5efb638222059063a0b82c7f6974d0c9ca5b2c18
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_2d.py
@@ -0,0 +1,346 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ..embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from .unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block
+
+
+@dataclass
+class UNet2DOutput(BaseOutput):
+ """
+ The output of [`UNet2DModel`].
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+ The hidden states output from the last layer of the model.
+ """
+
+ sample: torch.FloatTensor
+
+
+class UNet2DModel(ModelMixin, ConfigMixin):
+ r"""
+ A 2D UNet model that takes a noisy sample and a timestep and returns a sample shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+ Height and width of input/output sample. Dimensions must be a multiple of `2 ** (len(block_out_channels) -
+ 1)`.
+ in_channels (`int`, *optional*, defaults to 3): Number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 3): Number of channels in the output.
+ center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+ time_embedding_type (`str`, *optional*, defaults to `"positional"`): Type of time embedding to use.
+ freq_shift (`int`, *optional*, defaults to 0): Frequency shift for Fourier time embedding.
+ flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+ Whether to flip sin to cos for Fourier time embedding.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D")`):
+ Tuple of downsample block types.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2D"`):
+ Block type for middle of UNet, it can be either `UNetMidBlock2D` or `UnCLIPUNetMidBlock2D`.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D")`):
+ Tuple of upsample block types.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(224, 448, 672, 896)`):
+ Tuple of block output channels.
+ layers_per_block (`int`, *optional*, defaults to `2`): The number of layers per block.
+ mid_block_scale_factor (`float`, *optional*, defaults to `1`): The scale factor for the mid block.
+ downsample_padding (`int`, *optional*, defaults to `1`): The padding for the downsample convolution.
+ downsample_type (`str`, *optional*, defaults to `conv`):
+ The downsample type for downsampling layers. Choose between "conv" and "resnet"
+ upsample_type (`str`, *optional*, defaults to `conv`):
+ The upsample type for upsampling layers. Choose between "conv" and "resnet"
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ attention_head_dim (`int`, *optional*, defaults to `8`): The attention head dimension.
+ norm_num_groups (`int`, *optional*, defaults to `32`): The number of groups for normalization.
+ attn_norm_num_groups (`int`, *optional*, defaults to `None`):
+ If set to an integer, a group norm layer will be created in the mid block's [`Attention`] layer with the
+ given number of groups. If left as `None`, the group norm layer will only be created if
+ `resnet_time_scale_shift` is set to `default`, and if created will have `norm_num_groups` groups.
+ norm_eps (`float`, *optional*, defaults to `1e-5`): The epsilon for normalization.
+ resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+ for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+ `"timestep"`, or `"identity"`.
+ num_class_embeds (`int`, *optional*, defaults to `None`):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim` when performing class
+ conditioning with `class_embed_type` equal to `None`.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[Union[int, Tuple[int, int]]] = None,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ center_input_sample: bool = False,
+ time_embedding_type: str = "positional",
+ freq_shift: int = 0,
+ flip_sin_to_cos: bool = True,
+ down_block_types: Tuple[str, ...] = ("DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D"),
+ up_block_types: Tuple[str, ...] = ("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D"),
+ block_out_channels: Tuple[int, ...] = (224, 448, 672, 896),
+ layers_per_block: int = 2,
+ mid_block_scale_factor: float = 1,
+ downsample_padding: int = 1,
+ downsample_type: str = "conv",
+ upsample_type: str = "conv",
+ dropout: float = 0.0,
+ act_fn: str = "silu",
+ attention_head_dim: Optional[int] = 8,
+ norm_num_groups: int = 32,
+ attn_norm_num_groups: Optional[int] = None,
+ norm_eps: float = 1e-5,
+ resnet_time_scale_shift: str = "default",
+ add_attention: bool = True,
+ class_embed_type: Optional[str] = None,
+ num_class_embeds: Optional[int] = None,
+ num_train_timesteps: Optional[int] = None,
+ ):
+ super().__init__()
+
+ self.sample_size = sample_size
+ time_embed_dim = block_out_channels[0] * 4
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ # input
+ self.conv_in = nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
+
+ # time
+ if time_embedding_type == "fourier":
+ self.time_proj = GaussianFourierProjection(embedding_size=block_out_channels[0], scale=16)
+ timestep_input_dim = 2 * block_out_channels[0]
+ elif time_embedding_type == "positional":
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+ timestep_input_dim = block_out_channels[0]
+ elif time_embedding_type == "learned":
+ self.time_proj = nn.Embedding(num_train_timesteps, block_out_channels[0])
+ timestep_input_dim = block_out_channels[0]
+
+ self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+ # class embedding
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ else:
+ self.class_embedding = None
+
+ self.down_blocks = nn.ModuleList([])
+ self.mid_block = None
+ self.up_blocks = nn.ModuleList([])
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ attention_head_dim=attention_head_dim if attention_head_dim is not None else output_channel,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ downsample_type=downsample_type,
+ dropout=dropout,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ self.mid_block = UNetMidBlock2D(
+ in_channels=block_out_channels[-1],
+ temb_channels=time_embed_dim,
+ dropout=dropout,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_head_dim=attention_head_dim if attention_head_dim is not None else block_out_channels[-1],
+ resnet_groups=norm_num_groups,
+ attn_groups=attn_norm_num_groups,
+ add_attention=add_attention,
+ )
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ is_final_block = i == len(block_out_channels) - 1
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=layers_per_block + 1,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=time_embed_dim,
+ add_upsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ attention_head_dim=attention_head_dim if attention_head_dim is not None else output_channel,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ upsample_type=upsample_type,
+ dropout=dropout,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ num_groups_out = norm_num_groups if norm_num_groups is not None else min(block_out_channels[0] // 4, 32)
+ self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=num_groups_out, eps=norm_eps)
+ self.conv_act = nn.SiLU()
+ self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ class_labels: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ ) -> Union[UNet2DOutput, Tuple]:
+ r"""
+ The [`UNet2DModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, channel, height, width)`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ class_labels (`torch.FloatTensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unet_2d.UNet2DOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.unet_2d.UNet2DOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unet_2d.UNet2DOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is the sample tensor.
+ """
+ # 0. center input if necessary
+ if self.config.center_input_sample:
+ sample = 2 * sample - 1.0
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
+ elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps * torch.ones(sample.shape[0], dtype=timesteps.dtype, device=timesteps.device)
+
+ t_emb = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=self.dtype)
+ emb = self.time_embedding(t_emb)
+
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError("class_labels should be provided when doing class conditioning")
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+ emb = emb + class_emb
+ elif self.class_embedding is None and class_labels is not None:
+ raise ValueError("class_embedding needs to be initialized in order to use class conditioning")
+
+ # 2. pre-process
+ skip_sample = sample
+ sample = self.conv_in(sample)
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "skip_conv"):
+ sample, res_samples, skip_sample = downsample_block(
+ hidden_states=sample, temb=emb, skip_sample=skip_sample
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+ down_block_res_samples += res_samples
+
+ # 4. mid
+ sample = self.mid_block(sample, emb)
+
+ # 5. up
+ skip_sample = None
+ for upsample_block in self.up_blocks:
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ if hasattr(upsample_block, "skip_conv"):
+ sample, skip_sample = upsample_block(sample, res_samples, emb, skip_sample)
+ else:
+ sample = upsample_block(sample, res_samples, emb)
+
+ # 6. post-process
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ if skip_sample is not None:
+ sample += skip_sample
+
+ if self.config.time_embedding_type == "fourier":
+ timesteps = timesteps.reshape((sample.shape[0], *([1] * len(sample.shape[1:]))))
+ sample = sample / timesteps
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet2DOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/unet_2d_blocks.py b/model/diffusers_c/models/unets/unet_2d_blocks.py
new file mode 100644
index 0000000000000000000000000000000000000000..ab99a1da755f32b1cf4fbae47db1564a5e26a1c8
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_2d_blocks.py
@@ -0,0 +1,3815 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...utils import is_torch_version, logging
+from ...utils.torch_utils import apply_freeu
+from ..activations import get_activation
+from ..attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
+from ..normalization import AdaGroupNorm
+from ..resnet import (
+ Downsample2D,
+ FirDownsample2D,
+ FirUpsample2D,
+ KDownsample2D,
+ KUpsample2D,
+ ResnetBlock2D,
+ ResnetBlockCondNorm2D,
+ Upsample2D,
+)
+from ..transformers.dual_transformer_2d import DualTransformer2DModel
+from ..transformers.transformer_2d import Transformer2DModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def get_down_block(
+ down_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ add_downsample: bool,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ resnet_groups: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ downsample_padding: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: float = 1.0,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = None,
+ downsample_type: Optional[str] = None,
+ dropout: float = 0.0,
+):
+ # If attn head dim is not defined, we default it to the number of heads
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+ )
+ attention_head_dim = num_attention_heads
+
+ down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+ if down_block_type == "DownBlock2D":
+ return DownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "ResnetDownsampleBlock2D":
+ return ResnetDownsampleBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ output_scale_factor=resnet_out_scale_factor,
+ )
+ elif down_block_type == "AttnDownBlock2D":
+ if add_downsample is False:
+ downsample_type = None
+ else:
+ downsample_type = downsample_type or "conv" # default to 'conv'
+ return AttnDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ downsample_type=downsample_type,
+ )
+ elif down_block_type == "CrossAttnDownBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
+ return CrossAttnDownBlock2D(
+ num_layers=num_layers,
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ )
+ elif down_block_type == "SimpleCrossAttnDownBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
+ return SimpleCrossAttnDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ output_scale_factor=resnet_out_scale_factor,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+ elif down_block_type == "SkipDownBlock2D":
+ return SkipDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "AttnSkipDownBlock2D":
+ return AttnSkipDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "DownEncoderBlock2D":
+ return DownEncoderBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "AttnDownEncoderBlock2D":
+ return AttnDownEncoderBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "KDownBlock2D":
+ return KDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ )
+ elif down_block_type == "KCrossAttnDownBlock2D":
+ return KCrossAttnDownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ add_self_attention=True if not add_downsample else False,
+ )
+ raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_mid_block(
+ mid_block_type: str,
+ temb_channels: int,
+ in_channels: int,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ resnet_groups: int,
+ output_scale_factor: float = 1.0,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ mid_block_only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = 1,
+ dropout: float = 0.0,
+):
+ if mid_block_type == "UNetMidBlock2DCrossAttn":
+ return UNetMidBlock2DCrossAttn(
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ resnet_groups=resnet_groups,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+ return UNetMidBlock2DSimpleCrossAttn(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ only_cross_attention=mid_block_only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+ elif mid_block_type == "UNetMidBlock2D":
+ return UNetMidBlock2D(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ num_layers=0,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ add_attention=False,
+ )
+ elif mid_block_type == "MidBlock2D":
+ return MidBlock2D(
+ in_channels=in_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resnet_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ )
+ elif mid_block_type is None:
+ return None
+ else:
+ raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+
+def get_up_block(
+ up_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ add_upsample: bool,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ resolution_idx: Optional[int] = None,
+ transformer_layers_per_block: int = 1,
+ num_attention_heads: Optional[int] = None,
+ resnet_groups: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ attention_type: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: float = 1.0,
+ cross_attention_norm: Optional[str] = None,
+ attention_head_dim: Optional[int] = None,
+ upsample_type: Optional[str] = None,
+ dropout: float = 0.0,
+) -> nn.Module:
+ # If attn head dim is not defined, we default it to the number of heads
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+ )
+ attention_head_dim = num_attention_heads
+
+ up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+ if up_block_type == "UpBlock2D":
+ return UpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif up_block_type == "ResnetUpsampleBlock2D":
+ return ResnetUpsampleBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ output_scale_factor=resnet_out_scale_factor,
+ )
+ elif up_block_type == "CrossAttnUpBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
+ return CrossAttnUpBlock2D(
+ num_layers=num_layers,
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ )
+ elif up_block_type == "SimpleCrossAttnUpBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
+ return SimpleCrossAttnUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ output_scale_factor=resnet_out_scale_factor,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+ elif up_block_type == "AttnUpBlock2D":
+ if add_upsample is False:
+ upsample_type = None
+ else:
+ upsample_type = upsample_type or "conv" # default to 'conv'
+
+ return AttnUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ upsample_type=upsample_type,
+ )
+ elif up_block_type == "SkipUpBlock2D":
+ return SkipUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif up_block_type == "AttnSkipUpBlock2D":
+ return AttnSkipUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif up_block_type == "UpDecoderBlock2D":
+ return UpDecoderBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ temb_channels=temb_channels,
+ )
+ elif up_block_type == "AttnUpDecoderBlock2D":
+ return AttnUpDecoderBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ attention_head_dim=attention_head_dim,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ temb_channels=temb_channels,
+ )
+ elif up_block_type == "KUpBlock2D":
+ return KUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ )
+ elif up_block_type == "KCrossAttnUpBlock2D":
+ return KCrossAttnUpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ cross_attention_dim=cross_attention_dim,
+ attention_head_dim=attention_head_dim,
+ )
+
+ raise ValueError(f"{up_block_type} does not exist.")
+
+
+class AutoencoderTinyBlock(nn.Module):
+ """
+ Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU
+ blocks.
+
+ Args:
+ in_channels (`int`): The number of input channels.
+ out_channels (`int`): The number of output channels.
+ act_fn (`str`):
+ ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`.
+
+ Returns:
+ `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to
+ `out_channels`.
+ """
+
+ def __init__(self, in_channels: int, out_channels: int, act_fn: str):
+ super().__init__()
+ act_fn = get_activation(act_fn)
+ self.conv = nn.Sequential(
+ nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1),
+ act_fn,
+ nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+ act_fn,
+ nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1),
+ )
+ self.skip = (
+ nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
+ if in_channels != out_channels
+ else nn.Identity()
+ )
+ self.fuse = nn.ReLU()
+
+ def forward(self, x: torch.FloatTensor) -> torch.FloatTensor:
+ return self.fuse(self.conv(x) + self.skip(x))
+
+
+class UNetMidBlock2D(nn.Module):
+ """
+ A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks.
+
+ Args:
+ in_channels (`int`): The number of input channels.
+ temb_channels (`int`): The number of temporal embedding channels.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+ num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+ resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+ resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
+ The type of normalization to apply to the time embeddings. This can help to improve the performance of the
+ model on tasks with long-range temporal dependencies.
+ resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
+ resnet_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use in the group normalization layers of the resnet blocks.
+ attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
+ resnet_pre_norm (`bool`, *optional*, defaults to `True`):
+ Whether to use pre-normalization for the resnet blocks.
+ add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
+ attention_head_dim (`int`, *optional*, defaults to 1):
+ Dimension of a single attention head. The number of attention heads is determined based on this value and
+ the number of input channels.
+ output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
+
+ Returns:
+ `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+ in_channels, height, width)`.
+
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default", # default, spatial
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ attn_groups: Optional[int] = None,
+ resnet_pre_norm: bool = True,
+ add_attention: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ ):
+ super().__init__()
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+ self.add_attention = add_attention
+
+ if attn_groups is None:
+ attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None
+
+ # there is always at least one resnet
+ if resnet_time_scale_shift == "spatial":
+ resnets = [
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ ]
+ else:
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ attentions = []
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+ )
+ attention_head_dim = in_channels
+
+ for _ in range(num_layers):
+ if self.add_attention:
+ attentions.append(
+ Attention(
+ in_channels,
+ heads=in_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=attn_groups,
+ spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+ else:
+ attentions.append(None)
+
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = self.resnets[0](hidden_states, temb)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ if attn is not None:
+ hidden_states = attn(hidden_states, temb=temb)
+ hidden_states = resnet(hidden_states, temb)
+
+ return hidden_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ output_scale_factor: float = 1.0,
+ cross_attention_dim: int = 1280,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ # support for variable transformer layers per block
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ attentions = []
+
+ for i in range(num_layers):
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ else:
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class UNetMidBlock2DSimpleCrossAttn(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ cross_attention_dim: int = 1280,
+ skip_time_act: bool = False,
+ only_cross_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+
+ self.attention_head_dim = attention_head_dim
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ self.num_heads = in_channels // self.attention_head_dim
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ ]
+ attentions = []
+
+ for _ in range(num_layers):
+ processor = (
+ AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+ )
+
+ attentions.append(
+ Attention(
+ query_dim=in_channels,
+ cross_attention_dim=in_channels,
+ heads=self.num_heads,
+ dim_head=self.attention_head_dim,
+ added_kv_proj_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ bias=True,
+ upcast_softmax=True,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ processor=processor,
+ )
+ )
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+
+ if attention_mask is None:
+ # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
+ mask = None if encoder_hidden_states is None else encoder_attention_mask
+ else:
+ # when attention_mask is defined: we don't even check for encoder_attention_mask.
+ # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
+ # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
+ # then we can simplify this whole if/else block to:
+ # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
+ mask = attention_mask
+
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ # attn
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+
+ # resnet
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class MidBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ use_linear_projection: bool = False,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = False
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+
+ for i in range(num_layers):
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ lora_scale = 1.0
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+ for resnet in self.resnets[1:]:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class AttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ downsample_padding: int = 1,
+ downsample_type: str = "conv",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+ self.downsample_type = downsample_type
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=resnet_groups,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if downsample_type == "conv":
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ elif downsample_type == "resnet":
+ self.downsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ down=True,
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+
+ output_states = ()
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ cross_attention_kwargs.update({"scale": lora_scale})
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(hidden_states, **cross_attention_kwargs)
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ if self.downsample_type == "resnet":
+ hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale)
+ else:
+ hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class CrossAttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ downsample_padding: int = 1,
+ add_downsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ additional_residuals: Optional[torch.FloatTensor] = None,
+ down_block_add_samples: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+ blocks = list(zip(self.resnets, self.attentions))
+
+ for i, (resnet, attn) in enumerate(blocks):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+
+ # apply additional residuals to the output of the last pair of resnet and attention blocks
+ if i == len(blocks) - 1 and additional_residuals is not None:
+ hidden_states = hidden_states + additional_residuals
+
+ if down_block_add_samples is not None:
+ hidden_states = hidden_states + down_block_add_samples.pop(0)
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+ if down_block_add_samples is not None:
+ hidden_states = hidden_states + down_block_add_samples.pop(0) # todo: add before or after
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class DownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ down_block_add_samples: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if down_block_add_samples is not None:
+ hidden_states = hidden_states + down_block_add_samples.pop(0)
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale=scale)
+
+ if down_block_add_samples is not None:
+ hidden_states = hidden_states + down_block_add_samples.pop(0) # todo: add before or after
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class DownEncoderBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, temb=None, scale=scale)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale)
+
+ return hidden_states
+
+
+class AttnDownEncoderBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=resnet_groups,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor:
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states, temb=None, scale=scale)
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = attn(hidden_states, **cross_attention_kwargs)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale)
+
+ return hidden_states
+
+
+class AttnSkipDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = np.sqrt(2.0),
+ add_downsample: bool = True,
+ ):
+ super().__init__()
+ self.attentions = nn.ModuleList([])
+ self.resnets = nn.ModuleList([])
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ self.resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(in_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ self.attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=32,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ if add_downsample:
+ self.resnet_down = ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ use_in_shortcut=True,
+ down=True,
+ kernel="fir",
+ )
+ self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
+ self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
+ else:
+ self.resnet_down = None
+ self.downsamplers = None
+ self.skip_conv = None
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ skip_sample: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]:
+ output_states = ()
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = attn(hidden_states, **cross_attention_kwargs)
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ hidden_states = self.resnet_down(hidden_states, temb, scale=scale)
+ for downsampler in self.downsamplers:
+ skip_sample = downsampler(skip_sample)
+
+ hidden_states = self.skip_conv(skip_sample) + hidden_states
+
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states, skip_sample
+
+
+class SkipDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = np.sqrt(2.0),
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ self.resnets = nn.ModuleList([])
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ self.resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(in_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ if add_downsample:
+ self.resnet_down = ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ use_in_shortcut=True,
+ down=True,
+ kernel="fir",
+ )
+ self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)])
+ self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1))
+ else:
+ self.resnet_down = None
+ self.downsamplers = None
+ self.skip_conv = None
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ skip_sample: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, temb, scale)
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ hidden_states = self.resnet_down(hidden_states, temb, scale)
+ for downsampler in self.downsamplers:
+ skip_sample = downsampler(skip_sample)
+
+ hidden_states = self.skip_conv(skip_sample) + hidden_states
+
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states, skip_sample
+
+
+class ResnetDownsampleBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ skip_time_act: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ down=True,
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale)
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, temb, scale)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class SimpleCrossAttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ skip_time_act: bool = False,
+ only_cross_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+
+ resnets = []
+ attentions = []
+
+ self.attention_head_dim = attention_head_dim
+ self.num_heads = out_channels // self.attention_head_dim
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ processor = (
+ AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+ )
+
+ attentions.append(
+ Attention(
+ query_dim=out_channels,
+ cross_attention_dim=out_channels,
+ heads=self.num_heads,
+ dim_head=attention_head_dim,
+ added_kv_proj_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ bias=True,
+ upcast_softmax=True,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ processor=processor,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ down=True,
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+
+ if attention_mask is None:
+ # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
+ mask = None if encoder_hidden_states is None else encoder_attention_mask
+ else:
+ # when attention_mask is defined: we don't even check for encoder_attention_mask.
+ # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
+ # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
+ # then we can simplify this whole if/else block to:
+ # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
+ mask = attention_mask
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, temb, scale=lora_scale)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class KDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 4,
+ resnet_eps: float = 1e-5,
+ resnet_act_fn: str = "gelu",
+ resnet_group_size: int = 32,
+ add_downsample: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ groups = in_channels // resnet_group_size
+ groups_out = out_channels // resnet_group_size
+
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ dropout=dropout,
+ temb_channels=temb_channels,
+ groups=groups,
+ groups_out=groups_out,
+ eps=resnet_eps,
+ non_linearity=resnet_act_fn,
+ time_embedding_norm="ada_group",
+ conv_shortcut_bias=False,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ # YiYi's comments- might be able to use FirDownsample2D, look into details later
+ self.downsamplers = nn.ModuleList([KDownsample2D()])
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale)
+
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ return hidden_states, output_states
+
+
+class KCrossAttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ cross_attention_dim: int,
+ dropout: float = 0.0,
+ num_layers: int = 4,
+ resnet_group_size: int = 32,
+ add_downsample: bool = True,
+ attention_head_dim: int = 64,
+ add_self_attention: bool = False,
+ resnet_eps: float = 1e-5,
+ resnet_act_fn: str = "gelu",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ groups = in_channels // resnet_group_size
+ groups_out = out_channels // resnet_group_size
+
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ dropout=dropout,
+ temb_channels=temb_channels,
+ groups=groups,
+ groups_out=groups_out,
+ eps=resnet_eps,
+ non_linearity=resnet_act_fn,
+ time_embedding_norm="ada_group",
+ conv_shortcut_bias=False,
+ )
+ )
+ attentions.append(
+ KAttentionBlock(
+ out_channels,
+ out_channels // attention_head_dim,
+ attention_head_dim,
+ cross_attention_dim=cross_attention_dim,
+ temb_channels=temb_channels,
+ attention_bias=True,
+ add_self_attention=add_self_attention,
+ cross_attention_norm="layer_norm",
+ group_size=resnet_group_size,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.attentions = nn.ModuleList(attentions)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList([KDownsample2D()])
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ emb=temb,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ emb=temb,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+
+ if self.downsamplers is None:
+ output_states += (None,)
+ else:
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ return hidden_states, output_states
+
+
+class AttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: int = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ upsample_type: str = "conv",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.upsample_type = upsample_type
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=resnet_groups,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if upsample_type == "conv":
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ elif upsample_type == "resnet":
+ self.upsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ up=True,
+ )
+ ]
+ )
+ else:
+ self.upsamplers = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = attn(hidden_states, **cross_attention_kwargs)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ if self.upsample_type == "resnet":
+ hidden_states = upsampler(hidden_states, temb=temb, scale=scale)
+ else:
+ hidden_states = upsampler(hidden_states, scale=scale)
+
+ return hidden_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ return_res_samples: Optional[bool] = False,
+ up_block_add_samples: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+ if return_res_samples:
+ output_states = ()
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ if return_res_samples:
+ output_states = output_states + (hidden_states,)
+ if up_block_add_samples is not None:
+ hidden_states = hidden_states + up_block_add_samples.pop(0)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
+ if return_res_samples:
+ output_states = output_states + (hidden_states,)
+ if up_block_add_samples is not None:
+ hidden_states = hidden_states + up_block_add_samples.pop(0)
+
+ if return_res_samples:
+ return hidden_states, output_states
+ else:
+ return hidden_states
+
+
+class UpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ return_res_samples: Optional[bool] = False,
+ up_block_add_samples: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+ if return_res_samples:
+ output_states = ()
+
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if return_res_samples:
+ output_states = output_states + (hidden_states,)
+ if up_block_add_samples is not None:
+ hidden_states = hidden_states + up_block_add_samples.pop(0) # todo: add before or after
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+ if return_res_samples:
+ output_states = output_states + (hidden_states,)
+ if up_block_add_samples is not None:
+ hidden_states = hidden_states + up_block_add_samples.pop(0) # todo: add before or after
+
+ if return_res_samples:
+ return hidden_states, output_states
+ else:
+ return hidden_states
+
+
+class UpDecoderBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default", # default, spatial
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ temb_channels: Optional[int] = None,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ input_channels = in_channels if i == 0 else out_channels
+
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> torch.FloatTensor:
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, temb=temb, scale=scale)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
+
+
+class AttnUpDecoderBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ temb_channels: Optional[int] = None,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ for i in range(num_layers):
+ input_channels = in_channels if i == 0 else out_channels
+
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None,
+ spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> torch.FloatTensor:
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states, temb=temb, scale=scale)
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, scale=scale)
+
+ return hidden_states
+
+
+class AttnSkipUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = np.sqrt(2.0),
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ self.attentions = nn.ModuleList([])
+ self.resnets = nn.ModuleList([])
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ self.resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(resnet_in_channels + res_skip_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}."
+ )
+ attention_head_dim = out_channels
+
+ self.attentions.append(
+ Attention(
+ out_channels,
+ heads=out_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=32,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+
+ self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
+ if add_upsample:
+ self.resnet_up = ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(out_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ use_in_shortcut=True,
+ up=True,
+ kernel="fir",
+ )
+ self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+ self.skip_norm = torch.nn.GroupNorm(
+ num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
+ )
+ self.act = nn.SiLU()
+ else:
+ self.resnet_up = None
+ self.skip_conv = None
+ self.skip_norm = None
+ self.act = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ skip_sample=None,
+ scale: float = 1.0,
+ ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ cross_attention_kwargs = {"scale": scale}
+ hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs)
+
+ if skip_sample is not None:
+ skip_sample = self.upsampler(skip_sample)
+ else:
+ skip_sample = 0
+
+ if self.resnet_up is not None:
+ skip_sample_states = self.skip_norm(hidden_states)
+ skip_sample_states = self.act(skip_sample_states)
+ skip_sample_states = self.skip_conv(skip_sample_states)
+
+ skip_sample = skip_sample + skip_sample_states
+
+ hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
+
+ return hidden_states, skip_sample
+
+
+class SkipUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = np.sqrt(2.0),
+ add_upsample: bool = True,
+ upsample_padding: int = 1,
+ ):
+ super().__init__()
+ self.resnets = nn.ModuleList([])
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ self.resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min((resnet_in_channels + res_skip_channels) // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels)
+ if add_upsample:
+ self.resnet_up = ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=min(out_channels // 4, 32),
+ groups_out=min(out_channels // 4, 32),
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ use_in_shortcut=True,
+ up=True,
+ kernel="fir",
+ )
+ self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
+ self.skip_norm = torch.nn.GroupNorm(
+ num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True
+ )
+ self.act = nn.SiLU()
+ else:
+ self.resnet_up = None
+ self.skip_conv = None
+ self.skip_norm = None
+ self.act = None
+
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ skip_sample=None,
+ scale: float = 1.0,
+ ) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if skip_sample is not None:
+ skip_sample = self.upsampler(skip_sample)
+ else:
+ skip_sample = 0
+
+ if self.resnet_up is not None:
+ skip_sample_states = self.skip_norm(hidden_states)
+ skip_sample_states = self.act(skip_sample_states)
+ skip_sample_states = self.skip_conv(skip_sample_states)
+
+ skip_sample = skip_sample + skip_sample_states
+
+ hidden_states = self.resnet_up(hidden_states, temb, scale=scale)
+
+ return hidden_states, skip_sample
+
+
+class ResnetUpsampleBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ skip_time_act: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ up=True,
+ )
+ ]
+ )
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, temb, scale=scale)
+
+ return hidden_states
+
+
+class SimpleCrossAttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ skip_time_act: bool = False,
+ only_cross_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.attention_head_dim = attention_head_dim
+
+ self.num_heads = out_channels // self.attention_head_dim
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ processor = (
+ AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+ )
+
+ attentions.append(
+ Attention(
+ query_dim=out_channels,
+ cross_attention_dim=out_channels,
+ heads=self.num_heads,
+ dim_head=self.attention_head_dim,
+ added_kv_proj_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ bias=True,
+ upcast_softmax=True,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ processor=processor,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList(
+ [
+ ResnetBlock2D(
+ in_channels=out_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ up=True,
+ )
+ ]
+ )
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+ if attention_mask is None:
+ # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
+ mask = None if encoder_hidden_states is None else encoder_attention_mask
+ else:
+ # when attention_mask is defined: we don't even check for encoder_attention_mask.
+ # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
+ # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
+ # then we can simplify this whole if/else block to:
+ # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
+ mask = attention_mask
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # resnet
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class KUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: int,
+ dropout: float = 0.0,
+ num_layers: int = 5,
+ resnet_eps: float = 1e-5,
+ resnet_act_fn: str = "gelu",
+ resnet_group_size: Optional[int] = 32,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+ k_in_channels = 2 * out_channels
+ k_out_channels = in_channels
+ num_layers = num_layers - 1
+
+ for i in range(num_layers):
+ in_channels = k_in_channels if i == 0 else out_channels
+ groups = in_channels // resnet_group_size
+ groups_out = out_channels // resnet_group_size
+
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=k_out_channels if (i == num_layers - 1) else out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=groups,
+ groups_out=groups_out,
+ dropout=dropout,
+ non_linearity=resnet_act_fn,
+ time_embedding_norm="ada_group",
+ conv_shortcut_bias=False,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([KUpsample2D()])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ res_hidden_states_tuple = res_hidden_states_tuple[-1]
+ if res_hidden_states_tuple is not None:
+ hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
+
+
+class KCrossAttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: int,
+ dropout: float = 0.0,
+ num_layers: int = 4,
+ resnet_eps: float = 1e-5,
+ resnet_act_fn: str = "gelu",
+ resnet_group_size: int = 32,
+ attention_head_dim: int = 1, # attention dim_head
+ cross_attention_dim: int = 768,
+ add_upsample: bool = True,
+ upcast_attention: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ is_first_block = in_channels == out_channels == temb_channels
+ is_middle_block = in_channels != out_channels
+ add_self_attention = True if is_first_block else False
+
+ self.has_cross_attention = True
+ self.attention_head_dim = attention_head_dim
+
+ # in_channels, and out_channels for the block (k-unet)
+ k_in_channels = out_channels if is_first_block else 2 * out_channels
+ k_out_channels = in_channels
+
+ num_layers = num_layers - 1
+
+ for i in range(num_layers):
+ in_channels = k_in_channels if i == 0 else out_channels
+ groups = in_channels // resnet_group_size
+ groups_out = out_channels // resnet_group_size
+
+ if is_middle_block and (i == num_layers - 1):
+ conv_2d_out_channels = k_out_channels
+ else:
+ conv_2d_out_channels = None
+
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ conv_2d_out_channels=conv_2d_out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=groups,
+ groups_out=groups_out,
+ dropout=dropout,
+ non_linearity=resnet_act_fn,
+ time_embedding_norm="ada_group",
+ conv_shortcut_bias=False,
+ )
+ )
+ attentions.append(
+ KAttentionBlock(
+ k_out_channels if (i == num_layers - 1) else out_channels,
+ k_out_channels // attention_head_dim
+ if (i == num_layers - 1)
+ else out_channels // attention_head_dim,
+ attention_head_dim,
+ cross_attention_dim=cross_attention_dim,
+ temb_channels=temb_channels,
+ attention_bias=True,
+ add_self_attention=add_self_attention,
+ cross_attention_norm="layer_norm",
+ upcast_attention=upcast_attention,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.attentions = nn.ModuleList(attentions)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([KUpsample2D()])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ res_hidden_states_tuple = res_hidden_states_tuple[-1]
+ if res_hidden_states_tuple is not None:
+ hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1)
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ for resnet, attn in zip(self.resnets, self.attentions):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ emb=temb,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ emb=temb,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
+
+
+# can potentially later be renamed to `No-feed-forward` attention
+class KAttentionBlock(nn.Module):
+ r"""
+ A basic Transformer block.
+
+ Parameters:
+ dim (`int`): The number of channels in the input and output.
+ num_attention_heads (`int`): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`): The number of channels in each head.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+ attention_bias (`bool`, *optional*, defaults to `False`):
+ Configure if the attention layers should contain a bias parameter.
+ upcast_attention (`bool`, *optional*, defaults to `False`):
+ Set to `True` to upcast the attention computation to `float32`.
+ temb_channels (`int`, *optional*, defaults to 768):
+ The number of channels in the token embedding.
+ add_self_attention (`bool`, *optional*, defaults to `False`):
+ Set to `True` to add self-attention to the block.
+ cross_attention_norm (`str`, *optional*, defaults to `None`):
+ The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`.
+ group_size (`int`, *optional*, defaults to 32):
+ The number of groups to separate the channels into for group normalization.
+ """
+
+ def __init__(
+ self,
+ dim: int,
+ num_attention_heads: int,
+ attention_head_dim: int,
+ dropout: float = 0.0,
+ cross_attention_dim: Optional[int] = None,
+ attention_bias: bool = False,
+ upcast_attention: bool = False,
+ temb_channels: int = 768, # for ada_group_norm
+ add_self_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ group_size: int = 32,
+ ):
+ super().__init__()
+ self.add_self_attention = add_self_attention
+
+ # 1. Self-Attn
+ if add_self_attention:
+ self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
+ self.attn1 = Attention(
+ query_dim=dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ cross_attention_dim=None,
+ cross_attention_norm=None,
+ )
+
+ # 2. Cross-Attn
+ self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size))
+ self.attn2 = Attention(
+ query_dim=dim,
+ cross_attention_dim=cross_attention_dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ upcast_attention=upcast_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+
+ def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor:
+ return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1)
+
+ def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor:
+ return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ # TODO: mark emb as non-optional (self.norm2 requires it).
+ # requires assessing impact of change to positional param interface.
+ emb: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+ # 1. Self-Attention
+ if self.add_self_attention:
+ norm_hidden_states = self.norm1(hidden_states, emb)
+
+ height, weight = norm_hidden_states.shape[2:]
+ norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
+
+ attn_output = self.attn1(
+ norm_hidden_states,
+ encoder_hidden_states=None,
+ attention_mask=attention_mask,
+ **cross_attention_kwargs,
+ )
+ attn_output = self._to_4d(attn_output, height, weight)
+
+ hidden_states = attn_output + hidden_states
+
+ # 2. Cross-Attention/None
+ norm_hidden_states = self.norm2(hidden_states, emb)
+
+ height, weight = norm_hidden_states.shape[2:]
+ norm_hidden_states = self._to_3d(norm_hidden_states, height, weight)
+ attn_output = self.attn2(
+ norm_hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask,
+ **cross_attention_kwargs,
+ )
+ attn_output = self._to_4d(attn_output, height, weight)
+
+ hidden_states = attn_output + hidden_states
+
+ return hidden_states
diff --git a/model/diffusers_c/models/unets/unet_2d_blocks_flax.py b/model/diffusers_c/models/unets/unet_2d_blocks_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..a4585dbc8823185a19952a103bd911e1bd6fa590
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_2d_blocks_flax.py
@@ -0,0 +1,400 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import flax.linen as nn
+import jax.numpy as jnp
+
+from ..attention_flax import FlaxTransformer2DModel
+from ..resnet_flax import FlaxDownsample2D, FlaxResnetBlock2D, FlaxUpsample2D
+
+
+class FlaxCrossAttnDownBlock2D(nn.Module):
+ r"""
+ Cross Attention 2D Downsizing block - original architecture from Unet transformers:
+ https://arxiv.org/abs/2103.06104
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input channels
+ out_channels (:obj:`int`):
+ Output channels
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ num_layers (:obj:`int`, *optional*, defaults to 1):
+ Number of attention blocks layers
+ num_attention_heads (:obj:`int`, *optional*, defaults to 1):
+ Number of attention heads of each spatial transformer block
+ add_downsample (:obj:`bool`, *optional*, defaults to `True`):
+ Whether to add downsampling layer before each final output
+ use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+ enable memory efficient attention https://arxiv.org/abs/2112.05682
+ split_head_dim (`bool`, *optional*, defaults to `False`):
+ Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+ enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ out_channels: int
+ dropout: float = 0.0
+ num_layers: int = 1
+ num_attention_heads: int = 1
+ add_downsample: bool = True
+ use_linear_projection: bool = False
+ only_cross_attention: bool = False
+ use_memory_efficient_attention: bool = False
+ split_head_dim: bool = False
+ dtype: jnp.dtype = jnp.float32
+ transformer_layers_per_block: int = 1
+
+ def setup(self):
+ resnets = []
+ attentions = []
+
+ for i in range(self.num_layers):
+ in_channels = self.in_channels if i == 0 else self.out_channels
+
+ res_block = FlaxResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=self.out_channels,
+ dropout_prob=self.dropout,
+ dtype=self.dtype,
+ )
+ resnets.append(res_block)
+
+ attn_block = FlaxTransformer2DModel(
+ in_channels=self.out_channels,
+ n_heads=self.num_attention_heads,
+ d_head=self.out_channels // self.num_attention_heads,
+ depth=self.transformer_layers_per_block,
+ use_linear_projection=self.use_linear_projection,
+ only_cross_attention=self.only_cross_attention,
+ use_memory_efficient_attention=self.use_memory_efficient_attention,
+ split_head_dim=self.split_head_dim,
+ dtype=self.dtype,
+ )
+ attentions.append(attn_block)
+
+ self.resnets = resnets
+ self.attentions = attentions
+
+ if self.add_downsample:
+ self.downsamplers_0 = FlaxDownsample2D(self.out_channels, dtype=self.dtype)
+
+ def __call__(self, hidden_states, temb, encoder_hidden_states, deterministic=True):
+ output_states = ()
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ hidden_states = resnet(hidden_states, temb, deterministic=deterministic)
+ hidden_states = attn(hidden_states, encoder_hidden_states, deterministic=deterministic)
+ output_states += (hidden_states,)
+
+ if self.add_downsample:
+ hidden_states = self.downsamplers_0(hidden_states)
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class FlaxDownBlock2D(nn.Module):
+ r"""
+ Flax 2D downsizing block
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input channels
+ out_channels (:obj:`int`):
+ Output channels
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ num_layers (:obj:`int`, *optional*, defaults to 1):
+ Number of attention blocks layers
+ add_downsample (:obj:`bool`, *optional*, defaults to `True`):
+ Whether to add downsampling layer before each final output
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ out_channels: int
+ dropout: float = 0.0
+ num_layers: int = 1
+ add_downsample: bool = True
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ resnets = []
+
+ for i in range(self.num_layers):
+ in_channels = self.in_channels if i == 0 else self.out_channels
+
+ res_block = FlaxResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=self.out_channels,
+ dropout_prob=self.dropout,
+ dtype=self.dtype,
+ )
+ resnets.append(res_block)
+ self.resnets = resnets
+
+ if self.add_downsample:
+ self.downsamplers_0 = FlaxDownsample2D(self.out_channels, dtype=self.dtype)
+
+ def __call__(self, hidden_states, temb, deterministic=True):
+ output_states = ()
+
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, temb, deterministic=deterministic)
+ output_states += (hidden_states,)
+
+ if self.add_downsample:
+ hidden_states = self.downsamplers_0(hidden_states)
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class FlaxCrossAttnUpBlock2D(nn.Module):
+ r"""
+ Cross Attention 2D Upsampling block - original architecture from Unet transformers:
+ https://arxiv.org/abs/2103.06104
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input channels
+ out_channels (:obj:`int`):
+ Output channels
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ num_layers (:obj:`int`, *optional*, defaults to 1):
+ Number of attention blocks layers
+ num_attention_heads (:obj:`int`, *optional*, defaults to 1):
+ Number of attention heads of each spatial transformer block
+ add_upsample (:obj:`bool`, *optional*, defaults to `True`):
+ Whether to add upsampling layer before each final output
+ use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+ enable memory efficient attention https://arxiv.org/abs/2112.05682
+ split_head_dim (`bool`, *optional*, defaults to `False`):
+ Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+ enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ out_channels: int
+ prev_output_channel: int
+ dropout: float = 0.0
+ num_layers: int = 1
+ num_attention_heads: int = 1
+ add_upsample: bool = True
+ use_linear_projection: bool = False
+ only_cross_attention: bool = False
+ use_memory_efficient_attention: bool = False
+ split_head_dim: bool = False
+ dtype: jnp.dtype = jnp.float32
+ transformer_layers_per_block: int = 1
+
+ def setup(self):
+ resnets = []
+ attentions = []
+
+ for i in range(self.num_layers):
+ res_skip_channels = self.in_channels if (i == self.num_layers - 1) else self.out_channels
+ resnet_in_channels = self.prev_output_channel if i == 0 else self.out_channels
+
+ res_block = FlaxResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=self.out_channels,
+ dropout_prob=self.dropout,
+ dtype=self.dtype,
+ )
+ resnets.append(res_block)
+
+ attn_block = FlaxTransformer2DModel(
+ in_channels=self.out_channels,
+ n_heads=self.num_attention_heads,
+ d_head=self.out_channels // self.num_attention_heads,
+ depth=self.transformer_layers_per_block,
+ use_linear_projection=self.use_linear_projection,
+ only_cross_attention=self.only_cross_attention,
+ use_memory_efficient_attention=self.use_memory_efficient_attention,
+ split_head_dim=self.split_head_dim,
+ dtype=self.dtype,
+ )
+ attentions.append(attn_block)
+
+ self.resnets = resnets
+ self.attentions = attentions
+
+ if self.add_upsample:
+ self.upsamplers_0 = FlaxUpsample2D(self.out_channels, dtype=self.dtype)
+
+ def __call__(self, hidden_states, res_hidden_states_tuple, temb, encoder_hidden_states, deterministic=True):
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = jnp.concatenate((hidden_states, res_hidden_states), axis=-1)
+
+ hidden_states = resnet(hidden_states, temb, deterministic=deterministic)
+ hidden_states = attn(hidden_states, encoder_hidden_states, deterministic=deterministic)
+
+ if self.add_upsample:
+ hidden_states = self.upsamplers_0(hidden_states)
+
+ return hidden_states
+
+
+class FlaxUpBlock2D(nn.Module):
+ r"""
+ Flax 2D upsampling block
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input channels
+ out_channels (:obj:`int`):
+ Output channels
+ prev_output_channel (:obj:`int`):
+ Output channels from the previous block
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ num_layers (:obj:`int`, *optional*, defaults to 1):
+ Number of attention blocks layers
+ add_downsample (:obj:`bool`, *optional*, defaults to `True`):
+ Whether to add downsampling layer before each final output
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ out_channels: int
+ prev_output_channel: int
+ dropout: float = 0.0
+ num_layers: int = 1
+ add_upsample: bool = True
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ resnets = []
+
+ for i in range(self.num_layers):
+ res_skip_channels = self.in_channels if (i == self.num_layers - 1) else self.out_channels
+ resnet_in_channels = self.prev_output_channel if i == 0 else self.out_channels
+
+ res_block = FlaxResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=self.out_channels,
+ dropout_prob=self.dropout,
+ dtype=self.dtype,
+ )
+ resnets.append(res_block)
+
+ self.resnets = resnets
+
+ if self.add_upsample:
+ self.upsamplers_0 = FlaxUpsample2D(self.out_channels, dtype=self.dtype)
+
+ def __call__(self, hidden_states, res_hidden_states_tuple, temb, deterministic=True):
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = jnp.concatenate((hidden_states, res_hidden_states), axis=-1)
+
+ hidden_states = resnet(hidden_states, temb, deterministic=deterministic)
+
+ if self.add_upsample:
+ hidden_states = self.upsamplers_0(hidden_states)
+
+ return hidden_states
+
+
+class FlaxUNetMidBlock2DCrossAttn(nn.Module):
+ r"""
+ Cross Attention 2D Mid-level block - original architecture from Unet transformers: https://arxiv.org/abs/2103.06104
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input channels
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ num_layers (:obj:`int`, *optional*, defaults to 1):
+ Number of attention blocks layers
+ num_attention_heads (:obj:`int`, *optional*, defaults to 1):
+ Number of attention heads of each spatial transformer block
+ use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+ enable memory efficient attention https://arxiv.org/abs/2112.05682
+ split_head_dim (`bool`, *optional*, defaults to `False`):
+ Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+ enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ dropout: float = 0.0
+ num_layers: int = 1
+ num_attention_heads: int = 1
+ use_linear_projection: bool = False
+ use_memory_efficient_attention: bool = False
+ split_head_dim: bool = False
+ dtype: jnp.dtype = jnp.float32
+ transformer_layers_per_block: int = 1
+
+ def setup(self):
+ # there is always at least one resnet
+ resnets = [
+ FlaxResnetBlock2D(
+ in_channels=self.in_channels,
+ out_channels=self.in_channels,
+ dropout_prob=self.dropout,
+ dtype=self.dtype,
+ )
+ ]
+
+ attentions = []
+
+ for _ in range(self.num_layers):
+ attn_block = FlaxTransformer2DModel(
+ in_channels=self.in_channels,
+ n_heads=self.num_attention_heads,
+ d_head=self.in_channels // self.num_attention_heads,
+ depth=self.transformer_layers_per_block,
+ use_linear_projection=self.use_linear_projection,
+ use_memory_efficient_attention=self.use_memory_efficient_attention,
+ split_head_dim=self.split_head_dim,
+ dtype=self.dtype,
+ )
+ attentions.append(attn_block)
+
+ res_block = FlaxResnetBlock2D(
+ in_channels=self.in_channels,
+ out_channels=self.in_channels,
+ dropout_prob=self.dropout,
+ dtype=self.dtype,
+ )
+ resnets.append(res_block)
+
+ self.resnets = resnets
+ self.attentions = attentions
+
+ def __call__(self, hidden_states, temb, encoder_hidden_states, deterministic=True):
+ hidden_states = self.resnets[0](hidden_states, temb)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ hidden_states = attn(hidden_states, encoder_hidden_states, deterministic=deterministic)
+ hidden_states = resnet(hidden_states, temb, deterministic=deterministic)
+
+ return hidden_states
diff --git a/model/diffusers_c/models/unets/unet_2d_condition.py b/model/diffusers_c/models/unets/unet_2d_condition.py
new file mode 100644
index 0000000000000000000000000000000000000000..ab15cc085afdf67a4f63adf9f207b7b0d4a6926a
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_2d_condition.py
@@ -0,0 +1,1362 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
+from ...utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ..activations import get_activation
+from ..attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ Attention,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ..embeddings import (
+ GaussianFourierProjection,
+ GLIGENTextBoundingboxProjection,
+ ImageHintTimeEmbedding,
+ ImageProjection,
+ ImageTimeEmbedding,
+ TextImageProjection,
+ TextImageTimeEmbedding,
+ TextTimeEmbedding,
+ TimestepEmbedding,
+ Timesteps,
+)
+from ..modeling_utils import ModelMixin
+from .unet_2d_blocks import (
+ get_down_block,
+ get_mid_block,
+ get_up_block,
+)
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+ """
+ The output of [`UNet2DConditionModel`].
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+ The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+ """
+
+ sample: torch.FloatTensor = None
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin):
+ r"""
+ A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+ shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+ Height and width of input/output sample.
+ in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+ center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+ flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+ The tuple of downsample blocks to use.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+ Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+ `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+ The tuple of upsample blocks to use.
+ only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+ Whether to include self-attention in the basic transformer blocks, see
+ [`~models.attention.BasicTransformerBlock`].
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+ downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+ If `None`, normalization and activation layers is skipped in post-processing.
+ norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+ cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+ blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ encoder_hid_dim (`int`, *optional*, defaults to None):
+ If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+ dimension to `cross_attention_dim`.
+ encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+ If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+ embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+ attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+ num_attention_heads (`int`, *optional*):
+ The number of attention heads. If not defined, defaults to `attention_head_dim`
+ resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+ for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+ addition_embed_type (`str`, *optional*, defaults to `None`):
+ Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+ "text". "text" will use the `TextTimeEmbedding` layer.
+ addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+ Dimension for the timestep embeddings.
+ num_class_embeds (`int`, *optional*, defaults to `None`):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+ class conditioning with `class_embed_type` equal to `None`.
+ time_embedding_type (`str`, *optional*, defaults to `positional`):
+ The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+ time_embedding_dim (`int`, *optional*, defaults to `None`):
+ An optional override for the dimension of the projected time embedding.
+ time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+ Optional activation function to use only once on the time embeddings before they are passed to the rest of
+ the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+ timestep_post_act (`str`, *optional*, defaults to `None`):
+ The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+ time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+ The dimension of `cond_proj` layer in the timestep embedding.
+ conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+ *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+ *optional*): The dimension of the `class_labels` input when
+ `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+ class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+ embeddings with the class embeddings.
+ mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+ Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+ `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+ `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+ otherwise.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[int] = None,
+ in_channels: int = 4,
+ out_channels: int = 4,
+ center_input_sample: bool = False,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str] = (
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ),
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+ up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+ layers_per_block: Union[int, Tuple[int]] = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ dropout: float = 0.0,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: Union[int, Tuple[int]] = 1280,
+ transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+ reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+ encoder_hid_dim: Optional[int] = None,
+ encoder_hid_dim_type: Optional[str] = None,
+ attention_head_dim: Union[int, Tuple[int]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ class_embed_type: Optional[str] = None,
+ addition_embed_type: Optional[str] = None,
+ addition_time_embed_dim: Optional[int] = None,
+ num_class_embeds: Optional[int] = None,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: float = 1.0,
+ time_embedding_type: str = "positional",
+ time_embedding_dim: Optional[int] = None,
+ time_embedding_act_fn: Optional[str] = None,
+ timestep_post_act: Optional[str] = None,
+ time_cond_proj_dim: Optional[int] = None,
+ conv_in_kernel: int = 3,
+ conv_out_kernel: int = 3,
+ projection_class_embeddings_input_dim: Optional[int] = None,
+ attention_type: str = "default",
+ class_embeddings_concat: bool = False,
+ mid_block_only_cross_attention: Optional[bool] = None,
+ cross_attention_norm: Optional[str] = None,
+ addition_embed_type_num_heads: int = 64,
+ ):
+ super().__init__()
+
+ self.sample_size = sample_size
+
+ if num_attention_heads is not None:
+ raise ValueError(
+ "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+ )
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ self._check_config(
+ down_block_types=down_block_types,
+ up_block_types=up_block_types,
+ only_cross_attention=only_cross_attention,
+ block_out_channels=block_out_channels,
+ layers_per_block=layers_per_block,
+ cross_attention_dim=cross_attention_dim,
+ transformer_layers_per_block=transformer_layers_per_block,
+ reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
+ attention_head_dim=attention_head_dim,
+ num_attention_heads=num_attention_heads,
+ )
+
+ # input
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in = nn.Conv2d(
+ in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+ )
+
+ # time
+ time_embed_dim, timestep_input_dim = self._set_time_proj(
+ time_embedding_type,
+ block_out_channels=block_out_channels,
+ flip_sin_to_cos=flip_sin_to_cos,
+ freq_shift=freq_shift,
+ time_embedding_dim=time_embedding_dim,
+ )
+
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ post_act_fn=timestep_post_act,
+ cond_proj_dim=time_cond_proj_dim,
+ )
+
+ self._set_encoder_hid_proj(
+ encoder_hid_dim_type,
+ cross_attention_dim=cross_attention_dim,
+ encoder_hid_dim=encoder_hid_dim,
+ )
+
+ # class embedding
+ self._set_class_embedding(
+ class_embed_type,
+ act_fn=act_fn,
+ num_class_embeds=num_class_embeds,
+ projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+ time_embed_dim=time_embed_dim,
+ timestep_input_dim=timestep_input_dim,
+ )
+
+ self._set_add_embedding(
+ addition_embed_type,
+ addition_embed_type_num_heads=addition_embed_type_num_heads,
+ addition_time_embed_dim=addition_time_embed_dim,
+ cross_attention_dim=cross_attention_dim,
+ encoder_hid_dim=encoder_hid_dim,
+ flip_sin_to_cos=flip_sin_to_cos,
+ freq_shift=freq_shift,
+ projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+ time_embed_dim=time_embed_dim,
+ )
+
+ if time_embedding_act_fn is None:
+ self.time_embed_act = None
+ else:
+ self.time_embed_act = get_activation(time_embedding_act_fn)
+
+ self.down_blocks = nn.ModuleList([])
+ self.up_blocks = nn.ModuleList([])
+
+ if isinstance(only_cross_attention, bool):
+ if mid_block_only_cross_attention is None:
+ mid_block_only_cross_attention = only_cross_attention
+
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+ if mid_block_only_cross_attention is None:
+ mid_block_only_cross_attention = False
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ if isinstance(attention_head_dim, int):
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+ if isinstance(layers_per_block, int):
+ layers_per_block = [layers_per_block] * len(down_block_types)
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+ if class_embeddings_concat:
+ # The time embeddings are concatenated with the class embeddings. The dimension of the
+ # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+ # regular time embeddings
+ blocks_time_embed_dim = time_embed_dim * 2
+ else:
+ blocks_time_embed_dim = time_embed_dim
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block[i],
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim[i],
+ num_attention_heads=num_attention_heads[i],
+ downsample_padding=downsample_padding,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ resnet_out_scale_factor=resnet_out_scale_factor,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ dropout=dropout,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ self.mid_block = get_mid_block(
+ mid_block_type,
+ temb_channels=blocks_time_embed_dim,
+ in_channels=block_out_channels[-1],
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ output_scale_factor=mid_block_scale_factor,
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ num_attention_heads=num_attention_heads[-1],
+ cross_attention_dim=cross_attention_dim[-1],
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ mid_block_only_cross_attention=mid_block_only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim[-1],
+ dropout=dropout,
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ reversed_layers_per_block = list(reversed(layers_per_block))
+ reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+ reversed_transformer_layers_per_block = (
+ list(reversed(transformer_layers_per_block))
+ if reverse_transformer_layers_per_block is None
+ else reverse_transformer_layers_per_block
+ )
+ only_cross_attention = list(reversed(only_cross_attention))
+
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=reversed_layers_per_block[i] + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resolution_idx=i,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=reversed_cross_attention_dim[i],
+ num_attention_heads=reversed_num_attention_heads[i],
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ resnet_out_scale_factor=resnet_out_scale_factor,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ dropout=dropout,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ if norm_num_groups is not None:
+ self.conv_norm_out = nn.GroupNorm(
+ num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+ )
+
+ self.conv_act = get_activation(act_fn)
+
+ else:
+ self.conv_norm_out = None
+ self.conv_act = None
+
+ conv_out_padding = (conv_out_kernel - 1) // 2
+ self.conv_out = nn.Conv2d(
+ block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+ )
+
+ self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)
+
+ def _check_config(
+ self,
+ down_block_types: Tuple[str],
+ up_block_types: Tuple[str],
+ only_cross_attention: Union[bool, Tuple[bool]],
+ block_out_channels: Tuple[int],
+ layers_per_block: Union[int, Tuple[int]],
+ cross_attention_dim: Union[int, Tuple[int]],
+ transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
+ reverse_transformer_layers_per_block: bool,
+ attention_head_dim: int,
+ num_attention_heads: Optional[Union[int, Tuple[int]]],
+ ):
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+ )
+ if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+ for layer_number_per_block in transformer_layers_per_block:
+ if isinstance(layer_number_per_block, list):
+ raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+ def _set_time_proj(
+ self,
+ time_embedding_type: str,
+ block_out_channels: int,
+ flip_sin_to_cos: bool,
+ freq_shift: float,
+ time_embedding_dim: int,
+ ) -> Tuple[int, int]:
+ if time_embedding_type == "fourier":
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+ if time_embed_dim % 2 != 0:
+ raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+ self.time_proj = GaussianFourierProjection(
+ time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+ )
+ timestep_input_dim = time_embed_dim
+ elif time_embedding_type == "positional":
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+ timestep_input_dim = block_out_channels[0]
+ else:
+ raise ValueError(
+ f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+ )
+
+ return time_embed_dim, timestep_input_dim
+
+ def _set_encoder_hid_proj(
+ self,
+ encoder_hid_dim_type: Optional[str],
+ cross_attention_dim: Union[int, Tuple[int]],
+ encoder_hid_dim: Optional[int],
+ ):
+ if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+ encoder_hid_dim_type = "text_proj"
+ self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+ logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+ if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+ )
+
+ if encoder_hid_dim_type == "text_proj":
+ self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+ elif encoder_hid_dim_type == "text_image_proj":
+ # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+ self.encoder_hid_proj = TextImageProjection(
+ text_embed_dim=encoder_hid_dim,
+ image_embed_dim=cross_attention_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+ elif encoder_hid_dim_type == "image_proj":
+ # Kandinsky 2.2
+ self.encoder_hid_proj = ImageProjection(
+ image_embed_dim=encoder_hid_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+ elif encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+ )
+ else:
+ self.encoder_hid_proj = None
+
+ def _set_class_embedding(
+ self,
+ class_embed_type: Optional[str],
+ act_fn: str,
+ num_class_embeds: Optional[int],
+ projection_class_embeddings_input_dim: Optional[int],
+ time_embed_dim: int,
+ timestep_input_dim: int,
+ ):
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ elif class_embed_type == "projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+ # 2. it projects from an arbitrary input dimension.
+ #
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+ self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ elif class_embed_type == "simple_projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+ else:
+ self.class_embedding = None
+
+ def _set_add_embedding(
+ self,
+ addition_embed_type: str,
+ addition_embed_type_num_heads: int,
+ addition_time_embed_dim: Optional[int],
+ flip_sin_to_cos: bool,
+ freq_shift: float,
+ cross_attention_dim: Optional[int],
+ encoder_hid_dim: Optional[int],
+ projection_class_embeddings_input_dim: Optional[int],
+ time_embed_dim: int,
+ ):
+ if addition_embed_type == "text":
+ if encoder_hid_dim is not None:
+ text_time_embedding_from_dim = encoder_hid_dim
+ else:
+ text_time_embedding_from_dim = cross_attention_dim
+
+ self.add_embedding = TextTimeEmbedding(
+ text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+ )
+ elif addition_embed_type == "text_image":
+ # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+ self.add_embedding = TextImageTimeEmbedding(
+ text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+ )
+ elif addition_embed_type == "text_time":
+ self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+ self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ elif addition_embed_type == "image":
+ # Kandinsky 2.2
+ self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+ elif addition_embed_type == "image_hint":
+ # Kandinsky 2.2 ControlNet
+ self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+ elif addition_embed_type is not None:
+ raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+ def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
+ if attention_type in ["gated", "gated-text-image"]:
+ positive_len = 768
+ if isinstance(cross_attention_dim, int):
+ positive_len = cross_attention_dim
+ elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
+ positive_len = cross_attention_dim[0]
+
+ feature_type = "text-only" if attention_type == "gated" else "text-image"
+ self.position_net = GLIGENTextBoundingboxProjection(
+ positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+ )
+
+ @property
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ def set_attention_slice(self, slice_size: Union[str, int, List[int]] = "auto"):
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if hasattr(module, "gradient_checkpointing"):
+ module.gradient_checkpointing = value
+
+ def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+ r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+ The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+ Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+ are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+ Args:
+ s1 (`float`):
+ Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ s2 (`float`):
+ Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+ """
+ for i, upsample_block in enumerate(self.up_blocks):
+ setattr(upsample_block, "s1", s1)
+ setattr(upsample_block, "s2", s2)
+ setattr(upsample_block, "b1", b1)
+ setattr(upsample_block, "b2", b2)
+
+ def disable_freeu(self):
+ """Disables the FreeU mechanism."""
+ freeu_keys = {"s1", "s2", "b1", "b2"}
+ for i, upsample_block in enumerate(self.up_blocks):
+ for k in freeu_keys:
+ if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+ setattr(upsample_block, k, None)
+
+ def fuse_qkv_projections(self):
+ """
+ Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is 🧪 experimental.
+
+
+ """
+ self.original_attn_processors = None
+
+ for _, attn_processor in self.attn_processors.items():
+ if "Added" in str(attn_processor.__class__.__name__):
+ raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+ self.original_attn_processors = self.attn_processors
+
+ for module in self.modules():
+ if isinstance(module, Attention):
+ module.fuse_projections(fuse=True)
+
+ def unfuse_qkv_projections(self):
+ """Disables the fused QKV projection if enabled.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ """
+ if self.original_attn_processors is not None:
+ self.set_attn_processor(self.original_attn_processors)
+
+ def unload_lora(self):
+ """Unloads LoRA weights."""
+ deprecate(
+ "unload_lora",
+ "0.28.0",
+ "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().",
+ )
+ for module in self.modules():
+ if hasattr(module, "set_lora_layer"):
+ module.set_lora_layer(None)
+
+ def get_time_embed(
+ self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int]
+ ) -> Optional[torch.Tensor]:
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+ return t_emb
+
+ def get_class_embed(self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
+ class_emb = None
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # there might be better ways to encapsulate this.
+ class_labels = class_labels.to(dtype=sample.dtype)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+ return class_emb
+
+ def get_aug_embed(
+ self, emb: torch.Tensor, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+ ) -> Optional[torch.Tensor]:
+ aug_emb = None
+ if self.config.addition_embed_type == "text":
+ aug_emb = self.add_embedding(encoder_hidden_states)
+ elif self.config.addition_embed_type == "text_image":
+ # Kandinsky 2.1 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+ )
+
+ image_embs = added_cond_kwargs.get("image_embeds")
+ text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+ aug_emb = self.add_embedding(text_embs, image_embs)
+ elif self.config.addition_embed_type == "text_time":
+ # SDXL - style
+ if "text_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+ )
+ text_embeds = added_cond_kwargs.get("text_embeds")
+ if "time_ids" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+ )
+ time_ids = added_cond_kwargs.get("time_ids")
+ time_embeds = self.add_time_proj(time_ids.flatten())
+ time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+ add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+ add_embeds = add_embeds.to(emb.dtype)
+ aug_emb = self.add_embedding(add_embeds)
+ elif self.config.addition_embed_type == "image":
+ # Kandinsky 2.2 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+ )
+ image_embs = added_cond_kwargs.get("image_embeds")
+ aug_emb = self.add_embedding(image_embs)
+ elif self.config.addition_embed_type == "image_hint":
+ # Kandinsky 2.2 - style
+ if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+ )
+ image_embs = added_cond_kwargs.get("image_embeds")
+ hint = added_cond_kwargs.get("hint")
+ aug_emb = self.add_embedding(image_embs, hint)
+ return aug_emb
+
+ def process_encoder_hidden_states(
+ self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+ ) -> torch.Tensor:
+ if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+ encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+ elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+ # Kadinsky 2.1 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+ elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+ # Kandinsky 2.2 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+ elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ image_embeds = self.encoder_hid_proj(image_embeds)
+ encoder_hidden_states = (encoder_hidden_states, image_embeds)
+ return encoder_hidden_states
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+ mid_block_additional_residual: Optional[torch.Tensor] = None,
+ down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+ encoder_attention_mask: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ down_block_add_samples: Optional[Tuple[torch.Tensor]] = None,
+ mid_block_add_sample: Optional[Tuple[torch.Tensor]] = None,
+ up_block_add_samples: Optional[Tuple[torch.Tensor]] = None,
+ ) -> Union[UNet2DConditionOutput, Tuple]:
+ r"""
+ The [`UNet2DConditionModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, channel, height, width)`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.FloatTensor`):
+ The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+ Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+ through the `self.time_embedding` layer to obtain the timestep embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ added_cond_kwargs: (`dict`, *optional*):
+ A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+ are passed along to the UNet blocks.
+ down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+ A tuple of tensors that if specified are added to the residuals of down unet blocks.
+ mid_block_additional_residual: (`torch.Tensor`, *optional*):
+ A tensor that if specified is added to the residual of the middle unet block.
+ encoder_attention_mask (`torch.Tensor`):
+ A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+ `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+ which adds large negative values to the attention scores corresponding to "discard" tokens.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+ tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+ added_cond_kwargs: (`dict`, *optional*):
+ A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+ are passed along to the UNet blocks.
+ down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+ additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+ example from ControlNet side model(s)
+ mid_block_additional_residual (`torch.Tensor`, *optional*):
+ additional residual to be added to UNet mid block output, for example from ControlNet side model
+ down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+ additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+ Returns:
+ [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+ a `tuple` is returned where the first element is the sample tensor.
+ """
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
+ # on the fly if necessary.
+ default_overall_up_factor = 2**self.num_upsamplers
+
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+ forward_upsample_size = False
+ upsample_size = None
+
+ for dim in sample.shape[-2:]:
+ if dim % default_overall_up_factor != 0:
+ # Forward upsample size to force interpolation output size.
+ forward_upsample_size = True
+ break
+
+ # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+ # expects mask of shape:
+ # [batch, key_tokens]
+ # adds singleton query_tokens dimension:
+ # [batch, 1, key_tokens]
+ # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+ # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+ # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+ if attention_mask is not None:
+ # assume that mask is expressed as:
+ # (1 = keep, 0 = discard)
+ # convert mask into a bias that can be added to attention scores:
+ # (keep = +0, discard = -10000.0)
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # convert encoder_attention_mask to a bias the same way we do for attention_mask
+ if encoder_attention_mask is not None:
+ encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+ encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+ # 0. center input if necessary
+ if self.config.center_input_sample:
+ sample = 2 * sample - 1.0
+
+ # 1. time
+ t_emb = self.get_time_embed(sample=sample, timestep=timestep)
+ emb = self.time_embedding(t_emb, timestep_cond)
+ aug_emb = None
+
+ class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
+ if class_emb is not None:
+ if self.config.class_embeddings_concat:
+ emb = torch.cat([emb, class_emb], dim=-1)
+ else:
+ emb = emb + class_emb
+
+ aug_emb = self.get_aug_embed(
+ emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+ )
+ if self.config.addition_embed_type == "image_hint":
+ aug_emb, hint = aug_emb
+ sample = torch.cat([sample, hint], dim=1)
+
+ emb = emb + aug_emb if aug_emb is not None else emb
+
+ if self.time_embed_act is not None:
+ emb = self.time_embed_act(emb)
+
+ encoder_hidden_states = self.process_encoder_hidden_states(
+ encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+ )
+
+ # 2. pre-process
+ sample = self.conv_in(sample)
+
+ # 2.5 GLIGEN position net
+ if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+ cross_attention_kwargs = cross_attention_kwargs.copy()
+ gligen_args = cross_attention_kwargs.pop("gligen")
+ cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+ # 3. down
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ if USE_PEFT_BACKEND:
+ # weight the lora layers by setting `lora_scale` for each PEFT layer
+ scale_lora_layers(self, lora_scale)
+
+ is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+ # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+ is_adapter = down_intrablock_additional_residuals is not None
+ # maintain backward compatibility for legacy usage, where
+ # T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+ # but can only use one or the other
+ is_brushnet = (
+ down_block_add_samples is not None
+ and mid_block_add_sample is not None
+ and up_block_add_samples is not None
+ )
+ if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+ deprecate(
+ "T2I should not use down_block_additional_residuals",
+ "1.3.0",
+ "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+ and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \
+ for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+ standard_warn=False,
+ )
+ down_intrablock_additional_residuals = down_block_additional_residuals
+ is_adapter = True
+
+ down_block_res_samples = (sample,)
+
+ if is_brushnet:
+ sample = sample + down_block_add_samples.pop(0)
+
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ # For t2i-adapter CrossAttnDownBlock2D
+ additional_residuals = {}
+ if is_adapter and len(down_intrablock_additional_residuals) > 0:
+ additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+ if is_brushnet and len(down_block_add_samples) > 0:
+ additional_residuals["down_block_add_samples"] = [
+ down_block_add_samples.pop(0)
+ for _ in range(len(downsample_block.resnets) + (downsample_block.downsamplers != None))
+ ]
+
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ **additional_residuals,
+ )
+ else:
+ additional_residuals = {}
+ if is_brushnet and len(down_block_add_samples) > 0:
+ additional_residuals["down_block_add_samples"] = [
+ down_block_add_samples.pop(0)
+ for _ in range(len(downsample_block.resnets) + (downsample_block.downsamplers != None))
+ ]
+
+ sample, res_samples = downsample_block(
+ hidden_states=sample, temb=emb, scale=lora_scale, **additional_residuals
+ )
+ if is_adapter and len(down_intrablock_additional_residuals) > 0:
+ sample += down_intrablock_additional_residuals.pop(0)
+
+ down_block_res_samples += res_samples
+
+ if is_controlnet:
+ new_down_block_res_samples = ()
+
+ for down_block_res_sample, down_block_additional_residual in zip(
+ down_block_res_samples, down_block_additional_residuals
+ ):
+ down_block_res_sample = down_block_res_sample + down_block_additional_residual
+ new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+ down_block_res_samples = new_down_block_res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+ else:
+ sample = self.mid_block(sample, emb)
+
+ # To support T2I-Adapter-XL
+ if (
+ is_adapter
+ and len(down_intrablock_additional_residuals) > 0
+ and sample.shape == down_intrablock_additional_residuals[0].shape
+ ):
+ sample += down_intrablock_additional_residuals.pop(0)
+
+ if is_controlnet:
+ sample = sample + mid_block_additional_residual
+
+ if is_brushnet:
+ sample = sample + mid_block_add_sample
+
+ # 5. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block and forward_upsample_size:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ additional_residuals = {}
+ if is_brushnet and len(up_block_add_samples) > 0:
+ additional_residuals["up_block_add_samples"] = [
+ up_block_add_samples.pop(0)
+ for _ in range(len(upsample_block.resnets) + (upsample_block.upsamplers != None))
+ ]
+
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ **additional_residuals,
+ )
+ else:
+ additional_residuals = {}
+ if is_brushnet and len(up_block_add_samples) > 0:
+ additional_residuals["up_block_add_samples"] = [
+ up_block_add_samples.pop(0)
+ for _ in range(len(upsample_block.resnets) + (upsample_block.upsamplers != None))
+ ]
+
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ scale=lora_scale,
+ **additional_residuals,
+ )
+
+ # 6. post-process
+ if self.conv_norm_out:
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ if USE_PEFT_BACKEND:
+ # remove `lora_scale` from each PEFT layer
+ unscale_lora_layers(self, lora_scale)
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet2DConditionOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/unet_2d_condition_flax.py b/model/diffusers_c/models/unets/unet_2d_condition_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..a5ec2875ca0ea140678616bdc4fc021fc7e90b50
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_2d_condition_flax.py
@@ -0,0 +1,453 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Dict, Optional, Tuple, Union
+
+import flax
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict
+
+from ...configuration_utils import ConfigMixin, flax_register_to_config
+from ...utils import BaseOutput
+from ..embeddings_flax import FlaxTimestepEmbedding, FlaxTimesteps
+from ..modeling_flax_utils import FlaxModelMixin
+from .unet_2d_blocks_flax import (
+ FlaxCrossAttnDownBlock2D,
+ FlaxCrossAttnUpBlock2D,
+ FlaxDownBlock2D,
+ FlaxUNetMidBlock2DCrossAttn,
+ FlaxUpBlock2D,
+)
+
+
+@flax.struct.dataclass
+class FlaxUNet2DConditionOutput(BaseOutput):
+ """
+ The output of [`FlaxUNet2DConditionModel`].
+
+ Args:
+ sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)`):
+ The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+ """
+
+ sample: jnp.ndarray
+
+
+@flax_register_to_config
+class FlaxUNet2DConditionModel(nn.Module, FlaxModelMixin, ConfigMixin):
+ r"""
+ A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+ shaped output.
+
+ This model inherits from [`FlaxModelMixin`]. Check the superclass documentation for it's generic methods
+ implemented for all models (such as downloading or saving).
+
+ This model is also a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
+ subclass. Use it as a regular Flax Linen module and refer to the Flax documentation for all matters related to its
+ general usage and behavior.
+
+ Inherent JAX features such as the following are supported:
+ - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+ - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+ - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+ - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+ Parameters:
+ sample_size (`int`, *optional*):
+ The size of the input sample.
+ in_channels (`int`, *optional*, defaults to 4):
+ The number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 4):
+ The number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("FlaxCrossAttnDownBlock2D", "FlaxCrossAttnDownBlock2D", "FlaxCrossAttnDownBlock2D", "FlaxDownBlock2D")`):
+ The tuple of downsample blocks to use.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("FlaxUpBlock2D", "FlaxCrossAttnUpBlock2D", "FlaxCrossAttnUpBlock2D", "FlaxCrossAttnUpBlock2D")`):
+ The tuple of upsample blocks to use.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+ Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`. If `None`, the mid block layer is skipped.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2):
+ The number of layers per block.
+ attention_head_dim (`int` or `Tuple[int]`, *optional*, defaults to 8):
+ The dimension of the attention heads.
+ num_attention_heads (`int` or `Tuple[int]`, *optional*):
+ The number of attention heads.
+ cross_attention_dim (`int`, *optional*, defaults to 768):
+ The dimension of the cross attention features.
+ dropout (`float`, *optional*, defaults to 0):
+ Dropout probability for down, up and bottleneck blocks.
+ flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+ use_memory_efficient_attention (`bool`, *optional*, defaults to `False`):
+ Enable memory efficient attention as described [here](https://arxiv.org/abs/2112.05682).
+ split_head_dim (`bool`, *optional*, defaults to `False`):
+ Whether to split the head dimension into a new axis for the self-attention computation. In most cases,
+ enabling this flag should speed up the computation for Stable Diffusion 2.x and Stable Diffusion XL.
+ """
+
+ sample_size: int = 32
+ in_channels: int = 4
+ out_channels: int = 4
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ )
+ up_block_types: Tuple[str, ...] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn"
+ only_cross_attention: Union[bool, Tuple[bool]] = False
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280)
+ layers_per_block: int = 2
+ attention_head_dim: Union[int, Tuple[int, ...]] = 8
+ num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None
+ cross_attention_dim: int = 1280
+ dropout: float = 0.0
+ use_linear_projection: bool = False
+ dtype: jnp.dtype = jnp.float32
+ flip_sin_to_cos: bool = True
+ freq_shift: int = 0
+ use_memory_efficient_attention: bool = False
+ split_head_dim: bool = False
+ transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1
+ addition_embed_type: Optional[str] = None
+ addition_time_embed_dim: Optional[int] = None
+ addition_embed_type_num_heads: int = 64
+ projection_class_embeddings_input_dim: Optional[int] = None
+
+ def init_weights(self, rng: jax.Array) -> FrozenDict:
+ # init input tensors
+ sample_shape = (1, self.in_channels, self.sample_size, self.sample_size)
+ sample = jnp.zeros(sample_shape, dtype=jnp.float32)
+ timesteps = jnp.ones((1,), dtype=jnp.int32)
+ encoder_hidden_states = jnp.zeros((1, 1, self.cross_attention_dim), dtype=jnp.float32)
+
+ params_rng, dropout_rng = jax.random.split(rng)
+ rngs = {"params": params_rng, "dropout": dropout_rng}
+
+ added_cond_kwargs = None
+ if self.addition_embed_type == "text_time":
+ # we retrieve the expected `text_embeds_dim` by first checking if the architecture is a refiner
+ # or non-refiner architecture and then by "reverse-computing" from `projection_class_embeddings_input_dim`
+ is_refiner = (
+ 5 * self.config.addition_time_embed_dim + self.config.cross_attention_dim
+ == self.config.projection_class_embeddings_input_dim
+ )
+ num_micro_conditions = 5 if is_refiner else 6
+
+ text_embeds_dim = self.config.projection_class_embeddings_input_dim - (
+ num_micro_conditions * self.config.addition_time_embed_dim
+ )
+
+ time_ids_channels = self.projection_class_embeddings_input_dim - text_embeds_dim
+ time_ids_dims = time_ids_channels // self.addition_time_embed_dim
+ added_cond_kwargs = {
+ "text_embeds": jnp.zeros((1, text_embeds_dim), dtype=jnp.float32),
+ "time_ids": jnp.zeros((1, time_ids_dims), dtype=jnp.float32),
+ }
+ return self.init(rngs, sample, timesteps, encoder_hidden_states, added_cond_kwargs)["params"]
+
+ def setup(self) -> None:
+ block_out_channels = self.block_out_channels
+ time_embed_dim = block_out_channels[0] * 4
+
+ if self.num_attention_heads is not None:
+ raise ValueError(
+ "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+ )
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = self.num_attention_heads or self.attention_head_dim
+
+ # input
+ self.conv_in = nn.Conv(
+ block_out_channels[0],
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ # time
+ self.time_proj = FlaxTimesteps(
+ block_out_channels[0], flip_sin_to_cos=self.flip_sin_to_cos, freq_shift=self.config.freq_shift
+ )
+ self.time_embedding = FlaxTimestepEmbedding(time_embed_dim, dtype=self.dtype)
+
+ only_cross_attention = self.only_cross_attention
+ if isinstance(only_cross_attention, bool):
+ only_cross_attention = (only_cross_attention,) * len(self.down_block_types)
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(self.down_block_types)
+
+ # transformer layers per block
+ transformer_layers_per_block = self.transformer_layers_per_block
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(self.down_block_types)
+
+ # addition embed types
+ if self.addition_embed_type is None:
+ self.add_embedding = None
+ elif self.addition_embed_type == "text_time":
+ if self.addition_time_embed_dim is None:
+ raise ValueError(
+ f"addition_embed_type {self.addition_embed_type} requires `addition_time_embed_dim` to not be None"
+ )
+ self.add_time_proj = FlaxTimesteps(self.addition_time_embed_dim, self.flip_sin_to_cos, self.freq_shift)
+ self.add_embedding = FlaxTimestepEmbedding(time_embed_dim, dtype=self.dtype)
+ else:
+ raise ValueError(f"addition_embed_type: {self.addition_embed_type} must be None or `text_time`.")
+
+ # down
+ down_blocks = []
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(self.down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ if down_block_type == "CrossAttnDownBlock2D":
+ down_block = FlaxCrossAttnDownBlock2D(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ dropout=self.dropout,
+ num_layers=self.layers_per_block,
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ num_attention_heads=num_attention_heads[i],
+ add_downsample=not is_final_block,
+ use_linear_projection=self.use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ use_memory_efficient_attention=self.use_memory_efficient_attention,
+ split_head_dim=self.split_head_dim,
+ dtype=self.dtype,
+ )
+ else:
+ down_block = FlaxDownBlock2D(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ dropout=self.dropout,
+ num_layers=self.layers_per_block,
+ add_downsample=not is_final_block,
+ dtype=self.dtype,
+ )
+
+ down_blocks.append(down_block)
+ self.down_blocks = down_blocks
+
+ # mid
+ if self.config.mid_block_type == "UNetMidBlock2DCrossAttn":
+ self.mid_block = FlaxUNetMidBlock2DCrossAttn(
+ in_channels=block_out_channels[-1],
+ dropout=self.dropout,
+ num_attention_heads=num_attention_heads[-1],
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ use_linear_projection=self.use_linear_projection,
+ use_memory_efficient_attention=self.use_memory_efficient_attention,
+ split_head_dim=self.split_head_dim,
+ dtype=self.dtype,
+ )
+ elif self.config.mid_block_type is None:
+ self.mid_block = None
+ else:
+ raise ValueError(f"Unexpected mid_block_type {self.config.mid_block_type}")
+
+ # up
+ up_blocks = []
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ only_cross_attention = list(reversed(only_cross_attention))
+ output_channel = reversed_block_out_channels[0]
+ reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+ for i, up_block_type in enumerate(self.up_block_types):
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ is_final_block = i == len(block_out_channels) - 1
+
+ if up_block_type == "CrossAttnUpBlock2D":
+ up_block = FlaxCrossAttnUpBlock2D(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ num_layers=self.layers_per_block + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ num_attention_heads=reversed_num_attention_heads[i],
+ add_upsample=not is_final_block,
+ dropout=self.dropout,
+ use_linear_projection=self.use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ use_memory_efficient_attention=self.use_memory_efficient_attention,
+ split_head_dim=self.split_head_dim,
+ dtype=self.dtype,
+ )
+ else:
+ up_block = FlaxUpBlock2D(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ num_layers=self.layers_per_block + 1,
+ add_upsample=not is_final_block,
+ dropout=self.dropout,
+ dtype=self.dtype,
+ )
+
+ up_blocks.append(up_block)
+ prev_output_channel = output_channel
+ self.up_blocks = up_blocks
+
+ # out
+ self.conv_norm_out = nn.GroupNorm(num_groups=32, epsilon=1e-5)
+ self.conv_out = nn.Conv(
+ self.out_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ def __call__(
+ self,
+ sample: jnp.ndarray,
+ timesteps: Union[jnp.ndarray, float, int],
+ encoder_hidden_states: jnp.ndarray,
+ added_cond_kwargs: Optional[Union[Dict, FrozenDict]] = None,
+ down_block_additional_residuals: Optional[Tuple[jnp.ndarray, ...]] = None,
+ mid_block_additional_residual: Optional[jnp.ndarray] = None,
+ return_dict: bool = True,
+ train: bool = False,
+ ) -> Union[FlaxUNet2DConditionOutput, Tuple[jnp.ndarray]]:
+ r"""
+ Args:
+ sample (`jnp.ndarray`): (batch, channel, height, width) noisy inputs tensor
+ timestep (`jnp.ndarray` or `float` or `int`): timesteps
+ encoder_hidden_states (`jnp.ndarray`): (batch_size, sequence_length, hidden_size) encoder hidden states
+ added_cond_kwargs: (`dict`, *optional*):
+ A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+ are passed along to the UNet blocks.
+ down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+ A tuple of tensors that if specified are added to the residuals of down unet blocks.
+ mid_block_additional_residual: (`torch.Tensor`, *optional*):
+ A tensor that if specified is added to the residual of the middle unet block.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`models.unets.unet_2d_condition_flax.FlaxUNet2DConditionOutput`] instead of a
+ plain tuple.
+ train (`bool`, *optional*, defaults to `False`):
+ Use deterministic functions and disable dropout when not training.
+
+ Returns:
+ [`~models.unets.unet_2d_condition_flax.FlaxUNet2DConditionOutput`] or `tuple`:
+ [`~models.unets.unet_2d_condition_flax.FlaxUNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`.
+ When returning a tuple, the first element is the sample tensor.
+ """
+ # 1. time
+ if not isinstance(timesteps, jnp.ndarray):
+ timesteps = jnp.array([timesteps], dtype=jnp.int32)
+ elif isinstance(timesteps, jnp.ndarray) and len(timesteps.shape) == 0:
+ timesteps = timesteps.astype(dtype=jnp.float32)
+ timesteps = jnp.expand_dims(timesteps, 0)
+
+ t_emb = self.time_proj(timesteps)
+ t_emb = self.time_embedding(t_emb)
+
+ # additional embeddings
+ aug_emb = None
+ if self.addition_embed_type == "text_time":
+ if added_cond_kwargs is None:
+ raise ValueError(
+ f"Need to provide argument `added_cond_kwargs` for {self.__class__} when using `addition_embed_type={self.addition_embed_type}`"
+ )
+ text_embeds = added_cond_kwargs.get("text_embeds")
+ if text_embeds is None:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+ )
+ time_ids = added_cond_kwargs.get("time_ids")
+ if time_ids is None:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+ )
+ # compute time embeds
+ time_embeds = self.add_time_proj(jnp.ravel(time_ids)) # (1, 6) => (6,) => (6, 256)
+ time_embeds = jnp.reshape(time_embeds, (text_embeds.shape[0], -1))
+ add_embeds = jnp.concatenate([text_embeds, time_embeds], axis=-1)
+ aug_emb = self.add_embedding(add_embeds)
+
+ t_emb = t_emb + aug_emb if aug_emb is not None else t_emb
+
+ # 2. pre-process
+ sample = jnp.transpose(sample, (0, 2, 3, 1))
+ sample = self.conv_in(sample)
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for down_block in self.down_blocks:
+ if isinstance(down_block, FlaxCrossAttnDownBlock2D):
+ sample, res_samples = down_block(sample, t_emb, encoder_hidden_states, deterministic=not train)
+ else:
+ sample, res_samples = down_block(sample, t_emb, deterministic=not train)
+ down_block_res_samples += res_samples
+
+ if down_block_additional_residuals is not None:
+ new_down_block_res_samples = ()
+
+ for down_block_res_sample, down_block_additional_residual in zip(
+ down_block_res_samples, down_block_additional_residuals
+ ):
+ down_block_res_sample += down_block_additional_residual
+ new_down_block_res_samples += (down_block_res_sample,)
+
+ down_block_res_samples = new_down_block_res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ sample = self.mid_block(sample, t_emb, encoder_hidden_states, deterministic=not train)
+
+ if mid_block_additional_residual is not None:
+ sample += mid_block_additional_residual
+
+ # 5. up
+ for up_block in self.up_blocks:
+ res_samples = down_block_res_samples[-(self.layers_per_block + 1) :]
+ down_block_res_samples = down_block_res_samples[: -(self.layers_per_block + 1)]
+ if isinstance(up_block, FlaxCrossAttnUpBlock2D):
+ sample = up_block(
+ sample,
+ temb=t_emb,
+ encoder_hidden_states=encoder_hidden_states,
+ res_hidden_states_tuple=res_samples,
+ deterministic=not train,
+ )
+ else:
+ sample = up_block(sample, temb=t_emb, res_hidden_states_tuple=res_samples, deterministic=not train)
+
+ # 6. post-process
+ sample = self.conv_norm_out(sample)
+ sample = nn.silu(sample)
+ sample = self.conv_out(sample)
+ sample = jnp.transpose(sample, (0, 3, 1, 2))
+
+ if not return_dict:
+ return (sample,)
+
+ return FlaxUNet2DConditionOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/unet_3d_blocks.py b/model/diffusers_c/models/unets/unet_3d_blocks.py
new file mode 100644
index 0000000000000000000000000000000000000000..e1ada1021b3a425e810e5d504bbab8f47e8750d5
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_3d_blocks.py
@@ -0,0 +1,2384 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ...utils import is_torch_version
+from ...utils.torch_utils import apply_freeu
+from ..attention import Attention
+from ..resnet import (
+ Downsample2D,
+ ResnetBlock2D,
+ SpatioTemporalResBlock,
+ TemporalConvLayer,
+ Upsample2D,
+)
+from ..transformers.dual_transformer_2d import DualTransformer2DModel
+from ..transformers.transformer_2d import Transformer2DModel
+from ..transformers.transformer_temporal import (
+ TransformerSpatioTemporalModel,
+ TransformerTemporalModel,
+)
+
+
+def get_down_block(
+ down_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ add_downsample: bool,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ num_attention_heads: int,
+ resnet_groups: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ downsample_padding: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = True,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ temporal_num_attention_heads: int = 8,
+ temporal_max_seq_length: int = 32,
+ transformer_layers_per_block: int = 1,
+) -> Union[
+ "DownBlock3D",
+ "CrossAttnDownBlock3D",
+ "DownBlockMotion",
+ "CrossAttnDownBlockMotion",
+ "DownBlockSpatioTemporal",
+ "CrossAttnDownBlockSpatioTemporal",
+]:
+ if down_block_type == "DownBlock3D":
+ return DownBlock3D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "CrossAttnDownBlock3D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock3D")
+ return CrossAttnDownBlock3D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ if down_block_type == "DownBlockMotion":
+ return DownBlockMotion(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ temporal_num_attention_heads=temporal_num_attention_heads,
+ temporal_max_seq_length=temporal_max_seq_length,
+ )
+ elif down_block_type == "CrossAttnDownBlockMotion":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockMotion")
+ return CrossAttnDownBlockMotion(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ temporal_num_attention_heads=temporal_num_attention_heads,
+ temporal_max_seq_length=temporal_max_seq_length,
+ )
+ elif down_block_type == "DownBlockSpatioTemporal":
+ # added for SDV
+ return DownBlockSpatioTemporal(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ )
+ elif down_block_type == "CrossAttnDownBlockSpatioTemporal":
+ # added for SDV
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockSpatioTemporal")
+ return CrossAttnDownBlockSpatioTemporal(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ num_layers=num_layers,
+ transformer_layers_per_block=transformer_layers_per_block,
+ add_downsample=add_downsample,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ )
+
+ raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+ up_block_type: str,
+ num_layers: int,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ add_upsample: bool,
+ resnet_eps: float,
+ resnet_act_fn: str,
+ num_attention_heads: int,
+ resolution_idx: Optional[int] = None,
+ resnet_groups: Optional[int] = None,
+ cross_attention_dim: Optional[int] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = True,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ temporal_num_attention_heads: int = 8,
+ temporal_cross_attention_dim: Optional[int] = None,
+ temporal_max_seq_length: int = 32,
+ transformer_layers_per_block: int = 1,
+ dropout: float = 0.0,
+) -> Union[
+ "UpBlock3D",
+ "CrossAttnUpBlock3D",
+ "UpBlockMotion",
+ "CrossAttnUpBlockMotion",
+ "UpBlockSpatioTemporal",
+ "CrossAttnUpBlockSpatioTemporal",
+]:
+ if up_block_type == "UpBlock3D":
+ return UpBlock3D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resolution_idx=resolution_idx,
+ )
+ elif up_block_type == "CrossAttnUpBlock3D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock3D")
+ return CrossAttnUpBlock3D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resolution_idx=resolution_idx,
+ )
+ if up_block_type == "UpBlockMotion":
+ return UpBlockMotion(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resolution_idx=resolution_idx,
+ temporal_num_attention_heads=temporal_num_attention_heads,
+ temporal_max_seq_length=temporal_max_seq_length,
+ )
+ elif up_block_type == "CrossAttnUpBlockMotion":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockMotion")
+ return CrossAttnUpBlockMotion(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resolution_idx=resolution_idx,
+ temporal_num_attention_heads=temporal_num_attention_heads,
+ temporal_max_seq_length=temporal_max_seq_length,
+ )
+ elif up_block_type == "UpBlockSpatioTemporal":
+ # added for SDV
+ return UpBlockSpatioTemporal(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ resolution_idx=resolution_idx,
+ add_upsample=add_upsample,
+ )
+ elif up_block_type == "CrossAttnUpBlockSpatioTemporal":
+ # added for SDV
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockSpatioTemporal")
+ return CrossAttnUpBlockSpatioTemporal(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ num_layers=num_layers,
+ transformer_layers_per_block=transformer_layers_per_block,
+ add_upsample=add_upsample,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ resolution_idx=resolution_idx,
+ )
+
+ raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlock3DCrossAttn(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ output_scale_factor: float = 1.0,
+ cross_attention_dim: int = 1280,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = True,
+ upcast_attention: bool = False,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ temp_convs = [
+ TemporalConvLayer(
+ in_channels,
+ in_channels,
+ dropout=0.1,
+ norm_num_groups=resnet_groups,
+ )
+ ]
+ attentions = []
+ temp_attentions = []
+
+ for _ in range(num_layers):
+ attentions.append(
+ Transformer2DModel(
+ in_channels // num_attention_heads,
+ num_attention_heads,
+ in_channels=in_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ )
+ )
+ temp_attentions.append(
+ TransformerTemporalModel(
+ in_channels // num_attention_heads,
+ num_attention_heads,
+ in_channels=in_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ temp_convs.append(
+ TemporalConvLayer(
+ in_channels,
+ in_channels,
+ dropout=0.1,
+ norm_num_groups=resnet_groups,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.temp_convs = nn.ModuleList(temp_convs)
+ self.attentions = nn.ModuleList(attentions)
+ self.temp_attentions = nn.ModuleList(temp_attentions)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ num_frames: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ ) -> torch.FloatTensor:
+ hidden_states = self.resnets[0](hidden_states, temb)
+ hidden_states = self.temp_convs[0](hidden_states, num_frames=num_frames)
+ for attn, temp_attn, resnet, temp_conv in zip(
+ self.attentions, self.temp_attentions, self.resnets[1:], self.temp_convs[1:]
+ ):
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+ hidden_states = temp_attn(
+ hidden_states,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+ hidden_states = resnet(hidden_states, temb)
+ hidden_states = temp_conv(hidden_states, num_frames=num_frames)
+
+ return hidden_states
+
+
+class CrossAttnDownBlock3D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ downsample_padding: int = 1,
+ add_downsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+ temp_attentions = []
+ temp_convs = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ temp_convs.append(
+ TemporalConvLayer(
+ out_channels,
+ out_channels,
+ dropout=0.1,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ attentions.append(
+ Transformer2DModel(
+ out_channels // num_attention_heads,
+ num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ )
+ )
+ temp_attentions.append(
+ TransformerTemporalModel(
+ out_channels // num_attention_heads,
+ num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ self.resnets = nn.ModuleList(resnets)
+ self.temp_convs = nn.ModuleList(temp_convs)
+ self.attentions = nn.ModuleList(attentions)
+ self.temp_attentions = nn.ModuleList(temp_attentions)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels,
+ use_conv=True,
+ out_channels=out_channels,
+ padding=downsample_padding,
+ name="op",
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ num_frames: int = 1,
+ cross_attention_kwargs: Dict[str, Any] = None,
+ ) -> Union[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ # TODO(Patrick, William) - attention mask is not used
+ output_states = ()
+
+ for resnet, temp_conv, attn, temp_attn in zip(
+ self.resnets, self.temp_convs, self.attentions, self.temp_attentions
+ ):
+ hidden_states = resnet(hidden_states, temb)
+ hidden_states = temp_conv(hidden_states, num_frames=num_frames)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+ hidden_states = temp_attn(
+ hidden_states,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class DownBlock3D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ resnets = []
+ temp_convs = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ temp_convs.append(
+ TemporalConvLayer(
+ out_channels,
+ out_channels,
+ dropout=0.1,
+ norm_num_groups=resnet_groups,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.temp_convs = nn.ModuleList(temp_convs)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels,
+ use_conv=True,
+ out_channels=out_channels,
+ padding=downsample_padding,
+ name="op",
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ num_frames: int = 1,
+ ) -> Union[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ for resnet, temp_conv in zip(self.resnets, self.temp_convs):
+ hidden_states = resnet(hidden_states, temb)
+ hidden_states = temp_conv(hidden_states, num_frames=num_frames)
+
+ output_states += (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ output_states += (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class CrossAttnUpBlock3D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ resolution_idx: Optional[int] = None,
+ ):
+ super().__init__()
+ resnets = []
+ temp_convs = []
+ attentions = []
+ temp_attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ temp_convs.append(
+ TemporalConvLayer(
+ out_channels,
+ out_channels,
+ dropout=0.1,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ attentions.append(
+ Transformer2DModel(
+ out_channels // num_attention_heads,
+ num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ )
+ )
+ temp_attentions.append(
+ TransformerTemporalModel(
+ out_channels // num_attention_heads,
+ num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ self.resnets = nn.ModuleList(resnets)
+ self.temp_convs = nn.ModuleList(temp_convs)
+ self.attentions = nn.ModuleList(attentions)
+ self.temp_attentions = nn.ModuleList(temp_attentions)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ num_frames: int = 1,
+ cross_attention_kwargs: Dict[str, Any] = None,
+ ) -> torch.FloatTensor:
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+
+ # TODO(Patrick, William) - attention mask is not used
+ for resnet, temp_conv, attn, temp_attn in zip(
+ self.resnets, self.temp_convs, self.attentions, self.temp_attentions
+ ):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ hidden_states = resnet(hidden_states, temb)
+ hidden_states = temp_conv(hidden_states, num_frames=num_frames)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+ hidden_states = temp_attn(
+ hidden_states,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size)
+
+ return hidden_states
+
+
+class UpBlock3D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ resolution_idx: Optional[int] = None,
+ ):
+ super().__init__()
+ resnets = []
+ temp_convs = []
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ temp_convs.append(
+ TemporalConvLayer(
+ out_channels,
+ out_channels,
+ dropout=0.1,
+ norm_num_groups=resnet_groups,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.temp_convs = nn.ModuleList(temp_convs)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ num_frames: int = 1,
+ ) -> torch.FloatTensor:
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+ for resnet, temp_conv in zip(self.resnets, self.temp_convs):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ hidden_states = resnet(hidden_states, temb)
+ hidden_states = temp_conv(hidden_states, num_frames=num_frames)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size)
+
+ return hidden_states
+
+
+class DownBlockMotion(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ temporal_num_attention_heads: int = 1,
+ temporal_cross_attention_dim: Optional[int] = None,
+ temporal_max_seq_length: int = 32,
+ ):
+ super().__init__()
+ resnets = []
+ motion_modules = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ motion_modules.append(
+ TransformerTemporalModel(
+ num_attention_heads=temporal_num_attention_heads,
+ in_channels=out_channels,
+ norm_num_groups=resnet_groups,
+ cross_attention_dim=temporal_cross_attention_dim,
+ attention_bias=False,
+ activation_fn="geglu",
+ positional_embeddings="sinusoidal",
+ num_positional_embeddings=temporal_max_seq_length,
+ attention_head_dim=out_channels // temporal_num_attention_heads,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.motion_modules = nn.ModuleList(motion_modules)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels,
+ use_conv=True,
+ out_channels=out_channels,
+ padding=downsample_padding,
+ name="op",
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ scale: float = 1.0,
+ num_frames: int = 1,
+ ) -> Union[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ blocks = zip(self.resnets, self.motion_modules)
+ for resnet, motion_module in blocks:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ use_reentrant=False,
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, scale
+ )
+
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+ hidden_states = motion_module(hidden_states, num_frames=num_frames)[0]
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale=scale)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class CrossAttnDownBlockMotion(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ downsample_padding: int = 1,
+ add_downsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ temporal_cross_attention_dim: Optional[int] = None,
+ temporal_num_attention_heads: int = 8,
+ temporal_max_seq_length: int = 32,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+ motion_modules = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+
+ motion_modules.append(
+ TransformerTemporalModel(
+ num_attention_heads=temporal_num_attention_heads,
+ in_channels=out_channels,
+ norm_num_groups=resnet_groups,
+ cross_attention_dim=temporal_cross_attention_dim,
+ attention_bias=False,
+ activation_fn="geglu",
+ positional_embeddings="sinusoidal",
+ num_positional_embeddings=temporal_max_seq_length,
+ attention_head_dim=out_channels // temporal_num_attention_heads,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+ self.motion_modules = nn.ModuleList(motion_modules)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels,
+ use_conv=True,
+ out_channels=out_channels,
+ padding=downsample_padding,
+ name="op",
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ num_frames: int = 1,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ additional_residuals: Optional[torch.FloatTensor] = None,
+ ):
+ output_states = ()
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+ blocks = list(zip(self.resnets, self.attentions, self.motion_modules))
+ for i, (resnet, attn, motion_module) in enumerate(blocks):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = motion_module(
+ hidden_states,
+ num_frames=num_frames,
+ )[0]
+
+ # apply additional residuals to the output of the last pair of resnet and attention blocks
+ if i == len(blocks) - 1 and additional_residuals is not None:
+ hidden_states = hidden_states + additional_residuals
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class CrossAttnUpBlockMotion(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ temporal_cross_attention_dim: Optional[int] = None,
+ temporal_num_attention_heads: int = 8,
+ temporal_max_seq_length: int = 32,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+ motion_modules = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ motion_modules.append(
+ TransformerTemporalModel(
+ num_attention_heads=temporal_num_attention_heads,
+ in_channels=out_channels,
+ norm_num_groups=resnet_groups,
+ cross_attention_dim=temporal_cross_attention_dim,
+ attention_bias=False,
+ activation_fn="geglu",
+ positional_embeddings="sinusoidal",
+ num_positional_embeddings=temporal_max_seq_length,
+ attention_head_dim=out_channels // temporal_num_attention_heads,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+ self.motion_modules = nn.ModuleList(motion_modules)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ num_frames: int = 1,
+ ) -> torch.FloatTensor:
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+
+ blocks = zip(self.resnets, self.attentions, self.motion_modules)
+ for resnet, attn, motion_module in blocks:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = motion_module(
+ hidden_states,
+ num_frames=num_frames,
+ )[0]
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
+
+ return hidden_states
+
+
+class UpBlockMotion(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ temporal_norm_num_groups: int = 32,
+ temporal_cross_attention_dim: Optional[int] = None,
+ temporal_num_attention_heads: int = 8,
+ temporal_max_seq_length: int = 32,
+ ):
+ super().__init__()
+ resnets = []
+ motion_modules = []
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ motion_modules.append(
+ TransformerTemporalModel(
+ num_attention_heads=temporal_num_attention_heads,
+ in_channels=out_channels,
+ norm_num_groups=temporal_norm_num_groups,
+ cross_attention_dim=temporal_cross_attention_dim,
+ attention_bias=False,
+ activation_fn="geglu",
+ positional_embeddings="sinusoidal",
+ num_positional_embeddings=temporal_max_seq_length,
+ attention_head_dim=out_channels // temporal_num_attention_heads,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+ self.motion_modules = nn.ModuleList(motion_modules)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size=None,
+ scale: float = 1.0,
+ num_frames: int = 1,
+ ) -> torch.FloatTensor:
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+
+ blocks = zip(self.resnets, self.motion_modules)
+
+ for resnet, motion_module in blocks:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ use_reentrant=False,
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+ hidden_states = motion_module(hidden_states, num_frames=num_frames)[0]
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+ return hidden_states
+
+
+class UNetMidBlockCrossAttnMotion(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ output_scale_factor: float = 1.0,
+ cross_attention_dim: int = 1280,
+ dual_cross_attention: float = False,
+ use_linear_projection: float = False,
+ upcast_attention: float = False,
+ attention_type: str = "default",
+ temporal_num_attention_heads: int = 1,
+ temporal_cross_attention_dim: Optional[int] = None,
+ temporal_max_seq_length: int = 32,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ attentions = []
+ motion_modules = []
+
+ for _ in range(num_layers):
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=transformer_layers_per_block,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ motion_modules.append(
+ TransformerTemporalModel(
+ num_attention_heads=temporal_num_attention_heads,
+ attention_head_dim=in_channels // temporal_num_attention_heads,
+ in_channels=in_channels,
+ norm_num_groups=resnet_groups,
+ cross_attention_dim=temporal_cross_attention_dim,
+ attention_bias=False,
+ positional_embeddings="sinusoidal",
+ num_positional_embeddings=temporal_max_seq_length,
+ activation_fn="geglu",
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+ self.motion_modules = nn.ModuleList(motion_modules)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ num_frames: int = 1,
+ ) -> torch.FloatTensor:
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+
+ blocks = zip(self.attentions, self.resnets[1:], self.motion_modules)
+ for attn, resnet, motion_module in blocks:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(motion_module),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ else:
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = motion_module(
+ hidden_states,
+ num_frames=num_frames,
+ )[0]
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+class MidBlockTemporalDecoder(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ attention_head_dim: int = 512,
+ num_layers: int = 1,
+ upcast_attention: bool = False,
+ ):
+ super().__init__()
+
+ resnets = []
+ attentions = []
+ for i in range(num_layers):
+ input_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ SpatioTemporalResBlock(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=1e-6,
+ temporal_eps=1e-5,
+ merge_factor=0.0,
+ merge_strategy="learned",
+ switch_spatial_to_temporal_mix=True,
+ )
+ )
+
+ attentions.append(
+ Attention(
+ query_dim=in_channels,
+ heads=in_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ eps=1e-6,
+ upcast_attention=upcast_attention,
+ norm_num_groups=32,
+ bias=True,
+ residual_connection=True,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ image_only_indicator: torch.FloatTensor,
+ ):
+ hidden_states = self.resnets[0](
+ hidden_states,
+ image_only_indicator=image_only_indicator,
+ )
+ for resnet, attn in zip(self.resnets[1:], self.attentions):
+ hidden_states = attn(hidden_states)
+ hidden_states = resnet(
+ hidden_states,
+ image_only_indicator=image_only_indicator,
+ )
+
+ return hidden_states
+
+
+class UpBlockTemporalDecoder(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ num_layers: int = 1,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+ for i in range(num_layers):
+ input_channels = in_channels if i == 0 else out_channels
+
+ resnets.append(
+ SpatioTemporalResBlock(
+ in_channels=input_channels,
+ out_channels=out_channels,
+ temb_channels=None,
+ eps=1e-6,
+ temporal_eps=1e-5,
+ merge_factor=0.0,
+ merge_strategy="learned",
+ switch_spatial_to_temporal_mix=True,
+ )
+ )
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ image_only_indicator: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ for resnet in self.resnets:
+ hidden_states = resnet(
+ hidden_states,
+ image_only_indicator=image_only_indicator,
+ )
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
+
+
+class UNetMidBlockSpatioTemporal(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ # support for variable transformer layers per block
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ # there is always at least one resnet
+ resnets = [
+ SpatioTemporalResBlock(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=1e-5,
+ )
+ ]
+ attentions = []
+
+ for i in range(num_layers):
+ attentions.append(
+ TransformerSpatioTemporalModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ )
+ )
+
+ resnets.append(
+ SpatioTemporalResBlock(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=1e-5,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ image_only_indicator: Optional[torch.Tensor] = None,
+ ) -> torch.FloatTensor:
+ hidden_states = self.resnets[0](
+ hidden_states,
+ temb,
+ image_only_indicator=image_only_indicator,
+ )
+
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ if self.training and self.gradient_checkpointing: # TODO
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ return_dict=False,
+ )[0]
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ image_only_indicator,
+ **ckpt_kwargs,
+ )
+ else:
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ return_dict=False,
+ )[0]
+ hidden_states = resnet(
+ hidden_states,
+ temb,
+ image_only_indicator=image_only_indicator,
+ )
+
+ return hidden_states
+
+
+class DownBlockSpatioTemporal(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ num_layers: int = 1,
+ add_downsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ SpatioTemporalResBlock(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=1e-5,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels,
+ use_conv=True,
+ out_channels=out_channels,
+ name="op",
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ image_only_indicator: Optional[torch.Tensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ image_only_indicator,
+ use_reentrant=False,
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ image_only_indicator,
+ )
+ else:
+ hidden_states = resnet(
+ hidden_states,
+ temb,
+ image_only_indicator=image_only_indicator,
+ )
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class CrossAttnDownBlockSpatioTemporal(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ add_downsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ SpatioTemporalResBlock(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=1e-6,
+ )
+ )
+ attentions.append(
+ TransformerSpatioTemporalModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels,
+ use_conv=True,
+ out_channels=out_channels,
+ padding=1,
+ name="op",
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ image_only_indicator: Optional[torch.Tensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ blocks = list(zip(self.resnets, self.attentions))
+ for resnet, attn in blocks:
+ if self.training and self.gradient_checkpointing: # TODO
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ image_only_indicator,
+ **ckpt_kwargs,
+ )
+
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(
+ hidden_states,
+ temb,
+ image_only_indicator=image_only_indicator,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ return_dict=False,
+ )[0]
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class UpBlockSpatioTemporal(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ SpatioTemporalResBlock(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ image_only_indicator: Optional[torch.Tensor] = None,
+ ) -> torch.FloatTensor:
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ image_only_indicator,
+ use_reentrant=False,
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ image_only_indicator,
+ )
+ else:
+ hidden_states = resnet(
+ hidden_states,
+ temb,
+ image_only_indicator=image_only_indicator,
+ )
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
+
+
+class CrossAttnUpBlockSpatioTemporal(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ SpatioTemporalResBlock(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ )
+ )
+ attentions.append(
+ TransformerSpatioTemporalModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ image_only_indicator: Optional[torch.Tensor] = None,
+ ) -> torch.FloatTensor:
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing: # TODO
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ image_only_indicator,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(
+ hidden_states,
+ temb,
+ image_only_indicator=image_only_indicator,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ return_dict=False,
+ )[0]
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states)
+
+ return hidden_states
diff --git a/model/diffusers_c/models/unets/unet_3d_condition.py b/model/diffusers_c/models/unets/unet_3d_condition.py
new file mode 100644
index 0000000000000000000000000000000000000000..b7641a96a7a1c16b52a5629fe1e29710cdea632d
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_3d_condition.py
@@ -0,0 +1,753 @@
+# Copyright 2024 Alibaba DAMO-VILAB and The HuggingFace Team. All rights reserved.
+# Copyright 2024 The ModelScope Team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import UNet2DConditionLoadersMixin
+from ...utils import BaseOutput, deprecate, logging
+from ..activations import get_activation
+from ..attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ Attention,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from ..transformers.transformer_temporal import TransformerTemporalModel
+from .unet_3d_blocks import (
+ CrossAttnDownBlock3D,
+ CrossAttnUpBlock3D,
+ DownBlock3D,
+ UNetMidBlock3DCrossAttn,
+ UpBlock3D,
+ get_down_block,
+ get_up_block,
+)
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet3DConditionOutput(BaseOutput):
+ """
+ The output of [`UNet3DConditionModel`].
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size, num_channels, num_frames, height, width)`):
+ The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+ """
+
+ sample: torch.FloatTensor
+
+
+class UNet3DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+ r"""
+ A conditional 3D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+ shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+ Height and width of input/output sample.
+ in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock3D", "CrossAttnDownBlock3D", "CrossAttnDownBlock3D", "DownBlock3D")`):
+ The tuple of downsample blocks to use.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock3D", "CrossAttnUpBlock3D", "CrossAttnUpBlock3D", "CrossAttnUpBlock3D")`):
+ The tuple of upsample blocks to use.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+ downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+ If `None`, normalization and activation layers is skipped in post-processing.
+ norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+ cross_attention_dim (`int`, *optional*, defaults to 1024): The dimension of the cross attention features.
+ attention_head_dim (`int`, *optional*, defaults to 64): The dimension of the attention heads.
+ num_attention_heads (`int`, *optional*): The number of attention heads.
+ """
+
+ _supports_gradient_checkpointing = False
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[int] = None,
+ in_channels: int = 4,
+ out_channels: int = 4,
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlock3D",
+ "CrossAttnDownBlock3D",
+ "CrossAttnDownBlock3D",
+ "DownBlock3D",
+ ),
+ up_block_types: Tuple[str, ...] = (
+ "UpBlock3D",
+ "CrossAttnUpBlock3D",
+ "CrossAttnUpBlock3D",
+ "CrossAttnUpBlock3D",
+ ),
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+ layers_per_block: int = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: int = 1024,
+ attention_head_dim: Union[int, Tuple[int]] = 64,
+ num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+ ):
+ super().__init__()
+
+ self.sample_size = sample_size
+
+ if num_attention_heads is not None:
+ raise NotImplementedError(
+ "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+ )
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ # input
+ conv_in_kernel = 3
+ conv_out_kernel = 3
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in = nn.Conv2d(
+ in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+ )
+
+ # time
+ time_embed_dim = block_out_channels[0] * 4
+ self.time_proj = Timesteps(block_out_channels[0], True, 0)
+ timestep_input_dim = block_out_channels[0]
+
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ )
+
+ self.transformer_in = TransformerTemporalModel(
+ num_attention_heads=8,
+ attention_head_dim=attention_head_dim,
+ in_channels=block_out_channels[0],
+ num_layers=1,
+ norm_num_groups=norm_num_groups,
+ )
+
+ # class embedding
+ self.down_blocks = nn.ModuleList([])
+ self.up_blocks = nn.ModuleList([])
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[i],
+ downsample_padding=downsample_padding,
+ dual_cross_attention=False,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ self.mid_block = UNetMidBlock3DCrossAttn(
+ in_channels=block_out_channels[-1],
+ temb_channels=time_embed_dim,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ dual_cross_attention=False,
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=layers_per_block + 1,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=reversed_num_attention_heads[i],
+ dual_cross_attention=False,
+ resolution_idx=i,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ if norm_num_groups is not None:
+ self.conv_norm_out = nn.GroupNorm(
+ num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+ )
+ self.conv_act = get_activation("silu")
+ else:
+ self.conv_norm_out = None
+ self.conv_act = None
+
+ conv_out_padding = (conv_out_kernel - 1) // 2
+ self.conv_out = nn.Conv2d(
+ block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+ )
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+ def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+ """
+ Sets the attention processor to use [feed forward
+ chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+ Parameters:
+ chunk_size (`int`, *optional*):
+ The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+ over each tensor of dim=`dim`.
+ dim (`int`, *optional*, defaults to `0`):
+ The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+ or dim=1 (sequence length).
+ """
+ if dim not in [0, 1]:
+ raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+ # By default chunk size is 1
+ chunk_size = chunk_size or 1
+
+ def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+ if hasattr(module, "set_chunk_feed_forward"):
+ module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+ for child in module.children():
+ fn_recursive_feed_forward(child, chunk_size, dim)
+
+ for module in self.children():
+ fn_recursive_feed_forward(module, chunk_size, dim)
+
+ def disable_forward_chunking(self):
+ def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+ if hasattr(module, "set_chunk_feed_forward"):
+ module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+ for child in module.children():
+ fn_recursive_feed_forward(child, chunk_size, dim)
+
+ for module in self.children():
+ fn_recursive_feed_forward(module, None, 0)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (CrossAttnDownBlock3D, DownBlock3D, CrossAttnUpBlock3D, UpBlock3D)):
+ module.gradient_checkpointing = value
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.enable_freeu
+ def enable_freeu(self, s1, s2, b1, b2):
+ r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+ The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+ Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+ are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+ Args:
+ s1 (`float`):
+ Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ s2 (`float`):
+ Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+ """
+ for i, upsample_block in enumerate(self.up_blocks):
+ setattr(upsample_block, "s1", s1)
+ setattr(upsample_block, "s2", s2)
+ setattr(upsample_block, "b1", b1)
+ setattr(upsample_block, "b2", b2)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.disable_freeu
+ def disable_freeu(self):
+ """Disables the FreeU mechanism."""
+ freeu_keys = {"s1", "s2", "b1", "b2"}
+ for i, upsample_block in enumerate(self.up_blocks):
+ for k in freeu_keys:
+ if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+ setattr(upsample_block, k, None)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+ def fuse_qkv_projections(self):
+ """
+ Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is 🧪 experimental.
+
+
+ """
+ self.original_attn_processors = None
+
+ for _, attn_processor in self.attn_processors.items():
+ if "Added" in str(attn_processor.__class__.__name__):
+ raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+ self.original_attn_processors = self.attn_processors
+
+ for module in self.modules():
+ if isinstance(module, Attention):
+ module.fuse_projections(fuse=True)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+ def unfuse_qkv_projections(self):
+ """Disables the fused QKV projection if enabled.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ """
+ if self.original_attn_processors is not None:
+ self.set_attn_processor(self.original_attn_processors)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unload_lora
+ def unload_lora(self):
+ """Unloads LoRA weights."""
+ deprecate(
+ "unload_lora",
+ "0.28.0",
+ "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().",
+ )
+ for module in self.modules():
+ if hasattr(module, "set_lora_layer"):
+ module.set_lora_layer(None)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+ mid_block_additional_residual: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ ) -> Union[UNet3DConditionOutput, Tuple[torch.FloatTensor]]:
+ r"""
+ The [`UNet3DConditionModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, num_channels, num_frames, height, width`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.FloatTensor`):
+ The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+ Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+ through the `self.time_embedding` layer to obtain the timestep embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+ A tuple of tensors that if specified are added to the residuals of down unet blocks.
+ mid_block_additional_residual: (`torch.Tensor`, *optional*):
+ A tensor that if specified is added to the residual of the middle unet block.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unet_3d_condition.UNet3DConditionOutput`] instead of a plain
+ tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+
+ Returns:
+ [`~models.unet_3d_condition.UNet3DConditionOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unet_3d_condition.UNet3DConditionOutput`] is returned, otherwise
+ a `tuple` is returned where the first element is the sample tensor.
+ """
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
+ # on the fly if necessary.
+ default_overall_up_factor = 2**self.num_upsamplers
+
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+ forward_upsample_size = False
+ upsample_size = None
+
+ if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+ logger.info("Forward upsample size to force interpolation output size.")
+ forward_upsample_size = True
+
+ # prepare attention_mask
+ if attention_mask is not None:
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ num_frames = sample.shape[2]
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=self.dtype)
+
+ emb = self.time_embedding(t_emb, timestep_cond)
+ emb = emb.repeat_interleave(repeats=num_frames, dim=0)
+ encoder_hidden_states = encoder_hidden_states.repeat_interleave(repeats=num_frames, dim=0)
+
+ # 2. pre-process
+ sample = sample.permute(0, 2, 1, 3, 4).reshape((sample.shape[0] * num_frames, -1) + sample.shape[3:])
+ sample = self.conv_in(sample)
+
+ sample = self.transformer_in(
+ sample,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb, num_frames=num_frames)
+
+ down_block_res_samples += res_samples
+
+ if down_block_additional_residuals is not None:
+ new_down_block_res_samples = ()
+
+ for down_block_res_sample, down_block_additional_residual in zip(
+ down_block_res_samples, down_block_additional_residuals
+ ):
+ down_block_res_sample = down_block_res_sample + down_block_additional_residual
+ new_down_block_res_samples += (down_block_res_sample,)
+
+ down_block_res_samples = new_down_block_res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+
+ if mid_block_additional_residual is not None:
+ sample = sample + mid_block_additional_residual
+
+ # 5. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block and forward_upsample_size:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ num_frames=num_frames,
+ )
+
+ # 6. post-process
+ if self.conv_norm_out:
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+
+ sample = self.conv_out(sample)
+
+ # reshape to (batch, channel, framerate, width, height)
+ sample = sample[None, :].reshape((-1, num_frames) + sample.shape[1:]).permute(0, 2, 1, 3, 4)
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet3DConditionOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/unet_i2vgen_xl.py b/model/diffusers_c/models/unets/unet_i2vgen_xl.py
new file mode 100644
index 0000000000000000000000000000000000000000..a096f842ab6c0713249a9a71c04907f1a98579e2
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_i2vgen_xl.py
@@ -0,0 +1,724 @@
+# Copyright 2024 Alibaba DAMO-VILAB and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import UNet2DConditionLoadersMixin
+from ...utils import logging
+from ..activations import get_activation
+from ..attention import Attention, FeedForward
+from ..attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from ..transformers.transformer_temporal import TransformerTemporalModel
+from .unet_3d_blocks import (
+ CrossAttnDownBlock3D,
+ CrossAttnUpBlock3D,
+ DownBlock3D,
+ UNetMidBlock3DCrossAttn,
+ UpBlock3D,
+ get_down_block,
+ get_up_block,
+)
+from .unet_3d_condition import UNet3DConditionOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class I2VGenXLTransformerTemporalEncoder(nn.Module):
+ def __init__(
+ self,
+ dim: int,
+ num_attention_heads: int,
+ attention_head_dim: int,
+ activation_fn: str = "geglu",
+ upcast_attention: bool = False,
+ ff_inner_dim: Optional[int] = None,
+ dropout: int = 0.0,
+ ):
+ super().__init__()
+ self.norm1 = nn.LayerNorm(dim, elementwise_affine=True, eps=1e-5)
+ self.attn1 = Attention(
+ query_dim=dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=False,
+ upcast_attention=upcast_attention,
+ out_bias=True,
+ )
+ self.ff = FeedForward(
+ dim,
+ dropout=dropout,
+ activation_fn=activation_fn,
+ final_dropout=False,
+ inner_dim=ff_inner_dim,
+ bias=True,
+ )
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ norm_hidden_states = self.norm1(hidden_states)
+ attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+ hidden_states = attn_output + hidden_states
+ if hidden_states.ndim == 4:
+ hidden_states = hidden_states.squeeze(1)
+
+ ff_output = self.ff(hidden_states, scale=1.0)
+ hidden_states = ff_output + hidden_states
+ if hidden_states.ndim == 4:
+ hidden_states = hidden_states.squeeze(1)
+
+ return hidden_states
+
+
+class I2VGenXLUNet(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+ r"""
+ I2VGenXL UNet. It is a conditional 3D UNet model that takes a noisy sample, conditional state, and a timestep
+ and returns a sample-shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+ Height and width of input/output sample.
+ in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+ The tuple of downsample blocks to use.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+ The tuple of upsample blocks to use.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+ norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+ If `None`, normalization and activation layers is skipped in post-processing.
+ cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features.
+ attention_head_dim (`int`, *optional*, defaults to 64): Attention head dim.
+ num_attention_heads (`int`, *optional*): The number of attention heads.
+ """
+
+ _supports_gradient_checkpointing = False
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[int] = None,
+ in_channels: int = 4,
+ out_channels: int = 4,
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlock3D",
+ "CrossAttnDownBlock3D",
+ "CrossAttnDownBlock3D",
+ "DownBlock3D",
+ ),
+ up_block_types: Tuple[str, ...] = (
+ "UpBlock3D",
+ "CrossAttnUpBlock3D",
+ "CrossAttnUpBlock3D",
+ "CrossAttnUpBlock3D",
+ ),
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+ layers_per_block: int = 2,
+ norm_num_groups: Optional[int] = 32,
+ cross_attention_dim: int = 1024,
+ attention_head_dim: Union[int, Tuple[int]] = 64,
+ num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+ ):
+ super().__init__()
+
+ # When we first integrated the UNet into the library, we didn't have `attention_head_dim`. As a consequence
+ # of that, we used `num_attention_heads` for arguments that actually denote attention head dimension. This
+ # is why we ignore `num_attention_heads` and calculate it from `attention_head_dims` below.
+ # This is still an incorrect way of calculating `num_attention_heads` but we need to stick to it
+ # without running proper depcrecation cycles for the {down,mid,up} blocks which are a
+ # part of the public API.
+ num_attention_heads = attention_head_dim
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ # input
+ self.conv_in = nn.Conv2d(in_channels + in_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+ self.transformer_in = TransformerTemporalModel(
+ num_attention_heads=8,
+ attention_head_dim=num_attention_heads,
+ in_channels=block_out_channels[0],
+ num_layers=1,
+ norm_num_groups=norm_num_groups,
+ )
+
+ # image embedding
+ self.image_latents_proj_in = nn.Sequential(
+ nn.Conv2d(4, in_channels * 4, 3, padding=1),
+ nn.SiLU(),
+ nn.Conv2d(in_channels * 4, in_channels * 4, 3, stride=1, padding=1),
+ nn.SiLU(),
+ nn.Conv2d(in_channels * 4, in_channels, 3, stride=1, padding=1),
+ )
+ self.image_latents_temporal_encoder = I2VGenXLTransformerTemporalEncoder(
+ dim=in_channels,
+ num_attention_heads=2,
+ ff_inner_dim=in_channels * 4,
+ attention_head_dim=in_channels,
+ activation_fn="gelu",
+ )
+ self.image_latents_context_embedding = nn.Sequential(
+ nn.Conv2d(4, in_channels * 8, 3, padding=1),
+ nn.SiLU(),
+ nn.AdaptiveAvgPool2d((32, 32)),
+ nn.Conv2d(in_channels * 8, in_channels * 16, 3, stride=2, padding=1),
+ nn.SiLU(),
+ nn.Conv2d(in_channels * 16, cross_attention_dim, 3, stride=2, padding=1),
+ )
+
+ # other embeddings -- time, context, fps, etc.
+ time_embed_dim = block_out_channels[0] * 4
+ self.time_proj = Timesteps(block_out_channels[0], True, 0)
+ timestep_input_dim = block_out_channels[0]
+
+ self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn="silu")
+ self.context_embedding = nn.Sequential(
+ nn.Linear(cross_attention_dim, time_embed_dim),
+ nn.SiLU(),
+ nn.Linear(time_embed_dim, cross_attention_dim * in_channels),
+ )
+ self.fps_embedding = nn.Sequential(
+ nn.Linear(timestep_input_dim, time_embed_dim), nn.SiLU(), nn.Linear(time_embed_dim, time_embed_dim)
+ )
+
+ # blocks
+ self.down_blocks = nn.ModuleList([])
+ self.up_blocks = nn.ModuleList([])
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=1e-05,
+ resnet_act_fn="silu",
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[i],
+ downsample_padding=1,
+ dual_cross_attention=False,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ self.mid_block = UNetMidBlock3DCrossAttn(
+ in_channels=block_out_channels[-1],
+ temb_channels=time_embed_dim,
+ resnet_eps=1e-05,
+ resnet_act_fn="silu",
+ output_scale_factor=1,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ dual_cross_attention=False,
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=layers_per_block + 1,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=1e-05,
+ resnet_act_fn="silu",
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=reversed_num_attention_heads[i],
+ dual_cross_attention=False,
+ resolution_idx=i,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=1e-05)
+ self.conv_act = get_activation("silu")
+ self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+ def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+ """
+ Sets the attention processor to use [feed forward
+ chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+ Parameters:
+ chunk_size (`int`, *optional*):
+ The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+ over each tensor of dim=`dim`.
+ dim (`int`, *optional*, defaults to `0`):
+ The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+ or dim=1 (sequence length).
+ """
+ if dim not in [0, 1]:
+ raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+ # By default chunk size is 1
+ chunk_size = chunk_size or 1
+
+ def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+ if hasattr(module, "set_chunk_feed_forward"):
+ module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+ for child in module.children():
+ fn_recursive_feed_forward(child, chunk_size, dim)
+
+ for module in self.children():
+ fn_recursive_feed_forward(module, chunk_size, dim)
+
+ # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.disable_forward_chunking
+ def disable_forward_chunking(self):
+ def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+ if hasattr(module, "set_chunk_feed_forward"):
+ module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+ for child in module.children():
+ fn_recursive_feed_forward(child, chunk_size, dim)
+
+ for module in self.children():
+ fn_recursive_feed_forward(module, None, 0)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel._set_gradient_checkpointing
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (CrossAttnDownBlock3D, DownBlock3D, CrossAttnUpBlock3D, UpBlock3D)):
+ module.gradient_checkpointing = value
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.enable_freeu
+ def enable_freeu(self, s1, s2, b1, b2):
+ r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+ The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+ Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+ are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+ Args:
+ s1 (`float`):
+ Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ s2 (`float`):
+ Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+ """
+ for i, upsample_block in enumerate(self.up_blocks):
+ setattr(upsample_block, "s1", s1)
+ setattr(upsample_block, "s2", s2)
+ setattr(upsample_block, "b1", b1)
+ setattr(upsample_block, "b2", b2)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.disable_freeu
+ def disable_freeu(self):
+ """Disables the FreeU mechanism."""
+ freeu_keys = {"s1", "s2", "b1", "b2"}
+ for i, upsample_block in enumerate(self.up_blocks):
+ for k in freeu_keys:
+ if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+ setattr(upsample_block, k, None)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+ def fuse_qkv_projections(self):
+ """
+ Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is 🧪 experimental.
+
+
+ """
+ self.original_attn_processors = None
+
+ for _, attn_processor in self.attn_processors.items():
+ if "Added" in str(attn_processor.__class__.__name__):
+ raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+ self.original_attn_processors = self.attn_processors
+
+ for module in self.modules():
+ if isinstance(module, Attention):
+ module.fuse_projections(fuse=True)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+ def unfuse_qkv_projections(self):
+ """Disables the fused QKV projection if enabled.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ """
+ if self.original_attn_processors is not None:
+ self.set_attn_processor(self.original_attn_processors)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ fps: torch.Tensor,
+ image_latents: torch.Tensor,
+ image_embeddings: Optional[torch.Tensor] = None,
+ encoder_hidden_states: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ return_dict: bool = True,
+ ) -> Union[UNet3DConditionOutput, Tuple[torch.FloatTensor]]:
+ r"""
+ The [`I2VGenXLUNet`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, num_frames, channel, height, width`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ fps (`torch.Tensor`): Frames per second for the video being generated. Used as a "micro-condition".
+ image_latents (`torch.FloatTensor`): Image encodings from the VAE.
+ image_embeddings (`torch.FloatTensor`): Projection embeddings of the conditioning image computed with a vision encoder.
+ encoder_hidden_states (`torch.FloatTensor`):
+ The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unet_3d_condition.UNet3DConditionOutput`] instead of a plain
+ tuple.
+
+ Returns:
+ [`~models.unet_3d_condition.UNet3DConditionOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unet_3d_condition.UNet3DConditionOutput`] is returned, otherwise
+ a `tuple` is returned where the first element is the sample tensor.
+ """
+ batch_size, channels, num_frames, height, width = sample.shape
+
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
+ # on the fly if necessary.
+ default_overall_up_factor = 2**self.num_upsamplers
+
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+ forward_upsample_size = False
+ upsample_size = None
+
+ if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+ logger.info("Forward upsample size to force interpolation output size.")
+ forward_upsample_size = True
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass `timesteps` as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timesteps, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+ t_emb = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=self.dtype)
+ t_emb = self.time_embedding(t_emb, timestep_cond)
+
+ # 2. FPS
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ fps = fps.expand(fps.shape[0])
+ fps_emb = self.fps_embedding(self.time_proj(fps).to(dtype=self.dtype))
+
+ # 3. time + FPS embeddings.
+ emb = t_emb + fps_emb
+ emb = emb.repeat_interleave(repeats=num_frames, dim=0)
+
+ # 4. context embeddings.
+ # The context embeddings consist of both text embeddings from the input prompt
+ # AND the image embeddings from the input image. For images, both VAE encodings
+ # and the CLIP image embeddings are incorporated.
+ # So the final `context_embeddings` becomes the query for cross-attention.
+ context_emb = sample.new_zeros(batch_size, 0, self.config.cross_attention_dim)
+ context_emb = torch.cat([context_emb, encoder_hidden_states], dim=1)
+
+ image_latents_for_context_embds = image_latents[:, :, :1, :]
+ image_latents_context_embs = image_latents_for_context_embds.permute(0, 2, 1, 3, 4).reshape(
+ image_latents_for_context_embds.shape[0] * image_latents_for_context_embds.shape[2],
+ image_latents_for_context_embds.shape[1],
+ image_latents_for_context_embds.shape[3],
+ image_latents_for_context_embds.shape[4],
+ )
+ image_latents_context_embs = self.image_latents_context_embedding(image_latents_context_embs)
+
+ _batch_size, _channels, _height, _width = image_latents_context_embs.shape
+ image_latents_context_embs = image_latents_context_embs.permute(0, 2, 3, 1).reshape(
+ _batch_size, _height * _width, _channels
+ )
+ context_emb = torch.cat([context_emb, image_latents_context_embs], dim=1)
+
+ image_emb = self.context_embedding(image_embeddings)
+ image_emb = image_emb.view(-1, self.config.in_channels, self.config.cross_attention_dim)
+ context_emb = torch.cat([context_emb, image_emb], dim=1)
+ context_emb = context_emb.repeat_interleave(repeats=num_frames, dim=0)
+
+ image_latents = image_latents.permute(0, 2, 1, 3, 4).reshape(
+ image_latents.shape[0] * image_latents.shape[2],
+ image_latents.shape[1],
+ image_latents.shape[3],
+ image_latents.shape[4],
+ )
+ image_latents = self.image_latents_proj_in(image_latents)
+ image_latents = (
+ image_latents[None, :]
+ .reshape(batch_size, num_frames, channels, height, width)
+ .permute(0, 3, 4, 1, 2)
+ .reshape(batch_size * height * width, num_frames, channels)
+ )
+ image_latents = self.image_latents_temporal_encoder(image_latents)
+ image_latents = image_latents.reshape(batch_size, height, width, num_frames, channels).permute(0, 4, 3, 1, 2)
+
+ # 5. pre-process
+ sample = torch.cat([sample, image_latents], dim=1)
+ sample = sample.permute(0, 2, 1, 3, 4).reshape((sample.shape[0] * num_frames, -1) + sample.shape[3:])
+ sample = self.conv_in(sample)
+ sample = self.transformer_in(
+ sample,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # 6. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=context_emb,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb, num_frames=num_frames)
+
+ down_block_res_samples += res_samples
+
+ # 7. mid
+ if self.mid_block is not None:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=context_emb,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ # 8. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block and forward_upsample_size:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=context_emb,
+ upsample_size=upsample_size,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ num_frames=num_frames,
+ )
+
+ # 9. post-process
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+
+ sample = self.conv_out(sample)
+
+ # reshape to (batch, channel, framerate, width, height)
+ sample = sample[None, :].reshape((-1, num_frames) + sample.shape[1:]).permute(0, 2, 1, 3, 4)
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet3DConditionOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/unet_kandinsky3.py b/model/diffusers_c/models/unets/unet_kandinsky3.py
new file mode 100644
index 0000000000000000000000000000000000000000..b981c8e17eb93d99bdb16d2c89d3c7088d77cdb3
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_kandinsky3.py
@@ -0,0 +1,535 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Dict, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput, logging
+from ..attention_processor import Attention, AttentionProcessor, AttnProcessor
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class Kandinsky3UNetOutput(BaseOutput):
+ sample: torch.FloatTensor = None
+
+
+class Kandinsky3EncoderProj(nn.Module):
+ def __init__(self, encoder_hid_dim, cross_attention_dim):
+ super().__init__()
+ self.projection_linear = nn.Linear(encoder_hid_dim, cross_attention_dim, bias=False)
+ self.projection_norm = nn.LayerNorm(cross_attention_dim)
+
+ def forward(self, x):
+ x = self.projection_linear(x)
+ x = self.projection_norm(x)
+ return x
+
+
+class Kandinsky3UNet(ModelMixin, ConfigMixin):
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 4,
+ time_embedding_dim: int = 1536,
+ groups: int = 32,
+ attention_head_dim: int = 64,
+ layers_per_block: Union[int, Tuple[int]] = 3,
+ block_out_channels: Tuple[int] = (384, 768, 1536, 3072),
+ cross_attention_dim: Union[int, Tuple[int]] = 4096,
+ encoder_hid_dim: int = 4096,
+ ):
+ super().__init__()
+
+ # TOOD(Yiyi): Give better name and put into config for the following 4 parameters
+ expansion_ratio = 4
+ compression_ratio = 2
+ add_cross_attention = (False, True, True, True)
+ add_self_attention = (False, True, True, True)
+
+ out_channels = in_channels
+ init_channels = block_out_channels[0] // 2
+ self.time_proj = Timesteps(init_channels, flip_sin_to_cos=False, downscale_freq_shift=1)
+
+ self.time_embedding = TimestepEmbedding(
+ init_channels,
+ time_embedding_dim,
+ )
+
+ self.add_time_condition = Kandinsky3AttentionPooling(
+ time_embedding_dim, cross_attention_dim, attention_head_dim
+ )
+
+ self.conv_in = nn.Conv2d(in_channels, init_channels, kernel_size=3, padding=1)
+
+ self.encoder_hid_proj = Kandinsky3EncoderProj(encoder_hid_dim, cross_attention_dim)
+
+ hidden_dims = [init_channels] + list(block_out_channels)
+ in_out_dims = list(zip(hidden_dims[:-1], hidden_dims[1:]))
+ text_dims = [cross_attention_dim if is_exist else None for is_exist in add_cross_attention]
+ num_blocks = len(block_out_channels) * [layers_per_block]
+ layer_params = [num_blocks, text_dims, add_self_attention]
+ rev_layer_params = map(reversed, layer_params)
+
+ cat_dims = []
+ self.num_levels = len(in_out_dims)
+ self.down_blocks = nn.ModuleList([])
+ for level, ((in_dim, out_dim), res_block_num, text_dim, self_attention) in enumerate(
+ zip(in_out_dims, *layer_params)
+ ):
+ down_sample = level != (self.num_levels - 1)
+ cat_dims.append(out_dim if level != (self.num_levels - 1) else 0)
+ self.down_blocks.append(
+ Kandinsky3DownSampleBlock(
+ in_dim,
+ out_dim,
+ time_embedding_dim,
+ text_dim,
+ res_block_num,
+ groups,
+ attention_head_dim,
+ expansion_ratio,
+ compression_ratio,
+ down_sample,
+ self_attention,
+ )
+ )
+
+ self.up_blocks = nn.ModuleList([])
+ for level, ((out_dim, in_dim), res_block_num, text_dim, self_attention) in enumerate(
+ zip(reversed(in_out_dims), *rev_layer_params)
+ ):
+ up_sample = level != 0
+ self.up_blocks.append(
+ Kandinsky3UpSampleBlock(
+ in_dim,
+ cat_dims.pop(),
+ out_dim,
+ time_embedding_dim,
+ text_dim,
+ res_block_num,
+ groups,
+ attention_head_dim,
+ expansion_ratio,
+ compression_ratio,
+ up_sample,
+ self_attention,
+ )
+ )
+
+ self.conv_norm_out = nn.GroupNorm(groups, init_channels)
+ self.conv_act_out = nn.SiLU()
+ self.conv_out = nn.Conv2d(init_channels, out_channels, kernel_size=3, padding=1)
+
+ @property
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "set_processor"):
+ processors[f"{name}.processor"] = module.processor
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ self.set_attn_processor(AttnProcessor())
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if hasattr(module, "gradient_checkpointing"):
+ module.gradient_checkpointing = value
+
+ def forward(self, sample, timestep, encoder_hidden_states=None, encoder_attention_mask=None, return_dict=True):
+ if encoder_attention_mask is not None:
+ encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+ encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+ if not torch.is_tensor(timestep):
+ dtype = torch.float32 if isinstance(timestep, float) else torch.int32
+ timestep = torch.tensor([timestep], dtype=dtype, device=sample.device)
+ elif len(timestep.shape) == 0:
+ timestep = timestep[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timestep = timestep.expand(sample.shape[0])
+ time_embed_input = self.time_proj(timestep).to(sample.dtype)
+ time_embed = self.time_embedding(time_embed_input)
+
+ encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+
+ if encoder_hidden_states is not None:
+ time_embed = self.add_time_condition(time_embed, encoder_hidden_states, encoder_attention_mask)
+
+ hidden_states = []
+ sample = self.conv_in(sample)
+ for level, down_sample in enumerate(self.down_blocks):
+ sample = down_sample(sample, time_embed, encoder_hidden_states, encoder_attention_mask)
+ if level != self.num_levels - 1:
+ hidden_states.append(sample)
+
+ for level, up_sample in enumerate(self.up_blocks):
+ if level != 0:
+ sample = torch.cat([sample, hidden_states.pop()], dim=1)
+ sample = up_sample(sample, time_embed, encoder_hidden_states, encoder_attention_mask)
+
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act_out(sample)
+ sample = self.conv_out(sample)
+
+ if not return_dict:
+ return (sample,)
+ return Kandinsky3UNetOutput(sample=sample)
+
+
+class Kandinsky3UpSampleBlock(nn.Module):
+ def __init__(
+ self,
+ in_channels,
+ cat_dim,
+ out_channels,
+ time_embed_dim,
+ context_dim=None,
+ num_blocks=3,
+ groups=32,
+ head_dim=64,
+ expansion_ratio=4,
+ compression_ratio=2,
+ up_sample=True,
+ self_attention=True,
+ ):
+ super().__init__()
+ up_resolutions = [[None, True if up_sample else None, None, None]] + [[None] * 4] * (num_blocks - 1)
+ hidden_channels = (
+ [(in_channels + cat_dim, in_channels)]
+ + [(in_channels, in_channels)] * (num_blocks - 2)
+ + [(in_channels, out_channels)]
+ )
+ attentions = []
+ resnets_in = []
+ resnets_out = []
+
+ self.self_attention = self_attention
+ self.context_dim = context_dim
+
+ if self_attention:
+ attentions.append(
+ Kandinsky3AttentionBlock(out_channels, time_embed_dim, None, groups, head_dim, expansion_ratio)
+ )
+ else:
+ attentions.append(nn.Identity())
+
+ for (in_channel, out_channel), up_resolution in zip(hidden_channels, up_resolutions):
+ resnets_in.append(
+ Kandinsky3ResNetBlock(in_channel, in_channel, time_embed_dim, groups, compression_ratio, up_resolution)
+ )
+
+ if context_dim is not None:
+ attentions.append(
+ Kandinsky3AttentionBlock(
+ in_channel, time_embed_dim, context_dim, groups, head_dim, expansion_ratio
+ )
+ )
+ else:
+ attentions.append(nn.Identity())
+
+ resnets_out.append(
+ Kandinsky3ResNetBlock(in_channel, out_channel, time_embed_dim, groups, compression_ratio)
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets_in = nn.ModuleList(resnets_in)
+ self.resnets_out = nn.ModuleList(resnets_out)
+
+ def forward(self, x, time_embed, context=None, context_mask=None, image_mask=None):
+ for attention, resnet_in, resnet_out in zip(self.attentions[1:], self.resnets_in, self.resnets_out):
+ x = resnet_in(x, time_embed)
+ if self.context_dim is not None:
+ x = attention(x, time_embed, context, context_mask, image_mask)
+ x = resnet_out(x, time_embed)
+
+ if self.self_attention:
+ x = self.attentions[0](x, time_embed, image_mask=image_mask)
+ return x
+
+
+class Kandinsky3DownSampleBlock(nn.Module):
+ def __init__(
+ self,
+ in_channels,
+ out_channels,
+ time_embed_dim,
+ context_dim=None,
+ num_blocks=3,
+ groups=32,
+ head_dim=64,
+ expansion_ratio=4,
+ compression_ratio=2,
+ down_sample=True,
+ self_attention=True,
+ ):
+ super().__init__()
+ attentions = []
+ resnets_in = []
+ resnets_out = []
+
+ self.self_attention = self_attention
+ self.context_dim = context_dim
+
+ if self_attention:
+ attentions.append(
+ Kandinsky3AttentionBlock(in_channels, time_embed_dim, None, groups, head_dim, expansion_ratio)
+ )
+ else:
+ attentions.append(nn.Identity())
+
+ up_resolutions = [[None] * 4] * (num_blocks - 1) + [[None, None, False if down_sample else None, None]]
+ hidden_channels = [(in_channels, out_channels)] + [(out_channels, out_channels)] * (num_blocks - 1)
+ for (in_channel, out_channel), up_resolution in zip(hidden_channels, up_resolutions):
+ resnets_in.append(
+ Kandinsky3ResNetBlock(in_channel, out_channel, time_embed_dim, groups, compression_ratio)
+ )
+
+ if context_dim is not None:
+ attentions.append(
+ Kandinsky3AttentionBlock(
+ out_channel, time_embed_dim, context_dim, groups, head_dim, expansion_ratio
+ )
+ )
+ else:
+ attentions.append(nn.Identity())
+
+ resnets_out.append(
+ Kandinsky3ResNetBlock(
+ out_channel, out_channel, time_embed_dim, groups, compression_ratio, up_resolution
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets_in = nn.ModuleList(resnets_in)
+ self.resnets_out = nn.ModuleList(resnets_out)
+
+ def forward(self, x, time_embed, context=None, context_mask=None, image_mask=None):
+ if self.self_attention:
+ x = self.attentions[0](x, time_embed, image_mask=image_mask)
+
+ for attention, resnet_in, resnet_out in zip(self.attentions[1:], self.resnets_in, self.resnets_out):
+ x = resnet_in(x, time_embed)
+ if self.context_dim is not None:
+ x = attention(x, time_embed, context, context_mask, image_mask)
+ x = resnet_out(x, time_embed)
+ return x
+
+
+class Kandinsky3ConditionalGroupNorm(nn.Module):
+ def __init__(self, groups, normalized_shape, context_dim):
+ super().__init__()
+ self.norm = nn.GroupNorm(groups, normalized_shape, affine=False)
+ self.context_mlp = nn.Sequential(nn.SiLU(), nn.Linear(context_dim, 2 * normalized_shape))
+ self.context_mlp[1].weight.data.zero_()
+ self.context_mlp[1].bias.data.zero_()
+
+ def forward(self, x, context):
+ context = self.context_mlp(context)
+
+ for _ in range(len(x.shape[2:])):
+ context = context.unsqueeze(-1)
+
+ scale, shift = context.chunk(2, dim=1)
+ x = self.norm(x) * (scale + 1.0) + shift
+ return x
+
+
+class Kandinsky3Block(nn.Module):
+ def __init__(self, in_channels, out_channels, time_embed_dim, kernel_size=3, norm_groups=32, up_resolution=None):
+ super().__init__()
+ self.group_norm = Kandinsky3ConditionalGroupNorm(norm_groups, in_channels, time_embed_dim)
+ self.activation = nn.SiLU()
+ if up_resolution is not None and up_resolution:
+ self.up_sample = nn.ConvTranspose2d(in_channels, in_channels, kernel_size=2, stride=2)
+ else:
+ self.up_sample = nn.Identity()
+
+ padding = int(kernel_size > 1)
+ self.projection = nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, padding=padding)
+
+ if up_resolution is not None and not up_resolution:
+ self.down_sample = nn.Conv2d(out_channels, out_channels, kernel_size=2, stride=2)
+ else:
+ self.down_sample = nn.Identity()
+
+ def forward(self, x, time_embed):
+ x = self.group_norm(x, time_embed)
+ x = self.activation(x)
+ x = self.up_sample(x)
+ x = self.projection(x)
+ x = self.down_sample(x)
+ return x
+
+
+class Kandinsky3ResNetBlock(nn.Module):
+ def __init__(
+ self, in_channels, out_channels, time_embed_dim, norm_groups=32, compression_ratio=2, up_resolutions=4 * [None]
+ ):
+ super().__init__()
+ kernel_sizes = [1, 3, 3, 1]
+ hidden_channel = max(in_channels, out_channels) // compression_ratio
+ hidden_channels = (
+ [(in_channels, hidden_channel)] + [(hidden_channel, hidden_channel)] * 2 + [(hidden_channel, out_channels)]
+ )
+ self.resnet_blocks = nn.ModuleList(
+ [
+ Kandinsky3Block(in_channel, out_channel, time_embed_dim, kernel_size, norm_groups, up_resolution)
+ for (in_channel, out_channel), kernel_size, up_resolution in zip(
+ hidden_channels, kernel_sizes, up_resolutions
+ )
+ ]
+ )
+ self.shortcut_up_sample = (
+ nn.ConvTranspose2d(in_channels, in_channels, kernel_size=2, stride=2)
+ if True in up_resolutions
+ else nn.Identity()
+ )
+ self.shortcut_projection = (
+ nn.Conv2d(in_channels, out_channels, kernel_size=1) if in_channels != out_channels else nn.Identity()
+ )
+ self.shortcut_down_sample = (
+ nn.Conv2d(out_channels, out_channels, kernel_size=2, stride=2)
+ if False in up_resolutions
+ else nn.Identity()
+ )
+
+ def forward(self, x, time_embed):
+ out = x
+ for resnet_block in self.resnet_blocks:
+ out = resnet_block(out, time_embed)
+
+ x = self.shortcut_up_sample(x)
+ x = self.shortcut_projection(x)
+ x = self.shortcut_down_sample(x)
+ x = x + out
+ return x
+
+
+class Kandinsky3AttentionPooling(nn.Module):
+ def __init__(self, num_channels, context_dim, head_dim=64):
+ super().__init__()
+ self.attention = Attention(
+ context_dim,
+ context_dim,
+ dim_head=head_dim,
+ out_dim=num_channels,
+ out_bias=False,
+ )
+
+ def forward(self, x, context, context_mask=None):
+ context_mask = context_mask.to(dtype=context.dtype)
+ context = self.attention(context.mean(dim=1, keepdim=True), context, context_mask)
+ return x + context.squeeze(1)
+
+
+class Kandinsky3AttentionBlock(nn.Module):
+ def __init__(self, num_channels, time_embed_dim, context_dim=None, norm_groups=32, head_dim=64, expansion_ratio=4):
+ super().__init__()
+ self.in_norm = Kandinsky3ConditionalGroupNorm(norm_groups, num_channels, time_embed_dim)
+ self.attention = Attention(
+ num_channels,
+ context_dim or num_channels,
+ dim_head=head_dim,
+ out_dim=num_channels,
+ out_bias=False,
+ )
+
+ hidden_channels = expansion_ratio * num_channels
+ self.out_norm = Kandinsky3ConditionalGroupNorm(norm_groups, num_channels, time_embed_dim)
+ self.feed_forward = nn.Sequential(
+ nn.Conv2d(num_channels, hidden_channels, kernel_size=1, bias=False),
+ nn.SiLU(),
+ nn.Conv2d(hidden_channels, num_channels, kernel_size=1, bias=False),
+ )
+
+ def forward(self, x, time_embed, context=None, context_mask=None, image_mask=None):
+ height, width = x.shape[-2:]
+ out = self.in_norm(x, time_embed)
+ out = out.reshape(x.shape[0], -1, height * width).permute(0, 2, 1)
+ context = context if context is not None else out
+ if context_mask is not None:
+ context_mask = context_mask.to(dtype=context.dtype)
+
+ out = self.attention(out, context, context_mask)
+ out = out.permute(0, 2, 1).unsqueeze(-1).reshape(out.shape[0], -1, height, width)
+ x = x + out
+
+ out = self.out_norm(x, time_embed)
+ out = self.feed_forward(out)
+ x = x + out
+ return x
diff --git a/model/diffusers_c/models/unets/unet_motion_model.py b/model/diffusers_c/models/unets/unet_motion_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..ab2eac4c9a9ab6a6dbdb3f4166fd699e661bb6ef
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_motion_model.py
@@ -0,0 +1,948 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import UNet2DConditionLoadersMixin
+from ...utils import logging
+from ..attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ Attention,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from ..transformers.transformer_temporal import TransformerTemporalModel
+from .unet_2d_blocks import UNetMidBlock2DCrossAttn
+from .unet_2d_condition import UNet2DConditionModel
+from .unet_3d_blocks import (
+ CrossAttnDownBlockMotion,
+ CrossAttnUpBlockMotion,
+ DownBlockMotion,
+ UNetMidBlockCrossAttnMotion,
+ UpBlockMotion,
+ get_down_block,
+ get_up_block,
+)
+from .unet_3d_condition import UNet3DConditionOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class MotionModules(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ layers_per_block: int = 2,
+ num_attention_heads: int = 8,
+ attention_bias: bool = False,
+ cross_attention_dim: Optional[int] = None,
+ activation_fn: str = "geglu",
+ norm_num_groups: int = 32,
+ max_seq_length: int = 32,
+ ):
+ super().__init__()
+ self.motion_modules = nn.ModuleList([])
+
+ for i in range(layers_per_block):
+ self.motion_modules.append(
+ TransformerTemporalModel(
+ in_channels=in_channels,
+ norm_num_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ activation_fn=activation_fn,
+ attention_bias=attention_bias,
+ num_attention_heads=num_attention_heads,
+ attention_head_dim=in_channels // num_attention_heads,
+ positional_embeddings="sinusoidal",
+ num_positional_embeddings=max_seq_length,
+ )
+ )
+
+
+class MotionAdapter(ModelMixin, ConfigMixin):
+ @register_to_config
+ def __init__(
+ self,
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+ motion_layers_per_block: int = 2,
+ motion_mid_block_layers_per_block: int = 1,
+ motion_num_attention_heads: int = 8,
+ motion_norm_num_groups: int = 32,
+ motion_max_seq_length: int = 32,
+ use_motion_mid_block: bool = True,
+ conv_in_channels: Optional[int] = None,
+ ):
+ """Container to store AnimateDiff Motion Modules
+
+ Args:
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each UNet block.
+ motion_layers_per_block (`int`, *optional*, defaults to 2):
+ The number of motion layers per UNet block.
+ motion_mid_block_layers_per_block (`int`, *optional*, defaults to 1):
+ The number of motion layers in the middle UNet block.
+ motion_num_attention_heads (`int`, *optional*, defaults to 8):
+ The number of heads to use in each attention layer of the motion module.
+ motion_norm_num_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use in each group normalization layer of the motion module.
+ motion_max_seq_length (`int`, *optional*, defaults to 32):
+ The maximum sequence length to use in the motion module.
+ use_motion_mid_block (`bool`, *optional*, defaults to True):
+ Whether to use a motion module in the middle of the UNet.
+ """
+
+ super().__init__()
+ down_blocks = []
+ up_blocks = []
+
+ if conv_in_channels:
+ # input
+ self.conv_in = nn.Conv2d(conv_in_channels, block_out_channels[0], kernel_size=3, padding=1)
+ else:
+ self.conv_in = None
+
+ for i, channel in enumerate(block_out_channels):
+ output_channel = block_out_channels[i]
+ down_blocks.append(
+ MotionModules(
+ in_channels=output_channel,
+ norm_num_groups=motion_norm_num_groups,
+ cross_attention_dim=None,
+ activation_fn="geglu",
+ attention_bias=False,
+ num_attention_heads=motion_num_attention_heads,
+ max_seq_length=motion_max_seq_length,
+ layers_per_block=motion_layers_per_block,
+ )
+ )
+
+ if use_motion_mid_block:
+ self.mid_block = MotionModules(
+ in_channels=block_out_channels[-1],
+ norm_num_groups=motion_norm_num_groups,
+ cross_attention_dim=None,
+ activation_fn="geglu",
+ attention_bias=False,
+ num_attention_heads=motion_num_attention_heads,
+ layers_per_block=motion_mid_block_layers_per_block,
+ max_seq_length=motion_max_seq_length,
+ )
+ else:
+ self.mid_block = None
+
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ output_channel = reversed_block_out_channels[0]
+ for i, channel in enumerate(reversed_block_out_channels):
+ output_channel = reversed_block_out_channels[i]
+ up_blocks.append(
+ MotionModules(
+ in_channels=output_channel,
+ norm_num_groups=motion_norm_num_groups,
+ cross_attention_dim=None,
+ activation_fn="geglu",
+ attention_bias=False,
+ num_attention_heads=motion_num_attention_heads,
+ max_seq_length=motion_max_seq_length,
+ layers_per_block=motion_layers_per_block + 1,
+ )
+ )
+
+ self.down_blocks = nn.ModuleList(down_blocks)
+ self.up_blocks = nn.ModuleList(up_blocks)
+
+ def forward(self, sample):
+ pass
+
+
+class UNetMotionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+ r"""
+ A modified conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a
+ sample shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[int] = None,
+ in_channels: int = 4,
+ out_channels: int = 4,
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlockMotion",
+ "CrossAttnDownBlockMotion",
+ "CrossAttnDownBlockMotion",
+ "DownBlockMotion",
+ ),
+ up_block_types: Tuple[str, ...] = (
+ "UpBlockMotion",
+ "CrossAttnUpBlockMotion",
+ "CrossAttnUpBlockMotion",
+ "CrossAttnUpBlockMotion",
+ ),
+ block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+ layers_per_block: int = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ act_fn: str = "silu",
+ norm_num_groups: int = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: int = 1280,
+ use_linear_projection: bool = False,
+ num_attention_heads: Union[int, Tuple[int, ...]] = 8,
+ motion_max_seq_length: int = 32,
+ motion_num_attention_heads: int = 8,
+ use_motion_mid_block: int = True,
+ encoder_hid_dim: Optional[int] = None,
+ encoder_hid_dim_type: Optional[str] = None,
+ time_cond_proj_dim: Optional[int] = None,
+ ):
+ super().__init__()
+
+ self.sample_size = sample_size
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ # input
+ conv_in_kernel = 3
+ conv_out_kernel = 3
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in = nn.Conv2d(
+ in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+ )
+
+ # time
+ time_embed_dim = block_out_channels[0] * 4
+ self.time_proj = Timesteps(block_out_channels[0], True, 0)
+ timestep_input_dim = block_out_channels[0]
+
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim, time_embed_dim, act_fn=act_fn, cond_proj_dim=time_cond_proj_dim
+ )
+
+ if encoder_hid_dim_type is None:
+ self.encoder_hid_proj = None
+
+ # class embedding
+ self.down_blocks = nn.ModuleList([])
+ self.up_blocks = nn.ModuleList([])
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[i],
+ downsample_padding=downsample_padding,
+ use_linear_projection=use_linear_projection,
+ dual_cross_attention=False,
+ temporal_num_attention_heads=motion_num_attention_heads,
+ temporal_max_seq_length=motion_max_seq_length,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ if use_motion_mid_block:
+ self.mid_block = UNetMidBlockCrossAttnMotion(
+ in_channels=block_out_channels[-1],
+ temb_channels=time_embed_dim,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ dual_cross_attention=False,
+ use_linear_projection=use_linear_projection,
+ temporal_num_attention_heads=motion_num_attention_heads,
+ temporal_max_seq_length=motion_max_seq_length,
+ )
+
+ else:
+ self.mid_block = UNetMidBlock2DCrossAttn(
+ in_channels=block_out_channels[-1],
+ temb_channels=time_embed_dim,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ dual_cross_attention=False,
+ use_linear_projection=use_linear_projection,
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=layers_per_block + 1,
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=reversed_num_attention_heads[i],
+ dual_cross_attention=False,
+ resolution_idx=i,
+ use_linear_projection=use_linear_projection,
+ temporal_num_attention_heads=motion_num_attention_heads,
+ temporal_max_seq_length=motion_max_seq_length,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ if norm_num_groups is not None:
+ self.conv_norm_out = nn.GroupNorm(
+ num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+ )
+ self.conv_act = nn.SiLU()
+ else:
+ self.conv_norm_out = None
+ self.conv_act = None
+
+ conv_out_padding = (conv_out_kernel - 1) // 2
+ self.conv_out = nn.Conv2d(
+ block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+ )
+
+ @classmethod
+ def from_unet2d(
+ cls,
+ unet: UNet2DConditionModel,
+ motion_adapter: Optional[MotionAdapter] = None,
+ load_weights: bool = True,
+ ):
+ has_motion_adapter = motion_adapter is not None
+
+ # based on https://github.com/guoyww/AnimateDiff/blob/895f3220c06318ea0760131ec70408b466c49333/animatediff/models/unet.py#L459
+ config = unet.config
+ config["_class_name"] = cls.__name__
+
+ down_blocks = []
+ for down_blocks_type in config["down_block_types"]:
+ if "CrossAttn" in down_blocks_type:
+ down_blocks.append("CrossAttnDownBlockMotion")
+ else:
+ down_blocks.append("DownBlockMotion")
+ config["down_block_types"] = down_blocks
+
+ up_blocks = []
+ for down_blocks_type in config["up_block_types"]:
+ if "CrossAttn" in down_blocks_type:
+ up_blocks.append("CrossAttnUpBlockMotion")
+ else:
+ up_blocks.append("UpBlockMotion")
+
+ config["up_block_types"] = up_blocks
+
+ if has_motion_adapter:
+ config["motion_num_attention_heads"] = motion_adapter.config["motion_num_attention_heads"]
+ config["motion_max_seq_length"] = motion_adapter.config["motion_max_seq_length"]
+ config["use_motion_mid_block"] = motion_adapter.config["use_motion_mid_block"]
+
+ # For PIA UNets we need to set the number input channels to 9
+ if motion_adapter.config["conv_in_channels"]:
+ config["in_channels"] = motion_adapter.config["conv_in_channels"]
+
+ # Need this for backwards compatibility with UNet2DConditionModel checkpoints
+ if not config.get("num_attention_heads"):
+ config["num_attention_heads"] = config["attention_head_dim"]
+
+ model = cls.from_config(config)
+
+ if not load_weights:
+ return model
+
+ # Logic for loading PIA UNets which allow the first 4 channels to be any UNet2DConditionModel conv_in weight
+ # while the last 5 channels must be PIA conv_in weights.
+ if has_motion_adapter and motion_adapter.config["conv_in_channels"]:
+ model.conv_in = motion_adapter.conv_in
+ updated_conv_in_weight = torch.cat(
+ [unet.conv_in.weight, motion_adapter.conv_in.weight[:, 4:, :, :]], dim=1
+ )
+ model.conv_in.load_state_dict({"weight": updated_conv_in_weight, "bias": unet.conv_in.bias})
+ else:
+ model.conv_in.load_state_dict(unet.conv_in.state_dict())
+
+ model.time_proj.load_state_dict(unet.time_proj.state_dict())
+ model.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+ for i, down_block in enumerate(unet.down_blocks):
+ model.down_blocks[i].resnets.load_state_dict(down_block.resnets.state_dict())
+ if hasattr(model.down_blocks[i], "attentions"):
+ model.down_blocks[i].attentions.load_state_dict(down_block.attentions.state_dict())
+ if model.down_blocks[i].downsamplers:
+ model.down_blocks[i].downsamplers.load_state_dict(down_block.downsamplers.state_dict())
+
+ for i, up_block in enumerate(unet.up_blocks):
+ model.up_blocks[i].resnets.load_state_dict(up_block.resnets.state_dict())
+ if hasattr(model.up_blocks[i], "attentions"):
+ model.up_blocks[i].attentions.load_state_dict(up_block.attentions.state_dict())
+ if model.up_blocks[i].upsamplers:
+ model.up_blocks[i].upsamplers.load_state_dict(up_block.upsamplers.state_dict())
+
+ model.mid_block.resnets.load_state_dict(unet.mid_block.resnets.state_dict())
+ model.mid_block.attentions.load_state_dict(unet.mid_block.attentions.state_dict())
+
+ if unet.conv_norm_out is not None:
+ model.conv_norm_out.load_state_dict(unet.conv_norm_out.state_dict())
+ if unet.conv_act is not None:
+ model.conv_act.load_state_dict(unet.conv_act.state_dict())
+ model.conv_out.load_state_dict(unet.conv_out.state_dict())
+
+ if has_motion_adapter:
+ model.load_motion_modules(motion_adapter)
+
+ # ensure that the Motion UNet is the same dtype as the UNet2DConditionModel
+ model.to(unet.dtype)
+
+ return model
+
+ def freeze_unet2d_params(self) -> None:
+ """Freeze the weights of just the UNet2DConditionModel, and leave the motion modules
+ unfrozen for fine tuning.
+ """
+ # Freeze everything
+ for param in self.parameters():
+ param.requires_grad = False
+
+ # Unfreeze Motion Modules
+ for down_block in self.down_blocks:
+ motion_modules = down_block.motion_modules
+ for param in motion_modules.parameters():
+ param.requires_grad = True
+
+ for up_block in self.up_blocks:
+ motion_modules = up_block.motion_modules
+ for param in motion_modules.parameters():
+ param.requires_grad = True
+
+ if hasattr(self.mid_block, "motion_modules"):
+ motion_modules = self.mid_block.motion_modules
+ for param in motion_modules.parameters():
+ param.requires_grad = True
+
+ def load_motion_modules(self, motion_adapter: Optional[MotionAdapter]) -> None:
+ for i, down_block in enumerate(motion_adapter.down_blocks):
+ self.down_blocks[i].motion_modules.load_state_dict(down_block.motion_modules.state_dict())
+ for i, up_block in enumerate(motion_adapter.up_blocks):
+ self.up_blocks[i].motion_modules.load_state_dict(up_block.motion_modules.state_dict())
+
+ # to support older motion modules that don't have a mid_block
+ if hasattr(self.mid_block, "motion_modules"):
+ self.mid_block.motion_modules.load_state_dict(motion_adapter.mid_block.motion_modules.state_dict())
+
+ def save_motion_modules(
+ self,
+ save_directory: str,
+ is_main_process: bool = True,
+ safe_serialization: bool = True,
+ variant: Optional[str] = None,
+ push_to_hub: bool = False,
+ **kwargs,
+ ) -> None:
+ state_dict = self.state_dict()
+
+ # Extract all motion modules
+ motion_state_dict = {}
+ for k, v in state_dict.items():
+ if "motion_modules" in k:
+ motion_state_dict[k] = v
+
+ adapter = MotionAdapter(
+ block_out_channels=self.config["block_out_channels"],
+ motion_layers_per_block=self.config["layers_per_block"],
+ motion_norm_num_groups=self.config["norm_num_groups"],
+ motion_num_attention_heads=self.config["motion_num_attention_heads"],
+ motion_max_seq_length=self.config["motion_max_seq_length"],
+ use_motion_mid_block=self.config["use_motion_mid_block"],
+ )
+ adapter.load_state_dict(motion_state_dict)
+ adapter.save_pretrained(
+ save_directory=save_directory,
+ is_main_process=is_main_process,
+ safe_serialization=safe_serialization,
+ variant=variant,
+ push_to_hub=push_to_hub,
+ **kwargs,
+ )
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+ def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+ """
+ Sets the attention processor to use [feed forward
+ chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+ Parameters:
+ chunk_size (`int`, *optional*):
+ The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+ over each tensor of dim=`dim`.
+ dim (`int`, *optional*, defaults to `0`):
+ The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+ or dim=1 (sequence length).
+ """
+ if dim not in [0, 1]:
+ raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+ # By default chunk size is 1
+ chunk_size = chunk_size or 1
+
+ def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+ if hasattr(module, "set_chunk_feed_forward"):
+ module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+ for child in module.children():
+ fn_recursive_feed_forward(child, chunk_size, dim)
+
+ for module in self.children():
+ fn_recursive_feed_forward(module, chunk_size, dim)
+
+ # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.disable_forward_chunking
+ def disable_forward_chunking(self) -> None:
+ def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+ if hasattr(module, "set_chunk_feed_forward"):
+ module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+ for child in module.children():
+ fn_recursive_feed_forward(child, chunk_size, dim)
+
+ for module in self.children():
+ fn_recursive_feed_forward(module, None, 0)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self) -> None:
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (CrossAttnDownBlockMotion, DownBlockMotion, CrossAttnUpBlockMotion, UpBlockMotion)):
+ module.gradient_checkpointing = value
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.enable_freeu
+ def enable_freeu(self, s1: float, s2: float, b1: float, b2: float) -> None:
+ r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+ The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+ Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+ are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+ Args:
+ s1 (`float`):
+ Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ s2 (`float`):
+ Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+ """
+ for i, upsample_block in enumerate(self.up_blocks):
+ setattr(upsample_block, "s1", s1)
+ setattr(upsample_block, "s2", s2)
+ setattr(upsample_block, "b1", b1)
+ setattr(upsample_block, "b2", b2)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.disable_freeu
+ def disable_freeu(self) -> None:
+ """Disables the FreeU mechanism."""
+ freeu_keys = {"s1", "s2", "b1", "b2"}
+ for i, upsample_block in enumerate(self.up_blocks):
+ for k in freeu_keys:
+ if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+ setattr(upsample_block, k, None)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections
+ def fuse_qkv_projections(self):
+ """
+ Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is 🧪 experimental.
+
+
+ """
+ self.original_attn_processors = None
+
+ for _, attn_processor in self.attn_processors.items():
+ if "Added" in str(attn_processor.__class__.__name__):
+ raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+ self.original_attn_processors = self.attn_processors
+
+ for module in self.modules():
+ if isinstance(module, Attention):
+ module.fuse_projections(fuse=True)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
+ def unfuse_qkv_projections(self):
+ """Disables the fused QKV projection if enabled.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ """
+ if self.original_attn_processors is not None:
+ self.set_attn_processor(self.original_attn_processors)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+ mid_block_additional_residual: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ ) -> Union[UNet3DConditionOutput, Tuple[torch.Tensor]]:
+ r"""
+ The [`UNetMotionModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, num_frames, channel, height, width`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.FloatTensor`):
+ The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+ timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+ Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+ through the `self.time_embedding` layer to obtain the timestep embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+ A tuple of tensors that if specified are added to the residuals of down unet blocks.
+ mid_block_additional_residual: (`torch.Tensor`, *optional*):
+ A tensor that if specified is added to the residual of the middle unet block.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unet_3d_condition.UNet3DConditionOutput`] instead of a plain
+ tuple.
+
+ Returns:
+ [`~models.unet_3d_condition.UNet3DConditionOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unet_3d_condition.UNet3DConditionOutput`] is returned, otherwise
+ a `tuple` is returned where the first element is the sample tensor.
+ """
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
+ # on the fly if necessary.
+ default_overall_up_factor = 2**self.num_upsamplers
+
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+ forward_upsample_size = False
+ upsample_size = None
+
+ if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+ logger.info("Forward upsample size to force interpolation output size.")
+ forward_upsample_size = True
+
+ # prepare attention_mask
+ if attention_mask is not None:
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ num_frames = sample.shape[2]
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # timesteps does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=self.dtype)
+
+ emb = self.time_embedding(t_emb, timestep_cond)
+ emb = emb.repeat_interleave(repeats=num_frames, dim=0)
+ encoder_hidden_states = encoder_hidden_states.repeat_interleave(repeats=num_frames, dim=0)
+
+ if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ image_embeds = self.encoder_hid_proj(image_embeds)
+ image_embeds = [image_embed.repeat_interleave(repeats=num_frames, dim=0) for image_embed in image_embeds]
+ encoder_hidden_states = (encoder_hidden_states, image_embeds)
+
+ # 2. pre-process
+ sample = sample.permute(0, 2, 1, 3, 4).reshape((sample.shape[0] * num_frames, -1) + sample.shape[3:])
+ sample = self.conv_in(sample)
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb, num_frames=num_frames)
+
+ down_block_res_samples += res_samples
+
+ if down_block_additional_residuals is not None:
+ new_down_block_res_samples = ()
+
+ for down_block_res_sample, down_block_additional_residual in zip(
+ down_block_res_samples, down_block_additional_residuals
+ ):
+ down_block_res_sample = down_block_res_sample + down_block_additional_residual
+ new_down_block_res_samples += (down_block_res_sample,)
+
+ down_block_res_samples = new_down_block_res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ # To support older versions of motion modules that don't have a mid_block
+ if hasattr(self.mid_block, "motion_modules"):
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+
+ if mid_block_additional_residual is not None:
+ sample = sample + mid_block_additional_residual
+
+ # 5. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block and forward_upsample_size:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ num_frames=num_frames,
+ )
+
+ # 6. post-process
+ if self.conv_norm_out:
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+
+ sample = self.conv_out(sample)
+
+ # reshape to (batch, channel, framerate, width, height)
+ sample = sample[None, :].reshape((-1, num_frames) + sample.shape[1:]).permute(0, 2, 1, 3, 4)
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet3DConditionOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/unet_spatio_temporal_condition.py b/model/diffusers_c/models/unets/unet_spatio_temporal_condition.py
new file mode 100644
index 0000000000000000000000000000000000000000..5fe265e63fc507f8fa874acc1b69e9a3f3ed072f
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_spatio_temporal_condition.py
@@ -0,0 +1,489 @@
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import UNet2DConditionLoadersMixin
+from ...utils import BaseOutput, logging
+from ..attention_processor import CROSS_ATTENTION_PROCESSORS, AttentionProcessor, AttnProcessor
+from ..embeddings import TimestepEmbedding, Timesteps
+from ..modeling_utils import ModelMixin
+from .unet_3d_blocks import UNetMidBlockSpatioTemporal, get_down_block, get_up_block
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class UNetSpatioTemporalConditionOutput(BaseOutput):
+ """
+ The output of [`UNetSpatioTemporalConditionModel`].
+
+ Args:
+ sample (`torch.FloatTensor` of shape `(batch_size, num_frames, num_channels, height, width)`):
+ The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+ """
+
+ sample: torch.FloatTensor = None
+
+
+class UNetSpatioTemporalConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+ r"""
+ A conditional Spatio-Temporal UNet model that takes a noisy video frames, conditional state, and a timestep and returns a sample
+ shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+ Height and width of input/output sample.
+ in_channels (`int`, *optional*, defaults to 8): Number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlockSpatioTemporal", "CrossAttnDownBlockSpatioTemporal", "CrossAttnDownBlockSpatioTemporal", "DownBlockSpatioTemporal")`):
+ The tuple of downsample blocks to use.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal")`):
+ The tuple of upsample blocks to use.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ addition_time_embed_dim: (`int`, defaults to 256):
+ Dimension to to encode the additional time ids.
+ projection_class_embeddings_input_dim (`int`, defaults to 768):
+ The dimension of the projection of encoded `added_time_ids`.
+ layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+ cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_3d_blocks.CrossAttnDownBlockSpatioTemporal`], [`~models.unet_3d_blocks.CrossAttnUpBlockSpatioTemporal`],
+ [`~models.unet_3d_blocks.UNetMidBlockSpatioTemporal`].
+ num_attention_heads (`int`, `Tuple[int]`, defaults to `(5, 10, 10, 20)`):
+ The number of attention heads.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[int] = None,
+ in_channels: int = 8,
+ out_channels: int = 4,
+ down_block_types: Tuple[str] = (
+ "CrossAttnDownBlockSpatioTemporal",
+ "CrossAttnDownBlockSpatioTemporal",
+ "CrossAttnDownBlockSpatioTemporal",
+ "DownBlockSpatioTemporal",
+ ),
+ up_block_types: Tuple[str] = (
+ "UpBlockSpatioTemporal",
+ "CrossAttnUpBlockSpatioTemporal",
+ "CrossAttnUpBlockSpatioTemporal",
+ "CrossAttnUpBlockSpatioTemporal",
+ ),
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+ addition_time_embed_dim: int = 256,
+ projection_class_embeddings_input_dim: int = 768,
+ layers_per_block: Union[int, Tuple[int]] = 2,
+ cross_attention_dim: Union[int, Tuple[int]] = 1024,
+ transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+ num_attention_heads: Union[int, Tuple[int]] = (5, 10, 20, 20),
+ num_frames: int = 25,
+ ):
+ super().__init__()
+
+ self.sample_size = sample_size
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+ )
+
+ # input
+ self.conv_in = nn.Conv2d(
+ in_channels,
+ block_out_channels[0],
+ kernel_size=3,
+ padding=1,
+ )
+
+ # time
+ time_embed_dim = block_out_channels[0] * 4
+
+ self.time_proj = Timesteps(block_out_channels[0], True, downscale_freq_shift=0)
+ timestep_input_dim = block_out_channels[0]
+
+ self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+ self.add_time_proj = Timesteps(addition_time_embed_dim, True, downscale_freq_shift=0)
+ self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+
+ self.down_blocks = nn.ModuleList([])
+ self.up_blocks = nn.ModuleList([])
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+ if isinstance(layers_per_block, int):
+ layers_per_block = [layers_per_block] * len(down_block_types)
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+ blocks_time_embed_dim = time_embed_dim
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block[i],
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=1e-5,
+ cross_attention_dim=cross_attention_dim[i],
+ num_attention_heads=num_attention_heads[i],
+ resnet_act_fn="silu",
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ self.mid_block = UNetMidBlockSpatioTemporal(
+ block_out_channels[-1],
+ temb_channels=blocks_time_embed_dim,
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ cross_attention_dim=cross_attention_dim[-1],
+ num_attention_heads=num_attention_heads[-1],
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ reversed_layers_per_block = list(reversed(layers_per_block))
+ reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+ reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=reversed_layers_per_block[i] + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=1e-5,
+ resolution_idx=i,
+ cross_attention_dim=reversed_cross_attention_dim[i],
+ num_attention_heads=reversed_num_attention_heads[i],
+ resnet_act_fn="silu",
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=32, eps=1e-5)
+ self.conv_act = nn.SiLU()
+
+ self.conv_out = nn.Conv2d(
+ block_out_channels[0],
+ out_channels,
+ kernel_size=3,
+ padding=1,
+ )
+
+ @property
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(
+ name: str,
+ module: torch.nn.Module,
+ processors: Dict[str, AttentionProcessor],
+ ):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if hasattr(module, "gradient_checkpointing"):
+ module.gradient_checkpointing = value
+
+ # Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
+ def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
+ """
+ Sets the attention processor to use [feed forward
+ chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
+
+ Parameters:
+ chunk_size (`int`, *optional*):
+ The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
+ over each tensor of dim=`dim`.
+ dim (`int`, *optional*, defaults to `0`):
+ The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
+ or dim=1 (sequence length).
+ """
+ if dim not in [0, 1]:
+ raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
+
+ # By default chunk size is 1
+ chunk_size = chunk_size or 1
+
+ def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
+ if hasattr(module, "set_chunk_feed_forward"):
+ module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
+
+ for child in module.children():
+ fn_recursive_feed_forward(child, chunk_size, dim)
+
+ for module in self.children():
+ fn_recursive_feed_forward(module, chunk_size, dim)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ added_time_ids: torch.Tensor,
+ return_dict: bool = True,
+ ) -> Union[UNetSpatioTemporalConditionOutput, Tuple]:
+ r"""
+ The [`UNetSpatioTemporalConditionModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, num_frames, channel, height, width)`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.FloatTensor`):
+ The encoder hidden states with shape `(batch, sequence_length, cross_attention_dim)`.
+ added_time_ids: (`torch.FloatTensor`):
+ The additional time ids with shape `(batch, num_additional_ids)`. These are encoded with sinusoidal
+ embeddings and added to the time embeddings.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] instead of a plain
+ tuple.
+ Returns:
+ [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] is returned, otherwise
+ a `tuple` is returned where the first element is the sample tensor.
+ """
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ batch_size, num_frames = sample.shape[:2]
+ timesteps = timesteps.expand(batch_size)
+
+ t_emb = self.time_proj(timesteps)
+
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+
+ emb = self.time_embedding(t_emb)
+
+ time_embeds = self.add_time_proj(added_time_ids.flatten())
+ time_embeds = time_embeds.reshape((batch_size, -1))
+ time_embeds = time_embeds.to(emb.dtype)
+ aug_emb = self.add_embedding(time_embeds)
+ emb = emb + aug_emb
+
+ # Flatten the batch and frames dimensions
+ # sample: [batch, frames, channels, height, width] -> [batch * frames, channels, height, width]
+ sample = sample.flatten(0, 1)
+ # Repeat the embeddings num_video_frames times
+ # emb: [batch, channels] -> [batch * frames, channels]
+ emb = emb.repeat_interleave(num_frames, dim=0)
+ # encoder_hidden_states: [batch, 1, channels] -> [batch * frames, 1, channels]
+ encoder_hidden_states = encoder_hidden_states.repeat_interleave(num_frames, dim=0)
+
+ # 2. pre-process
+ sample = self.conv_in(sample)
+
+ image_only_indicator = torch.zeros(batch_size, num_frames, dtype=sample.dtype, device=sample.device)
+
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ )
+ else:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ image_only_indicator=image_only_indicator,
+ )
+
+ down_block_res_samples += res_samples
+
+ # 4. mid
+ sample = self.mid_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ )
+
+ # 5. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ image_only_indicator=image_only_indicator,
+ )
+ else:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ image_only_indicator=image_only_indicator,
+ )
+
+ # 6. post-process
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ # 7. Reshape back to original shape
+ sample = sample.reshape(batch_size, num_frames, *sample.shape[1:])
+
+ if not return_dict:
+ return (sample,)
+
+ return UNetSpatioTemporalConditionOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/unet_stable_cascade.py b/model/diffusers_c/models/unets/unet_stable_cascade.py
new file mode 100644
index 0000000000000000000000000000000000000000..b8833779ba9f6f4a8b41638af5dd0b09cd393718
--- /dev/null
+++ b/model/diffusers_c/models/unets/unet_stable_cascade.py
@@ -0,0 +1,609 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ..attention_processor import Attention
+from ..modeling_utils import ModelMixin
+
+
+# Copied from diffusers.pipelines.wuerstchen.modeling_wuerstchen_common.WuerstchenLayerNorm with WuerstchenLayerNorm -> SDCascadeLayerNorm
+class SDCascadeLayerNorm(nn.LayerNorm):
+ def __init__(self, *args, **kwargs):
+ super().__init__(*args, **kwargs)
+
+ def forward(self, x):
+ x = x.permute(0, 2, 3, 1)
+ x = super().forward(x)
+ return x.permute(0, 3, 1, 2)
+
+
+class SDCascadeTimestepBlock(nn.Module):
+ def __init__(self, c, c_timestep, conds=[]):
+ super().__init__()
+ linear_cls = nn.Linear
+ self.mapper = linear_cls(c_timestep, c * 2)
+ self.conds = conds
+ for cname in conds:
+ setattr(self, f"mapper_{cname}", linear_cls(c_timestep, c * 2))
+
+ def forward(self, x, t):
+ t = t.chunk(len(self.conds) + 1, dim=1)
+ a, b = self.mapper(t[0])[:, :, None, None].chunk(2, dim=1)
+ for i, c in enumerate(self.conds):
+ ac, bc = getattr(self, f"mapper_{c}")(t[i + 1])[:, :, None, None].chunk(2, dim=1)
+ a, b = a + ac, b + bc
+ return x * (1 + a) + b
+
+
+class SDCascadeResBlock(nn.Module):
+ def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0):
+ super().__init__()
+ self.depthwise = nn.Conv2d(c, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c)
+ self.norm = SDCascadeLayerNorm(c, elementwise_affine=False, eps=1e-6)
+ self.channelwise = nn.Sequential(
+ nn.Linear(c + c_skip, c * 4),
+ nn.GELU(),
+ GlobalResponseNorm(c * 4),
+ nn.Dropout(dropout),
+ nn.Linear(c * 4, c),
+ )
+
+ def forward(self, x, x_skip=None):
+ x_res = x
+ x = self.norm(self.depthwise(x))
+ if x_skip is not None:
+ x = torch.cat([x, x_skip], dim=1)
+ x = self.channelwise(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+ return x + x_res
+
+
+# from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105
+class GlobalResponseNorm(nn.Module):
+ def __init__(self, dim):
+ super().__init__()
+ self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim))
+ self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim))
+
+ def forward(self, x):
+ agg_norm = torch.norm(x, p=2, dim=(1, 2), keepdim=True)
+ stand_div_norm = agg_norm / (agg_norm.mean(dim=-1, keepdim=True) + 1e-6)
+ return self.gamma * (x * stand_div_norm) + self.beta + x
+
+
+class SDCascadeAttnBlock(nn.Module):
+ def __init__(self, c, c_cond, nhead, self_attn=True, dropout=0.0):
+ super().__init__()
+ linear_cls = nn.Linear
+
+ self.self_attn = self_attn
+ self.norm = SDCascadeLayerNorm(c, elementwise_affine=False, eps=1e-6)
+ self.attention = Attention(query_dim=c, heads=nhead, dim_head=c // nhead, dropout=dropout, bias=True)
+ self.kv_mapper = nn.Sequential(nn.SiLU(), linear_cls(c_cond, c))
+
+ def forward(self, x, kv):
+ kv = self.kv_mapper(kv)
+ norm_x = self.norm(x)
+ if self.self_attn:
+ batch_size, channel, _, _ = x.shape
+ kv = torch.cat([norm_x.view(batch_size, channel, -1).transpose(1, 2), kv], dim=1)
+ x = x + self.attention(norm_x, encoder_hidden_states=kv)
+ return x
+
+
+class UpDownBlock2d(nn.Module):
+ def __init__(self, in_channels, out_channels, mode, enabled=True):
+ super().__init__()
+ if mode not in ["up", "down"]:
+ raise ValueError(f"{mode} not supported")
+ interpolation = (
+ nn.Upsample(scale_factor=2 if mode == "up" else 0.5, mode="bilinear", align_corners=True)
+ if enabled
+ else nn.Identity()
+ )
+ mapping = nn.Conv2d(in_channels, out_channels, kernel_size=1)
+ self.blocks = nn.ModuleList([interpolation, mapping] if mode == "up" else [mapping, interpolation])
+
+ def forward(self, x):
+ for block in self.blocks:
+ x = block(x)
+ return x
+
+
+@dataclass
+class StableCascadeUNetOutput(BaseOutput):
+ sample: torch.FloatTensor = None
+
+
+class StableCascadeUNet(ModelMixin, ConfigMixin):
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 16,
+ out_channels: int = 16,
+ timestep_ratio_embedding_dim: int = 64,
+ patch_size: int = 1,
+ conditioning_dim: int = 2048,
+ block_out_channels: Tuple[int] = (2048, 2048),
+ num_attention_heads: Tuple[int] = (32, 32),
+ down_num_layers_per_block: Tuple[int] = (8, 24),
+ up_num_layers_per_block: Tuple[int] = (24, 8),
+ down_blocks_repeat_mappers: Optional[Tuple[int]] = (
+ 1,
+ 1,
+ ),
+ up_blocks_repeat_mappers: Optional[Tuple[int]] = (1, 1),
+ block_types_per_layer: Tuple[Tuple[str]] = (
+ ("SDCascadeResBlock", "SDCascadeTimestepBlock", "SDCascadeAttnBlock"),
+ ("SDCascadeResBlock", "SDCascadeTimestepBlock", "SDCascadeAttnBlock"),
+ ),
+ clip_text_in_channels: Optional[int] = None,
+ clip_text_pooled_in_channels=1280,
+ clip_image_in_channels: Optional[int] = None,
+ clip_seq=4,
+ effnet_in_channels: Optional[int] = None,
+ pixel_mapper_in_channels: Optional[int] = None,
+ kernel_size=3,
+ dropout: Union[float, Tuple[float]] = (0.1, 0.1),
+ self_attn: Union[bool, Tuple[bool]] = True,
+ timestep_conditioning_type: Tuple[str] = ("sca", "crp"),
+ switch_level: Optional[Tuple[bool]] = None,
+ ):
+ """
+
+ Parameters:
+ in_channels (`int`, defaults to 16):
+ Number of channels in the input sample.
+ out_channels (`int`, defaults to 16):
+ Number of channels in the output sample.
+ timestep_ratio_embedding_dim (`int`, defaults to 64):
+ Dimension of the projected time embedding.
+ patch_size (`int`, defaults to 1):
+ Patch size to use for pixel unshuffling layer
+ conditioning_dim (`int`, defaults to 2048):
+ Dimension of the image and text conditional embedding.
+ block_out_channels (Tuple[int], defaults to (2048, 2048)):
+ Tuple of output channels for each block.
+ num_attention_heads (Tuple[int], defaults to (32, 32)):
+ Number of attention heads in each attention block. Set to -1 to if block types in a layer do not have attention.
+ down_num_layers_per_block (Tuple[int], defaults to [8, 24]):
+ Number of layers in each down block.
+ up_num_layers_per_block (Tuple[int], defaults to [24, 8]):
+ Number of layers in each up block.
+ down_blocks_repeat_mappers (Tuple[int], optional, defaults to [1, 1]):
+ Number of 1x1 Convolutional layers to repeat in each down block.
+ up_blocks_repeat_mappers (Tuple[int], optional, defaults to [1, 1]):
+ Number of 1x1 Convolutional layers to repeat in each up block.
+ block_types_per_layer (Tuple[Tuple[str]], optional,
+ defaults to (
+ ("SDCascadeResBlock", "SDCascadeTimestepBlock", "SDCascadeAttnBlock"),
+ ("SDCascadeResBlock", "SDCascadeTimestepBlock", "SDCascadeAttnBlock")
+ ):
+ Block types used in each layer of the up/down blocks.
+ clip_text_in_channels (`int`, *optional*, defaults to `None`):
+ Number of input channels for CLIP based text conditioning.
+ clip_text_pooled_in_channels (`int`, *optional*, defaults to 1280):
+ Number of input channels for pooled CLIP text embeddings.
+ clip_image_in_channels (`int`, *optional*):
+ Number of input channels for CLIP based image conditioning.
+ clip_seq (`int`, *optional*, defaults to 4):
+ effnet_in_channels (`int`, *optional*, defaults to `None`):
+ Number of input channels for effnet conditioning.
+ pixel_mapper_in_channels (`int`, defaults to `None`):
+ Number of input channels for pixel mapper conditioning.
+ kernel_size (`int`, *optional*, defaults to 3):
+ Kernel size to use in the block convolutional layers.
+ dropout (Tuple[float], *optional*, defaults to (0.1, 0.1)):
+ Dropout to use per block.
+ self_attn (Union[bool, Tuple[bool]]):
+ Tuple of booleans that determine whether to use self attention in a block or not.
+ timestep_conditioning_type (Tuple[str], defaults to ("sca", "crp")):
+ Timestep conditioning type.
+ switch_level (Optional[Tuple[bool]], *optional*, defaults to `None`):
+ Tuple that indicates whether upsampling or downsampling should be applied in a block
+ """
+
+ super().__init__()
+
+ if len(block_out_channels) != len(down_num_layers_per_block):
+ raise ValueError(
+ f"Number of elements in `down_num_layers_per_block` must match the length of `block_out_channels`: {len(block_out_channels)}"
+ )
+
+ elif len(block_out_channels) != len(up_num_layers_per_block):
+ raise ValueError(
+ f"Number of elements in `up_num_layers_per_block` must match the length of `block_out_channels`: {len(block_out_channels)}"
+ )
+
+ elif len(block_out_channels) != len(down_blocks_repeat_mappers):
+ raise ValueError(
+ f"Number of elements in `down_blocks_repeat_mappers` must match the length of `block_out_channels`: {len(block_out_channels)}"
+ )
+
+ elif len(block_out_channels) != len(up_blocks_repeat_mappers):
+ raise ValueError(
+ f"Number of elements in `up_blocks_repeat_mappers` must match the length of `block_out_channels`: {len(block_out_channels)}"
+ )
+
+ elif len(block_out_channels) != len(block_types_per_layer):
+ raise ValueError(
+ f"Number of elements in `block_types_per_layer` must match the length of `block_out_channels`: {len(block_out_channels)}"
+ )
+
+ if isinstance(dropout, float):
+ dropout = (dropout,) * len(block_out_channels)
+ if isinstance(self_attn, bool):
+ self_attn = (self_attn,) * len(block_out_channels)
+
+ # CONDITIONING
+ if effnet_in_channels is not None:
+ self.effnet_mapper = nn.Sequential(
+ nn.Conv2d(effnet_in_channels, block_out_channels[0] * 4, kernel_size=1),
+ nn.GELU(),
+ nn.Conv2d(block_out_channels[0] * 4, block_out_channels[0], kernel_size=1),
+ SDCascadeLayerNorm(block_out_channels[0], elementwise_affine=False, eps=1e-6),
+ )
+ if pixel_mapper_in_channels is not None:
+ self.pixels_mapper = nn.Sequential(
+ nn.Conv2d(pixel_mapper_in_channels, block_out_channels[0] * 4, kernel_size=1),
+ nn.GELU(),
+ nn.Conv2d(block_out_channels[0] * 4, block_out_channels[0], kernel_size=1),
+ SDCascadeLayerNorm(block_out_channels[0], elementwise_affine=False, eps=1e-6),
+ )
+
+ self.clip_txt_pooled_mapper = nn.Linear(clip_text_pooled_in_channels, conditioning_dim * clip_seq)
+ if clip_text_in_channels is not None:
+ self.clip_txt_mapper = nn.Linear(clip_text_in_channels, conditioning_dim)
+ if clip_image_in_channels is not None:
+ self.clip_img_mapper = nn.Linear(clip_image_in_channels, conditioning_dim * clip_seq)
+ self.clip_norm = nn.LayerNorm(conditioning_dim, elementwise_affine=False, eps=1e-6)
+
+ self.embedding = nn.Sequential(
+ nn.PixelUnshuffle(patch_size),
+ nn.Conv2d(in_channels * (patch_size**2), block_out_channels[0], kernel_size=1),
+ SDCascadeLayerNorm(block_out_channels[0], elementwise_affine=False, eps=1e-6),
+ )
+
+ def get_block(block_type, in_channels, nhead, c_skip=0, dropout=0, self_attn=True):
+ if block_type == "SDCascadeResBlock":
+ return SDCascadeResBlock(in_channels, c_skip, kernel_size=kernel_size, dropout=dropout)
+ elif block_type == "SDCascadeAttnBlock":
+ return SDCascadeAttnBlock(in_channels, conditioning_dim, nhead, self_attn=self_attn, dropout=dropout)
+ elif block_type == "SDCascadeTimestepBlock":
+ return SDCascadeTimestepBlock(
+ in_channels, timestep_ratio_embedding_dim, conds=timestep_conditioning_type
+ )
+ else:
+ raise ValueError(f"Block type {block_type} not supported")
+
+ # BLOCKS
+ # -- down blocks
+ self.down_blocks = nn.ModuleList()
+ self.down_downscalers = nn.ModuleList()
+ self.down_repeat_mappers = nn.ModuleList()
+ for i in range(len(block_out_channels)):
+ if i > 0:
+ self.down_downscalers.append(
+ nn.Sequential(
+ SDCascadeLayerNorm(block_out_channels[i - 1], elementwise_affine=False, eps=1e-6),
+ UpDownBlock2d(
+ block_out_channels[i - 1], block_out_channels[i], mode="down", enabled=switch_level[i - 1]
+ )
+ if switch_level is not None
+ else nn.Conv2d(block_out_channels[i - 1], block_out_channels[i], kernel_size=2, stride=2),
+ )
+ )
+ else:
+ self.down_downscalers.append(nn.Identity())
+
+ down_block = nn.ModuleList()
+ for _ in range(down_num_layers_per_block[i]):
+ for block_type in block_types_per_layer[i]:
+ block = get_block(
+ block_type,
+ block_out_channels[i],
+ num_attention_heads[i],
+ dropout=dropout[i],
+ self_attn=self_attn[i],
+ )
+ down_block.append(block)
+ self.down_blocks.append(down_block)
+
+ if down_blocks_repeat_mappers is not None:
+ block_repeat_mappers = nn.ModuleList()
+ for _ in range(down_blocks_repeat_mappers[i] - 1):
+ block_repeat_mappers.append(nn.Conv2d(block_out_channels[i], block_out_channels[i], kernel_size=1))
+ self.down_repeat_mappers.append(block_repeat_mappers)
+
+ # -- up blocks
+ self.up_blocks = nn.ModuleList()
+ self.up_upscalers = nn.ModuleList()
+ self.up_repeat_mappers = nn.ModuleList()
+ for i in reversed(range(len(block_out_channels))):
+ if i > 0:
+ self.up_upscalers.append(
+ nn.Sequential(
+ SDCascadeLayerNorm(block_out_channels[i], elementwise_affine=False, eps=1e-6),
+ UpDownBlock2d(
+ block_out_channels[i], block_out_channels[i - 1], mode="up", enabled=switch_level[i - 1]
+ )
+ if switch_level is not None
+ else nn.ConvTranspose2d(
+ block_out_channels[i], block_out_channels[i - 1], kernel_size=2, stride=2
+ ),
+ )
+ )
+ else:
+ self.up_upscalers.append(nn.Identity())
+
+ up_block = nn.ModuleList()
+ for j in range(up_num_layers_per_block[::-1][i]):
+ for k, block_type in enumerate(block_types_per_layer[i]):
+ c_skip = block_out_channels[i] if i < len(block_out_channels) - 1 and j == k == 0 else 0
+ block = get_block(
+ block_type,
+ block_out_channels[i],
+ num_attention_heads[i],
+ c_skip=c_skip,
+ dropout=dropout[i],
+ self_attn=self_attn[i],
+ )
+ up_block.append(block)
+ self.up_blocks.append(up_block)
+
+ if up_blocks_repeat_mappers is not None:
+ block_repeat_mappers = nn.ModuleList()
+ for _ in range(up_blocks_repeat_mappers[::-1][i] - 1):
+ block_repeat_mappers.append(nn.Conv2d(block_out_channels[i], block_out_channels[i], kernel_size=1))
+ self.up_repeat_mappers.append(block_repeat_mappers)
+
+ # OUTPUT
+ self.clf = nn.Sequential(
+ SDCascadeLayerNorm(block_out_channels[0], elementwise_affine=False, eps=1e-6),
+ nn.Conv2d(block_out_channels[0], out_channels * (patch_size**2), kernel_size=1),
+ nn.PixelShuffle(patch_size),
+ )
+
+ self.gradient_checkpointing = False
+
+ def _set_gradient_checkpointing(self, value=False):
+ self.gradient_checkpointing = value
+
+ def _init_weights(self, m):
+ if isinstance(m, (nn.Conv2d, nn.Linear)):
+ torch.nn.init.xavier_uniform_(m.weight)
+ if m.bias is not None:
+ nn.init.constant_(m.bias, 0)
+
+ nn.init.normal_(self.clip_txt_pooled_mapper.weight, std=0.02)
+ nn.init.normal_(self.clip_txt_mapper.weight, std=0.02) if hasattr(self, "clip_txt_mapper") else None
+ nn.init.normal_(self.clip_img_mapper.weight, std=0.02) if hasattr(self, "clip_img_mapper") else None
+
+ if hasattr(self, "effnet_mapper"):
+ nn.init.normal_(self.effnet_mapper[0].weight, std=0.02) # conditionings
+ nn.init.normal_(self.effnet_mapper[2].weight, std=0.02) # conditionings
+
+ if hasattr(self, "pixels_mapper"):
+ nn.init.normal_(self.pixels_mapper[0].weight, std=0.02) # conditionings
+ nn.init.normal_(self.pixels_mapper[2].weight, std=0.02) # conditionings
+
+ torch.nn.init.xavier_uniform_(self.embedding[1].weight, 0.02) # inputs
+ nn.init.constant_(self.clf[1].weight, 0) # outputs
+
+ # blocks
+ for level_block in self.down_blocks + self.up_blocks:
+ for block in level_block:
+ if isinstance(block, SDCascadeResBlock):
+ block.channelwise[-1].weight.data *= np.sqrt(1 / sum(self.config.blocks[0]))
+ elif isinstance(block, SDCascadeTimestepBlock):
+ nn.init.constant_(block.mapper.weight, 0)
+
+ def get_timestep_ratio_embedding(self, timestep_ratio, max_positions=10000):
+ r = timestep_ratio * max_positions
+ half_dim = self.config.timestep_ratio_embedding_dim // 2
+
+ emb = math.log(max_positions) / (half_dim - 1)
+ emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp()
+ emb = r[:, None] * emb[None, :]
+ emb = torch.cat([emb.sin(), emb.cos()], dim=1)
+
+ if self.config.timestep_ratio_embedding_dim % 2 == 1: # zero pad
+ emb = nn.functional.pad(emb, (0, 1), mode="constant")
+
+ return emb.to(dtype=r.dtype)
+
+ def get_clip_embeddings(self, clip_txt_pooled, clip_txt=None, clip_img=None):
+ if len(clip_txt_pooled.shape) == 2:
+ clip_txt_pool = clip_txt_pooled.unsqueeze(1)
+ clip_txt_pool = self.clip_txt_pooled_mapper(clip_txt_pooled).view(
+ clip_txt_pooled.size(0), clip_txt_pooled.size(1) * self.config.clip_seq, -1
+ )
+ if clip_txt is not None and clip_img is not None:
+ clip_txt = self.clip_txt_mapper(clip_txt)
+ if len(clip_img.shape) == 2:
+ clip_img = clip_img.unsqueeze(1)
+ clip_img = self.clip_img_mapper(clip_img).view(
+ clip_img.size(0), clip_img.size(1) * self.config.clip_seq, -1
+ )
+ clip = torch.cat([clip_txt, clip_txt_pool, clip_img], dim=1)
+ else:
+ clip = clip_txt_pool
+ return self.clip_norm(clip)
+
+ def _down_encode(self, x, r_embed, clip):
+ level_outputs = []
+ block_group = zip(self.down_blocks, self.down_downscalers, self.down_repeat_mappers)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ for down_block, downscaler, repmap in block_group:
+ x = downscaler(x)
+ for i in range(len(repmap) + 1):
+ for block in down_block:
+ if isinstance(block, SDCascadeResBlock):
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(block), x, use_reentrant=False)
+ elif isinstance(block, SDCascadeAttnBlock):
+ x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block), x, clip, use_reentrant=False
+ )
+ elif isinstance(block, SDCascadeTimestepBlock):
+ x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block), x, r_embed, use_reentrant=False
+ )
+ else:
+ x = x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block), use_reentrant=False
+ )
+ if i < len(repmap):
+ x = repmap[i](x)
+ level_outputs.insert(0, x)
+ else:
+ for down_block, downscaler, repmap in block_group:
+ x = downscaler(x)
+ for i in range(len(repmap) + 1):
+ for block in down_block:
+ if isinstance(block, SDCascadeResBlock):
+ x = block(x)
+ elif isinstance(block, SDCascadeAttnBlock):
+ x = block(x, clip)
+ elif isinstance(block, SDCascadeTimestepBlock):
+ x = block(x, r_embed)
+ else:
+ x = block(x)
+ if i < len(repmap):
+ x = repmap[i](x)
+ level_outputs.insert(0, x)
+ return level_outputs
+
+ def _up_decode(self, level_outputs, r_embed, clip):
+ x = level_outputs[0]
+ block_group = zip(self.up_blocks, self.up_upscalers, self.up_repeat_mappers)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ for i, (up_block, upscaler, repmap) in enumerate(block_group):
+ for j in range(len(repmap) + 1):
+ for k, block in enumerate(up_block):
+ if isinstance(block, SDCascadeResBlock):
+ skip = level_outputs[i] if k == 0 and i > 0 else None
+ if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
+ x = torch.nn.functional.interpolate(
+ x.float(), skip.shape[-2:], mode="bilinear", align_corners=True
+ )
+ x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block), x, skip, use_reentrant=False
+ )
+ elif isinstance(block, SDCascadeAttnBlock):
+ x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block), x, clip, use_reentrant=False
+ )
+ elif isinstance(block, SDCascadeTimestepBlock):
+ x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block), x, r_embed, use_reentrant=False
+ )
+ else:
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(block), x, use_reentrant=False)
+ if j < len(repmap):
+ x = repmap[j](x)
+ x = upscaler(x)
+ else:
+ for i, (up_block, upscaler, repmap) in enumerate(block_group):
+ for j in range(len(repmap) + 1):
+ for k, block in enumerate(up_block):
+ if isinstance(block, SDCascadeResBlock):
+ skip = level_outputs[i] if k == 0 and i > 0 else None
+ if skip is not None and (x.size(-1) != skip.size(-1) or x.size(-2) != skip.size(-2)):
+ x = torch.nn.functional.interpolate(
+ x.float(), skip.shape[-2:], mode="bilinear", align_corners=True
+ )
+ x = block(x, skip)
+ elif isinstance(block, SDCascadeAttnBlock):
+ x = block(x, clip)
+ elif isinstance(block, SDCascadeTimestepBlock):
+ x = block(x, r_embed)
+ else:
+ x = block(x)
+ if j < len(repmap):
+ x = repmap[j](x)
+ x = upscaler(x)
+ return x
+
+ def forward(
+ self,
+ sample,
+ timestep_ratio,
+ clip_text_pooled,
+ clip_text=None,
+ clip_img=None,
+ effnet=None,
+ pixels=None,
+ sca=None,
+ crp=None,
+ return_dict=True,
+ ):
+ if pixels is None:
+ pixels = sample.new_zeros(sample.size(0), 3, 8, 8)
+
+ # Process the conditioning embeddings
+ timestep_ratio_embed = self.get_timestep_ratio_embedding(timestep_ratio)
+ for c in self.config.timestep_conditioning_type:
+ if c == "sca":
+ cond = sca
+ elif c == "crp":
+ cond = crp
+ else:
+ cond = None
+ t_cond = cond or torch.zeros_like(timestep_ratio)
+ timestep_ratio_embed = torch.cat([timestep_ratio_embed, self.get_timestep_ratio_embedding(t_cond)], dim=1)
+ clip = self.get_clip_embeddings(clip_txt_pooled=clip_text_pooled, clip_txt=clip_text, clip_img=clip_img)
+
+ # Model Blocks
+ x = self.embedding(sample)
+ if hasattr(self, "effnet_mapper") and effnet is not None:
+ x = x + self.effnet_mapper(
+ nn.functional.interpolate(effnet, size=x.shape[-2:], mode="bilinear", align_corners=True)
+ )
+ if hasattr(self, "pixels_mapper"):
+ x = x + nn.functional.interpolate(
+ self.pixels_mapper(pixels), size=x.shape[-2:], mode="bilinear", align_corners=True
+ )
+ level_outputs = self._down_encode(x, timestep_ratio_embed, clip)
+ x = self._up_decode(level_outputs, timestep_ratio_embed, clip)
+ sample = self.clf(x)
+
+ if not return_dict:
+ return (sample,)
+ return StableCascadeUNetOutput(sample=sample)
diff --git a/model/diffusers_c/models/unets/uvit_2d.py b/model/diffusers_c/models/unets/uvit_2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..1e2d8cc0c9a7619662d6f69dab7648013a979c2e
--- /dev/null
+++ b/model/diffusers_c/models/unets/uvit_2d.py
@@ -0,0 +1,469 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Dict, Union
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.utils.checkpoint import checkpoint
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin
+from ..attention import BasicTransformerBlock, SkipFFTransformerBlock
+from ..attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ..embeddings import TimestepEmbedding, get_timestep_embedding
+from ..modeling_utils import ModelMixin
+from ..normalization import GlobalResponseNorm, RMSNorm
+from ..resnet import Downsample2D, Upsample2D
+
+
+class UVit2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ # global config
+ hidden_size: int = 1024,
+ use_bias: bool = False,
+ hidden_dropout: float = 0.0,
+ # conditioning dimensions
+ cond_embed_dim: int = 768,
+ micro_cond_encode_dim: int = 256,
+ micro_cond_embed_dim: int = 1280,
+ encoder_hidden_size: int = 768,
+ # num tokens
+ vocab_size: int = 8256, # codebook_size + 1 (for the mask token) rounded
+ codebook_size: int = 8192,
+ # `UVit2DConvEmbed`
+ in_channels: int = 768,
+ block_out_channels: int = 768,
+ num_res_blocks: int = 3,
+ downsample: bool = False,
+ upsample: bool = False,
+ block_num_heads: int = 12,
+ # `TransformerLayer`
+ num_hidden_layers: int = 22,
+ num_attention_heads: int = 16,
+ # `Attention`
+ attention_dropout: float = 0.0,
+ # `FeedForward`
+ intermediate_size: int = 2816,
+ # `Norm`
+ layer_norm_eps: float = 1e-6,
+ ln_elementwise_affine: bool = True,
+ sample_size: int = 64,
+ ):
+ super().__init__()
+
+ self.encoder_proj = nn.Linear(encoder_hidden_size, hidden_size, bias=use_bias)
+ self.encoder_proj_layer_norm = RMSNorm(hidden_size, layer_norm_eps, ln_elementwise_affine)
+
+ self.embed = UVit2DConvEmbed(
+ in_channels, block_out_channels, vocab_size, ln_elementwise_affine, layer_norm_eps, use_bias
+ )
+
+ self.cond_embed = TimestepEmbedding(
+ micro_cond_embed_dim + cond_embed_dim, hidden_size, sample_proj_bias=use_bias
+ )
+
+ self.down_block = UVitBlock(
+ block_out_channels,
+ num_res_blocks,
+ hidden_size,
+ hidden_dropout,
+ ln_elementwise_affine,
+ layer_norm_eps,
+ use_bias,
+ block_num_heads,
+ attention_dropout,
+ downsample,
+ False,
+ )
+
+ self.project_to_hidden_norm = RMSNorm(block_out_channels, layer_norm_eps, ln_elementwise_affine)
+ self.project_to_hidden = nn.Linear(block_out_channels, hidden_size, bias=use_bias)
+
+ self.transformer_layers = nn.ModuleList(
+ [
+ BasicTransformerBlock(
+ dim=hidden_size,
+ num_attention_heads=num_attention_heads,
+ attention_head_dim=hidden_size // num_attention_heads,
+ dropout=hidden_dropout,
+ cross_attention_dim=hidden_size,
+ attention_bias=use_bias,
+ norm_type="ada_norm_continuous",
+ ada_norm_continous_conditioning_embedding_dim=hidden_size,
+ norm_elementwise_affine=ln_elementwise_affine,
+ norm_eps=layer_norm_eps,
+ ada_norm_bias=use_bias,
+ ff_inner_dim=intermediate_size,
+ ff_bias=use_bias,
+ attention_out_bias=use_bias,
+ )
+ for _ in range(num_hidden_layers)
+ ]
+ )
+
+ self.project_from_hidden_norm = RMSNorm(hidden_size, layer_norm_eps, ln_elementwise_affine)
+ self.project_from_hidden = nn.Linear(hidden_size, block_out_channels, bias=use_bias)
+
+ self.up_block = UVitBlock(
+ block_out_channels,
+ num_res_blocks,
+ hidden_size,
+ hidden_dropout,
+ ln_elementwise_affine,
+ layer_norm_eps,
+ use_bias,
+ block_num_heads,
+ attention_dropout,
+ downsample=False,
+ upsample=upsample,
+ )
+
+ self.mlm_layer = ConvMlmLayer(
+ block_out_channels, in_channels, use_bias, ln_elementwise_affine, layer_norm_eps, codebook_size
+ )
+
+ self.gradient_checkpointing = False
+
+ def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ pass
+
+ def forward(self, input_ids, encoder_hidden_states, pooled_text_emb, micro_conds, cross_attention_kwargs=None):
+ encoder_hidden_states = self.encoder_proj(encoder_hidden_states)
+ encoder_hidden_states = self.encoder_proj_layer_norm(encoder_hidden_states)
+
+ micro_cond_embeds = get_timestep_embedding(
+ micro_conds.flatten(), self.config.micro_cond_encode_dim, flip_sin_to_cos=True, downscale_freq_shift=0
+ )
+
+ micro_cond_embeds = micro_cond_embeds.reshape((input_ids.shape[0], -1))
+
+ pooled_text_emb = torch.cat([pooled_text_emb, micro_cond_embeds], dim=1)
+ pooled_text_emb = pooled_text_emb.to(dtype=self.dtype)
+ pooled_text_emb = self.cond_embed(pooled_text_emb).to(encoder_hidden_states.dtype)
+
+ hidden_states = self.embed(input_ids)
+
+ hidden_states = self.down_block(
+ hidden_states,
+ pooled_text_emb=pooled_text_emb,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+
+ batch_size, channels, height, width = hidden_states.shape
+ hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch_size, height * width, channels)
+
+ hidden_states = self.project_to_hidden_norm(hidden_states)
+ hidden_states = self.project_to_hidden(hidden_states)
+
+ for layer in self.transformer_layers:
+ if self.training and self.gradient_checkpointing:
+
+ def layer_(*args):
+ return checkpoint(layer, *args)
+
+ else:
+ layer_ = layer
+
+ hidden_states = layer_(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs={"pooled_text_emb": pooled_text_emb},
+ )
+
+ hidden_states = self.project_from_hidden_norm(hidden_states)
+ hidden_states = self.project_from_hidden(hidden_states)
+
+ hidden_states = hidden_states.reshape(batch_size, height, width, channels).permute(0, 3, 1, 2)
+
+ hidden_states = self.up_block(
+ hidden_states,
+ pooled_text_emb=pooled_text_emb,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+
+ logits = self.mlm_layer(hidden_states)
+
+ return logits
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+
+class UVit2DConvEmbed(nn.Module):
+ def __init__(self, in_channels, block_out_channels, vocab_size, elementwise_affine, eps, bias):
+ super().__init__()
+ self.embeddings = nn.Embedding(vocab_size, in_channels)
+ self.layer_norm = RMSNorm(in_channels, eps, elementwise_affine)
+ self.conv = nn.Conv2d(in_channels, block_out_channels, kernel_size=1, bias=bias)
+
+ def forward(self, input_ids):
+ embeddings = self.embeddings(input_ids)
+ embeddings = self.layer_norm(embeddings)
+ embeddings = embeddings.permute(0, 3, 1, 2)
+ embeddings = self.conv(embeddings)
+ return embeddings
+
+
+class UVitBlock(nn.Module):
+ def __init__(
+ self,
+ channels,
+ num_res_blocks: int,
+ hidden_size,
+ hidden_dropout,
+ ln_elementwise_affine,
+ layer_norm_eps,
+ use_bias,
+ block_num_heads,
+ attention_dropout,
+ downsample: bool,
+ upsample: bool,
+ ):
+ super().__init__()
+
+ if downsample:
+ self.downsample = Downsample2D(
+ channels,
+ use_conv=True,
+ padding=0,
+ name="Conv2d_0",
+ kernel_size=2,
+ norm_type="rms_norm",
+ eps=layer_norm_eps,
+ elementwise_affine=ln_elementwise_affine,
+ bias=use_bias,
+ )
+ else:
+ self.downsample = None
+
+ self.res_blocks = nn.ModuleList(
+ [
+ ConvNextBlock(
+ channels,
+ layer_norm_eps,
+ ln_elementwise_affine,
+ use_bias,
+ hidden_dropout,
+ hidden_size,
+ )
+ for i in range(num_res_blocks)
+ ]
+ )
+
+ self.attention_blocks = nn.ModuleList(
+ [
+ SkipFFTransformerBlock(
+ channels,
+ block_num_heads,
+ channels // block_num_heads,
+ hidden_size,
+ use_bias,
+ attention_dropout,
+ channels,
+ attention_bias=use_bias,
+ attention_out_bias=use_bias,
+ )
+ for _ in range(num_res_blocks)
+ ]
+ )
+
+ if upsample:
+ self.upsample = Upsample2D(
+ channels,
+ use_conv_transpose=True,
+ kernel_size=2,
+ padding=0,
+ name="conv",
+ norm_type="rms_norm",
+ eps=layer_norm_eps,
+ elementwise_affine=ln_elementwise_affine,
+ bias=use_bias,
+ interpolate=False,
+ )
+ else:
+ self.upsample = None
+
+ def forward(self, x, pooled_text_emb, encoder_hidden_states, cross_attention_kwargs):
+ if self.downsample is not None:
+ x = self.downsample(x)
+
+ for res_block, attention_block in zip(self.res_blocks, self.attention_blocks):
+ x = res_block(x, pooled_text_emb)
+
+ batch_size, channels, height, width = x.shape
+ x = x.view(batch_size, channels, height * width).permute(0, 2, 1)
+ x = attention_block(
+ x, encoder_hidden_states=encoder_hidden_states, cross_attention_kwargs=cross_attention_kwargs
+ )
+ x = x.permute(0, 2, 1).view(batch_size, channels, height, width)
+
+ if self.upsample is not None:
+ x = self.upsample(x)
+
+ return x
+
+
+class ConvNextBlock(nn.Module):
+ def __init__(
+ self, channels, layer_norm_eps, ln_elementwise_affine, use_bias, hidden_dropout, hidden_size, res_ffn_factor=4
+ ):
+ super().__init__()
+ self.depthwise = nn.Conv2d(
+ channels,
+ channels,
+ kernel_size=3,
+ padding=1,
+ groups=channels,
+ bias=use_bias,
+ )
+ self.norm = RMSNorm(channels, layer_norm_eps, ln_elementwise_affine)
+ self.channelwise_linear_1 = nn.Linear(channels, int(channels * res_ffn_factor), bias=use_bias)
+ self.channelwise_act = nn.GELU()
+ self.channelwise_norm = GlobalResponseNorm(int(channels * res_ffn_factor))
+ self.channelwise_linear_2 = nn.Linear(int(channels * res_ffn_factor), channels, bias=use_bias)
+ self.channelwise_dropout = nn.Dropout(hidden_dropout)
+ self.cond_embeds_mapper = nn.Linear(hidden_size, channels * 2, use_bias)
+
+ def forward(self, x, cond_embeds):
+ x_res = x
+
+ x = self.depthwise(x)
+
+ x = x.permute(0, 2, 3, 1)
+ x = self.norm(x)
+
+ x = self.channelwise_linear_1(x)
+ x = self.channelwise_act(x)
+ x = self.channelwise_norm(x)
+ x = self.channelwise_linear_2(x)
+ x = self.channelwise_dropout(x)
+
+ x = x.permute(0, 3, 1, 2)
+
+ x = x + x_res
+
+ scale, shift = self.cond_embeds_mapper(F.silu(cond_embeds)).chunk(2, dim=1)
+ x = x * (1 + scale[:, :, None, None]) + shift[:, :, None, None]
+
+ return x
+
+
+class ConvMlmLayer(nn.Module):
+ def __init__(
+ self,
+ block_out_channels: int,
+ in_channels: int,
+ use_bias: bool,
+ ln_elementwise_affine: bool,
+ layer_norm_eps: float,
+ codebook_size: int,
+ ):
+ super().__init__()
+ self.conv1 = nn.Conv2d(block_out_channels, in_channels, kernel_size=1, bias=use_bias)
+ self.layer_norm = RMSNorm(in_channels, layer_norm_eps, ln_elementwise_affine)
+ self.conv2 = nn.Conv2d(in_channels, codebook_size, kernel_size=1, bias=use_bias)
+
+ def forward(self, hidden_states):
+ hidden_states = self.conv1(hidden_states)
+ hidden_states = self.layer_norm(hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+ logits = self.conv2(hidden_states)
+ return logits
diff --git a/model/diffusers_c/models/upsampling.py b/model/diffusers_c/models/upsampling.py
new file mode 100644
index 0000000000000000000000000000000000000000..080103504c53765a74b92ec605ee8f6d4cfbf428
--- /dev/null
+++ b/model/diffusers_c/models/upsampling.py
@@ -0,0 +1,454 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..utils import USE_PEFT_BACKEND
+from .lora import LoRACompatibleConv
+from .normalization import RMSNorm
+
+
+class Upsample1D(nn.Module):
+ """A 1D upsampling layer with an optional convolution.
+
+ Parameters:
+ channels (`int`):
+ number of channels in the inputs and outputs.
+ use_conv (`bool`, default `False`):
+ option to use a convolution.
+ use_conv_transpose (`bool`, default `False`):
+ option to use a convolution transpose.
+ out_channels (`int`, optional):
+ number of output channels. Defaults to `channels`.
+ name (`str`, default `conv`):
+ name of the upsampling 1D layer.
+ """
+
+ def __init__(
+ self,
+ channels: int,
+ use_conv: bool = False,
+ use_conv_transpose: bool = False,
+ out_channels: Optional[int] = None,
+ name: str = "conv",
+ ):
+ super().__init__()
+ self.channels = channels
+ self.out_channels = out_channels or channels
+ self.use_conv = use_conv
+ self.use_conv_transpose = use_conv_transpose
+ self.name = name
+
+ self.conv = None
+ if use_conv_transpose:
+ self.conv = nn.ConvTranspose1d(channels, self.out_channels, 4, 2, 1)
+ elif use_conv:
+ self.conv = nn.Conv1d(self.channels, self.out_channels, 3, padding=1)
+
+ def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+ assert inputs.shape[1] == self.channels
+ if self.use_conv_transpose:
+ return self.conv(inputs)
+
+ outputs = F.interpolate(inputs, scale_factor=2.0, mode="nearest")
+
+ if self.use_conv:
+ outputs = self.conv(outputs)
+
+ return outputs
+
+
+class Upsample2D(nn.Module):
+ """A 2D upsampling layer with an optional convolution.
+
+ Parameters:
+ channels (`int`):
+ number of channels in the inputs and outputs.
+ use_conv (`bool`, default `False`):
+ option to use a convolution.
+ use_conv_transpose (`bool`, default `False`):
+ option to use a convolution transpose.
+ out_channels (`int`, optional):
+ number of output channels. Defaults to `channels`.
+ name (`str`, default `conv`):
+ name of the upsampling 2D layer.
+ """
+
+ def __init__(
+ self,
+ channels: int,
+ use_conv: bool = False,
+ use_conv_transpose: bool = False,
+ out_channels: Optional[int] = None,
+ name: str = "conv",
+ kernel_size: Optional[int] = None,
+ padding=1,
+ norm_type=None,
+ eps=None,
+ elementwise_affine=None,
+ bias=True,
+ interpolate=True,
+ ):
+ super().__init__()
+ self.channels = channels
+ self.out_channels = out_channels or channels
+ self.use_conv = use_conv
+ self.use_conv_transpose = use_conv_transpose
+ self.name = name
+ self.interpolate = interpolate
+ conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+
+ if norm_type == "ln_norm":
+ self.norm = nn.LayerNorm(channels, eps, elementwise_affine)
+ elif norm_type == "rms_norm":
+ self.norm = RMSNorm(channels, eps, elementwise_affine)
+ elif norm_type is None:
+ self.norm = None
+ else:
+ raise ValueError(f"unknown norm_type: {norm_type}")
+
+ conv = None
+ if use_conv_transpose:
+ if kernel_size is None:
+ kernel_size = 4
+ conv = nn.ConvTranspose2d(
+ channels, self.out_channels, kernel_size=kernel_size, stride=2, padding=padding, bias=bias
+ )
+ elif use_conv:
+ if kernel_size is None:
+ kernel_size = 3
+ conv = conv_cls(self.channels, self.out_channels, kernel_size=kernel_size, padding=padding, bias=bias)
+
+ # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+ if name == "conv":
+ self.conv = conv
+ else:
+ self.Conv2d_0 = conv
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ output_size: Optional[int] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ assert hidden_states.shape[1] == self.channels
+
+ if self.norm is not None:
+ hidden_states = self.norm(hidden_states.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+
+ if self.use_conv_transpose:
+ return self.conv(hidden_states)
+
+ # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+ # TODO(Suraj): Remove this cast once the issue is fixed in PyTorch
+ # https://github.com/pytorch/pytorch/issues/86679
+ dtype = hidden_states.dtype
+ if dtype == torch.bfloat16:
+ hidden_states = hidden_states.to(torch.float32)
+
+ # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+ if hidden_states.shape[0] >= 64:
+ hidden_states = hidden_states.contiguous()
+
+ # if `output_size` is passed we force the interpolation output
+ # size and do not make use of `scale_factor=2`
+ if self.interpolate:
+ if output_size is None:
+ hidden_states = F.interpolate(hidden_states, scale_factor=2.0, mode="nearest")
+ else:
+ hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
+
+ # If the input is bfloat16, we cast back to bfloat16
+ if dtype == torch.bfloat16:
+ hidden_states = hidden_states.to(dtype)
+
+ # TODO(Suraj, Patrick) - clean up after weight dicts are correctly renamed
+ if self.use_conv:
+ if self.name == "conv":
+ if isinstance(self.conv, LoRACompatibleConv) and not USE_PEFT_BACKEND:
+ hidden_states = self.conv(hidden_states, scale)
+ else:
+ hidden_states = self.conv(hidden_states)
+ else:
+ if isinstance(self.Conv2d_0, LoRACompatibleConv) and not USE_PEFT_BACKEND:
+ hidden_states = self.Conv2d_0(hidden_states, scale)
+ else:
+ hidden_states = self.Conv2d_0(hidden_states)
+
+ return hidden_states
+
+
+class FirUpsample2D(nn.Module):
+ """A 2D FIR upsampling layer with an optional convolution.
+
+ Parameters:
+ channels (`int`, optional):
+ number of channels in the inputs and outputs.
+ use_conv (`bool`, default `False`):
+ option to use a convolution.
+ out_channels (`int`, optional):
+ number of output channels. Defaults to `channels`.
+ fir_kernel (`tuple`, default `(1, 3, 3, 1)`):
+ kernel for the FIR filter.
+ """
+
+ def __init__(
+ self,
+ channels: Optional[int] = None,
+ out_channels: Optional[int] = None,
+ use_conv: bool = False,
+ fir_kernel: Tuple[int, int, int, int] = (1, 3, 3, 1),
+ ):
+ super().__init__()
+ out_channels = out_channels if out_channels else channels
+ if use_conv:
+ self.Conv2d_0 = nn.Conv2d(channels, out_channels, kernel_size=3, stride=1, padding=1)
+ self.use_conv = use_conv
+ self.fir_kernel = fir_kernel
+ self.out_channels = out_channels
+
+ def _upsample_2d(
+ self,
+ hidden_states: torch.FloatTensor,
+ weight: Optional[torch.FloatTensor] = None,
+ kernel: Optional[torch.FloatTensor] = None,
+ factor: int = 2,
+ gain: float = 1,
+ ) -> torch.FloatTensor:
+ """Fused `upsample_2d()` followed by `Conv2d()`.
+
+ Padding is performed only once at the beginning, not between the operations. The fused op is considerably more
+ efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of
+ arbitrary order.
+
+ Args:
+ hidden_states (`torch.FloatTensor`):
+ Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+ weight (`torch.FloatTensor`, *optional*):
+ Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be
+ performed by `inChannels = x.shape[0] // numGroups`.
+ kernel (`torch.FloatTensor`, *optional*):
+ FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+ corresponds to nearest-neighbor upsampling.
+ factor (`int`, *optional*): Integer upsampling factor (default: 2).
+ gain (`float`, *optional*): Scaling factor for signal magnitude (default: 1.0).
+
+ Returns:
+ output (`torch.FloatTensor`):
+ Tensor of the shape `[N, C, H * factor, W * factor]` or `[N, H * factor, W * factor, C]`, and same
+ datatype as `hidden_states`.
+ """
+
+ assert isinstance(factor, int) and factor >= 1
+
+ # Setup filter kernel.
+ if kernel is None:
+ kernel = [1] * factor
+
+ # setup kernel
+ kernel = torch.tensor(kernel, dtype=torch.float32)
+ if kernel.ndim == 1:
+ kernel = torch.outer(kernel, kernel)
+ kernel /= torch.sum(kernel)
+
+ kernel = kernel * (gain * (factor**2))
+
+ if self.use_conv:
+ convH = weight.shape[2]
+ convW = weight.shape[3]
+ inC = weight.shape[1]
+
+ pad_value = (kernel.shape[0] - factor) - (convW - 1)
+
+ stride = (factor, factor)
+ # Determine data dimensions.
+ output_shape = (
+ (hidden_states.shape[2] - 1) * factor + convH,
+ (hidden_states.shape[3] - 1) * factor + convW,
+ )
+ output_padding = (
+ output_shape[0] - (hidden_states.shape[2] - 1) * stride[0] - convH,
+ output_shape[1] - (hidden_states.shape[3] - 1) * stride[1] - convW,
+ )
+ assert output_padding[0] >= 0 and output_padding[1] >= 0
+ num_groups = hidden_states.shape[1] // inC
+
+ # Transpose weights.
+ weight = torch.reshape(weight, (num_groups, -1, inC, convH, convW))
+ weight = torch.flip(weight, dims=[3, 4]).permute(0, 2, 1, 3, 4)
+ weight = torch.reshape(weight, (num_groups * inC, -1, convH, convW))
+
+ inverse_conv = F.conv_transpose2d(
+ hidden_states,
+ weight,
+ stride=stride,
+ output_padding=output_padding,
+ padding=0,
+ )
+
+ output = upfirdn2d_native(
+ inverse_conv,
+ torch.tensor(kernel, device=inverse_conv.device),
+ pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2 + 1),
+ )
+ else:
+ pad_value = kernel.shape[0] - factor
+ output = upfirdn2d_native(
+ hidden_states,
+ torch.tensor(kernel, device=hidden_states.device),
+ up=factor,
+ pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
+ )
+
+ return output
+
+ def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
+ if self.use_conv:
+ height = self._upsample_2d(hidden_states, self.Conv2d_0.weight, kernel=self.fir_kernel)
+ height = height + self.Conv2d_0.bias.reshape(1, -1, 1, 1)
+ else:
+ height = self._upsample_2d(hidden_states, kernel=self.fir_kernel, factor=2)
+
+ return height
+
+
+class KUpsample2D(nn.Module):
+ r"""A 2D K-upsampling layer.
+
+ Parameters:
+ pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use.
+ """
+
+ def __init__(self, pad_mode: str = "reflect"):
+ super().__init__()
+ self.pad_mode = pad_mode
+ kernel_1d = torch.tensor([[1 / 8, 3 / 8, 3 / 8, 1 / 8]]) * 2
+ self.pad = kernel_1d.shape[1] // 2 - 1
+ self.register_buffer("kernel", kernel_1d.T @ kernel_1d, persistent=False)
+
+ def forward(self, inputs: torch.Tensor) -> torch.Tensor:
+ inputs = F.pad(inputs, ((self.pad + 1) // 2,) * 4, self.pad_mode)
+ weight = inputs.new_zeros(
+ [
+ inputs.shape[1],
+ inputs.shape[1],
+ self.kernel.shape[0],
+ self.kernel.shape[1],
+ ]
+ )
+ indices = torch.arange(inputs.shape[1], device=inputs.device)
+ kernel = self.kernel.to(weight)[None, :].expand(inputs.shape[1], -1, -1)
+ weight[indices, indices] = kernel
+ return F.conv_transpose2d(inputs, weight, stride=2, padding=self.pad * 2 + 1)
+
+
+def upfirdn2d_native(
+ tensor: torch.Tensor,
+ kernel: torch.Tensor,
+ up: int = 1,
+ down: int = 1,
+ pad: Tuple[int, int] = (0, 0),
+) -> torch.Tensor:
+ up_x = up_y = up
+ down_x = down_y = down
+ pad_x0 = pad_y0 = pad[0]
+ pad_x1 = pad_y1 = pad[1]
+
+ _, channel, in_h, in_w = tensor.shape
+ tensor = tensor.reshape(-1, in_h, in_w, 1)
+
+ _, in_h, in_w, minor = tensor.shape
+ kernel_h, kernel_w = kernel.shape
+
+ out = tensor.view(-1, in_h, 1, in_w, 1, minor)
+ out = F.pad(out, [0, 0, 0, up_x - 1, 0, 0, 0, up_y - 1])
+ out = out.view(-1, in_h * up_y, in_w * up_x, minor)
+
+ out = F.pad(out, [0, 0, max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
+ out = out.to(tensor.device) # Move back to mps if necessary
+ out = out[
+ :,
+ max(-pad_y0, 0) : out.shape[1] - max(-pad_y1, 0),
+ max(-pad_x0, 0) : out.shape[2] - max(-pad_x1, 0),
+ :,
+ ]
+
+ out = out.permute(0, 3, 1, 2)
+ out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
+ w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
+ out = F.conv2d(out, w)
+ out = out.reshape(
+ -1,
+ minor,
+ in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
+ in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1,
+ )
+ out = out.permute(0, 2, 3, 1)
+ out = out[:, ::down_y, ::down_x, :]
+
+ out_h = (in_h * up_y + pad_y0 + pad_y1 - kernel_h) // down_y + 1
+ out_w = (in_w * up_x + pad_x0 + pad_x1 - kernel_w) // down_x + 1
+
+ return out.view(-1, channel, out_h, out_w)
+
+
+def upsample_2d(
+ hidden_states: torch.FloatTensor,
+ kernel: Optional[torch.FloatTensor] = None,
+ factor: int = 2,
+ gain: float = 1,
+) -> torch.FloatTensor:
+ r"""Upsample2D a batch of 2D images with the given filter.
+ Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and upsamples each image with the given
+ filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the specified
+ `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its shape is
+ a: multiple of the upsampling factor.
+
+ Args:
+ hidden_states (`torch.FloatTensor`):
+ Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`.
+ kernel (`torch.FloatTensor`, *optional*):
+ FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which
+ corresponds to nearest-neighbor upsampling.
+ factor (`int`, *optional*, default to `2`):
+ Integer upsampling factor.
+ gain (`float`, *optional*, default to `1.0`):
+ Scaling factor for signal magnitude (default: 1.0).
+
+ Returns:
+ output (`torch.FloatTensor`):
+ Tensor of the shape `[N, C, H * factor, W * factor]`
+ """
+ assert isinstance(factor, int) and factor >= 1
+ if kernel is None:
+ kernel = [1] * factor
+
+ kernel = torch.tensor(kernel, dtype=torch.float32)
+ if kernel.ndim == 1:
+ kernel = torch.outer(kernel, kernel)
+ kernel /= torch.sum(kernel)
+
+ kernel = kernel * (gain * (factor**2))
+ pad_value = kernel.shape[0] - factor
+ output = upfirdn2d_native(
+ hidden_states,
+ kernel.to(device=hidden_states.device),
+ up=factor,
+ pad=((pad_value + 1) // 2 + factor - 1, pad_value // 2),
+ )
+ return output
diff --git a/model/diffusers_c/models/vae_flax.py b/model/diffusers_c/models/vae_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..5027f4230e3b44875b1a2289be81b8ffbf92afef
--- /dev/null
+++ b/model/diffusers_c/models/vae_flax.py
@@ -0,0 +1,876 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# JAX implementation of VQGAN from taming-transformers https://github.com/CompVis/taming-transformers
+
+import math
+from functools import partial
+from typing import Tuple
+
+import flax
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict
+
+from ..configuration_utils import ConfigMixin, flax_register_to_config
+from ..utils import BaseOutput
+from .modeling_flax_utils import FlaxModelMixin
+
+
+@flax.struct.dataclass
+class FlaxDecoderOutput(BaseOutput):
+ """
+ Output of decoding method.
+
+ Args:
+ sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)`):
+ The decoded output sample from the last layer of the model.
+ dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`):
+ The `dtype` of the parameters.
+ """
+
+ sample: jnp.ndarray
+
+
+@flax.struct.dataclass
+class FlaxAutoencoderKLOutput(BaseOutput):
+ """
+ Output of AutoencoderKL encoding method.
+
+ Args:
+ latent_dist (`FlaxDiagonalGaussianDistribution`):
+ Encoded outputs of `Encoder` represented as the mean and logvar of `FlaxDiagonalGaussianDistribution`.
+ `FlaxDiagonalGaussianDistribution` allows for sampling latents from the distribution.
+ """
+
+ latent_dist: "FlaxDiagonalGaussianDistribution"
+
+
+class FlaxUpsample2D(nn.Module):
+ """
+ Flax implementation of 2D Upsample layer
+
+ Args:
+ in_channels (`int`):
+ Input channels
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ self.conv = nn.Conv(
+ self.in_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ def __call__(self, hidden_states):
+ batch, height, width, channels = hidden_states.shape
+ hidden_states = jax.image.resize(
+ hidden_states,
+ shape=(batch, height * 2, width * 2, channels),
+ method="nearest",
+ )
+ hidden_states = self.conv(hidden_states)
+ return hidden_states
+
+
+class FlaxDownsample2D(nn.Module):
+ """
+ Flax implementation of 2D Downsample layer
+
+ Args:
+ in_channels (`int`):
+ Input channels
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ self.conv = nn.Conv(
+ self.in_channels,
+ kernel_size=(3, 3),
+ strides=(2, 2),
+ padding="VALID",
+ dtype=self.dtype,
+ )
+
+ def __call__(self, hidden_states):
+ pad = ((0, 0), (0, 1), (0, 1), (0, 0)) # pad height and width dim
+ hidden_states = jnp.pad(hidden_states, pad_width=pad)
+ hidden_states = self.conv(hidden_states)
+ return hidden_states
+
+
+class FlaxResnetBlock2D(nn.Module):
+ """
+ Flax implementation of 2D Resnet Block.
+
+ Args:
+ in_channels (`int`):
+ Input channels
+ out_channels (`int`):
+ Output channels
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ groups (:obj:`int`, *optional*, defaults to `32`):
+ The number of groups to use for group norm.
+ use_nin_shortcut (:obj:`bool`, *optional*, defaults to `None`):
+ Whether to use `nin_shortcut`. This activates a new layer inside ResNet block
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ out_channels: int = None
+ dropout: float = 0.0
+ groups: int = 32
+ use_nin_shortcut: bool = None
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ out_channels = self.in_channels if self.out_channels is None else self.out_channels
+
+ self.norm1 = nn.GroupNorm(num_groups=self.groups, epsilon=1e-6)
+ self.conv1 = nn.Conv(
+ out_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ self.norm2 = nn.GroupNorm(num_groups=self.groups, epsilon=1e-6)
+ self.dropout_layer = nn.Dropout(self.dropout)
+ self.conv2 = nn.Conv(
+ out_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ use_nin_shortcut = self.in_channels != out_channels if self.use_nin_shortcut is None else self.use_nin_shortcut
+
+ self.conv_shortcut = None
+ if use_nin_shortcut:
+ self.conv_shortcut = nn.Conv(
+ out_channels,
+ kernel_size=(1, 1),
+ strides=(1, 1),
+ padding="VALID",
+ dtype=self.dtype,
+ )
+
+ def __call__(self, hidden_states, deterministic=True):
+ residual = hidden_states
+ hidden_states = self.norm1(hidden_states)
+ hidden_states = nn.swish(hidden_states)
+ hidden_states = self.conv1(hidden_states)
+
+ hidden_states = self.norm2(hidden_states)
+ hidden_states = nn.swish(hidden_states)
+ hidden_states = self.dropout_layer(hidden_states, deterministic)
+ hidden_states = self.conv2(hidden_states)
+
+ if self.conv_shortcut is not None:
+ residual = self.conv_shortcut(residual)
+
+ return hidden_states + residual
+
+
+class FlaxAttentionBlock(nn.Module):
+ r"""
+ Flax Convolutional based multi-head attention block for diffusion-based VAE.
+
+ Parameters:
+ channels (:obj:`int`):
+ Input channels
+ num_head_channels (:obj:`int`, *optional*, defaults to `None`):
+ Number of attention heads
+ num_groups (:obj:`int`, *optional*, defaults to `32`):
+ The number of groups to use for group norm
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+
+ """
+
+ channels: int
+ num_head_channels: int = None
+ num_groups: int = 32
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ self.num_heads = self.channels // self.num_head_channels if self.num_head_channels is not None else 1
+
+ dense = partial(nn.Dense, self.channels, dtype=self.dtype)
+
+ self.group_norm = nn.GroupNorm(num_groups=self.num_groups, epsilon=1e-6)
+ self.query, self.key, self.value = dense(), dense(), dense()
+ self.proj_attn = dense()
+
+ def transpose_for_scores(self, projection):
+ new_projection_shape = projection.shape[:-1] + (self.num_heads, -1)
+ # move heads to 2nd position (B, T, H * D) -> (B, T, H, D)
+ new_projection = projection.reshape(new_projection_shape)
+ # (B, T, H, D) -> (B, H, T, D)
+ new_projection = jnp.transpose(new_projection, (0, 2, 1, 3))
+ return new_projection
+
+ def __call__(self, hidden_states):
+ residual = hidden_states
+ batch, height, width, channels = hidden_states.shape
+
+ hidden_states = self.group_norm(hidden_states)
+
+ hidden_states = hidden_states.reshape((batch, height * width, channels))
+
+ query = self.query(hidden_states)
+ key = self.key(hidden_states)
+ value = self.value(hidden_states)
+
+ # transpose
+ query = self.transpose_for_scores(query)
+ key = self.transpose_for_scores(key)
+ value = self.transpose_for_scores(value)
+
+ # compute attentions
+ scale = 1 / math.sqrt(math.sqrt(self.channels / self.num_heads))
+ attn_weights = jnp.einsum("...qc,...kc->...qk", query * scale, key * scale)
+ attn_weights = nn.softmax(attn_weights, axis=-1)
+
+ # attend to values
+ hidden_states = jnp.einsum("...kc,...qk->...qc", value, attn_weights)
+
+ hidden_states = jnp.transpose(hidden_states, (0, 2, 1, 3))
+ new_hidden_states_shape = hidden_states.shape[:-2] + (self.channels,)
+ hidden_states = hidden_states.reshape(new_hidden_states_shape)
+
+ hidden_states = self.proj_attn(hidden_states)
+ hidden_states = hidden_states.reshape((batch, height, width, channels))
+ hidden_states = hidden_states + residual
+ return hidden_states
+
+
+class FlaxDownEncoderBlock2D(nn.Module):
+ r"""
+ Flax Resnet blocks-based Encoder block for diffusion-based VAE.
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input channels
+ out_channels (:obj:`int`):
+ Output channels
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ num_layers (:obj:`int`, *optional*, defaults to 1):
+ Number of Resnet layer block
+ resnet_groups (:obj:`int`, *optional*, defaults to `32`):
+ The number of groups to use for the Resnet block group norm
+ add_downsample (:obj:`bool`, *optional*, defaults to `True`):
+ Whether to add downsample layer
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ out_channels: int
+ dropout: float = 0.0
+ num_layers: int = 1
+ resnet_groups: int = 32
+ add_downsample: bool = True
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ resnets = []
+ for i in range(self.num_layers):
+ in_channels = self.in_channels if i == 0 else self.out_channels
+
+ res_block = FlaxResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=self.out_channels,
+ dropout=self.dropout,
+ groups=self.resnet_groups,
+ dtype=self.dtype,
+ )
+ resnets.append(res_block)
+ self.resnets = resnets
+
+ if self.add_downsample:
+ self.downsamplers_0 = FlaxDownsample2D(self.out_channels, dtype=self.dtype)
+
+ def __call__(self, hidden_states, deterministic=True):
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, deterministic=deterministic)
+
+ if self.add_downsample:
+ hidden_states = self.downsamplers_0(hidden_states)
+
+ return hidden_states
+
+
+class FlaxUpDecoderBlock2D(nn.Module):
+ r"""
+ Flax Resnet blocks-based Decoder block for diffusion-based VAE.
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input channels
+ out_channels (:obj:`int`):
+ Output channels
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ num_layers (:obj:`int`, *optional*, defaults to 1):
+ Number of Resnet layer block
+ resnet_groups (:obj:`int`, *optional*, defaults to `32`):
+ The number of groups to use for the Resnet block group norm
+ add_upsample (:obj:`bool`, *optional*, defaults to `True`):
+ Whether to add upsample layer
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ out_channels: int
+ dropout: float = 0.0
+ num_layers: int = 1
+ resnet_groups: int = 32
+ add_upsample: bool = True
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ resnets = []
+ for i in range(self.num_layers):
+ in_channels = self.in_channels if i == 0 else self.out_channels
+ res_block = FlaxResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=self.out_channels,
+ dropout=self.dropout,
+ groups=self.resnet_groups,
+ dtype=self.dtype,
+ )
+ resnets.append(res_block)
+
+ self.resnets = resnets
+
+ if self.add_upsample:
+ self.upsamplers_0 = FlaxUpsample2D(self.out_channels, dtype=self.dtype)
+
+ def __call__(self, hidden_states, deterministic=True):
+ for resnet in self.resnets:
+ hidden_states = resnet(hidden_states, deterministic=deterministic)
+
+ if self.add_upsample:
+ hidden_states = self.upsamplers_0(hidden_states)
+
+ return hidden_states
+
+
+class FlaxUNetMidBlock2D(nn.Module):
+ r"""
+ Flax Unet Mid-Block module.
+
+ Parameters:
+ in_channels (:obj:`int`):
+ Input channels
+ dropout (:obj:`float`, *optional*, defaults to 0.0):
+ Dropout rate
+ num_layers (:obj:`int`, *optional*, defaults to 1):
+ Number of Resnet layer block
+ resnet_groups (:obj:`int`, *optional*, defaults to `32`):
+ The number of groups to use for the Resnet and Attention block group norm
+ num_attention_heads (:obj:`int`, *optional*, defaults to `1`):
+ Number of attention heads for each attention block
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int
+ dropout: float = 0.0
+ num_layers: int = 1
+ resnet_groups: int = 32
+ num_attention_heads: int = 1
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ resnet_groups = self.resnet_groups if self.resnet_groups is not None else min(self.in_channels // 4, 32)
+
+ # there is always at least one resnet
+ resnets = [
+ FlaxResnetBlock2D(
+ in_channels=self.in_channels,
+ out_channels=self.in_channels,
+ dropout=self.dropout,
+ groups=resnet_groups,
+ dtype=self.dtype,
+ )
+ ]
+
+ attentions = []
+
+ for _ in range(self.num_layers):
+ attn_block = FlaxAttentionBlock(
+ channels=self.in_channels,
+ num_head_channels=self.num_attention_heads,
+ num_groups=resnet_groups,
+ dtype=self.dtype,
+ )
+ attentions.append(attn_block)
+
+ res_block = FlaxResnetBlock2D(
+ in_channels=self.in_channels,
+ out_channels=self.in_channels,
+ dropout=self.dropout,
+ groups=resnet_groups,
+ dtype=self.dtype,
+ )
+ resnets.append(res_block)
+
+ self.resnets = resnets
+ self.attentions = attentions
+
+ def __call__(self, hidden_states, deterministic=True):
+ hidden_states = self.resnets[0](hidden_states, deterministic=deterministic)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ hidden_states = attn(hidden_states)
+ hidden_states = resnet(hidden_states, deterministic=deterministic)
+
+ return hidden_states
+
+
+class FlaxEncoder(nn.Module):
+ r"""
+ Flax Implementation of VAE Encoder.
+
+ This model is a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
+ subclass. Use it as a regular Flax linen Module and refer to the Flax documentation for all matter related to
+ general usage and behavior.
+
+ Finally, this model supports inherent JAX features such as:
+ - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+ - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+ - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+ - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+ Parameters:
+ in_channels (:obj:`int`, *optional*, defaults to 3):
+ Input channels
+ out_channels (:obj:`int`, *optional*, defaults to 3):
+ Output channels
+ down_block_types (:obj:`Tuple[str]`, *optional*, defaults to `(DownEncoderBlock2D)`):
+ DownEncoder block type
+ block_out_channels (:obj:`Tuple[str]`, *optional*, defaults to `(64,)`):
+ Tuple containing the number of output channels for each block
+ layers_per_block (:obj:`int`, *optional*, defaults to `2`):
+ Number of Resnet layer for each block
+ norm_num_groups (:obj:`int`, *optional*, defaults to `32`):
+ norm num group
+ act_fn (:obj:`str`, *optional*, defaults to `silu`):
+ Activation function
+ double_z (:obj:`bool`, *optional*, defaults to `False`):
+ Whether to double the last output channels
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ Parameters `dtype`
+ """
+
+ in_channels: int = 3
+ out_channels: int = 3
+ down_block_types: Tuple[str] = ("DownEncoderBlock2D",)
+ block_out_channels: Tuple[int] = (64,)
+ layers_per_block: int = 2
+ norm_num_groups: int = 32
+ act_fn: str = "silu"
+ double_z: bool = False
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ block_out_channels = self.block_out_channels
+ # in
+ self.conv_in = nn.Conv(
+ block_out_channels[0],
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ # downsampling
+ down_blocks = []
+ output_channel = block_out_channels[0]
+ for i, _ in enumerate(self.down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = FlaxDownEncoderBlock2D(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ num_layers=self.layers_per_block,
+ resnet_groups=self.norm_num_groups,
+ add_downsample=not is_final_block,
+ dtype=self.dtype,
+ )
+ down_blocks.append(down_block)
+ self.down_blocks = down_blocks
+
+ # middle
+ self.mid_block = FlaxUNetMidBlock2D(
+ in_channels=block_out_channels[-1],
+ resnet_groups=self.norm_num_groups,
+ num_attention_heads=None,
+ dtype=self.dtype,
+ )
+
+ # end
+ conv_out_channels = 2 * self.out_channels if self.double_z else self.out_channels
+ self.conv_norm_out = nn.GroupNorm(num_groups=self.norm_num_groups, epsilon=1e-6)
+ self.conv_out = nn.Conv(
+ conv_out_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ def __call__(self, sample, deterministic: bool = True):
+ # in
+ sample = self.conv_in(sample)
+
+ # downsampling
+ for block in self.down_blocks:
+ sample = block(sample, deterministic=deterministic)
+
+ # middle
+ sample = self.mid_block(sample, deterministic=deterministic)
+
+ # end
+ sample = self.conv_norm_out(sample)
+ sample = nn.swish(sample)
+ sample = self.conv_out(sample)
+
+ return sample
+
+
+class FlaxDecoder(nn.Module):
+ r"""
+ Flax Implementation of VAE Decoder.
+
+ This model is a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
+ subclass. Use it as a regular Flax linen Module and refer to the Flax documentation for all matter related to
+ general usage and behavior.
+
+ Finally, this model supports inherent JAX features such as:
+ - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+ - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+ - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+ - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+ Parameters:
+ in_channels (:obj:`int`, *optional*, defaults to 3):
+ Input channels
+ out_channels (:obj:`int`, *optional*, defaults to 3):
+ Output channels
+ up_block_types (:obj:`Tuple[str]`, *optional*, defaults to `(UpDecoderBlock2D)`):
+ UpDecoder block type
+ block_out_channels (:obj:`Tuple[str]`, *optional*, defaults to `(64,)`):
+ Tuple containing the number of output channels for each block
+ layers_per_block (:obj:`int`, *optional*, defaults to `2`):
+ Number of Resnet layer for each block
+ norm_num_groups (:obj:`int`, *optional*, defaults to `32`):
+ norm num group
+ act_fn (:obj:`str`, *optional*, defaults to `silu`):
+ Activation function
+ double_z (:obj:`bool`, *optional*, defaults to `False`):
+ Whether to double the last output channels
+ dtype (:obj:`jnp.dtype`, *optional*, defaults to jnp.float32):
+ parameters `dtype`
+ """
+
+ in_channels: int = 3
+ out_channels: int = 3
+ up_block_types: Tuple[str] = ("UpDecoderBlock2D",)
+ block_out_channels: int = (64,)
+ layers_per_block: int = 2
+ norm_num_groups: int = 32
+ act_fn: str = "silu"
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ block_out_channels = self.block_out_channels
+
+ # z to block_in
+ self.conv_in = nn.Conv(
+ block_out_channels[-1],
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ # middle
+ self.mid_block = FlaxUNetMidBlock2D(
+ in_channels=block_out_channels[-1],
+ resnet_groups=self.norm_num_groups,
+ num_attention_heads=None,
+ dtype=self.dtype,
+ )
+
+ # upsampling
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ output_channel = reversed_block_out_channels[0]
+ up_blocks = []
+ for i, _ in enumerate(self.up_block_types):
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+
+ is_final_block = i == len(block_out_channels) - 1
+
+ up_block = FlaxUpDecoderBlock2D(
+ in_channels=prev_output_channel,
+ out_channels=output_channel,
+ num_layers=self.layers_per_block + 1,
+ resnet_groups=self.norm_num_groups,
+ add_upsample=not is_final_block,
+ dtype=self.dtype,
+ )
+ up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ self.up_blocks = up_blocks
+
+ # end
+ self.conv_norm_out = nn.GroupNorm(num_groups=self.norm_num_groups, epsilon=1e-6)
+ self.conv_out = nn.Conv(
+ self.out_channels,
+ kernel_size=(3, 3),
+ strides=(1, 1),
+ padding=((1, 1), (1, 1)),
+ dtype=self.dtype,
+ )
+
+ def __call__(self, sample, deterministic: bool = True):
+ # z to block_in
+ sample = self.conv_in(sample)
+
+ # middle
+ sample = self.mid_block(sample, deterministic=deterministic)
+
+ # upsampling
+ for block in self.up_blocks:
+ sample = block(sample, deterministic=deterministic)
+
+ sample = self.conv_norm_out(sample)
+ sample = nn.swish(sample)
+ sample = self.conv_out(sample)
+
+ return sample
+
+
+class FlaxDiagonalGaussianDistribution(object):
+ def __init__(self, parameters, deterministic=False):
+ # Last axis to account for channels-last
+ self.mean, self.logvar = jnp.split(parameters, 2, axis=-1)
+ self.logvar = jnp.clip(self.logvar, -30.0, 20.0)
+ self.deterministic = deterministic
+ self.std = jnp.exp(0.5 * self.logvar)
+ self.var = jnp.exp(self.logvar)
+ if self.deterministic:
+ self.var = self.std = jnp.zeros_like(self.mean)
+
+ def sample(self, key):
+ return self.mean + self.std * jax.random.normal(key, self.mean.shape)
+
+ def kl(self, other=None):
+ if self.deterministic:
+ return jnp.array([0.0])
+
+ if other is None:
+ return 0.5 * jnp.sum(self.mean**2 + self.var - 1.0 - self.logvar, axis=[1, 2, 3])
+
+ return 0.5 * jnp.sum(
+ jnp.square(self.mean - other.mean) / other.var + self.var / other.var - 1.0 - self.logvar + other.logvar,
+ axis=[1, 2, 3],
+ )
+
+ def nll(self, sample, axis=[1, 2, 3]):
+ if self.deterministic:
+ return jnp.array([0.0])
+
+ logtwopi = jnp.log(2.0 * jnp.pi)
+ return 0.5 * jnp.sum(logtwopi + self.logvar + jnp.square(sample - self.mean) / self.var, axis=axis)
+
+ def mode(self):
+ return self.mean
+
+
+@flax_register_to_config
+class FlaxAutoencoderKL(nn.Module, FlaxModelMixin, ConfigMixin):
+ r"""
+ Flax implementation of a VAE model with KL loss for decoding latent representations.
+
+ This model inherits from [`FlaxModelMixin`]. Check the superclass documentation for it's generic methods
+ implemented for all models (such as downloading or saving).
+
+ This model is a Flax Linen [flax.linen.Module](https://flax.readthedocs.io/en/latest/flax.linen.html#module)
+ subclass. Use it as a regular Flax Linen module and refer to the Flax documentation for all matter related to its
+ general usage and behavior.
+
+ Inherent JAX features such as the following are supported:
+
+ - [Just-In-Time (JIT) compilation](https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit)
+ - [Automatic Differentiation](https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation)
+ - [Vectorization](https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap)
+ - [Parallelization](https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap)
+
+ Parameters:
+ in_channels (`int`, *optional*, defaults to 3):
+ Number of channels in the input image.
+ out_channels (`int`, *optional*, defaults to 3):
+ Number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `(DownEncoderBlock2D)`):
+ Tuple of downsample block types.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `(UpDecoderBlock2D)`):
+ Tuple of upsample block types.
+ block_out_channels (`Tuple[str]`, *optional*, defaults to `(64,)`):
+ Tuple of block output channels.
+ layers_per_block (`int`, *optional*, defaults to `2`):
+ Number of ResNet layer for each block.
+ act_fn (`str`, *optional*, defaults to `silu`):
+ The activation function to use.
+ latent_channels (`int`, *optional*, defaults to `4`):
+ Number of channels in the latent space.
+ norm_num_groups (`int`, *optional*, defaults to `32`):
+ The number of groups for normalization.
+ sample_size (`int`, *optional*, defaults to 32):
+ Sample input size.
+ scaling_factor (`float`, *optional*, defaults to 0.18215):
+ The component-wise standard deviation of the trained latent space computed using the first batch of the
+ training set. This is used to scale the latent space to have unit variance when training the diffusion
+ model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+ diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+ / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+ Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+ dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`):
+ The `dtype` of the parameters.
+ """
+
+ in_channels: int = 3
+ out_channels: int = 3
+ down_block_types: Tuple[str] = ("DownEncoderBlock2D",)
+ up_block_types: Tuple[str] = ("UpDecoderBlock2D",)
+ block_out_channels: Tuple[int] = (64,)
+ layers_per_block: int = 1
+ act_fn: str = "silu"
+ latent_channels: int = 4
+ norm_num_groups: int = 32
+ sample_size: int = 32
+ scaling_factor: float = 0.18215
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ self.encoder = FlaxEncoder(
+ in_channels=self.config.in_channels,
+ out_channels=self.config.latent_channels,
+ down_block_types=self.config.down_block_types,
+ block_out_channels=self.config.block_out_channels,
+ layers_per_block=self.config.layers_per_block,
+ act_fn=self.config.act_fn,
+ norm_num_groups=self.config.norm_num_groups,
+ double_z=True,
+ dtype=self.dtype,
+ )
+ self.decoder = FlaxDecoder(
+ in_channels=self.config.latent_channels,
+ out_channels=self.config.out_channels,
+ up_block_types=self.config.up_block_types,
+ block_out_channels=self.config.block_out_channels,
+ layers_per_block=self.config.layers_per_block,
+ norm_num_groups=self.config.norm_num_groups,
+ act_fn=self.config.act_fn,
+ dtype=self.dtype,
+ )
+ self.quant_conv = nn.Conv(
+ 2 * self.config.latent_channels,
+ kernel_size=(1, 1),
+ strides=(1, 1),
+ padding="VALID",
+ dtype=self.dtype,
+ )
+ self.post_quant_conv = nn.Conv(
+ self.config.latent_channels,
+ kernel_size=(1, 1),
+ strides=(1, 1),
+ padding="VALID",
+ dtype=self.dtype,
+ )
+
+ def init_weights(self, rng: jax.Array) -> FrozenDict:
+ # init input tensors
+ sample_shape = (1, self.in_channels, self.sample_size, self.sample_size)
+ sample = jnp.zeros(sample_shape, dtype=jnp.float32)
+
+ params_rng, dropout_rng, gaussian_rng = jax.random.split(rng, 3)
+ rngs = {"params": params_rng, "dropout": dropout_rng, "gaussian": gaussian_rng}
+
+ return self.init(rngs, sample)["params"]
+
+ def encode(self, sample, deterministic: bool = True, return_dict: bool = True):
+ sample = jnp.transpose(sample, (0, 2, 3, 1))
+
+ hidden_states = self.encoder(sample, deterministic=deterministic)
+ moments = self.quant_conv(hidden_states)
+ posterior = FlaxDiagonalGaussianDistribution(moments)
+
+ if not return_dict:
+ return (posterior,)
+
+ return FlaxAutoencoderKLOutput(latent_dist=posterior)
+
+ def decode(self, latents, deterministic: bool = True, return_dict: bool = True):
+ if latents.shape[-1] != self.config.latent_channels:
+ latents = jnp.transpose(latents, (0, 2, 3, 1))
+
+ hidden_states = self.post_quant_conv(latents)
+ hidden_states = self.decoder(hidden_states, deterministic=deterministic)
+
+ hidden_states = jnp.transpose(hidden_states, (0, 3, 1, 2))
+
+ if not return_dict:
+ return (hidden_states,)
+
+ return FlaxDecoderOutput(sample=hidden_states)
+
+ def __call__(self, sample, sample_posterior=False, deterministic: bool = True, return_dict: bool = True):
+ posterior = self.encode(sample, deterministic=deterministic, return_dict=return_dict)
+ if sample_posterior:
+ rng = self.make_rng("gaussian")
+ hidden_states = posterior.latent_dist.sample(rng)
+ else:
+ hidden_states = posterior.latent_dist.mode()
+
+ sample = self.decode(hidden_states, return_dict=return_dict).sample
+
+ if not return_dict:
+ return (sample,)
+
+ return FlaxDecoderOutput(sample=sample)
diff --git a/model/diffusers_c/models/vq_model.py b/model/diffusers_c/models/vq_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..e5184446cee7022f752cb003adcc68080ee438f2
--- /dev/null
+++ b/model/diffusers_c/models/vq_model.py
@@ -0,0 +1,181 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.accelerate_utils import apply_forward_hook
+from .autoencoders.vae import Decoder, DecoderOutput, Encoder, VectorQuantizer
+from .modeling_utils import ModelMixin
+
+
+@dataclass
+class VQEncoderOutput(BaseOutput):
+ """
+ Output of VQModel encoding method.
+
+ Args:
+ latents (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+ The encoded output sample from the last layer of the model.
+ """
+
+ latents: torch.FloatTensor
+
+
+class VQModel(ModelMixin, ConfigMixin):
+ r"""
+ A VQ-VAE model for decoding latent representations.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+ out_channels (int, *optional*, defaults to 3): Number of channels in the output.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`):
+ Tuple of downsample block types.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`):
+ Tuple of upsample block types.
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`):
+ Tuple of block output channels.
+ layers_per_block (`int`, *optional*, defaults to `1`): Number of layers per block.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ latent_channels (`int`, *optional*, defaults to `3`): Number of channels in the latent space.
+ sample_size (`int`, *optional*, defaults to `32`): Sample input size.
+ num_vq_embeddings (`int`, *optional*, defaults to `256`): Number of codebook vectors in the VQ-VAE.
+ norm_num_groups (`int`, *optional*, defaults to `32`): Number of groups for normalization layers.
+ vq_embed_dim (`int`, *optional*): Hidden dim of codebook vectors in the VQ-VAE.
+ scaling_factor (`float`, *optional*, defaults to `0.18215`):
+ The component-wise standard deviation of the trained latent space computed using the first batch of the
+ training set. This is used to scale the latent space to have unit variance when training the diffusion
+ model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the
+ diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1
+ / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image
+ Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper.
+ norm_type (`str`, *optional*, defaults to `"group"`):
+ Type of normalization layer to use. Can be one of `"group"` or `"spatial"`.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ down_block_types: Tuple[str, ...] = ("DownEncoderBlock2D",),
+ up_block_types: Tuple[str, ...] = ("UpDecoderBlock2D",),
+ block_out_channels: Tuple[int, ...] = (64,),
+ layers_per_block: int = 1,
+ act_fn: str = "silu",
+ latent_channels: int = 3,
+ sample_size: int = 32,
+ num_vq_embeddings: int = 256,
+ norm_num_groups: int = 32,
+ vq_embed_dim: Optional[int] = None,
+ scaling_factor: float = 0.18215,
+ norm_type: str = "group", # group, spatial
+ mid_block_add_attention=True,
+ lookup_from_codebook=False,
+ force_upcast=False,
+ ):
+ super().__init__()
+
+ # pass init params to Encoder
+ self.encoder = Encoder(
+ in_channels=in_channels,
+ out_channels=latent_channels,
+ down_block_types=down_block_types,
+ block_out_channels=block_out_channels,
+ layers_per_block=layers_per_block,
+ act_fn=act_fn,
+ norm_num_groups=norm_num_groups,
+ double_z=False,
+ mid_block_add_attention=mid_block_add_attention,
+ )
+
+ vq_embed_dim = vq_embed_dim if vq_embed_dim is not None else latent_channels
+
+ self.quant_conv = nn.Conv2d(latent_channels, vq_embed_dim, 1)
+ self.quantize = VectorQuantizer(num_vq_embeddings, vq_embed_dim, beta=0.25, remap=None, sane_index_shape=False)
+ self.post_quant_conv = nn.Conv2d(vq_embed_dim, latent_channels, 1)
+
+ # pass init params to Decoder
+ self.decoder = Decoder(
+ in_channels=latent_channels,
+ out_channels=out_channels,
+ up_block_types=up_block_types,
+ block_out_channels=block_out_channels,
+ layers_per_block=layers_per_block,
+ act_fn=act_fn,
+ norm_num_groups=norm_num_groups,
+ norm_type=norm_type,
+ mid_block_add_attention=mid_block_add_attention,
+ )
+
+ @apply_forward_hook
+ def encode(self, x: torch.FloatTensor, return_dict: bool = True) -> VQEncoderOutput:
+ h = self.encoder(x)
+ h = self.quant_conv(h)
+
+ if not return_dict:
+ return (h,)
+
+ return VQEncoderOutput(latents=h)
+
+ @apply_forward_hook
+ def decode(
+ self, h: torch.FloatTensor, force_not_quantize: bool = False, return_dict: bool = True, shape=None
+ ) -> Union[DecoderOutput, torch.FloatTensor]:
+ # also go through quantization layer
+ if not force_not_quantize:
+ quant, _, _ = self.quantize(h)
+ elif self.config.lookup_from_codebook:
+ quant = self.quantize.get_codebook_entry(h, shape)
+ else:
+ quant = h
+ quant2 = self.post_quant_conv(quant)
+ dec = self.decoder(quant2, quant if self.config.norm_type == "spatial" else None)
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
+
+ def forward(
+ self, sample: torch.FloatTensor, return_dict: bool = True
+ ) -> Union[DecoderOutput, Tuple[torch.FloatTensor, ...]]:
+ r"""
+ The [`VQModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`): Input sample.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`models.vq_model.VQEncoderOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~models.vq_model.VQEncoderOutput`] or `tuple`:
+ If return_dict is True, a [`~models.vq_model.VQEncoderOutput`] is returned, otherwise a plain `tuple`
+ is returned.
+ """
+
+ h = self.encode(sample).latents
+ dec = self.decode(h).sample
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
diff --git a/model/diffusers_c/optimization.py b/model/diffusers_c/optimization.py
new file mode 100644
index 0000000000000000000000000000000000000000..81f9e697870e8e65465adbf208e7326946581c37
--- /dev/null
+++ b/model/diffusers_c/optimization.py
@@ -0,0 +1,360 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch optimization for diffusion models."""
+
+import math
+from enum import Enum
+from typing import Optional, Union
+
+from torch.optim import Optimizer
+from torch.optim.lr_scheduler import LambdaLR
+
+from .utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class SchedulerType(Enum):
+ LINEAR = "linear"
+ COSINE = "cosine"
+ COSINE_WITH_RESTARTS = "cosine_with_restarts"
+ POLYNOMIAL = "polynomial"
+ CONSTANT = "constant"
+ CONSTANT_WITH_WARMUP = "constant_with_warmup"
+ PIECEWISE_CONSTANT = "piecewise_constant"
+
+
+def get_constant_schedule(optimizer: Optimizer, last_epoch: int = -1) -> LambdaLR:
+ """
+ Create a schedule with a constant learning rate, using the learning rate set in optimizer.
+
+ Args:
+ optimizer ([`~torch.optim.Optimizer`]):
+ The optimizer for which to schedule the learning rate.
+ last_epoch (`int`, *optional*, defaults to -1):
+ The index of the last epoch when resuming training.
+
+ Return:
+ `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
+ """
+ return LambdaLR(optimizer, lambda _: 1, last_epoch=last_epoch)
+
+
+def get_constant_schedule_with_warmup(optimizer: Optimizer, num_warmup_steps: int, last_epoch: int = -1) -> LambdaLR:
+ """
+ Create a schedule with a constant learning rate preceded by a warmup period during which the learning rate
+ increases linearly between 0 and the initial lr set in the optimizer.
+
+ Args:
+ optimizer ([`~torch.optim.Optimizer`]):
+ The optimizer for which to schedule the learning rate.
+ num_warmup_steps (`int`):
+ The number of steps for the warmup phase.
+ last_epoch (`int`, *optional*, defaults to -1):
+ The index of the last epoch when resuming training.
+
+ Return:
+ `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
+ """
+
+ def lr_lambda(current_step: int):
+ if current_step < num_warmup_steps:
+ return float(current_step) / float(max(1.0, num_warmup_steps))
+ return 1.0
+
+ return LambdaLR(optimizer, lr_lambda, last_epoch=last_epoch)
+
+
+def get_piecewise_constant_schedule(optimizer: Optimizer, step_rules: str, last_epoch: int = -1) -> LambdaLR:
+ """
+ Create a schedule with a constant learning rate, using the learning rate set in optimizer.
+
+ Args:
+ optimizer ([`~torch.optim.Optimizer`]):
+ The optimizer for which to schedule the learning rate.
+ step_rules (`string`):
+ The rules for the learning rate. ex: rule_steps="1:10,0.1:20,0.01:30,0.005" it means that the learning rate
+ if multiple 1 for the first 10 steps, mutiple 0.1 for the next 20 steps, multiple 0.01 for the next 30
+ steps and multiple 0.005 for the other steps.
+ last_epoch (`int`, *optional*, defaults to -1):
+ The index of the last epoch when resuming training.
+
+ Return:
+ `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
+ """
+
+ rules_dict = {}
+ rule_list = step_rules.split(",")
+ for rule_str in rule_list[:-1]:
+ value_str, steps_str = rule_str.split(":")
+ steps = int(steps_str)
+ value = float(value_str)
+ rules_dict[steps] = value
+ last_lr_multiple = float(rule_list[-1])
+
+ def create_rules_function(rules_dict, last_lr_multiple):
+ def rule_func(steps: int) -> float:
+ sorted_steps = sorted(rules_dict.keys())
+ for i, sorted_step in enumerate(sorted_steps):
+ if steps < sorted_step:
+ return rules_dict[sorted_steps[i]]
+ return last_lr_multiple
+
+ return rule_func
+
+ rules_func = create_rules_function(rules_dict, last_lr_multiple)
+
+ return LambdaLR(optimizer, rules_func, last_epoch=last_epoch)
+
+
+def get_linear_schedule_with_warmup(
+ optimizer: Optimizer, num_warmup_steps: int, num_training_steps: int, last_epoch: int = -1
+) -> LambdaLR:
+ """
+ Create a schedule with a learning rate that decreases linearly from the initial lr set in the optimizer to 0, after
+ a warmup period during which it increases linearly from 0 to the initial lr set in the optimizer.
+
+ Args:
+ optimizer ([`~torch.optim.Optimizer`]):
+ The optimizer for which to schedule the learning rate.
+ num_warmup_steps (`int`):
+ The number of steps for the warmup phase.
+ num_training_steps (`int`):
+ The total number of training steps.
+ last_epoch (`int`, *optional*, defaults to -1):
+ The index of the last epoch when resuming training.
+
+ Return:
+ `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
+ """
+
+ def lr_lambda(current_step: int):
+ if current_step < num_warmup_steps:
+ return float(current_step) / float(max(1, num_warmup_steps))
+ return max(
+ 0.0, float(num_training_steps - current_step) / float(max(1, num_training_steps - num_warmup_steps))
+ )
+
+ return LambdaLR(optimizer, lr_lambda, last_epoch)
+
+
+def get_cosine_schedule_with_warmup(
+ optimizer: Optimizer, num_warmup_steps: int, num_training_steps: int, num_cycles: float = 0.5, last_epoch: int = -1
+) -> LambdaLR:
+ """
+ Create a schedule with a learning rate that decreases following the values of the cosine function between the
+ initial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the
+ initial lr set in the optimizer.
+
+ Args:
+ optimizer ([`~torch.optim.Optimizer`]):
+ The optimizer for which to schedule the learning rate.
+ num_warmup_steps (`int`):
+ The number of steps for the warmup phase.
+ num_training_steps (`int`):
+ The total number of training steps.
+ num_periods (`float`, *optional*, defaults to 0.5):
+ The number of periods of the cosine function in a schedule (the default is to just decrease from the max
+ value to 0 following a half-cosine).
+ last_epoch (`int`, *optional*, defaults to -1):
+ The index of the last epoch when resuming training.
+
+ Return:
+ `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
+ """
+
+ def lr_lambda(current_step):
+ if current_step < num_warmup_steps:
+ return float(current_step) / float(max(1, num_warmup_steps))
+ progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps))
+ return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress)))
+
+ return LambdaLR(optimizer, lr_lambda, last_epoch)
+
+
+def get_cosine_with_hard_restarts_schedule_with_warmup(
+ optimizer: Optimizer, num_warmup_steps: int, num_training_steps: int, num_cycles: int = 1, last_epoch: int = -1
+) -> LambdaLR:
+ """
+ Create a schedule with a learning rate that decreases following the values of the cosine function between the
+ initial lr set in the optimizer to 0, with several hard restarts, after a warmup period during which it increases
+ linearly between 0 and the initial lr set in the optimizer.
+
+ Args:
+ optimizer ([`~torch.optim.Optimizer`]):
+ The optimizer for which to schedule the learning rate.
+ num_warmup_steps (`int`):
+ The number of steps for the warmup phase.
+ num_training_steps (`int`):
+ The total number of training steps.
+ num_cycles (`int`, *optional*, defaults to 1):
+ The number of hard restarts to use.
+ last_epoch (`int`, *optional*, defaults to -1):
+ The index of the last epoch when resuming training.
+
+ Return:
+ `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
+ """
+
+ def lr_lambda(current_step):
+ if current_step < num_warmup_steps:
+ return float(current_step) / float(max(1, num_warmup_steps))
+ progress = float(current_step - num_warmup_steps) / float(max(1, num_training_steps - num_warmup_steps))
+ if progress >= 1.0:
+ return 0.0
+ return max(0.0, 0.5 * (1.0 + math.cos(math.pi * ((float(num_cycles) * progress) % 1.0))))
+
+ return LambdaLR(optimizer, lr_lambda, last_epoch)
+
+
+def get_polynomial_decay_schedule_with_warmup(
+ optimizer: Optimizer,
+ num_warmup_steps: int,
+ num_training_steps: int,
+ lr_end: float = 1e-7,
+ power: float = 1.0,
+ last_epoch: int = -1,
+) -> LambdaLR:
+ """
+ Create a schedule with a learning rate that decreases as a polynomial decay from the initial lr set in the
+ optimizer to end lr defined by *lr_end*, after a warmup period during which it increases linearly from 0 to the
+ initial lr set in the optimizer.
+
+ Args:
+ optimizer ([`~torch.optim.Optimizer`]):
+ The optimizer for which to schedule the learning rate.
+ num_warmup_steps (`int`):
+ The number of steps for the warmup phase.
+ num_training_steps (`int`):
+ The total number of training steps.
+ lr_end (`float`, *optional*, defaults to 1e-7):
+ The end LR.
+ power (`float`, *optional*, defaults to 1.0):
+ Power factor.
+ last_epoch (`int`, *optional*, defaults to -1):
+ The index of the last epoch when resuming training.
+
+ Note: *power* defaults to 1.0 as in the fairseq implementation, which in turn is based on the original BERT
+ implementation at
+ https://github.com/google-research/bert/blob/f39e881b169b9d53bea03d2d341b31707a6c052b/optimization.py#L37
+
+ Return:
+ `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
+
+ """
+
+ lr_init = optimizer.defaults["lr"]
+ if not (lr_init > lr_end):
+ raise ValueError(f"lr_end ({lr_end}) must be be smaller than initial lr ({lr_init})")
+
+ def lr_lambda(current_step: int):
+ if current_step < num_warmup_steps:
+ return float(current_step) / float(max(1, num_warmup_steps))
+ elif current_step > num_training_steps:
+ return lr_end / lr_init # as LambdaLR multiplies by lr_init
+ else:
+ lr_range = lr_init - lr_end
+ decay_steps = num_training_steps - num_warmup_steps
+ pct_remaining = 1 - (current_step - num_warmup_steps) / decay_steps
+ decay = lr_range * pct_remaining**power + lr_end
+ return decay / lr_init # as LambdaLR multiplies by lr_init
+
+ return LambdaLR(optimizer, lr_lambda, last_epoch)
+
+
+TYPE_TO_SCHEDULER_FUNCTION = {
+ SchedulerType.LINEAR: get_linear_schedule_with_warmup,
+ SchedulerType.COSINE: get_cosine_schedule_with_warmup,
+ SchedulerType.COSINE_WITH_RESTARTS: get_cosine_with_hard_restarts_schedule_with_warmup,
+ SchedulerType.POLYNOMIAL: get_polynomial_decay_schedule_with_warmup,
+ SchedulerType.CONSTANT: get_constant_schedule,
+ SchedulerType.CONSTANT_WITH_WARMUP: get_constant_schedule_with_warmup,
+ SchedulerType.PIECEWISE_CONSTANT: get_piecewise_constant_schedule,
+}
+
+
+def get_scheduler(
+ name: Union[str, SchedulerType],
+ optimizer: Optimizer,
+ step_rules: Optional[str] = None,
+ num_warmup_steps: Optional[int] = None,
+ num_training_steps: Optional[int] = None,
+ num_cycles: int = 1,
+ power: float = 1.0,
+ last_epoch: int = -1,
+) -> LambdaLR:
+ """
+ Unified API to get any scheduler from its name.
+
+ Args:
+ name (`str` or `SchedulerType`):
+ The name of the scheduler to use.
+ optimizer (`torch.optim.Optimizer`):
+ The optimizer that will be used during training.
+ step_rules (`str`, *optional*):
+ A string representing the step rules to use. This is only used by the `PIECEWISE_CONSTANT` scheduler.
+ num_warmup_steps (`int`, *optional*):
+ The number of warmup steps to do. This is not required by all schedulers (hence the argument being
+ optional), the function will raise an error if it's unset and the scheduler type requires it.
+ num_training_steps (`int``, *optional*):
+ The number of training steps to do. This is not required by all schedulers (hence the argument being
+ optional), the function will raise an error if it's unset and the scheduler type requires it.
+ num_cycles (`int`, *optional*):
+ The number of hard restarts used in `COSINE_WITH_RESTARTS` scheduler.
+ power (`float`, *optional*, defaults to 1.0):
+ Power factor. See `POLYNOMIAL` scheduler
+ last_epoch (`int`, *optional*, defaults to -1):
+ The index of the last epoch when resuming training.
+ """
+ name = SchedulerType(name)
+ schedule_func = TYPE_TO_SCHEDULER_FUNCTION[name]
+ if name == SchedulerType.CONSTANT:
+ return schedule_func(optimizer, last_epoch=last_epoch)
+
+ if name == SchedulerType.PIECEWISE_CONSTANT:
+ return schedule_func(optimizer, step_rules=step_rules, last_epoch=last_epoch)
+
+ # All other schedulers require `num_warmup_steps`
+ if num_warmup_steps is None:
+ raise ValueError(f"{name} requires `num_warmup_steps`, please provide that argument.")
+
+ if name == SchedulerType.CONSTANT_WITH_WARMUP:
+ return schedule_func(optimizer, num_warmup_steps=num_warmup_steps, last_epoch=last_epoch)
+
+ # All other schedulers require `num_training_steps`
+ if num_training_steps is None:
+ raise ValueError(f"{name} requires `num_training_steps`, please provide that argument.")
+
+ if name == SchedulerType.COSINE_WITH_RESTARTS:
+ return schedule_func(
+ optimizer,
+ num_warmup_steps=num_warmup_steps,
+ num_training_steps=num_training_steps,
+ num_cycles=num_cycles,
+ last_epoch=last_epoch,
+ )
+
+ if name == SchedulerType.POLYNOMIAL:
+ return schedule_func(
+ optimizer,
+ num_warmup_steps=num_warmup_steps,
+ num_training_steps=num_training_steps,
+ power=power,
+ last_epoch=last_epoch,
+ )
+
+ return schedule_func(
+ optimizer, num_warmup_steps=num_warmup_steps, num_training_steps=num_training_steps, last_epoch=last_epoch
+ )
diff --git a/model/diffusers_c/pipelines/README.md b/model/diffusers_c/pipelines/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..d5125ae5caf244425a0a372f054b069c98f27670
--- /dev/null
+++ b/model/diffusers_c/pipelines/README.md
@@ -0,0 +1,171 @@
+# 🧨 Diffusers Pipelines
+
+Pipelines provide a simple way to run state-of-the-art diffusion models in inference.
+Most diffusion systems consist of multiple independently-trained models and highly adaptable scheduler
+components - all of which are needed to have a functioning end-to-end diffusion system.
+
+As an example, [Stable Diffusion](https://huggingface.co/blog/stable_diffusion) has three independently trained models:
+- [Autoencoder](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/models/vae.py#L392)
+- [Conditional Unet](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/models/unet_2d_condition.py#L12)
+- [CLIP text encoder](https://huggingface.co/docs/transformers/main/en/model_doc/clip#transformers.CLIPTextModel)
+- a scheduler component, [scheduler](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_pndm.py),
+- a [CLIPImageProcessor](https://huggingface.co/docs/transformers/main/en/model_doc/clip#transformers.CLIPImageProcessor),
+- as well as a [safety checker](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/safety_checker.py).
+All of these components are necessary to run stable diffusion in inference even though they were trained
+or created independently from each other.
+
+To that end, we strive to offer all open-sourced, state-of-the-art diffusion system under a unified API.
+More specifically, we strive to provide pipelines that
+- 1. can load the officially published weights and yield 1-to-1 the same outputs as the original implementation according to the corresponding paper (*e.g.* [LDMTextToImagePipeline](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/latent_diffusion), uses the officially released weights of [High-Resolution Image Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752)),
+- 2. have a simple user interface to run the model in inference (see the [Pipelines API](#pipelines-api) section),
+- 3. are easy to understand with code that is self-explanatory and can be read along-side the official paper (see [Pipelines summary](#pipelines-summary)),
+- 4. can easily be contributed by the community (see the [Contribution](#contribution) section).
+
+**Note** that pipelines do not (and should not) offer any training functionality.
+If you are looking for *official* training examples, please have a look at [examples](https://github.com/huggingface/diffusers/tree/main/examples).
+
+
+## Pipelines Summary
+
+The following table summarizes all officially supported pipelines, their corresponding paper, and if
+available a colab notebook to directly try them out.
+
+| Pipeline | Source | Tasks | Colab
+|-------------------------------------------------------------------------------------------------------------------------------|------------------------------------------------------------------------------------------------------------------------------|:---:|:---:|
+| [dance diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/dance_diffusion) | [**Dance Diffusion**](https://github.com/Harmonai-org/sample-generator) | *Unconditional Audio Generation* |
+| [ddpm](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/ddpm) | [**Denoising Diffusion Probabilistic Models**](https://arxiv.org/abs/2006.11239) | *Unconditional Image Generation* |
+| [ddim](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/ddim) | [**Denoising Diffusion Implicit Models**](https://arxiv.org/abs/2010.02502) | *Unconditional Image Generation* | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
+| [latent_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752) | *Text-to-Image Generation* |
+| [latent_diffusion_uncond](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/latent_diffusion_uncond) | [**High-Resolution Image Synthesis with Latent Diffusion Models**](https://arxiv.org/abs/2112.10752) | *Unconditional Image Generation* |
+| [pndm](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pndm) | [**Pseudo Numerical Methods for Diffusion Models on Manifolds**](https://arxiv.org/abs/2202.09778) | *Unconditional Image Generation* |
+| [score_sde_ve](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/score_sde_ve) | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | *Unconditional Image Generation* |
+| [score_sde_vp](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/score_sde_vp) | [**Score-Based Generative Modeling through Stochastic Differential Equations**](https://openreview.net/forum?id=PxTIG12RRHS) | *Unconditional Image Generation* |
+| [stable_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | *Text-to-Image Generation* | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_diffusion.ipynb)
+| [stable_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | *Image-to-Image Text-Guided Generation* | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb)
+| [stable_diffusion](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion) | [**Stable Diffusion**](https://stability.ai/blog/stable-diffusion-public-release) | *Text-Guided Image Inpainting* | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
+| [stochastic_karras_ve](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stochastic_karras_ve) | [**Elucidating the Design Space of Diffusion-Based Generative Models**](https://arxiv.org/abs/2206.00364) | *Unconditional Image Generation* |
+
+**Note**: Pipelines are simple examples of how to play around with the diffusion systems as described in the corresponding papers.
+However, most of them can be adapted to use different scheduler components or even different model components. Some pipeline examples are shown in the [Examples](#examples) below.
+
+## Pipelines API
+
+Diffusion models often consist of multiple independently-trained models or other previously existing components.
+
+
+Each model has been trained independently on a different task and the scheduler can easily be swapped out and replaced with a different one.
+During inference, we however want to be able to easily load all components and use them in inference - even if one component, *e.g.* CLIP's text encoder, originates from a different library, such as [Transformers](https://github.com/huggingface/transformers). To that end, all pipelines provide the following functionality:
+
+- [`from_pretrained` method](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L139) that accepts a Hugging Face Hub repository id, *e.g.* [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5) or a path to a local directory, *e.g.*
+"./stable-diffusion". To correctly retrieve which models and components should be loaded, one has to provide a `model_index.json` file, *e.g.* [runwayml/stable-diffusion-v1-5/model_index.json](https://huggingface.co/runwayml/stable-diffusion-v1-5/blob/main/model_index.json), which defines all components that should be
+loaded into the pipelines. More specifically, for each model/component one needs to define the format `: ["", ""]`. `` is the attribute name given to the loaded instance of `` which can be found in the library or pipeline folder called `""`.
+- [`save_pretrained`](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L90) that accepts a local path, *e.g.* `./stable-diffusion` under which all models/components of the pipeline will be saved. For each component/model a folder is created inside the local path that is named after the given attribute name, *e.g.* `./stable_diffusion/unet`.
+In addition, a `model_index.json` file is created at the root of the local path, *e.g.* `./stable_diffusion/model_index.json` so that the complete pipeline can again be instantiated
+from the local path.
+- [`to`](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L118) which accepts a `string` or `torch.device` to move all models that are of type `torch.nn.Module` to the passed device. The behavior is fully analogous to [PyTorch's `to` method](https://pytorch.org/docs/stable/generated/torch.nn.Module.html#torch.nn.Module.to).
+- [`__call__`] method to use the pipeline in inference. `__call__` defines inference logic of the pipeline and should ideally encompass all aspects of it, from pre-processing to forwarding tensors to the different models and schedulers, as well as post-processing. The API of the `__call__` method can strongly vary from pipeline to pipeline. *E.g.* a text-to-image pipeline, such as [`StableDiffusionPipeline`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) should accept among other things the text prompt to generate the image. A pure image generation pipeline, such as [DDPMPipeline](https://github.com/huggingface/diffusers/tree/main/src/diffusers/pipelines/ddpm) on the other hand can be run without providing any inputs. To better understand what inputs can be adapted for
+each pipeline, one should look directly into the respective pipeline.
+
+**Note**: All pipelines have PyTorch's autograd disabled by decorating the `__call__` method with a [`torch.no_grad`](https://pytorch.org/docs/stable/generated/torch.no_grad.html) decorator because pipelines should
+not be used for training. If you want to store the gradients during the forward pass, we recommend writing your own pipeline, see also our [community-examples](https://github.com/huggingface/diffusers/tree/main/examples/community)
+
+## Contribution
+
+We are more than happy about any contribution to the officially supported pipelines 🤗. We aspire
+all of our pipelines to be **self-contained**, **easy-to-tweak**, **beginner-friendly** and for **one-purpose-only**.
+
+- **Self-contained**: A pipeline shall be as self-contained as possible. More specifically, this means that all functionality should be either directly defined in the pipeline file itself, should be inherited from (and only from) the [`DiffusionPipeline` class](https://github.com/huggingface/diffusers/blob/5cbed8e0d157f65d3ddc2420dfd09f2df630e978/src/diffusers/pipeline_utils.py#L56) or be directly attached to the model and scheduler components of the pipeline.
+- **Easy-to-use**: Pipelines should be extremely easy to use - one should be able to load the pipeline and
+use it for its designated task, *e.g.* text-to-image generation, in just a couple of lines of code. Most
+logic including pre-processing, an unrolled diffusion loop, and post-processing should all happen inside the `__call__` method.
+- **Easy-to-tweak**: Certain pipelines will not be able to handle all use cases and tasks that you might like them to. If you want to use a certain pipeline for a specific use case that is not yet supported, you might have to copy the pipeline file and tweak the code to your needs. We try to make the pipeline code as readable as possible so that each part –from pre-processing to diffusing to post-processing– can easily be adapted. If you would like the community to benefit from your customized pipeline, we would love to see a contribution to our [community-examples](https://github.com/huggingface/diffusers/tree/main/examples/community). If you feel that an important pipeline should be part of the official pipelines but isn't, a contribution to the [official pipelines](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines) would be even better.
+- **One-purpose-only**: Pipelines should be used for one task and one task only. Even if two tasks are very similar from a modeling point of view, *e.g.* image2image translation and in-painting, pipelines shall be used for one task only to keep them *easy-to-tweak* and *readable*.
+
+## Examples
+
+### Text-to-Image generation with Stable Diffusion
+
+```python
+# make sure you're logged in with `huggingface-cli login`
+from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler
+
+pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
+pipe = pipe.to("cuda")
+
+prompt = "a photo of an astronaut riding a horse on mars"
+image = pipe(prompt).images[0]
+
+image.save("astronaut_rides_horse.png")
+```
+
+### Image-to-Image text-guided generation with Stable Diffusion
+
+The `StableDiffusionImg2ImgPipeline` lets you pass a text prompt and an initial image to condition the generation of new images.
+
+```python
+import requests
+from PIL import Image
+from io import BytesIO
+
+from diffusers import StableDiffusionImg2ImgPipeline
+
+# load the pipeline
+device = "cuda"
+pipe = StableDiffusionImg2ImgPipeline.from_pretrained(
+ "runwayml/stable-diffusion-v1-5",
+ torch_dtype=torch.float16,
+).to(device)
+
+# let's download an initial image
+url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+
+response = requests.get(url)
+init_image = Image.open(BytesIO(response.content)).convert("RGB")
+init_image = init_image.resize((768, 512))
+
+prompt = "A fantasy landscape, trending on artstation"
+
+images = pipe(prompt=prompt, image=init_image, strength=0.75, guidance_scale=7.5).images
+
+images[0].save("fantasy_landscape.png")
+```
+You can also run this example on colab [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb)
+
+### Tweak prompts reusing seeds and latents
+
+You can generate your own latents to reproduce results, or tweak your prompt on a specific result you liked. [This notebook](https://github.com/pcuenca/diffusers-examples/blob/main/notebooks/stable-diffusion-seeds.ipynb) shows how to do it step by step. You can also run it in Google Colab [](https://colab.research.google.com/github/pcuenca/diffusers-examples/blob/main/notebooks/stable-diffusion-seeds.ipynb).
+
+
+### In-painting using Stable Diffusion
+
+The `StableDiffusionInpaintPipeline` lets you edit specific parts of an image by providing a mask and text prompt.
+
+```python
+import PIL
+import requests
+import torch
+from io import BytesIO
+
+from diffusers import StableDiffusionInpaintPipeline
+
+def download_image(url):
+ response = requests.get(url)
+ return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+init_image = download_image(img_url).resize((512, 512))
+mask_image = download_image(mask_url).resize((512, 512))
+
+pipe = StableDiffusionInpaintPipeline.from_pretrained(
+ "runwayml/stable-diffusion-inpainting",
+ torch_dtype=torch.float16,
+)
+pipe = pipe.to("cuda")
+
+prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
+```
+
+You can also run this example on colab [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
diff --git a/model/diffusers_c/pipelines/__init__.py b/model/diffusers_c/pipelines/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..3368ea954959c89e5fb57e28ef66119ffb830dd7
--- /dev/null
+++ b/model/diffusers_c/pipelines/__init__.py
@@ -0,0 +1,573 @@
+from typing import TYPE_CHECKING
+
+from ..utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_flax_available,
+ is_k_diffusion_available,
+ is_librosa_available,
+ is_note_seq_available,
+ is_onnx_available,
+ is_torch_available,
+ is_torch_npu_available,
+ is_transformers_available,
+)
+
+
+# These modules contain pipelines from multiple libraries/frameworks
+_dummy_objects = {}
+_import_structure = {
+ "brushnet": [],
+ "controlnet": [],
+ "controlnet_xs": [],
+ "deprecated": [],
+ "latent_diffusion": [],
+ "stable_diffusion": [],
+ "stable_diffusion_xl": [],
+}
+
+try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_pt_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_pt_objects))
+else:
+ _import_structure["auto_pipeline"] = [
+ "AutoPipelineForImage2Image",
+ "AutoPipelineForInpainting",
+ "AutoPipelineForText2Image",
+ ]
+ _import_structure["consistency_models"] = ["ConsistencyModelPipeline"]
+ _import_structure["dance_diffusion"] = ["DanceDiffusionPipeline"]
+ _import_structure["ddim"] = ["DDIMPipeline"]
+ _import_structure["ddpm"] = ["DDPMPipeline"]
+ _import_structure["dit"] = ["DiTPipeline"]
+ _import_structure["latent_diffusion"].extend(["LDMSuperResolutionPipeline"])
+ _import_structure["pipeline_utils"] = [
+ "AudioPipelineOutput",
+ "DiffusionPipeline",
+ "StableDiffusionMixin",
+ "ImagePipelineOutput",
+ ]
+ _import_structure["deprecated"].extend(
+ [
+ "PNDMPipeline",
+ "LDMPipeline",
+ "RePaintPipeline",
+ "ScoreSdeVePipeline",
+ "KarrasVePipeline",
+ ]
+ )
+try:
+ if not (is_torch_available() and is_librosa_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_torch_and_librosa_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_librosa_objects))
+else:
+ _import_structure["deprecated"].extend(["AudioDiffusionPipeline", "Mel"])
+
+try:
+ if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_transformers_and_torch_and_note_seq_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_transformers_and_torch_and_note_seq_objects))
+else:
+ _import_structure["deprecated"].extend(
+ [
+ "MidiProcessor",
+ "SpectrogramDiffusionPipeline",
+ ]
+ )
+
+try:
+ if not (is_torch_available() and is_transformers_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["deprecated"].extend(
+ [
+ "VQDiffusionPipeline",
+ "AltDiffusionPipeline",
+ "AltDiffusionImg2ImgPipeline",
+ "CycleDiffusionPipeline",
+ "StableDiffusionInpaintPipelineLegacy",
+ "StableDiffusionPix2PixZeroPipeline",
+ "StableDiffusionParadigmsPipeline",
+ "StableDiffusionModelEditingPipeline",
+ "VersatileDiffusionDualGuidedPipeline",
+ "VersatileDiffusionImageVariationPipeline",
+ "VersatileDiffusionPipeline",
+ "VersatileDiffusionTextToImagePipeline",
+ ]
+ )
+ _import_structure["amused"] = ["AmusedImg2ImgPipeline", "AmusedInpaintPipeline", "AmusedPipeline"]
+ _import_structure["animatediff"] = [
+ "AnimateDiffPipeline",
+ "AnimateDiffVideoToVideoPipeline",
+ ]
+ _import_structure["audioldm"] = ["AudioLDMPipeline"]
+ _import_structure["audioldm2"] = [
+ "AudioLDM2Pipeline",
+ "AudioLDM2ProjectionModel",
+ "AudioLDM2UNet2DConditionModel",
+ ]
+ _import_structure["blip_diffusion"] = ["BlipDiffusionPipeline"]
+ _import_structure["brushnet"].extend(
+ ["StableDiffusionBrushNetPipeline", "StableDiffusionBrushNetPowerPaintPipeline"]
+ )
+ _import_structure["controlnet"].extend(
+ [
+ "BlipDiffusionControlNetPipeline",
+ "StableDiffusionControlNetImg2ImgPipeline",
+ "StableDiffusionControlNetInpaintPipeline",
+ "StableDiffusionControlNetPipeline",
+ "StableDiffusionXLControlNetImg2ImgPipeline",
+ "StableDiffusionXLControlNetInpaintPipeline",
+ "StableDiffusionXLControlNetPipeline",
+ ]
+ )
+ _import_structure["deepfloyd_if"] = [
+ "IFImg2ImgPipeline",
+ "IFImg2ImgSuperResolutionPipeline",
+ "IFInpaintingPipeline",
+ "IFInpaintingSuperResolutionPipeline",
+ "IFPipeline",
+ "IFSuperResolutionPipeline",
+ ]
+ _import_structure["kandinsky"] = [
+ "KandinskyCombinedPipeline",
+ "KandinskyImg2ImgCombinedPipeline",
+ "KandinskyImg2ImgPipeline",
+ "KandinskyInpaintCombinedPipeline",
+ "KandinskyInpaintPipeline",
+ "KandinskyPipeline",
+ "KandinskyPriorPipeline",
+ ]
+ _import_structure["kandinsky2_2"] = [
+ "KandinskyV22CombinedPipeline",
+ "KandinskyV22ControlnetImg2ImgPipeline",
+ "KandinskyV22ControlnetPipeline",
+ "KandinskyV22Img2ImgCombinedPipeline",
+ "KandinskyV22Img2ImgPipeline",
+ "KandinskyV22InpaintCombinedPipeline",
+ "KandinskyV22InpaintPipeline",
+ "KandinskyV22Pipeline",
+ "KandinskyV22PriorEmb2EmbPipeline",
+ "KandinskyV22PriorPipeline",
+ ]
+ _import_structure["kandinsky3"] = [
+ "Kandinsky3Img2ImgPipeline",
+ "Kandinsky3Pipeline",
+ ]
+ _import_structure["latent_consistency_models"] = [
+ "LatentConsistencyModelImg2ImgPipeline",
+ "LatentConsistencyModelPipeline",
+ ]
+ _import_structure["latent_diffusion"].extend(["LDMTextToImagePipeline"])
+ _import_structure["musicldm"] = ["MusicLDMPipeline"]
+ _import_structure["paint_by_example"] = ["PaintByExamplePipeline"]
+ _import_structure["pia"] = ["PIAPipeline"]
+ _import_structure["pixart_alpha"] = ["PixArtAlphaPipeline"]
+ _import_structure["semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
+ _import_structure["shap_e"] = ["ShapEImg2ImgPipeline", "ShapEPipeline"]
+ _import_structure["stable_cascade"] = [
+ "StableCascadeCombinedPipeline",
+ "StableCascadeDecoderPipeline",
+ "StableCascadePriorPipeline",
+ ]
+ _import_structure["stable_diffusion"].extend(
+ [
+ "CLIPImageProjection",
+ "StableDiffusionDepth2ImgPipeline",
+ "StableDiffusionImageVariationPipeline",
+ "StableDiffusionImg2ImgPipeline",
+ "StableDiffusionInpaintPipeline",
+ "StableDiffusionInstructPix2PixPipeline",
+ "StableDiffusionLatentUpscalePipeline",
+ "StableDiffusionPipeline",
+ "StableDiffusionUpscalePipeline",
+ "StableUnCLIPImg2ImgPipeline",
+ "StableUnCLIPPipeline",
+ "StableDiffusionLDM3DPipeline",
+ ]
+ )
+ _import_structure["stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"]
+ _import_structure["stable_diffusion_safe"] = ["StableDiffusionPipelineSafe"]
+ _import_structure["stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"]
+ _import_structure["stable_diffusion_gligen"] = [
+ "StableDiffusionGLIGENPipeline",
+ "StableDiffusionGLIGENTextImagePipeline",
+ ]
+ _import_structure["stable_video_diffusion"] = ["StableVideoDiffusionPipeline"]
+ _import_structure["stable_diffusion_xl"].extend(
+ [
+ "StableDiffusionXLImg2ImgPipeline",
+ "StableDiffusionXLInpaintPipeline",
+ "StableDiffusionXLInstructPix2PixPipeline",
+ "StableDiffusionXLPipeline",
+ ]
+ )
+ _import_structure["stable_diffusion_diffedit"] = ["StableDiffusionDiffEditPipeline"]
+ _import_structure["stable_diffusion_ldm3d"] = ["StableDiffusionLDM3DPipeline"]
+ _import_structure["stable_diffusion_panorama"] = ["StableDiffusionPanoramaPipeline"]
+ _import_structure["t2i_adapter"] = [
+ "StableDiffusionAdapterPipeline",
+ "StableDiffusionXLAdapterPipeline",
+ ]
+ _import_structure["text_to_video_synthesis"] = [
+ "TextToVideoSDPipeline",
+ "TextToVideoZeroPipeline",
+ "TextToVideoZeroSDXLPipeline",
+ "VideoToVideoSDPipeline",
+ ]
+ _import_structure["i2vgen_xl"] = ["I2VGenXLPipeline"]
+ _import_structure["unclip"] = ["UnCLIPImageVariationPipeline", "UnCLIPPipeline"]
+ _import_structure["unidiffuser"] = [
+ "ImageTextPipelineOutput",
+ "UniDiffuserModel",
+ "UniDiffuserPipeline",
+ "UniDiffuserTextDecoder",
+ ]
+ _import_structure["wuerstchen"] = [
+ "WuerstchenCombinedPipeline",
+ "WuerstchenDecoderPipeline",
+ "WuerstchenPriorPipeline",
+ ]
+try:
+ if not is_onnx_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_onnx_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_onnx_objects))
+else:
+ _import_structure["onnx_utils"] = ["OnnxRuntimeModel"]
+try:
+ if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_torch_and_transformers_and_onnx_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_onnx_objects))
+else:
+ _import_structure["stable_diffusion"].extend(
+ [
+ "OnnxStableDiffusionImg2ImgPipeline",
+ "OnnxStableDiffusionInpaintPipeline",
+ "OnnxStableDiffusionPipeline",
+ "OnnxStableDiffusionUpscalePipeline",
+ "StableDiffusionOnnxPipeline",
+ ]
+ )
+
+try:
+ if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import (
+ dummy_torch_and_transformers_and_k_diffusion_objects,
+ )
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_k_diffusion_objects))
+else:
+ _import_structure["stable_diffusion_k_diffusion"] = [
+ "StableDiffusionKDiffusionPipeline",
+ "StableDiffusionXLKDiffusionPipeline",
+ ]
+try:
+ if not is_flax_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_flax_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_flax_objects))
+else:
+ _import_structure["pipeline_flax_utils"] = ["FlaxDiffusionPipeline"]
+try:
+ if not (is_flax_available() and is_transformers_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_flax_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects))
+else:
+ _import_structure["controlnet"].extend(["FlaxStableDiffusionControlNetPipeline"])
+ _import_structure["stable_diffusion"].extend(
+ [
+ "FlaxStableDiffusionImg2ImgPipeline",
+ "FlaxStableDiffusionInpaintPipeline",
+ "FlaxStableDiffusionPipeline",
+ ]
+ )
+ _import_structure["stable_diffusion_xl"].extend(
+ [
+ "FlaxStableDiffusionXLPipeline",
+ ]
+ )
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_pt_objects import * # noqa F403
+
+ else:
+ from .auto_pipeline import (
+ AutoPipelineForImage2Image,
+ AutoPipelineForInpainting,
+ AutoPipelineForText2Image,
+ )
+ from .consistency_models import ConsistencyModelPipeline
+ from .dance_diffusion import DanceDiffusionPipeline
+ from .ddim import DDIMPipeline
+ from .ddpm import DDPMPipeline
+ from .deprecated import KarrasVePipeline, LDMPipeline, PNDMPipeline, RePaintPipeline, ScoreSdeVePipeline
+ from .dit import DiTPipeline
+ from .latent_diffusion import LDMSuperResolutionPipeline
+ from .pipeline_utils import (
+ AudioPipelineOutput,
+ DiffusionPipeline,
+ ImagePipelineOutput,
+ StableDiffusionMixin,
+ )
+
+ try:
+ if not (is_torch_available() and is_librosa_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_torch_and_librosa_objects import *
+ else:
+ from .deprecated import AudioDiffusionPipeline, Mel
+
+ try:
+ if not (is_torch_available() and is_transformers_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .amused import AmusedImg2ImgPipeline, AmusedInpaintPipeline, AmusedPipeline
+ from .animatediff import AnimateDiffPipeline, AnimateDiffVideoToVideoPipeline
+ from .audioldm import AudioLDMPipeline
+ from .audioldm2 import (
+ AudioLDM2Pipeline,
+ AudioLDM2ProjectionModel,
+ AudioLDM2UNet2DConditionModel,
+ )
+ from .blip_diffusion import BlipDiffusionPipeline
+ from .brushnet import StableDiffusionBrushNetPipeline, StableDiffusionBrushNetPowerPaintPipeline
+ from .controlnet import (
+ BlipDiffusionControlNetPipeline,
+ StableDiffusionControlNetImg2ImgPipeline,
+ StableDiffusionControlNetInpaintPipeline,
+ StableDiffusionControlNetPipeline,
+ StableDiffusionXLControlNetImg2ImgPipeline,
+ StableDiffusionXLControlNetInpaintPipeline,
+ StableDiffusionXLControlNetPipeline,
+ )
+ from .deepfloyd_if import (
+ IFImg2ImgPipeline,
+ IFImg2ImgSuperResolutionPipeline,
+ IFInpaintingPipeline,
+ IFInpaintingSuperResolutionPipeline,
+ IFPipeline,
+ IFSuperResolutionPipeline,
+ )
+ from .deprecated import (
+ AltDiffusionImg2ImgPipeline,
+ AltDiffusionPipeline,
+ CycleDiffusionPipeline,
+ StableDiffusionInpaintPipelineLegacy,
+ StableDiffusionModelEditingPipeline,
+ StableDiffusionParadigmsPipeline,
+ StableDiffusionPix2PixZeroPipeline,
+ VersatileDiffusionDualGuidedPipeline,
+ VersatileDiffusionImageVariationPipeline,
+ VersatileDiffusionPipeline,
+ VersatileDiffusionTextToImagePipeline,
+ VQDiffusionPipeline,
+ )
+ from .i2vgen_xl import I2VGenXLPipeline
+ from .kandinsky import (
+ KandinskyCombinedPipeline,
+ KandinskyImg2ImgCombinedPipeline,
+ KandinskyImg2ImgPipeline,
+ KandinskyInpaintCombinedPipeline,
+ KandinskyInpaintPipeline,
+ KandinskyPipeline,
+ KandinskyPriorPipeline,
+ )
+ from .kandinsky2_2 import (
+ KandinskyV22CombinedPipeline,
+ KandinskyV22ControlnetImg2ImgPipeline,
+ KandinskyV22ControlnetPipeline,
+ KandinskyV22Img2ImgCombinedPipeline,
+ KandinskyV22Img2ImgPipeline,
+ KandinskyV22InpaintCombinedPipeline,
+ KandinskyV22InpaintPipeline,
+ KandinskyV22Pipeline,
+ KandinskyV22PriorEmb2EmbPipeline,
+ KandinskyV22PriorPipeline,
+ )
+ from .kandinsky3 import (
+ Kandinsky3Img2ImgPipeline,
+ Kandinsky3Pipeline,
+ )
+ from .latent_consistency_models import (
+ LatentConsistencyModelImg2ImgPipeline,
+ LatentConsistencyModelPipeline,
+ )
+ from .latent_diffusion import LDMTextToImagePipeline
+ from .musicldm import MusicLDMPipeline
+ from .paint_by_example import PaintByExamplePipeline
+ from .pia import PIAPipeline
+ from .pixart_alpha import PixArtAlphaPipeline
+ from .semantic_stable_diffusion import SemanticStableDiffusionPipeline
+ from .shap_e import ShapEImg2ImgPipeline, ShapEPipeline
+ from .stable_cascade import (
+ StableCascadeCombinedPipeline,
+ StableCascadeDecoderPipeline,
+ StableCascadePriorPipeline,
+ )
+ from .stable_diffusion import (
+ CLIPImageProjection,
+ StableDiffusionDepth2ImgPipeline,
+ StableDiffusionImageVariationPipeline,
+ StableDiffusionImg2ImgPipeline,
+ StableDiffusionInpaintPipeline,
+ StableDiffusionInstructPix2PixPipeline,
+ StableDiffusionLatentUpscalePipeline,
+ StableDiffusionPipeline,
+ StableDiffusionUpscalePipeline,
+ StableUnCLIPImg2ImgPipeline,
+ StableUnCLIPPipeline,
+ )
+ from .stable_diffusion_attend_and_excite import StableDiffusionAttendAndExcitePipeline
+ from .stable_diffusion_diffedit import StableDiffusionDiffEditPipeline
+ from .stable_diffusion_gligen import StableDiffusionGLIGENPipeline, StableDiffusionGLIGENTextImagePipeline
+ from .stable_diffusion_ldm3d import StableDiffusionLDM3DPipeline
+ from .stable_diffusion_panorama import StableDiffusionPanoramaPipeline
+ from .stable_diffusion_safe import StableDiffusionPipelineSafe
+ from .stable_diffusion_sag import StableDiffusionSAGPipeline
+ from .stable_diffusion_xl import (
+ StableDiffusionXLImg2ImgPipeline,
+ StableDiffusionXLInpaintPipeline,
+ StableDiffusionXLInstructPix2PixPipeline,
+ StableDiffusionXLPipeline,
+ )
+ from .stable_video_diffusion import StableVideoDiffusionPipeline
+ from .t2i_adapter import (
+ StableDiffusionAdapterPipeline,
+ StableDiffusionXLAdapterPipeline,
+ )
+ from .text_to_video_synthesis import (
+ TextToVideoSDPipeline,
+ TextToVideoZeroPipeline,
+ TextToVideoZeroSDXLPipeline,
+ VideoToVideoSDPipeline,
+ )
+ from .unclip import UnCLIPImageVariationPipeline, UnCLIPPipeline
+ from .unidiffuser import (
+ ImageTextPipelineOutput,
+ UniDiffuserModel,
+ UniDiffuserPipeline,
+ UniDiffuserTextDecoder,
+ )
+ from .wuerstchen import (
+ WuerstchenCombinedPipeline,
+ WuerstchenDecoderPipeline,
+ WuerstchenPriorPipeline,
+ )
+
+ try:
+ if not is_onnx_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_onnx_objects import * # noqa F403
+
+ else:
+ from .onnx_utils import OnnxRuntimeModel
+
+ try:
+ if not (is_torch_available() and is_transformers_available() and is_onnx_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_torch_and_transformers_and_onnx_objects import *
+ else:
+ from .stable_diffusion import (
+ OnnxStableDiffusionImg2ImgPipeline,
+ OnnxStableDiffusionInpaintPipeline,
+ OnnxStableDiffusionPipeline,
+ OnnxStableDiffusionUpscalePipeline,
+ StableDiffusionOnnxPipeline,
+ )
+
+ try:
+ if not (is_torch_available() and is_transformers_available() and is_k_diffusion_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_torch_and_transformers_and_k_diffusion_objects import *
+ else:
+ from .stable_diffusion_k_diffusion import (
+ StableDiffusionKDiffusionPipeline,
+ StableDiffusionXLKDiffusionPipeline,
+ )
+
+ try:
+ if not is_flax_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_flax_objects import * # noqa F403
+ else:
+ from .pipeline_flax_utils import FlaxDiffusionPipeline
+
+ try:
+ if not (is_flax_available() and is_transformers_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_flax_and_transformers_objects import *
+ else:
+ from .controlnet import FlaxStableDiffusionControlNetPipeline
+ from .stable_diffusion import (
+ FlaxStableDiffusionImg2ImgPipeline,
+ FlaxStableDiffusionInpaintPipeline,
+ FlaxStableDiffusionPipeline,
+ )
+ from .stable_diffusion_xl import (
+ FlaxStableDiffusionXLPipeline,
+ )
+
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_transformers_and_torch_and_note_seq_objects import * # noqa F403
+
+ else:
+ from .deprecated import (
+ MidiProcessor,
+ SpectrogramDiffusionPipeline,
+ )
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/pipelines/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..707f0141f6e9decc8da757001c84250a81c8b99f
Binary files /dev/null and b/model/diffusers_c/pipelines/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/pipelines/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..63ca34a9450d73d5ac2c888d9a136df73562db02
Binary files /dev/null and b/model/diffusers_c/pipelines/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/__pycache__/auto_pipeline.cpython-39.pyc b/model/diffusers_c/pipelines/__pycache__/auto_pipeline.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..16b718cac56a4b2f8364a9bfbea02761a8fb7f57
Binary files /dev/null and b/model/diffusers_c/pipelines/__pycache__/auto_pipeline.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/__pycache__/free_init_utils.cpython-39.pyc b/model/diffusers_c/pipelines/__pycache__/free_init_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..abb9c89d645ed86697c9e5da1d722865be2ddbcc
Binary files /dev/null and b/model/diffusers_c/pipelines/__pycache__/free_init_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/__pycache__/pipeline_loading_utils.cpython-310.pyc b/model/diffusers_c/pipelines/__pycache__/pipeline_loading_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9c25d2f352375f3c61c3d7e33af0b0938ab297ea
Binary files /dev/null and b/model/diffusers_c/pipelines/__pycache__/pipeline_loading_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/pipelines/__pycache__/pipeline_loading_utils.cpython-39.pyc b/model/diffusers_c/pipelines/__pycache__/pipeline_loading_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..344ed8d71b2494fe7ddd1d2f8e3f6957d8da87f8
Binary files /dev/null and b/model/diffusers_c/pipelines/__pycache__/pipeline_loading_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/__pycache__/pipeline_utils.cpython-310.pyc b/model/diffusers_c/pipelines/__pycache__/pipeline_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3c93882d20b97cdf825e888b95e9a15db59a4278
Binary files /dev/null and b/model/diffusers_c/pipelines/__pycache__/pipeline_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/pipelines/__pycache__/pipeline_utils.cpython-39.pyc b/model/diffusers_c/pipelines/__pycache__/pipeline_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..37fd7451a980b2c81a0373bc8eb447b1fbc301ec
Binary files /dev/null and b/model/diffusers_c/pipelines/__pycache__/pipeline_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/amused/__init__.py b/model/diffusers_c/pipelines/amused/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..3c4d07a426b54fabfcdf35bfb8e4486cd828b3b3
--- /dev/null
+++ b/model/diffusers_c/pipelines/amused/__init__.py
@@ -0,0 +1,62 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ AmusedImg2ImgPipeline,
+ AmusedInpaintPipeline,
+ AmusedPipeline,
+ )
+
+ _dummy_objects.update(
+ {
+ "AmusedPipeline": AmusedPipeline,
+ "AmusedImg2ImgPipeline": AmusedImg2ImgPipeline,
+ "AmusedInpaintPipeline": AmusedInpaintPipeline,
+ }
+ )
+else:
+ _import_structure["pipeline_amused"] = ["AmusedPipeline"]
+ _import_structure["pipeline_amused_img2img"] = ["AmusedImg2ImgPipeline"]
+ _import_structure["pipeline_amused_inpaint"] = ["AmusedInpaintPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ AmusedPipeline,
+ )
+ else:
+ from .pipeline_amused import AmusedPipeline
+ from .pipeline_amused_img2img import AmusedImg2ImgPipeline
+ from .pipeline_amused_inpaint import AmusedInpaintPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/amused/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/amused/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..36d0486b7e06931d55af86b82baab6d260c18b51
Binary files /dev/null and b/model/diffusers_c/pipelines/amused/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused.cpython-39.pyc b/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0135b5596bfd14dc3184392a25747c483304685f
Binary files /dev/null and b/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b4bf2345b50d881c56797e97895b51bddc588631
Binary files /dev/null and b/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused_inpaint.cpython-39.pyc b/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused_inpaint.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4bd6de000d7f029a74a16f6929882f9810a49bdc
Binary files /dev/null and b/model/diffusers_c/pipelines/amused/__pycache__/pipeline_amused_inpaint.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/amused/pipeline_amused.py b/model/diffusers_c/pipelines/amused/pipeline_amused.py
new file mode 100644
index 0000000000000000000000000000000000000000..aa682b46fe7030351553e5af965dad71425ae8f8
--- /dev/null
+++ b/model/diffusers_c/pipelines/amused/pipeline_amused.py
@@ -0,0 +1,328 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...models import UVit2DModel, VQModel
+from ...schedulers import AmusedScheduler
+from ...utils import replace_example_docstring
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import AmusedPipeline
+
+ >>> pipe = AmusedPipeline.from_pretrained(
+ ... "amused/amused-512", variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> image = pipe(prompt).images[0]
+ ```
+"""
+
+
+class AmusedPipeline(DiffusionPipeline):
+ image_processor: VaeImageProcessor
+ vqvae: VQModel
+ tokenizer: CLIPTokenizer
+ text_encoder: CLIPTextModelWithProjection
+ transformer: UVit2DModel
+ scheduler: AmusedScheduler
+
+ model_cpu_offload_seq = "text_encoder->transformer->vqvae"
+
+ def __init__(
+ self,
+ vqvae: VQModel,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModelWithProjection,
+ transformer: UVit2DModel,
+ scheduler: AmusedScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vqvae=vqvae,
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ transformer=transformer,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vqvae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_normalize=False)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[List[str], str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 12,
+ guidance_scale: float = 10.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[torch.Generator] = None,
+ latents: Optional[torch.IntTensor] = None,
+ prompt_embeds: Optional[torch.Tensor] = None,
+ encoder_hidden_states: Optional[torch.Tensor] = None,
+ negative_prompt_embeds: Optional[torch.Tensor] = None,
+ negative_encoder_hidden_states: Optional[torch.Tensor] = None,
+ output_type="pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ micro_conditioning_aesthetic_score: int = 6,
+ micro_conditioning_crop_coord: Tuple[int, int] = (0, 0),
+ temperature: Union[int, Tuple[int, int], List[int]] = (2, 0),
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.transformer.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 16):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 10.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.IntTensor`, *optional*):
+ Pre-generated tokens representing latent vectors in `self.vqvae`, to be used as inputs for image
+ gneration. If not provided, the starting latents will be completely masked.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument. A single vector from the
+ pooled and projected final hidden states.
+ encoder_hidden_states (`torch.FloatTensor`, *optional*):
+ Pre-generated penultimate hidden states from the text encoder providing additional text conditioning.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ negative_encoder_hidden_states (`torch.FloatTensor`, *optional*):
+ Analogous to `encoder_hidden_states` for the positive prompt.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ micro_conditioning_aesthetic_score (`int`, *optional*, defaults to 6):
+ The targeted aesthetic score according to the laion aesthetic classifier. See https://laion.ai/blog/laion-aesthetics/
+ and the micro-conditioning section of https://arxiv.org/abs/2307.01952.
+ micro_conditioning_crop_coord (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ The targeted height, width crop coordinates. See the micro-conditioning section of https://arxiv.org/abs/2307.01952.
+ temperature (`Union[int, Tuple[int, int], List[int]]`, *optional*, defaults to (2, 0)):
+ Configures the temperature scheduler on `self.scheduler` see `AmusedScheduler#set_timesteps`.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.pipeline_utils.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.pipeline_utils.ImagePipelineOutput`] is returned, otherwise a
+ `tuple` is returned where the first element is a list with the generated images.
+ """
+ if (prompt_embeds is not None and encoder_hidden_states is None) or (
+ prompt_embeds is None and encoder_hidden_states is not None
+ ):
+ raise ValueError("pass either both `prompt_embeds` and `encoder_hidden_states` or neither")
+
+ if (negative_prompt_embeds is not None and negative_encoder_hidden_states is None) or (
+ negative_prompt_embeds is None and negative_encoder_hidden_states is not None
+ ):
+ raise ValueError(
+ "pass either both `negatve_prompt_embeds` and `negative_encoder_hidden_states` or neither"
+ )
+
+ if (prompt is None and prompt_embeds is None) or (prompt is not None and prompt_embeds is not None):
+ raise ValueError("pass only one of `prompt` or `prompt_embeds`")
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if height is None:
+ height = self.transformer.config.sample_size * self.vae_scale_factor
+
+ if width is None:
+ width = self.transformer.config.sample_size * self.vae_scale_factor
+
+ if prompt_embeds is None:
+ input_ids = self.tokenizer(
+ prompt,
+ return_tensors="pt",
+ padding="max_length",
+ truncation=True,
+ max_length=self.tokenizer.model_max_length,
+ ).input_ids.to(self._execution_device)
+
+ outputs = self.text_encoder(input_ids, return_dict=True, output_hidden_states=True)
+ prompt_embeds = outputs.text_embeds
+ encoder_hidden_states = outputs.hidden_states[-2]
+
+ prompt_embeds = prompt_embeds.repeat(num_images_per_prompt, 1)
+ encoder_hidden_states = encoder_hidden_states.repeat(num_images_per_prompt, 1, 1)
+
+ if guidance_scale > 1.0:
+ if negative_prompt_embeds is None:
+ if negative_prompt is None:
+ negative_prompt = [""] * len(prompt)
+
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+
+ input_ids = self.tokenizer(
+ negative_prompt,
+ return_tensors="pt",
+ padding="max_length",
+ truncation=True,
+ max_length=self.tokenizer.model_max_length,
+ ).input_ids.to(self._execution_device)
+
+ outputs = self.text_encoder(input_ids, return_dict=True, output_hidden_states=True)
+ negative_prompt_embeds = outputs.text_embeds
+ negative_encoder_hidden_states = outputs.hidden_states[-2]
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(num_images_per_prompt, 1)
+ negative_encoder_hidden_states = negative_encoder_hidden_states.repeat(num_images_per_prompt, 1, 1)
+
+ prompt_embeds = torch.concat([negative_prompt_embeds, prompt_embeds])
+ encoder_hidden_states = torch.concat([negative_encoder_hidden_states, encoder_hidden_states])
+
+ # Note that the micro conditionings _do_ flip the order of width, height for the original size
+ # and the crop coordinates. This is how it was done in the original code base
+ micro_conds = torch.tensor(
+ [
+ width,
+ height,
+ micro_conditioning_crop_coord[0],
+ micro_conditioning_crop_coord[1],
+ micro_conditioning_aesthetic_score,
+ ],
+ device=self._execution_device,
+ dtype=encoder_hidden_states.dtype,
+ )
+ micro_conds = micro_conds.unsqueeze(0)
+ micro_conds = micro_conds.expand(2 * batch_size if guidance_scale > 1.0 else batch_size, -1)
+
+ shape = (batch_size, height // self.vae_scale_factor, width // self.vae_scale_factor)
+
+ if latents is None:
+ latents = torch.full(
+ shape, self.scheduler.config.mask_token_id, dtype=torch.long, device=self._execution_device
+ )
+
+ self.scheduler.set_timesteps(num_inference_steps, temperature, self._execution_device)
+
+ num_warmup_steps = len(self.scheduler.timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, timestep in enumerate(self.scheduler.timesteps):
+ if guidance_scale > 1.0:
+ model_input = torch.cat([latents] * 2)
+ else:
+ model_input = latents
+
+ model_output = self.transformer(
+ model_input,
+ micro_conds=micro_conds,
+ pooled_text_emb=prompt_embeds,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+
+ if guidance_scale > 1.0:
+ uncond_logits, cond_logits = model_output.chunk(2)
+ model_output = uncond_logits + guidance_scale * (cond_logits - uncond_logits)
+
+ latents = self.scheduler.step(
+ model_output=model_output,
+ timestep=timestep,
+ sample=latents,
+ generator=generator,
+ ).prev_sample
+
+ if i == len(self.scheduler.timesteps) - 1 or (
+ (i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0
+ ):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, timestep, latents)
+
+ if output_type == "latent":
+ output = latents
+ else:
+ needs_upcasting = self.vqvae.dtype == torch.float16 and self.vqvae.config.force_upcast
+
+ if needs_upcasting:
+ self.vqvae.float()
+
+ output = self.vqvae.decode(
+ latents,
+ force_not_quantize=True,
+ shape=(
+ batch_size,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ self.vqvae.config.latent_channels,
+ ),
+ ).sample.clip(0, 1)
+ output = self.image_processor.postprocess(output, output_type)
+
+ if needs_upcasting:
+ self.vqvae.half()
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (output,)
+
+ return ImagePipelineOutput(output)
diff --git a/model/diffusers_c/pipelines/amused/pipeline_amused_img2img.py b/model/diffusers_c/pipelines/amused/pipeline_amused_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..444d6354b7daea889778ae262a233fbe7f73e6e9
--- /dev/null
+++ b/model/diffusers_c/pipelines/amused/pipeline_amused_img2img.py
@@ -0,0 +1,347 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...models import UVit2DModel, VQModel
+from ...schedulers import AmusedScheduler
+from ...utils import replace_example_docstring
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import AmusedImg2ImgPipeline
+ >>> from diffusers.utils import load_image
+
+ >>> pipe = AmusedImg2ImgPipeline.from_pretrained(
+ ... "amused/amused-512", variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "winter mountains"
+ >>> input_image = (
+ ... load_image(
+ ... "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/open_muse/mountains.jpg"
+ ... )
+ ... .resize((512, 512))
+ ... .convert("RGB")
+ ... )
+ >>> image = pipe(prompt, input_image).images[0]
+ ```
+"""
+
+
+class AmusedImg2ImgPipeline(DiffusionPipeline):
+ image_processor: VaeImageProcessor
+ vqvae: VQModel
+ tokenizer: CLIPTokenizer
+ text_encoder: CLIPTextModelWithProjection
+ transformer: UVit2DModel
+ scheduler: AmusedScheduler
+
+ model_cpu_offload_seq = "text_encoder->transformer->vqvae"
+
+ # TODO - when calling self.vqvae.quantize, it uses self.vqvae.quantize.embedding.weight before
+ # the forward method of self.vqvae.quantize, so the hook doesn't get called to move the parameter
+ # off the meta device. There should be a way to fix this instead of just not offloading it
+ _exclude_from_cpu_offload = ["vqvae"]
+
+ def __init__(
+ self,
+ vqvae: VQModel,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModelWithProjection,
+ transformer: UVit2DModel,
+ scheduler: AmusedScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vqvae=vqvae,
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ transformer=transformer,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vqvae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_normalize=False)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[List[str], str]] = None,
+ image: PipelineImageInput = None,
+ strength: float = 0.5,
+ num_inference_steps: int = 12,
+ guidance_scale: float = 10.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[torch.Generator] = None,
+ prompt_embeds: Optional[torch.Tensor] = None,
+ encoder_hidden_states: Optional[torch.Tensor] = None,
+ negative_prompt_embeds: Optional[torch.Tensor] = None,
+ negative_encoder_hidden_states: Optional[torch.Tensor] = None,
+ output_type="pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ micro_conditioning_aesthetic_score: int = 6,
+ micro_conditioning_crop_coord: Tuple[int, int] = (0, 0),
+ temperature: Union[int, Tuple[int, int], List[int]] = (2, 0),
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
+ numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
+ or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
+ list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
+ latents as `image`, but if passing latents directly it is not encoded again.
+ strength (`float`, *optional*, defaults to 0.5):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 16):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 10.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument. A single vector from the
+ pooled and projected final hidden states.
+ encoder_hidden_states (`torch.FloatTensor`, *optional*):
+ Pre-generated penultimate hidden states from the text encoder providing additional text conditioning.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ negative_encoder_hidden_states (`torch.FloatTensor`, *optional*):
+ Analogous to `encoder_hidden_states` for the positive prompt.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ micro_conditioning_aesthetic_score (`int`, *optional*, defaults to 6):
+ The targeted aesthetic score according to the laion aesthetic classifier. See https://laion.ai/blog/laion-aesthetics/
+ and the micro-conditioning section of https://arxiv.org/abs/2307.01952.
+ micro_conditioning_crop_coord (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ The targeted height, width crop coordinates. See the micro-conditioning section of https://arxiv.org/abs/2307.01952.
+ temperature (`Union[int, Tuple[int, int], List[int]]`, *optional*, defaults to (2, 0)):
+ Configures the temperature scheduler on `self.scheduler` see `AmusedScheduler#set_timesteps`.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.pipeline_utils.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.pipeline_utils.ImagePipelineOutput`] is returned, otherwise a
+ `tuple` is returned where the first element is a list with the generated images.
+ """
+
+ if (prompt_embeds is not None and encoder_hidden_states is None) or (
+ prompt_embeds is None and encoder_hidden_states is not None
+ ):
+ raise ValueError("pass either both `prompt_embeds` and `encoder_hidden_states` or neither")
+
+ if (negative_prompt_embeds is not None and negative_encoder_hidden_states is None) or (
+ negative_prompt_embeds is None and negative_encoder_hidden_states is not None
+ ):
+ raise ValueError(
+ "pass either both `negatve_prompt_embeds` and `negative_encoder_hidden_states` or neither"
+ )
+
+ if (prompt is None and prompt_embeds is None) or (prompt is not None and prompt_embeds is not None):
+ raise ValueError("pass only one of `prompt` or `prompt_embeds`")
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if prompt_embeds is None:
+ input_ids = self.tokenizer(
+ prompt,
+ return_tensors="pt",
+ padding="max_length",
+ truncation=True,
+ max_length=self.tokenizer.model_max_length,
+ ).input_ids.to(self._execution_device)
+
+ outputs = self.text_encoder(input_ids, return_dict=True, output_hidden_states=True)
+ prompt_embeds = outputs.text_embeds
+ encoder_hidden_states = outputs.hidden_states[-2]
+
+ prompt_embeds = prompt_embeds.repeat(num_images_per_prompt, 1)
+ encoder_hidden_states = encoder_hidden_states.repeat(num_images_per_prompt, 1, 1)
+
+ if guidance_scale > 1.0:
+ if negative_prompt_embeds is None:
+ if negative_prompt is None:
+ negative_prompt = [""] * len(prompt)
+
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+
+ input_ids = self.tokenizer(
+ negative_prompt,
+ return_tensors="pt",
+ padding="max_length",
+ truncation=True,
+ max_length=self.tokenizer.model_max_length,
+ ).input_ids.to(self._execution_device)
+
+ outputs = self.text_encoder(input_ids, return_dict=True, output_hidden_states=True)
+ negative_prompt_embeds = outputs.text_embeds
+ negative_encoder_hidden_states = outputs.hidden_states[-2]
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(num_images_per_prompt, 1)
+ negative_encoder_hidden_states = negative_encoder_hidden_states.repeat(num_images_per_prompt, 1, 1)
+
+ prompt_embeds = torch.concat([negative_prompt_embeds, prompt_embeds])
+ encoder_hidden_states = torch.concat([negative_encoder_hidden_states, encoder_hidden_states])
+
+ image = self.image_processor.preprocess(image)
+
+ height, width = image.shape[-2:]
+
+ # Note that the micro conditionings _do_ flip the order of width, height for the original size
+ # and the crop coordinates. This is how it was done in the original code base
+ micro_conds = torch.tensor(
+ [
+ width,
+ height,
+ micro_conditioning_crop_coord[0],
+ micro_conditioning_crop_coord[1],
+ micro_conditioning_aesthetic_score,
+ ],
+ device=self._execution_device,
+ dtype=encoder_hidden_states.dtype,
+ )
+
+ micro_conds = micro_conds.unsqueeze(0)
+ micro_conds = micro_conds.expand(2 * batch_size if guidance_scale > 1.0 else batch_size, -1)
+
+ self.scheduler.set_timesteps(num_inference_steps, temperature, self._execution_device)
+ num_inference_steps = int(len(self.scheduler.timesteps) * strength)
+ start_timestep_idx = len(self.scheduler.timesteps) - num_inference_steps
+
+ needs_upcasting = self.vqvae.dtype == torch.float16 and self.vqvae.config.force_upcast
+
+ if needs_upcasting:
+ self.vqvae.float()
+
+ latents = self.vqvae.encode(image.to(dtype=self.vqvae.dtype, device=self._execution_device)).latents
+ latents_bsz, channels, latents_height, latents_width = latents.shape
+ latents = self.vqvae.quantize(latents)[2][2].reshape(latents_bsz, latents_height, latents_width)
+ latents = self.scheduler.add_noise(
+ latents, self.scheduler.timesteps[start_timestep_idx - 1], generator=generator
+ )
+ latents = latents.repeat(num_images_per_prompt, 1, 1)
+
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i in range(start_timestep_idx, len(self.scheduler.timesteps)):
+ timestep = self.scheduler.timesteps[i]
+
+ if guidance_scale > 1.0:
+ model_input = torch.cat([latents] * 2)
+ else:
+ model_input = latents
+
+ model_output = self.transformer(
+ model_input,
+ micro_conds=micro_conds,
+ pooled_text_emb=prompt_embeds,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+
+ if guidance_scale > 1.0:
+ uncond_logits, cond_logits = model_output.chunk(2)
+ model_output = uncond_logits + guidance_scale * (cond_logits - uncond_logits)
+
+ latents = self.scheduler.step(
+ model_output=model_output,
+ timestep=timestep,
+ sample=latents,
+ generator=generator,
+ ).prev_sample
+
+ if i == len(self.scheduler.timesteps) - 1 or ((i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, timestep, latents)
+
+ if output_type == "latent":
+ output = latents
+ else:
+ output = self.vqvae.decode(
+ latents,
+ force_not_quantize=True,
+ shape=(
+ batch_size,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ self.vqvae.config.latent_channels,
+ ),
+ ).sample.clip(0, 1)
+ output = self.image_processor.postprocess(output, output_type)
+
+ if needs_upcasting:
+ self.vqvae.half()
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (output,)
+
+ return ImagePipelineOutput(output)
diff --git a/model/diffusers_c/pipelines/amused/pipeline_amused_inpaint.py b/model/diffusers_c/pipelines/amused/pipeline_amused_inpaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..423f5734b4781e33951ce53d3405b93432675828
--- /dev/null
+++ b/model/diffusers_c/pipelines/amused/pipeline_amused_inpaint.py
@@ -0,0 +1,378 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...models import UVit2DModel, VQModel
+from ...schedulers import AmusedScheduler
+from ...utils import replace_example_docstring
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import AmusedInpaintPipeline
+ >>> from diffusers.utils import load_image
+
+ >>> pipe = AmusedInpaintPipeline.from_pretrained(
+ ... "amused/amused-512", variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "fall mountains"
+ >>> input_image = (
+ ... load_image(
+ ... "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/open_muse/mountains_1.jpg"
+ ... )
+ ... .resize((512, 512))
+ ... .convert("RGB")
+ ... )
+ >>> mask = (
+ ... load_image(
+ ... "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/open_muse/mountains_1_mask.png"
+ ... )
+ ... .resize((512, 512))
+ ... .convert("L")
+ ... )
+ >>> pipe(prompt, input_image, mask).images[0].save("out.png")
+ ```
+"""
+
+
+class AmusedInpaintPipeline(DiffusionPipeline):
+ image_processor: VaeImageProcessor
+ vqvae: VQModel
+ tokenizer: CLIPTokenizer
+ text_encoder: CLIPTextModelWithProjection
+ transformer: UVit2DModel
+ scheduler: AmusedScheduler
+
+ model_cpu_offload_seq = "text_encoder->transformer->vqvae"
+
+ # TODO - when calling self.vqvae.quantize, it uses self.vqvae.quantize.embedding.weight before
+ # the forward method of self.vqvae.quantize, so the hook doesn't get called to move the parameter
+ # off the meta device. There should be a way to fix this instead of just not offloading it
+ _exclude_from_cpu_offload = ["vqvae"]
+
+ def __init__(
+ self,
+ vqvae: VQModel,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModelWithProjection,
+ transformer: UVit2DModel,
+ scheduler: AmusedScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vqvae=vqvae,
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ transformer=transformer,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vqvae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_normalize=False)
+ self.mask_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor,
+ do_normalize=False,
+ do_binarize=True,
+ do_convert_grayscale=True,
+ do_resize=True,
+ )
+ self.scheduler.register_to_config(masking_schedule="linear")
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[List[str], str]] = None,
+ image: PipelineImageInput = None,
+ mask_image: PipelineImageInput = None,
+ strength: float = 1.0,
+ num_inference_steps: int = 12,
+ guidance_scale: float = 10.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[torch.Generator] = None,
+ prompt_embeds: Optional[torch.Tensor] = None,
+ encoder_hidden_states: Optional[torch.Tensor] = None,
+ negative_prompt_embeds: Optional[torch.Tensor] = None,
+ negative_encoder_hidden_states: Optional[torch.Tensor] = None,
+ output_type="pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ micro_conditioning_aesthetic_score: int = 6,
+ micro_conditioning_crop_coord: Tuple[int, int] = (0, 0),
+ temperature: Union[int, Tuple[int, int], List[int]] = (2, 0),
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
+ numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
+ or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
+ list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
+ latents as `image`, but if passing latents directly it is not encoded again.
+ mask_image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask
+ are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a
+ single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one
+ color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B,
+ H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W,
+ 1)`, or `(H, W)`.
+ strength (`float`, *optional*, defaults to 1.0):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 16):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 10.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument. A single vector from the
+ pooled and projected final hidden states.
+ encoder_hidden_states (`torch.FloatTensor`, *optional*):
+ Pre-generated penultimate hidden states from the text encoder providing additional text conditioning.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ negative_encoder_hidden_states (`torch.FloatTensor`, *optional*):
+ Analogous to `encoder_hidden_states` for the positive prompt.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ micro_conditioning_aesthetic_score (`int`, *optional*, defaults to 6):
+ The targeted aesthetic score according to the laion aesthetic classifier. See https://laion.ai/blog/laion-aesthetics/
+ and the micro-conditioning section of https://arxiv.org/abs/2307.01952.
+ micro_conditioning_crop_coord (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ The targeted height, width crop coordinates. See the micro-conditioning section of https://arxiv.org/abs/2307.01952.
+ temperature (`Union[int, Tuple[int, int], List[int]]`, *optional*, defaults to (2, 0)):
+ Configures the temperature scheduler on `self.scheduler` see `AmusedScheduler#set_timesteps`.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.pipeline_utils.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.pipeline_utils.ImagePipelineOutput`] is returned, otherwise a
+ `tuple` is returned where the first element is a list with the generated images.
+ """
+
+ if (prompt_embeds is not None and encoder_hidden_states is None) or (
+ prompt_embeds is None and encoder_hidden_states is not None
+ ):
+ raise ValueError("pass either both `prompt_embeds` and `encoder_hidden_states` or neither")
+
+ if (negative_prompt_embeds is not None and negative_encoder_hidden_states is None) or (
+ negative_prompt_embeds is None and negative_encoder_hidden_states is not None
+ ):
+ raise ValueError(
+ "pass either both `negatve_prompt_embeds` and `negative_encoder_hidden_states` or neither"
+ )
+
+ if (prompt is None and prompt_embeds is None) or (prompt is not None and prompt_embeds is not None):
+ raise ValueError("pass only one of `prompt` or `prompt_embeds`")
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if prompt_embeds is None:
+ input_ids = self.tokenizer(
+ prompt,
+ return_tensors="pt",
+ padding="max_length",
+ truncation=True,
+ max_length=self.tokenizer.model_max_length,
+ ).input_ids.to(self._execution_device)
+
+ outputs = self.text_encoder(input_ids, return_dict=True, output_hidden_states=True)
+ prompt_embeds = outputs.text_embeds
+ encoder_hidden_states = outputs.hidden_states[-2]
+
+ prompt_embeds = prompt_embeds.repeat(num_images_per_prompt, 1)
+ encoder_hidden_states = encoder_hidden_states.repeat(num_images_per_prompt, 1, 1)
+
+ if guidance_scale > 1.0:
+ if negative_prompt_embeds is None:
+ if negative_prompt is None:
+ negative_prompt = [""] * len(prompt)
+
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+
+ input_ids = self.tokenizer(
+ negative_prompt,
+ return_tensors="pt",
+ padding="max_length",
+ truncation=True,
+ max_length=self.tokenizer.model_max_length,
+ ).input_ids.to(self._execution_device)
+
+ outputs = self.text_encoder(input_ids, return_dict=True, output_hidden_states=True)
+ negative_prompt_embeds = outputs.text_embeds
+ negative_encoder_hidden_states = outputs.hidden_states[-2]
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(num_images_per_prompt, 1)
+ negative_encoder_hidden_states = negative_encoder_hidden_states.repeat(num_images_per_prompt, 1, 1)
+
+ prompt_embeds = torch.concat([negative_prompt_embeds, prompt_embeds])
+ encoder_hidden_states = torch.concat([negative_encoder_hidden_states, encoder_hidden_states])
+
+ image = self.image_processor.preprocess(image)
+
+ height, width = image.shape[-2:]
+
+ # Note that the micro conditionings _do_ flip the order of width, height for the original size
+ # and the crop coordinates. This is how it was done in the original code base
+ micro_conds = torch.tensor(
+ [
+ width,
+ height,
+ micro_conditioning_crop_coord[0],
+ micro_conditioning_crop_coord[1],
+ micro_conditioning_aesthetic_score,
+ ],
+ device=self._execution_device,
+ dtype=encoder_hidden_states.dtype,
+ )
+
+ micro_conds = micro_conds.unsqueeze(0)
+ micro_conds = micro_conds.expand(2 * batch_size if guidance_scale > 1.0 else batch_size, -1)
+
+ self.scheduler.set_timesteps(num_inference_steps, temperature, self._execution_device)
+ num_inference_steps = int(len(self.scheduler.timesteps) * strength)
+ start_timestep_idx = len(self.scheduler.timesteps) - num_inference_steps
+
+ needs_upcasting = self.vqvae.dtype == torch.float16 and self.vqvae.config.force_upcast
+
+ if needs_upcasting:
+ self.vqvae.float()
+
+ latents = self.vqvae.encode(image.to(dtype=self.vqvae.dtype, device=self._execution_device)).latents
+ latents_bsz, channels, latents_height, latents_width = latents.shape
+ latents = self.vqvae.quantize(latents)[2][2].reshape(latents_bsz, latents_height, latents_width)
+
+ mask = self.mask_processor.preprocess(
+ mask_image, height // self.vae_scale_factor, width // self.vae_scale_factor
+ )
+ mask = mask.reshape(mask.shape[0], latents_height, latents_width).bool().to(latents.device)
+ latents[mask] = self.scheduler.config.mask_token_id
+
+ starting_mask_ratio = mask.sum() / latents.numel()
+
+ latents = latents.repeat(num_images_per_prompt, 1, 1)
+
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i in range(start_timestep_idx, len(self.scheduler.timesteps)):
+ timestep = self.scheduler.timesteps[i]
+
+ if guidance_scale > 1.0:
+ model_input = torch.cat([latents] * 2)
+ else:
+ model_input = latents
+
+ model_output = self.transformer(
+ model_input,
+ micro_conds=micro_conds,
+ pooled_text_emb=prompt_embeds,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+
+ if guidance_scale > 1.0:
+ uncond_logits, cond_logits = model_output.chunk(2)
+ model_output = uncond_logits + guidance_scale * (cond_logits - uncond_logits)
+
+ latents = self.scheduler.step(
+ model_output=model_output,
+ timestep=timestep,
+ sample=latents,
+ generator=generator,
+ starting_mask_ratio=starting_mask_ratio,
+ ).prev_sample
+
+ if i == len(self.scheduler.timesteps) - 1 or ((i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, timestep, latents)
+
+ if output_type == "latent":
+ output = latents
+ else:
+ output = self.vqvae.decode(
+ latents,
+ force_not_quantize=True,
+ shape=(
+ batch_size,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ self.vqvae.config.latent_channels,
+ ),
+ ).sample.clip(0, 1)
+ output = self.image_processor.postprocess(output, output_type)
+
+ if needs_upcasting:
+ self.vqvae.half()
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (output,)
+
+ return ImagePipelineOutput(output)
diff --git a/model/diffusers_c/pipelines/animatediff/__init__.py b/model/diffusers_c/pipelines/animatediff/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..35b99a76fd2107176cc345b20048ec7dd3acf6ca
--- /dev/null
+++ b/model/diffusers_c/pipelines/animatediff/__init__.py
@@ -0,0 +1,49 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {"pipeline_output": ["AnimateDiffPipelineOutput"]}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_animatediff"] = ["AnimateDiffPipeline"]
+ _import_structure["pipeline_animatediff_video2video"] = ["AnimateDiffVideoToVideoPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+
+ else:
+ from .pipeline_animatediff import AnimateDiffPipeline
+ from .pipeline_animatediff_video2video import AnimateDiffVideoToVideoPipeline
+ from .pipeline_output import AnimateDiffPipelineOutput
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/animatediff/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/animatediff/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9885c2da2138f4f1ec957abdd53587ed67b5d64b
Binary files /dev/null and b/model/diffusers_c/pipelines/animatediff/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_animatediff.cpython-39.pyc b/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_animatediff.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..21411a36834b5cf9ac38b4958345ed5fce493a9a
Binary files /dev/null and b/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_animatediff.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_animatediff_video2video.cpython-39.pyc b/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_animatediff_video2video.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f384aecc42827f06bf8b425839fec2a9961b4c17
Binary files /dev/null and b/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_animatediff_video2video.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_output.cpython-39.pyc b/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_output.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5a95b5624d04f5c914b1484eec290bc8aabc622b
Binary files /dev/null and b/model/diffusers_c/pipelines/animatediff/__pycache__/pipeline_output.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/animatediff/pipeline_animatediff.py b/model/diffusers_c/pipelines/animatediff/pipeline_animatediff.py
new file mode 100644
index 0000000000000000000000000000000000000000..cd7f0a283b63d052aceee3a05f47451be730447b
--- /dev/null
+++ b/model/diffusers_c/pipelines/animatediff/pipeline_animatediff.py
@@ -0,0 +1,847 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...models.unets.unet_motion_model import MotionAdapter
+from ...schedulers import (
+ DDIMScheduler,
+ DPMSolverMultistepScheduler,
+ EulerAncestralDiscreteScheduler,
+ EulerDiscreteScheduler,
+ LMSDiscreteScheduler,
+ PNDMScheduler,
+)
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..free_init_utils import FreeInitMixin
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import AnimateDiffPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import MotionAdapter, AnimateDiffPipeline, DDIMScheduler
+ >>> from diffusers.utils import export_to_gif
+
+ >>> adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-2")
+ >>> pipe = AnimateDiffPipeline.from_pretrained("frankjoshua/toonyou_beta6", motion_adapter=adapter)
+ >>> pipe.scheduler = DDIMScheduler(beta_schedule="linear", steps_offset=1, clip_sample=False)
+ >>> output = pipe(prompt="A corgi walking in the park")
+ >>> frames = output.frames[0]
+ >>> export_to_gif(frames, "animation.gif")
+ ```
+"""
+
+
+def tensor2vid(video: torch.Tensor, processor: "VaeImageProcessor", output_type: str = "np"):
+ batch_size, channels, num_frames, height, width = video.shape
+ outputs = []
+ for batch_idx in range(batch_size):
+ batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+ batch_output = processor.postprocess(batch_vid, output_type)
+
+ outputs.append(batch_output)
+
+ if output_type == "np":
+ outputs = np.stack(outputs)
+
+ elif output_type == "pt":
+ outputs = torch.stack(outputs)
+
+ elif not output_type == "pil":
+ raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+ return outputs
+
+
+class AnimateDiffPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FreeInitMixin,
+):
+ r"""
+ Pipeline for text-to-video generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`] to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A [`UNet2DConditionModel`] used to create a UNetMotionModel to denoise the encoded video latents.
+ motion_adapter ([`MotionAdapter`]):
+ A [`MotionAdapter`] to be used in combination with `unet` to denoise the encoded video latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["feature_extractor", "image_encoder", "motion_adapter"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ motion_adapter: MotionAdapter,
+ scheduler: Union[
+ DDIMScheduler,
+ PNDMScheduler,
+ LMSDiscreteScheduler,
+ EulerDiscreteScheduler,
+ EulerAncestralDiscreteScheduler,
+ DPMSolverMultistepScheduler,
+ ],
+ feature_extractor: CLIPImageProcessor = None,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ ):
+ super().__init__()
+ if isinstance(unet, UNet2DConditionModel):
+ unet = UNetMotionModel.from_unet2d(unet, motion_adapter)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ motion_adapter=motion_adapter,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with num_images_per_prompt -> num_videos_per_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis/pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
+ def decode_latents(self, latents):
+ latents = 1 / self.vae.config.scaling_factor * latents
+
+ batch_size, channels, num_frames, height, width = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+ image = self.vae.decode(latents).sample
+ video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ video = video.float()
+ return video
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
+ def prepare_latents(
+ self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+ ):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ num_frames,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ num_frames: Optional[int] = 16,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_videos_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated video.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated video.
+ num_frames (`int`, *optional*, defaults to 16):
+ The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds
+ amounts to 2 seconds of video.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`. Latents should be of shape
+ `(batch_size, num_channel, num_frames, height, width)`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated video. Choose between `torch.FloatTensor`, `PIL.Image` or
+ `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead
+ of a plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.animatediff.pipeline_output.AnimateDiffPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.animatediff.pipeline_output.AnimateDiffPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ num_videos_per_prompt = 1
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_videos_per_prompt,
+ num_channels_latents,
+ num_frames,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ num_free_init_iters = self._free_init_num_iters if self.free_init_enabled else 1
+ for free_init_iter in range(num_free_init_iters):
+ if self.free_init_enabled:
+ latents, timesteps = self._apply_free_init(
+ latents, free_init_iter, num_inference_steps, device, latents.dtype, generator
+ )
+
+ self._num_timesteps = len(timesteps)
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+
+ # 8. Denoising loop
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ ).sample
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, latents)
+
+ # 9. Post processing
+ if output_type == "latent":
+ video = latents
+ else:
+ video_tensor = self.decode_latents(latents)
+ video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)
+
+ # 10. Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (video,)
+
+ return AnimateDiffPipelineOutput(frames=video)
diff --git a/model/diffusers_c/pipelines/animatediff/pipeline_animatediff_video2video.py b/model/diffusers_c/pipelines/animatediff/pipeline_animatediff_video2video.py
new file mode 100644
index 0000000000000000000000000000000000000000..cb6b71351faf098481f8d62b1ed316b0fe899795
--- /dev/null
+++ b/model/diffusers_c/pipelines/animatediff/pipeline_animatediff_video2video.py
@@ -0,0 +1,997 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...models.unets.unet_motion_model import MotionAdapter
+from ...schedulers import (
+ DDIMScheduler,
+ DPMSolverMultistepScheduler,
+ EulerAncestralDiscreteScheduler,
+ EulerDiscreteScheduler,
+ LMSDiscreteScheduler,
+ PNDMScheduler,
+)
+from ...utils import USE_PEFT_BACKEND, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import randn_tensor
+from ..free_init_utils import FreeInitMixin
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import AnimateDiffPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import imageio
+ >>> import requests
+ >>> import torch
+ >>> from diffusers import AnimateDiffVideoToVideoPipeline, DDIMScheduler, MotionAdapter
+ >>> from diffusers.utils import export_to_gif
+ >>> from io import BytesIO
+ >>> from PIL import Image
+
+ >>> adapter = MotionAdapter.from_pretrained("guoyww/animatediff-motion-adapter-v1-5-2", torch_dtype=torch.float16)
+ >>> pipe = AnimateDiffVideoToVideoPipeline.from_pretrained("SG161222/Realistic_Vision_V5.1_noVAE", motion_adapter=adapter).to("cuda")
+ >>> pipe.scheduler = DDIMScheduler(beta_schedule="linear", steps_offset=1, clip_sample=False, timespace_spacing="linspace")
+
+ >>> def load_video(file_path: str):
+ ... images = []
+ ...
+ ... if file_path.startswith(('http://', 'https://')):
+ ... # If the file_path is a URL
+ ... response = requests.get(file_path)
+ ... response.raise_for_status()
+ ... content = BytesIO(response.content)
+ ... vid = imageio.get_reader(content)
+ ... else:
+ ... # Assuming it's a local file path
+ ... vid = imageio.get_reader(file_path)
+ ...
+ ... for frame in vid:
+ ... pil_image = Image.fromarray(frame)
+ ... images.append(pil_image)
+ ...
+ ... return images
+
+ >>> video = load_video("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-vid2vid-input-1.gif")
+ >>> output = pipe(video=video, prompt="panda playing a guitar, on a boat, in the ocean, high quality", strength=0.5)
+ >>> frames = output.frames[0]
+ >>> export_to_gif(frames, "animation.gif")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor, output_type="np"):
+ batch_size, channels, num_frames, height, width = video.shape
+ outputs = []
+ for batch_idx in range(batch_size):
+ batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+ batch_output = processor.postprocess(batch_vid, output_type)
+
+ outputs.append(batch_output)
+
+ if output_type == "np":
+ outputs = np.stack(outputs)
+
+ elif output_type == "pt":
+ outputs = torch.stack(outputs)
+
+ elif not output_type == "pil":
+ raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+ return outputs
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class AnimateDiffVideoToVideoPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FreeInitMixin,
+):
+ r"""
+ Pipeline for video-to-video generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`] to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A [`UNet2DConditionModel`] used to create a UNetMotionModel to denoise the encoded video latents.
+ motion_adapter ([`MotionAdapter`]):
+ A [`MotionAdapter`] to be used in combination with `unet` to denoise the encoded video latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["feature_extractor", "image_encoder", "motion_adapter"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ motion_adapter: MotionAdapter,
+ scheduler: Union[
+ DDIMScheduler,
+ PNDMScheduler,
+ LMSDiscreteScheduler,
+ EulerDiscreteScheduler,
+ EulerAncestralDiscreteScheduler,
+ DPMSolverMultistepScheduler,
+ ],
+ feature_extractor: CLIPImageProcessor = None,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ ):
+ super().__init__()
+ if isinstance(unet, UNet2DConditionModel):
+ unet = UNetMotionModel.from_unet2d(unet, motion_adapter)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ motion_adapter=motion_adapter,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with num_images_per_prompt -> num_videos_per_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis/pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
+ def decode_latents(self, latents):
+ latents = 1 / self.vae.config.scaling_factor * latents
+
+ batch_size, channels, num_frames, height, width = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+ image = self.vae.decode(latents).sample
+ video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ video = video.float()
+ return video
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ strength,
+ height,
+ width,
+ video=None,
+ latents=None,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if video is not None and latents is not None:
+ raise ValueError("Only one of `video` or `latents` should be provided")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def get_timesteps(self, num_inference_steps, timesteps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = timesteps[t_start * self.scheduler.order :]
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(
+ self,
+ video,
+ height,
+ width,
+ num_channels_latents,
+ batch_size,
+ timestep,
+ dtype,
+ device,
+ generator,
+ latents=None,
+ ):
+ # video must be a list of list of images
+ # the outer list denotes having multiple videos as input, whereas inner list means the frames of the video
+ # as a list of images
+ if not isinstance(video[0], list):
+ video = [video]
+ if latents is None:
+ video = torch.cat(
+ [self.image_processor.preprocess(vid, height=height, width=width).unsqueeze(0) for vid in video], dim=0
+ )
+ video = video.to(device=device, dtype=dtype)
+ num_frames = video.shape[1]
+ else:
+ num_frames = latents.shape[2]
+
+ shape = (
+ batch_size,
+ num_channels_latents,
+ num_frames,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ if self.vae.config.force_upcast:
+ video = video.float()
+ self.vae.to(dtype=torch.float32)
+
+ if isinstance(generator, list):
+ if len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ init_latents = [
+ retrieve_latents(self.vae.encode(video[i]), generator=generator[i]).unsqueeze(0)
+ for i in range(batch_size)
+ ]
+ else:
+ init_latents = [
+ retrieve_latents(self.vae.encode(vid), generator=generator).unsqueeze(0) for vid in video
+ ]
+
+ init_latents = torch.cat(init_latents, dim=0)
+
+ # restore vae to original dtype
+ if self.vae.config.force_upcast:
+ self.vae.to(dtype)
+
+ init_latents = init_latents.to(dtype)
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ error_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
+ " images (`image`). Please make sure to update your script to pass as many initial images as text prompts"
+ )
+ raise ValueError(error_message)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ noise = randn_tensor(init_latents.shape, generator=generator, device=device, dtype=dtype)
+ latents = self.scheduler.add_noise(init_latents, noise, timestep).permute(0, 2, 1, 3, 4)
+ else:
+ if shape != latents.shape:
+ # [B, C, F, H, W]
+ raise ValueError(f"`latents` expected to have {shape=}, but found {latents.shape=}")
+ latents = latents.to(device, dtype=dtype)
+
+ return latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ video: List[List[PipelineImageInput]] = None,
+ prompt: Optional[Union[str, List[str]]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: Optional[List[int]] = None,
+ guidance_scale: float = 7.5,
+ strength: float = 0.8,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_videos_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ video (`List[PipelineImageInput]`):
+ The input video to condition the generation on. Must be a list of images/frames of the video.
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated video.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated video.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
+ expense of slower inference.
+ strength (`float`, *optional*, defaults to 0.8):
+ Higher strength leads to more differences between original video and generated video.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`. Latents should be of shape
+ `(batch_size, num_channel, num_frames, height, width)`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated video. Choose between `torch.FloatTensor`, `PIL.Image` or
+ `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`AnimateDiffPipelineOutput`] instead
+ of a plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`pipelines.animatediff.pipeline_output.AnimateDiffPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`pipelines.animatediff.pipeline_output.AnimateDiffPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
+ """
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ num_videos_per_prompt = 1
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt=prompt,
+ strength=strength,
+ height=height,
+ width=width,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ video=video,
+ latents=latents,
+ ip_adapter_image=ip_adapter_image,
+ ip_adapter_image_embeds=ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, timesteps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ video=video,
+ height=height,
+ width=width,
+ num_channels_latents=num_channels_latents,
+ batch_size=batch_size * num_videos_per_prompt,
+ timestep=latent_timestep,
+ dtype=prompt_embeds.dtype,
+ device=device,
+ generator=generator,
+ latents=latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ num_free_init_iters = self._free_init_num_iters if self.free_init_enabled else 1
+ for free_init_iter in range(num_free_init_iters):
+ if self.free_init_enabled:
+ latents, timesteps = self._apply_free_init(
+ latents, free_init_iter, num_inference_steps, device, latents.dtype, generator
+ )
+ num_inference_steps = len(timesteps)
+ # make sure to readjust timesteps based on strength
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, timesteps, strength, device)
+
+ self._num_timesteps = len(timesteps)
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+
+ # 8. Denoising loop
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ ).sample
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+
+ # 9. Post-processing
+ if output_type == "latent":
+ video = latents
+ else:
+ video_tensor = self.decode_latents(latents)
+ video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)
+
+ # 10. Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (video,)
+
+ return AnimateDiffPipelineOutput(frames=video)
diff --git a/model/diffusers_c/pipelines/animatediff/pipeline_output.py b/model/diffusers_c/pipelines/animatediff/pipeline_output.py
new file mode 100644
index 0000000000000000000000000000000000000000..184a45848a37c4458b9c16ef7f144dac4ed3e14e
--- /dev/null
+++ b/model/diffusers_c/pipelines/animatediff/pipeline_output.py
@@ -0,0 +1,23 @@
+from dataclasses import dataclass
+from typing import List, Union
+
+import numpy as np
+import PIL.Image
+import torch
+
+from ...utils import BaseOutput
+
+
+@dataclass
+class AnimateDiffPipelineOutput(BaseOutput):
+ r"""
+ Output class for AnimateDiff pipelines.
+
+ Args:
+ frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+ List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing denoised
+ PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+ `(batch_size, num_frames, channels, height, width)`
+ """
+
+ frames: Union[torch.Tensor, np.ndarray, List[List[PIL.Image.Image]]]
diff --git a/model/diffusers_c/pipelines/audioldm/__init__.py b/model/diffusers_c/pipelines/audioldm/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..a002b4aa72e0a180c7042c406667d37122d6e4cc
--- /dev/null
+++ b/model/diffusers_c/pipelines/audioldm/__init__.py
@@ -0,0 +1,51 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ is_torch_available,
+ is_transformers_available,
+ is_transformers_version,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ AudioLDMPipeline,
+ )
+
+ _dummy_objects.update({"AudioLDMPipeline": AudioLDMPipeline})
+else:
+ _import_structure["pipeline_audioldm"] = ["AudioLDMPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ AudioLDMPipeline,
+ )
+
+ else:
+ from .pipeline_audioldm import AudioLDMPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/audioldm/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/audioldm/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e49a82133167fd841808d011925d7d22a1b12a7e
Binary files /dev/null and b/model/diffusers_c/pipelines/audioldm/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/audioldm/__pycache__/pipeline_audioldm.cpython-39.pyc b/model/diffusers_c/pipelines/audioldm/__pycache__/pipeline_audioldm.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5b1d02b8268e48ac70ca0cd2f18f4214a0c12455
Binary files /dev/null and b/model/diffusers_c/pipelines/audioldm/__pycache__/pipeline_audioldm.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/audioldm/pipeline_audioldm.py b/model/diffusers_c/pipelines/audioldm/pipeline_audioldm.py
new file mode 100644
index 0000000000000000000000000000000000000000..69bebdd0dc4fb7bf3a4867e0637bb27c125233d1
--- /dev/null
+++ b/model/diffusers_c/pipelines/audioldm/pipeline_audioldm.py
@@ -0,0 +1,546 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from transformers import ClapTextModelWithProjection, RobertaTokenizer, RobertaTokenizerFast, SpeechT5HifiGan
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline, StableDiffusionMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import AudioLDMPipeline
+ >>> import torch
+ >>> import scipy
+
+ >>> repo_id = "cvssp/audioldm-s-full-v2"
+ >>> pipe = AudioLDMPipeline.from_pretrained(repo_id, torch_dtype=torch.float16)
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "Techno music with a strong, upbeat tempo and high melodic riffs"
+ >>> audio = pipe(prompt, num_inference_steps=10, audio_length_in_s=5.0).audios[0]
+
+ >>> # save the audio sample as a .wav file
+ >>> scipy.io.wavfile.write("techno.wav", rate=16000, data=audio)
+ ```
+"""
+
+
+class AudioLDMPipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+ Pipeline for text-to-audio generation using AudioLDM.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.ClapTextModelWithProjection`]):
+ Frozen text-encoder (`ClapTextModelWithProjection`, specifically the
+ [laion/clap-htsat-unfused](https://huggingface.co/laion/clap-htsat-unfused) variant.
+ tokenizer ([`PreTrainedTokenizer`]):
+ A [`~transformers.RobertaTokenizer`] to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded audio latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded audio latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ vocoder ([`~transformers.SpeechT5HifiGan`]):
+ Vocoder of class `SpeechT5HifiGan`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: ClapTextModelWithProjection,
+ tokenizer: Union[RobertaTokenizer, RobertaTokenizerFast],
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ vocoder: SpeechT5HifiGan,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ vocoder=vocoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_waveforms_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device (`torch.device`):
+ torch device
+ num_waveforms_per_prompt (`int`):
+ number of waveforms that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the audio generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ attention_mask = text_inputs.attention_mask
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLAP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask.to(device),
+ )
+ prompt_embeds = prompt_embeds.text_embeds
+ # additional L_2 normalization over each hidden-state
+ prompt_embeds = F.normalize(prompt_embeds, dim=-1)
+
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+ (
+ bs_embed,
+ seq_len,
+ ) = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_waveforms_per_prompt)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_waveforms_per_prompt, seq_len)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ uncond_input_ids = uncond_input.input_ids.to(device)
+ attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input_ids,
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds.text_embeds
+ # additional L_2 normalization over each hidden-state
+ negative_prompt_embeds = F.normalize(negative_prompt_embeds, dim=-1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_waveforms_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_waveforms_per_prompt, seq_len)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ def decode_latents(self, latents):
+ latents = 1 / self.vae.config.scaling_factor * latents
+ mel_spectrogram = self.vae.decode(latents).sample
+ return mel_spectrogram
+
+ def mel_spectrogram_to_waveform(self, mel_spectrogram):
+ if mel_spectrogram.dim() == 4:
+ mel_spectrogram = mel_spectrogram.squeeze(1)
+
+ waveform = self.vocoder(mel_spectrogram)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ waveform = waveform.cpu().float()
+ return waveform
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ audio_length_in_s,
+ vocoder_upsample_factor,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ min_audio_length_in_s = vocoder_upsample_factor * self.vae_scale_factor
+ if audio_length_in_s < min_audio_length_in_s:
+ raise ValueError(
+ f"`audio_length_in_s` has to be a positive value greater than or equal to {min_audio_length_in_s}, but "
+ f"is {audio_length_in_s}."
+ )
+
+ if self.vocoder.config.model_in_dim % self.vae_scale_factor != 0:
+ raise ValueError(
+ f"The number of frequency bins in the vocoder's log-mel spectrogram has to be divisible by the "
+ f"VAE scale factor, but got {self.vocoder.config.model_in_dim} bins and a scale factor of "
+ f"{self.vae_scale_factor}."
+ )
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents with width->self.vocoder.config.model_in_dim
+ def prepare_latents(self, batch_size, num_channels_latents, height, dtype, device, generator, latents=None):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ height // self.vae_scale_factor,
+ self.vocoder.config.model_in_dim // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ audio_length_in_s: Optional[float] = None,
+ num_inference_steps: int = 10,
+ guidance_scale: float = 2.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_waveforms_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ output_type: Optional[str] = "np",
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide audio generation. If not defined, you need to pass `prompt_embeds`.
+ audio_length_in_s (`int`, *optional*, defaults to 5.12):
+ The length of the generated audio sample in seconds.
+ num_inference_steps (`int`, *optional*, defaults to 10):
+ The number of denoising steps. More denoising steps usually lead to a higher quality audio at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 2.5):
+ A higher guidance scale value encourages the model to generate audio that is closely linked to the text
+ `prompt` at the expense of lower sound quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in audio generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_waveforms_per_prompt (`int`, *optional*, defaults to 1):
+ The number of waveforms to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.AudioPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ output_type (`str`, *optional*, defaults to `"np"`):
+ The output format of the generated image. Choose between `"np"` to return a NumPy `np.ndarray` or
+ `"pt"` to return a PyTorch `torch.Tensor` object.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.AudioPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.AudioPipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated audio.
+ """
+ # 0. Convert audio input length from seconds to spectrogram height
+ vocoder_upsample_factor = np.prod(self.vocoder.config.upsample_rates) / self.vocoder.config.sampling_rate
+
+ if audio_length_in_s is None:
+ audio_length_in_s = self.unet.config.sample_size * self.vae_scale_factor * vocoder_upsample_factor
+
+ height = int(audio_length_in_s / vocoder_upsample_factor)
+
+ original_waveform_length = int(audio_length_in_s * self.vocoder.config.sampling_rate)
+ if height % self.vae_scale_factor != 0:
+ height = int(np.ceil(height / self.vae_scale_factor)) * self.vae_scale_factor
+ logger.info(
+ f"Audio length in seconds {audio_length_in_s} is increased to {height * vocoder_upsample_factor} "
+ f"so that it can be handled by the model. It will be cut to {audio_length_in_s} after the "
+ f"denoising process."
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ audio_length_in_s,
+ vocoder_upsample_factor,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds = self._encode_prompt(
+ prompt,
+ device,
+ num_waveforms_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_waveforms_per_prompt,
+ num_channels_latents,
+ height,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=None,
+ class_labels=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 8. Post-processing
+ mel_spectrogram = self.decode_latents(latents)
+
+ audio = self.mel_spectrogram_to_waveform(mel_spectrogram)
+
+ audio = audio[:, :original_waveform_length]
+
+ if output_type == "np":
+ audio = audio.numpy()
+
+ if not return_dict:
+ return (audio,)
+
+ return AudioPipelineOutput(audios=audio)
diff --git a/model/diffusers_c/pipelines/audioldm2/__init__.py b/model/diffusers_c/pipelines/audioldm2/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..23cd0e44f89217b8391d0ce236070271db9aaf83
--- /dev/null
+++ b/model/diffusers_c/pipelines/audioldm2/__init__.py
@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+ is_transformers_version,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["modeling_audioldm2"] = ["AudioLDM2ProjectionModel", "AudioLDM2UNet2DConditionModel"]
+ _import_structure["pipeline_audioldm2"] = ["AudioLDM2Pipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+
+ else:
+ from .modeling_audioldm2 import AudioLDM2ProjectionModel, AudioLDM2UNet2DConditionModel
+ from .pipeline_audioldm2 import AudioLDM2Pipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/audioldm2/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/audioldm2/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8bd7c2d619de9d11a8255aed7e30c85b4eaf590d
Binary files /dev/null and b/model/diffusers_c/pipelines/audioldm2/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/audioldm2/__pycache__/modeling_audioldm2.cpython-39.pyc b/model/diffusers_c/pipelines/audioldm2/__pycache__/modeling_audioldm2.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2b174bb0a1c8c3f31b5e58394cde7a5765d93105
Binary files /dev/null and b/model/diffusers_c/pipelines/audioldm2/__pycache__/modeling_audioldm2.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/audioldm2/__pycache__/pipeline_audioldm2.cpython-39.pyc b/model/diffusers_c/pipelines/audioldm2/__pycache__/pipeline_audioldm2.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5c0df8c7306410eb9709ee222f2148abc3a50412
Binary files /dev/null and b/model/diffusers_c/pipelines/audioldm2/__pycache__/pipeline_audioldm2.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/audioldm2/modeling_audioldm2.py b/model/diffusers_c/pipelines/audioldm2/modeling_audioldm2.py
new file mode 100644
index 0000000000000000000000000000000000000000..c0b85e4db5f69795fdc6850fc249c050144f0a84
--- /dev/null
+++ b/model/diffusers_c/pipelines/audioldm2/modeling_audioldm2.py
@@ -0,0 +1,1511 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import UNet2DConditionLoadersMixin
+from ...models.activations import get_activation
+from ...models.attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ...models.embeddings import (
+ TimestepEmbedding,
+ Timesteps,
+)
+from ...models.modeling_utils import ModelMixin
+from ...models.resnet import Downsample2D, ResnetBlock2D, Upsample2D
+from ...models.transformers.transformer_2d import Transformer2DModel
+from ...models.unets.unet_2d_blocks import DownBlock2D, UpBlock2D
+from ...models.unets.unet_2d_condition import UNet2DConditionOutput
+from ...utils import BaseOutput, is_torch_version, logging
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def add_special_tokens(hidden_states, attention_mask, sos_token, eos_token):
+ batch_size = hidden_states.shape[0]
+
+ if attention_mask is not None:
+ # Add two more steps to attn mask
+ new_attn_mask_step = attention_mask.new_ones((batch_size, 1))
+ attention_mask = torch.concat([new_attn_mask_step, attention_mask, new_attn_mask_step], dim=-1)
+
+ # Add the SOS / EOS tokens at the start / end of the sequence respectively
+ sos_token = sos_token.expand(batch_size, 1, -1)
+ eos_token = eos_token.expand(batch_size, 1, -1)
+ hidden_states = torch.concat([sos_token, hidden_states, eos_token], dim=1)
+ return hidden_states, attention_mask
+
+
+@dataclass
+class AudioLDM2ProjectionModelOutput(BaseOutput):
+ """
+ Args:
+ Class for AudioLDM2 projection layer's outputs.
+ hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+ Sequence of hidden-states obtained by linearly projecting the hidden-states for each of the text
+ encoders and subsequently concatenating them together.
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+ Mask to avoid performing attention on padding token indices, formed by concatenating the attention masks
+ for the two text encoders together. Mask values selected in `[0, 1]`:
+
+ - 1 for tokens that are **not masked**,
+ - 0 for tokens that are **masked**.
+ """
+
+ hidden_states: torch.FloatTensor
+ attention_mask: Optional[torch.LongTensor] = None
+
+
+class AudioLDM2ProjectionModel(ModelMixin, ConfigMixin):
+ """
+ A simple linear projection model to map two text embeddings to a shared latent space. It also inserts learned
+ embedding vectors at the start and end of each text embedding sequence respectively. Each variable appended with
+ `_1` refers to that corresponding to the second text encoder. Otherwise, it is from the first.
+
+ Args:
+ text_encoder_dim (`int`):
+ Dimensionality of the text embeddings from the first text encoder (CLAP).
+ text_encoder_1_dim (`int`):
+ Dimensionality of the text embeddings from the second text encoder (T5 or VITS).
+ langauge_model_dim (`int`):
+ Dimensionality of the text embeddings from the language model (GPT2).
+ """
+
+ @register_to_config
+ def __init__(self, text_encoder_dim, text_encoder_1_dim, langauge_model_dim):
+ super().__init__()
+ # additional projection layers for each text encoder
+ self.projection = nn.Linear(text_encoder_dim, langauge_model_dim)
+ self.projection_1 = nn.Linear(text_encoder_1_dim, langauge_model_dim)
+
+ # learnable SOS / EOS token embeddings for each text encoder
+ self.sos_embed = nn.Parameter(torch.ones(langauge_model_dim))
+ self.eos_embed = nn.Parameter(torch.ones(langauge_model_dim))
+
+ self.sos_embed_1 = nn.Parameter(torch.ones(langauge_model_dim))
+ self.eos_embed_1 = nn.Parameter(torch.ones(langauge_model_dim))
+
+ def forward(
+ self,
+ hidden_states: Optional[torch.FloatTensor] = None,
+ hidden_states_1: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.LongTensor] = None,
+ attention_mask_1: Optional[torch.LongTensor] = None,
+ ):
+ hidden_states = self.projection(hidden_states)
+ hidden_states, attention_mask = add_special_tokens(
+ hidden_states, attention_mask, sos_token=self.sos_embed, eos_token=self.eos_embed
+ )
+
+ hidden_states_1 = self.projection_1(hidden_states_1)
+ hidden_states_1, attention_mask_1 = add_special_tokens(
+ hidden_states_1, attention_mask_1, sos_token=self.sos_embed_1, eos_token=self.eos_embed_1
+ )
+
+ # concatenate clap and t5 text encoding
+ hidden_states = torch.cat([hidden_states, hidden_states_1], dim=1)
+
+ # concatenate attention masks
+ if attention_mask is None and attention_mask_1 is not None:
+ attention_mask = attention_mask_1.new_ones((hidden_states[:2]))
+ elif attention_mask is not None and attention_mask_1 is None:
+ attention_mask_1 = attention_mask.new_ones((hidden_states_1[:2]))
+
+ if attention_mask is not None and attention_mask_1 is not None:
+ attention_mask = torch.cat([attention_mask, attention_mask_1], dim=-1)
+ else:
+ attention_mask = None
+
+ return AudioLDM2ProjectionModelOutput(
+ hidden_states=hidden_states,
+ attention_mask=attention_mask,
+ )
+
+
+class AudioLDM2UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
+ r"""
+ A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+ shaped output. Compared to the vanilla [`UNet2DConditionModel`], this variant optionally includes an additional
+ self-attention layer in each Transformer block, as well as multiple cross-attention layers. It also allows for up
+ to two cross-attention embeddings, `encoder_hidden_states` and `encoder_hidden_states_1`.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+ Height and width of input/output sample.
+ in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+ flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+ The tuple of downsample blocks to use.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+ Block type for middle of UNet, it can only be `UNetMidBlock2DCrossAttn` for AudioLDM2.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+ The tuple of upsample blocks to use.
+ only_cross_attention (`bool` or `Tuple[bool]`, *optional*, default to `False`):
+ Whether to include self-attention in the basic transformer blocks, see
+ [`~models.attention.BasicTransformerBlock`].
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+ downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+ If `None`, normalization and activation layers is skipped in post-processing.
+ norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+ cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+ [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+ attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+ num_attention_heads (`int`, *optional*):
+ The number of attention heads. If not defined, defaults to `attention_head_dim`
+ resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+ for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+ num_class_embeds (`int`, *optional*, defaults to `None`):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+ class conditioning with `class_embed_type` equal to `None`.
+ time_embedding_type (`str`, *optional*, defaults to `positional`):
+ The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+ time_embedding_dim (`int`, *optional*, defaults to `None`):
+ An optional override for the dimension of the projected time embedding.
+ time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+ Optional activation function to use only once on the time embeddings before they are passed to the rest of
+ the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+ timestep_post_act (`str`, *optional*, defaults to `None`):
+ The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+ time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+ The dimension of `cond_proj` layer in the timestep embedding.
+ conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
+ conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
+ projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
+ `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+ class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+ embeddings with the class embeddings.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[int] = None,
+ in_channels: int = 4,
+ out_channels: int = 4,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str] = (
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "CrossAttnDownBlock2D",
+ "DownBlock2D",
+ ),
+ mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+ up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+ layers_per_block: Union[int, Tuple[int]] = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: Union[int, Tuple[int]] = 1280,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ attention_head_dim: Union[int, Tuple[int]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+ use_linear_projection: bool = False,
+ class_embed_type: Optional[str] = None,
+ num_class_embeds: Optional[int] = None,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ time_embedding_type: str = "positional",
+ time_embedding_dim: Optional[int] = None,
+ time_embedding_act_fn: Optional[str] = None,
+ timestep_post_act: Optional[str] = None,
+ time_cond_proj_dim: Optional[int] = None,
+ conv_in_kernel: int = 3,
+ conv_out_kernel: int = 3,
+ projection_class_embeddings_input_dim: Optional[int] = None,
+ class_embeddings_concat: bool = False,
+ ):
+ super().__init__()
+
+ self.sample_size = sample_size
+
+ if num_attention_heads is not None:
+ raise ValueError(
+ "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+ )
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+ )
+
+ # input
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in = nn.Conv2d(
+ in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+ )
+
+ # time
+ if time_embedding_type == "positional":
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+ timestep_input_dim = block_out_channels[0]
+ else:
+ raise ValueError(f"{time_embedding_type} does not exist. Please make sure to use `positional`.")
+
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ post_act_fn=timestep_post_act,
+ cond_proj_dim=time_cond_proj_dim,
+ )
+
+ # class embedding
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ elif class_embed_type == "projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+ # 2. it projects from an arbitrary input dimension.
+ #
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+ self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ elif class_embed_type == "simple_projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+ else:
+ self.class_embedding = None
+
+ if time_embedding_act_fn is None:
+ self.time_embed_act = None
+ else:
+ self.time_embed_act = get_activation(time_embedding_act_fn)
+
+ self.down_blocks = nn.ModuleList([])
+ self.up_blocks = nn.ModuleList([])
+
+ if isinstance(only_cross_attention, bool):
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+ if isinstance(layers_per_block, int):
+ layers_per_block = [layers_per_block] * len(down_block_types)
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+ if class_embeddings_concat:
+ # The time embeddings are concatenated with the class embeddings. The dimension of the
+ # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+ # regular time embeddings
+ blocks_time_embed_dim = time_embed_dim * 2
+ else:
+ blocks_time_embed_dim = time_embed_dim
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block[i],
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim[i],
+ num_attention_heads=num_attention_heads[i],
+ downsample_padding=downsample_padding,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ if mid_block_type == "UNetMidBlock2DCrossAttn":
+ self.mid_block = UNetMidBlock2DCrossAttn(
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ in_channels=block_out_channels[-1],
+ temb_channels=blocks_time_embed_dim,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim[-1],
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ )
+ else:
+ raise ValueError(
+ f"unknown mid_block_type : {mid_block_type}. Should be `UNetMidBlock2DCrossAttn` for AudioLDM2."
+ )
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ reversed_layers_per_block = list(reversed(layers_per_block))
+ reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+ reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block))
+ only_cross_attention = list(reversed(only_cross_attention))
+
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=reversed_layers_per_block[i] + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=reversed_cross_attention_dim[i],
+ num_attention_heads=reversed_num_attention_heads[i],
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ if norm_num_groups is not None:
+ self.conv_norm_out = nn.GroupNorm(
+ num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+ )
+
+ self.conv_act = get_activation(act_fn)
+
+ else:
+ self.conv_norm_out = None
+ self.conv_act = None
+
+ conv_out_padding = (conv_out_kernel - 1) // 2
+ self.conv_out = nn.Conv2d(
+ block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+ )
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+ def set_attention_slice(self, slice_size):
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel._set_gradient_checkpointing
+ def _set_gradient_checkpointing(self, module, value=False):
+ if hasattr(module, "gradient_checkpointing"):
+ module.gradient_checkpointing = value
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ encoder_hidden_states_1: Optional[torch.Tensor] = None,
+ encoder_attention_mask_1: Optional[torch.Tensor] = None,
+ ) -> Union[UNet2DConditionOutput, Tuple]:
+ r"""
+ The [`AudioLDM2UNet2DConditionModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, channel, height, width)`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.FloatTensor`):
+ The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+ encoder_attention_mask (`torch.Tensor`):
+ A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+ `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+ which adds large negative values to the attention scores corresponding to "discard" tokens.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+ tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+ encoder_hidden_states_1 (`torch.FloatTensor`, *optional*):
+ A second set of encoder hidden states with shape `(batch, sequence_length_2, feature_dim_2)`. Can be
+ used to condition the model on a different set of embeddings to `encoder_hidden_states`.
+ encoder_attention_mask_1 (`torch.Tensor`, *optional*):
+ A cross-attention mask of shape `(batch, sequence_length_2)` is applied to `encoder_hidden_states_1`.
+ If `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+ which adds large negative values to the attention scores corresponding to "discard" tokens.
+
+ Returns:
+ [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+ a `tuple` is returned where the first element is the sample tensor.
+ """
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
+ # on the fly if necessary.
+ default_overall_up_factor = 2**self.num_upsamplers
+
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+ forward_upsample_size = False
+ upsample_size = None
+
+ if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+ logger.info("Forward upsample size to force interpolation output size.")
+ forward_upsample_size = True
+
+ # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+ # expects mask of shape:
+ # [batch, key_tokens]
+ # adds singleton query_tokens dimension:
+ # [batch, 1, key_tokens]
+ # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+ # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+ # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+ if attention_mask is not None:
+ # assume that mask is expressed as:
+ # (1 = keep, 0 = discard)
+ # convert mask into a bias that can be added to attention scores:
+ # (keep = +0, discard = -10000.0)
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # convert encoder_attention_mask to a bias the same way we do for attention_mask
+ if encoder_attention_mask is not None:
+ encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+ encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+ if encoder_attention_mask_1 is not None:
+ encoder_attention_mask_1 = (1 - encoder_attention_mask_1.to(sample.dtype)) * -10000.0
+ encoder_attention_mask_1 = encoder_attention_mask_1.unsqueeze(1)
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+
+ emb = self.time_embedding(t_emb, timestep_cond)
+ aug_emb = None
+
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # there might be better ways to encapsulate this.
+ class_labels = class_labels.to(dtype=sample.dtype)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+ if self.config.class_embeddings_concat:
+ emb = torch.cat([emb, class_emb], dim=-1)
+ else:
+ emb = emb + class_emb
+
+ emb = emb + aug_emb if aug_emb is not None else emb
+
+ if self.time_embed_act is not None:
+ emb = self.time_embed_act(emb)
+
+ # 2. pre-process
+ sample = self.conv_in(sample)
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ encoder_hidden_states_1=encoder_hidden_states_1,
+ encoder_attention_mask_1=encoder_attention_mask_1,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+ down_block_res_samples += res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ encoder_hidden_states_1=encoder_hidden_states_1,
+ encoder_attention_mask_1=encoder_attention_mask_1,
+ )
+
+ # 5. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block and forward_upsample_size:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ encoder_hidden_states_1=encoder_hidden_states_1,
+ encoder_attention_mask_1=encoder_attention_mask_1,
+ )
+ else:
+ sample = upsample_block(
+ hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
+ )
+
+ # 6. post-process
+ if self.conv_norm_out:
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet2DConditionOutput(sample=sample)
+
+
+def get_down_block(
+ down_block_type,
+ num_layers,
+ in_channels,
+ out_channels,
+ temb_channels,
+ add_downsample,
+ resnet_eps,
+ resnet_act_fn,
+ transformer_layers_per_block=1,
+ num_attention_heads=None,
+ resnet_groups=None,
+ cross_attention_dim=None,
+ downsample_padding=None,
+ use_linear_projection=False,
+ only_cross_attention=False,
+ upcast_attention=False,
+ resnet_time_scale_shift="default",
+):
+ down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+ if down_block_type == "DownBlock2D":
+ return DownBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "CrossAttnDownBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
+ return CrossAttnDownBlock2D(
+ num_layers=num_layers,
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+ up_block_type,
+ num_layers,
+ in_channels,
+ out_channels,
+ prev_output_channel,
+ temb_channels,
+ add_upsample,
+ resnet_eps,
+ resnet_act_fn,
+ transformer_layers_per_block=1,
+ num_attention_heads=None,
+ resnet_groups=None,
+ cross_attention_dim=None,
+ use_linear_projection=False,
+ only_cross_attention=False,
+ upcast_attention=False,
+ resnet_time_scale_shift="default",
+):
+ up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+ if up_block_type == "UpBlock2D":
+ return UpBlock2D(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif up_block_type == "CrossAttnUpBlock2D":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
+ return CrossAttnUpBlock2D(
+ num_layers=num_layers,
+ transformer_layers_per_block=transformer_layers_per_block,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ raise ValueError(f"{up_block_type} does not exist.")
+
+
+class CrossAttnDownBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads=1,
+ cross_attention_dim=1280,
+ output_scale_factor=1.0,
+ downsample_padding=1,
+ add_downsample=True,
+ use_linear_projection=False,
+ only_cross_attention=False,
+ upcast_attention=False,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,)
+ if isinstance(cross_attention_dim, (list, tuple)) and len(cross_attention_dim) > 4:
+ raise ValueError(
+ "Only up to 4 cross-attention layers are supported. Ensure that the length of cross-attention "
+ f"dims is less than or equal to 4. Got cross-attention dims {cross_attention_dim} of length {len(cross_attention_dim)}"
+ )
+ self.cross_attention_dim = cross_attention_dim
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ for j in range(len(cross_attention_dim)):
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block,
+ cross_attention_dim=cross_attention_dim[j],
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ double_self_attention=True if cross_attention_dim[j] is None else False,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ Downsample2D(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states_1: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask_1: Optional[torch.FloatTensor] = None,
+ ):
+ output_states = ()
+ num_layers = len(self.resnets)
+ num_attention_per_layer = len(self.attentions) // num_layers
+
+ encoder_hidden_states_1 = (
+ encoder_hidden_states_1 if encoder_hidden_states_1 is not None else encoder_hidden_states
+ )
+ encoder_attention_mask_1 = (
+ encoder_attention_mask_1 if encoder_hidden_states_1 is not None else encoder_attention_mask
+ )
+
+ for i in range(num_layers):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.resnets[i]),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ for idx, cross_attention_dim in enumerate(self.cross_attention_dim):
+ if cross_attention_dim is not None and idx <= 1:
+ forward_encoder_hidden_states = encoder_hidden_states
+ forward_encoder_attention_mask = encoder_attention_mask
+ elif cross_attention_dim is not None and idx > 1:
+ forward_encoder_hidden_states = encoder_hidden_states_1
+ forward_encoder_attention_mask = encoder_attention_mask_1
+ else:
+ forward_encoder_hidden_states = None
+ forward_encoder_attention_mask = None
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.attentions[i * num_attention_per_layer + idx], return_dict=False),
+ hidden_states,
+ forward_encoder_hidden_states,
+ None, # timestep
+ None, # class_labels
+ cross_attention_kwargs,
+ attention_mask,
+ forward_encoder_attention_mask,
+ **ckpt_kwargs,
+ )[0]
+ else:
+ hidden_states = self.resnets[i](hidden_states, temb)
+ for idx, cross_attention_dim in enumerate(self.cross_attention_dim):
+ if cross_attention_dim is not None and idx <= 1:
+ forward_encoder_hidden_states = encoder_hidden_states
+ forward_encoder_attention_mask = encoder_attention_mask
+ elif cross_attention_dim is not None and idx > 1:
+ forward_encoder_hidden_states = encoder_hidden_states_1
+ forward_encoder_attention_mask = encoder_attention_mask_1
+ else:
+ forward_encoder_hidden_states = None
+ forward_encoder_attention_mask = None
+ hidden_states = self.attentions[i * num_attention_per_layer + idx](
+ hidden_states,
+ attention_mask=attention_mask,
+ encoder_hidden_states=forward_encoder_hidden_states,
+ encoder_attention_mask=forward_encoder_attention_mask,
+ return_dict=False,
+ )[0]
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class UNetMidBlock2DCrossAttn(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads=1,
+ output_scale_factor=1.0,
+ cross_attention_dim=1280,
+ use_linear_projection=False,
+ upcast_attention=False,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,)
+ if isinstance(cross_attention_dim, (list, tuple)) and len(cross_attention_dim) > 4:
+ raise ValueError(
+ "Only up to 4 cross-attention layers are supported. Ensure that the length of cross-attention "
+ f"dims is less than or equal to 4. Got cross-attention dims {cross_attention_dim} of length {len(cross_attention_dim)}"
+ )
+ self.cross_attention_dim = cross_attention_dim
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ attentions = []
+
+ for i in range(num_layers):
+ for j in range(len(cross_attention_dim)):
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=transformer_layers_per_block,
+ cross_attention_dim=cross_attention_dim[j],
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ double_self_attention=True if cross_attention_dim[j] is None else False,
+ )
+ )
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states_1: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask_1: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ hidden_states = self.resnets[0](hidden_states, temb)
+ num_attention_per_layer = len(self.attentions) // (len(self.resnets) - 1)
+
+ encoder_hidden_states_1 = (
+ encoder_hidden_states_1 if encoder_hidden_states_1 is not None else encoder_hidden_states
+ )
+ encoder_attention_mask_1 = (
+ encoder_attention_mask_1 if encoder_hidden_states_1 is not None else encoder_attention_mask
+ )
+
+ for i in range(len(self.resnets[1:])):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ for idx, cross_attention_dim in enumerate(self.cross_attention_dim):
+ if cross_attention_dim is not None and idx <= 1:
+ forward_encoder_hidden_states = encoder_hidden_states
+ forward_encoder_attention_mask = encoder_attention_mask
+ elif cross_attention_dim is not None and idx > 1:
+ forward_encoder_hidden_states = encoder_hidden_states_1
+ forward_encoder_attention_mask = encoder_attention_mask_1
+ else:
+ forward_encoder_hidden_states = None
+ forward_encoder_attention_mask = None
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.attentions[i * num_attention_per_layer + idx], return_dict=False),
+ hidden_states,
+ forward_encoder_hidden_states,
+ None, # timestep
+ None, # class_labels
+ cross_attention_kwargs,
+ attention_mask,
+ forward_encoder_attention_mask,
+ **ckpt_kwargs,
+ )[0]
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.resnets[i + 1]),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ else:
+ for idx, cross_attention_dim in enumerate(self.cross_attention_dim):
+ if cross_attention_dim is not None and idx <= 1:
+ forward_encoder_hidden_states = encoder_hidden_states
+ forward_encoder_attention_mask = encoder_attention_mask
+ elif cross_attention_dim is not None and idx > 1:
+ forward_encoder_hidden_states = encoder_hidden_states_1
+ forward_encoder_attention_mask = encoder_attention_mask_1
+ else:
+ forward_encoder_hidden_states = None
+ forward_encoder_attention_mask = None
+ hidden_states = self.attentions[i * num_attention_per_layer + idx](
+ hidden_states,
+ attention_mask=attention_mask,
+ encoder_hidden_states=forward_encoder_hidden_states,
+ encoder_attention_mask=forward_encoder_attention_mask,
+ return_dict=False,
+ )[0]
+
+ hidden_states = self.resnets[i + 1](hidden_states, temb)
+
+ return hidden_states
+
+
+class CrossAttnUpBlock2D(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads=1,
+ cross_attention_dim=1280,
+ output_scale_factor=1.0,
+ add_upsample=True,
+ use_linear_projection=False,
+ only_cross_attention=False,
+ upcast_attention=False,
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,)
+ if isinstance(cross_attention_dim, (list, tuple)) and len(cross_attention_dim) > 4:
+ raise ValueError(
+ "Only up to 4 cross-attention layers are supported. Ensure that the length of cross-attention "
+ f"dims is less than or equal to 4. Got cross-attention dims {cross_attention_dim} of length {len(cross_attention_dim)}"
+ )
+ self.cross_attention_dim = cross_attention_dim
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlock2D(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ for j in range(len(cross_attention_dim)):
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block,
+ cross_attention_dim=cross_attention_dim[j],
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ double_self_attention=True if cross_attention_dim[j] is None else False,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states_1: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask_1: Optional[torch.FloatTensor] = None,
+ ):
+ num_layers = len(self.resnets)
+ num_attention_per_layer = len(self.attentions) // num_layers
+
+ encoder_hidden_states_1 = (
+ encoder_hidden_states_1 if encoder_hidden_states_1 is not None else encoder_hidden_states
+ )
+ encoder_attention_mask_1 = (
+ encoder_attention_mask_1 if encoder_hidden_states_1 is not None else encoder_attention_mask
+ )
+
+ for i in range(num_layers):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.resnets[i]),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ for idx, cross_attention_dim in enumerate(self.cross_attention_dim):
+ if cross_attention_dim is not None and idx <= 1:
+ forward_encoder_hidden_states = encoder_hidden_states
+ forward_encoder_attention_mask = encoder_attention_mask
+ elif cross_attention_dim is not None and idx > 1:
+ forward_encoder_hidden_states = encoder_hidden_states_1
+ forward_encoder_attention_mask = encoder_attention_mask_1
+ else:
+ forward_encoder_hidden_states = None
+ forward_encoder_attention_mask = None
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(self.attentions[i * num_attention_per_layer + idx], return_dict=False),
+ hidden_states,
+ forward_encoder_hidden_states,
+ None, # timestep
+ None, # class_labels
+ cross_attention_kwargs,
+ attention_mask,
+ forward_encoder_attention_mask,
+ **ckpt_kwargs,
+ )[0]
+ else:
+ hidden_states = self.resnets[i](hidden_states, temb)
+ for idx, cross_attention_dim in enumerate(self.cross_attention_dim):
+ if cross_attention_dim is not None and idx <= 1:
+ forward_encoder_hidden_states = encoder_hidden_states
+ forward_encoder_attention_mask = encoder_attention_mask
+ elif cross_attention_dim is not None and idx > 1:
+ forward_encoder_hidden_states = encoder_hidden_states_1
+ forward_encoder_attention_mask = encoder_attention_mask_1
+ else:
+ forward_encoder_hidden_states = None
+ forward_encoder_attention_mask = None
+ hidden_states = self.attentions[i * num_attention_per_layer + idx](
+ hidden_states,
+ attention_mask=attention_mask,
+ encoder_hidden_states=forward_encoder_hidden_states,
+ encoder_attention_mask=forward_encoder_attention_mask,
+ return_dict=False,
+ )[0]
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size)
+
+ return hidden_states
diff --git a/model/diffusers_c/pipelines/audioldm2/pipeline_audioldm2.py b/model/diffusers_c/pipelines/audioldm2/pipeline_audioldm2.py
new file mode 100644
index 0000000000000000000000000000000000000000..e01aa9929dd839ae4ea502f233d3cecb107b4a81
--- /dev/null
+++ b/model/diffusers_c/pipelines/audioldm2/pipeline_audioldm2.py
@@ -0,0 +1,980 @@
+# Copyright 2024 CVSSP, ByteDance and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import (
+ ClapFeatureExtractor,
+ ClapModel,
+ GPT2Model,
+ RobertaTokenizer,
+ RobertaTokenizerFast,
+ SpeechT5HifiGan,
+ T5EncoderModel,
+ T5Tokenizer,
+ T5TokenizerFast,
+)
+
+from ...models import AutoencoderKL
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ is_accelerate_available,
+ is_accelerate_version,
+ is_librosa_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline
+from .modeling_audioldm2 import AudioLDM2ProjectionModel, AudioLDM2UNet2DConditionModel
+
+
+if is_librosa_available():
+ import librosa
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import scipy
+ >>> import torch
+ >>> from diffusers import AudioLDM2Pipeline
+
+ >>> repo_id = "cvssp/audioldm2"
+ >>> pipe = AudioLDM2Pipeline.from_pretrained(repo_id, torch_dtype=torch.float16)
+ >>> pipe = pipe.to("cuda")
+
+ >>> # define the prompts
+ >>> prompt = "The sound of a hammer hitting a wooden surface."
+ >>> negative_prompt = "Low quality."
+
+ >>> # set the seed for generator
+ >>> generator = torch.Generator("cuda").manual_seed(0)
+
+ >>> # run the generation
+ >>> audio = pipe(
+ ... prompt,
+ ... negative_prompt=negative_prompt,
+ ... num_inference_steps=200,
+ ... audio_length_in_s=10.0,
+ ... num_waveforms_per_prompt=3,
+ ... generator=generator,
+ ... ).audios
+
+ >>> # save the best audio sample (index 0) as a .wav file
+ >>> scipy.io.wavfile.write("techno.wav", rate=16000, data=audio[0])
+ ```
+"""
+
+
+def prepare_inputs_for_generation(
+ inputs_embeds,
+ attention_mask=None,
+ past_key_values=None,
+ **kwargs,
+):
+ if past_key_values is not None:
+ # only last token for inputs_embeds if past is defined in kwargs
+ inputs_embeds = inputs_embeds[:, -1:]
+
+ return {
+ "inputs_embeds": inputs_embeds,
+ "attention_mask": attention_mask,
+ "past_key_values": past_key_values,
+ "use_cache": kwargs.get("use_cache"),
+ }
+
+
+class AudioLDM2Pipeline(DiffusionPipeline):
+ r"""
+ Pipeline for text-to-audio generation using AudioLDM2.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.ClapModel`]):
+ First frozen text-encoder. AudioLDM2 uses the joint audio-text embedding model
+ [CLAP](https://huggingface.co/docs/transformers/model_doc/clap#transformers.CLAPTextModelWithProjection),
+ specifically the [laion/clap-htsat-unfused](https://huggingface.co/laion/clap-htsat-unfused) variant. The
+ text branch is used to encode the text prompt to a prompt embedding. The full audio-text model is used to
+ rank generated waveforms against the text prompt by computing similarity scores.
+ text_encoder_2 ([`~transformers.T5EncoderModel`]):
+ Second frozen text-encoder. AudioLDM2 uses the encoder of
+ [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+ [google/flan-t5-large](https://huggingface.co/google/flan-t5-large) variant.
+ projection_model ([`AudioLDM2ProjectionModel`]):
+ A trained model used to linearly project the hidden-states from the first and second text encoder models
+ and insert learned SOS and EOS token embeddings. The projected hidden-states from the two text encoders are
+ concatenated to give the input to the language model.
+ language_model ([`~transformers.GPT2Model`]):
+ An auto-regressive language model used to generate a sequence of hidden-states conditioned on the projected
+ outputs from the two text encoders.
+ tokenizer ([`~transformers.RobertaTokenizer`]):
+ Tokenizer to tokenize text for the first frozen text-encoder.
+ tokenizer_2 ([`~transformers.T5Tokenizer`]):
+ Tokenizer to tokenize text for the second frozen text-encoder.
+ feature_extractor ([`~transformers.ClapFeatureExtractor`]):
+ Feature extractor to pre-process generated audio waveforms to log-mel spectrograms for automatic scoring.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded audio latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded audio latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ vocoder ([`~transformers.SpeechT5HifiGan`]):
+ Vocoder of class `SpeechT5HifiGan` to convert the mel-spectrogram latents to the final audio waveform.
+ """
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: ClapModel,
+ text_encoder_2: T5EncoderModel,
+ projection_model: AudioLDM2ProjectionModel,
+ language_model: GPT2Model,
+ tokenizer: Union[RobertaTokenizer, RobertaTokenizerFast],
+ tokenizer_2: Union[T5Tokenizer, T5TokenizerFast],
+ feature_extractor: ClapFeatureExtractor,
+ unet: AudioLDM2UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ vocoder: SpeechT5HifiGan,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ projection_model=projection_model,
+ language_model=language_model,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ feature_extractor=feature_extractor,
+ unet=unet,
+ scheduler=scheduler,
+ vocoder=vocoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.enable_vae_slicing
+ def enable_vae_slicing(self):
+ r"""
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+ """
+ self.vae.enable_slicing()
+
+ # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.disable_vae_slicing
+ def disable_vae_slicing(self):
+ r"""
+ Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+ computing decoding in one step.
+ """
+ self.vae.disable_slicing()
+
+ def enable_model_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+ """
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+ from accelerate import cpu_offload_with_hook
+ else:
+ raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+ device = torch.device(f"cuda:{gpu_id}")
+
+ if self.device.type != "cpu":
+ self.to("cpu", silence_dtype_warnings=True)
+ torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
+
+ model_sequence = [
+ self.text_encoder.text_model,
+ self.text_encoder.text_projection,
+ self.text_encoder_2,
+ self.projection_model,
+ self.language_model,
+ self.unet,
+ self.vae,
+ self.vocoder,
+ self.text_encoder,
+ ]
+
+ hook = None
+ for cpu_offloaded_model in model_sequence:
+ _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+ # We'll offload the last model manually.
+ self.final_offload_hook = hook
+
+ def generate_language_model(
+ self,
+ inputs_embeds: torch.Tensor = None,
+ max_new_tokens: int = 8,
+ **model_kwargs,
+ ):
+ """
+
+ Generates a sequence of hidden-states from the language model, conditioned on the embedding inputs.
+
+ Parameters:
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+ The sequence used as a prompt for the generation.
+ max_new_tokens (`int`):
+ Number of new tokens to generate.
+ model_kwargs (`Dict[str, Any]`, *optional*):
+ Ad hoc parametrization of additional model-specific kwargs that will be forwarded to the `forward`
+ function of the model.
+
+ Return:
+ `inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+ The sequence of generated hidden-states.
+ """
+ max_new_tokens = max_new_tokens if max_new_tokens is not None else self.language_model.config.max_new_tokens
+ for _ in range(max_new_tokens):
+ # prepare model inputs
+ model_inputs = prepare_inputs_for_generation(inputs_embeds, **model_kwargs)
+
+ # forward pass to get next hidden states
+ output = self.language_model(**model_inputs, return_dict=True)
+
+ next_hidden_states = output.last_hidden_state
+
+ # Update the model input
+ inputs_embeds = torch.cat([inputs_embeds, next_hidden_states[:, -1:, :]], dim=1)
+
+ # Update generated hidden states, model inputs, and length for next step
+ model_kwargs = self.language_model._update_model_kwargs_for_generation(output, model_kwargs)
+
+ return inputs_embeds[:, -max_new_tokens:, :]
+
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_waveforms_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ generated_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_generated_prompt_embeds: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.LongTensor] = None,
+ negative_attention_mask: Optional[torch.LongTensor] = None,
+ max_new_tokens: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device (`torch.device`):
+ torch device
+ num_waveforms_per_prompt (`int`):
+ number of waveforms that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the audio generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-computed text embeddings from the Flan T5 model. Can be used to easily tweak text inputs, *e.g.*
+ prompt weighting. If not provided, text embeddings will be computed from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-computed negative text embeddings from the Flan T5 model. Can be used to easily tweak text inputs,
+ *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be computed from
+ `negative_prompt` input argument.
+ generated_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings from the GPT2 langauge model. Can be used to easily tweak text inputs,
+ *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input
+ argument.
+ negative_generated_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings from the GPT2 language model. Can be used to easily tweak text
+ inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be computed from
+ `negative_prompt` input argument.
+ attention_mask (`torch.LongTensor`, *optional*):
+ Pre-computed attention mask to be applied to the `prompt_embeds`. If not provided, attention mask will
+ be computed from `prompt` input argument.
+ negative_attention_mask (`torch.LongTensor`, *optional*):
+ Pre-computed attention mask to be applied to the `negative_prompt_embeds`. If not provided, attention
+ mask will be computed from `negative_prompt` input argument.
+ max_new_tokens (`int`, *optional*, defaults to None):
+ The number of new tokens to generate with the GPT2 language model.
+ Returns:
+ prompt_embeds (`torch.FloatTensor`):
+ Text embeddings from the Flan T5 model.
+ attention_mask (`torch.LongTensor`):
+ Attention mask to be applied to the `prompt_embeds`.
+ generated_prompt_embeds (`torch.FloatTensor`):
+ Text embeddings generated from the GPT2 langauge model.
+
+ Example:
+
+ ```python
+ >>> import scipy
+ >>> import torch
+ >>> from diffusers import AudioLDM2Pipeline
+
+ >>> repo_id = "cvssp/audioldm2"
+ >>> pipe = AudioLDM2Pipeline.from_pretrained(repo_id, torch_dtype=torch.float16)
+ >>> pipe = pipe.to("cuda")
+
+ >>> # Get text embedding vectors
+ >>> prompt_embeds, attention_mask, generated_prompt_embeds = pipe.encode_prompt(
+ ... prompt="Techno music with a strong, upbeat tempo and high melodic riffs",
+ ... device="cuda",
+ ... do_classifier_free_guidance=True,
+ ... )
+
+ >>> # Pass text embeddings to pipeline for text-conditional audio generation
+ >>> audio = pipe(
+ ... prompt_embeds=prompt_embeds,
+ ... attention_mask=attention_mask,
+ ... generated_prompt_embeds=generated_prompt_embeds,
+ ... num_inference_steps=200,
+ ... audio_length_in_s=10.0,
+ ... ).audios[0]
+
+ >>> # save generated audio sample
+ >>> scipy.io.wavfile.write("techno.wav", rate=16000, data=audio)
+ ```"""
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2]
+ text_encoders = [self.text_encoder, self.text_encoder_2]
+
+ if prompt_embeds is None:
+ prompt_embeds_list = []
+ attention_mask_list = []
+
+ for tokenizer, text_encoder in zip(tokenizers, text_encoders):
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length" if isinstance(tokenizer, (RobertaTokenizer, RobertaTokenizerFast)) else True,
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ attention_mask = text_inputs.attention_mask
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ f"The following part of your input was truncated because {text_encoder.config.model_type} can "
+ f"only handle sequences up to {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ text_input_ids = text_input_ids.to(device)
+ attention_mask = attention_mask.to(device)
+
+ if text_encoder.config.model_type == "clap":
+ prompt_embeds = text_encoder.get_text_features(
+ text_input_ids,
+ attention_mask=attention_mask,
+ )
+ # append the seq-len dim: (bs, hidden_size) -> (bs, seq_len, hidden_size)
+ prompt_embeds = prompt_embeds[:, None, :]
+ # make sure that we attend to this single hidden-state
+ attention_mask = attention_mask.new_ones((batch_size, 1))
+ else:
+ prompt_embeds = text_encoder(
+ text_input_ids,
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+
+ prompt_embeds_list.append(prompt_embeds)
+ attention_mask_list.append(attention_mask)
+
+ projection_output = self.projection_model(
+ hidden_states=prompt_embeds_list[0],
+ hidden_states_1=prompt_embeds_list[1],
+ attention_mask=attention_mask_list[0],
+ attention_mask_1=attention_mask_list[1],
+ )
+ projected_prompt_embeds = projection_output.hidden_states
+ projected_attention_mask = projection_output.attention_mask
+
+ generated_prompt_embeds = self.generate_language_model(
+ projected_prompt_embeds,
+ attention_mask=projected_attention_mask,
+ max_new_tokens=max_new_tokens,
+ )
+
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ attention_mask = (
+ attention_mask.to(device=device)
+ if attention_mask is not None
+ else torch.ones(prompt_embeds.shape[:2], dtype=torch.long, device=device)
+ )
+ generated_prompt_embeds = generated_prompt_embeds.to(dtype=self.language_model.dtype, device=device)
+
+ bs_embed, seq_len, hidden_size = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_waveforms_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_waveforms_per_prompt, seq_len, hidden_size)
+
+ # duplicate attention mask for each generation per prompt
+ attention_mask = attention_mask.repeat(1, num_waveforms_per_prompt)
+ attention_mask = attention_mask.view(bs_embed * num_waveforms_per_prompt, seq_len)
+
+ bs_embed, seq_len, hidden_size = generated_prompt_embeds.shape
+ # duplicate generated embeddings for each generation per prompt, using mps friendly method
+ generated_prompt_embeds = generated_prompt_embeds.repeat(1, num_waveforms_per_prompt, 1)
+ generated_prompt_embeds = generated_prompt_embeds.view(
+ bs_embed * num_waveforms_per_prompt, seq_len, hidden_size
+ )
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ negative_prompt_embeds_list = []
+ negative_attention_mask_list = []
+ max_length = prompt_embeds.shape[1]
+ for tokenizer, text_encoder in zip(tokenizers, text_encoders):
+ uncond_input = tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=tokenizer.model_max_length
+ if isinstance(tokenizer, (RobertaTokenizer, RobertaTokenizerFast))
+ else max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ uncond_input_ids = uncond_input.input_ids.to(device)
+ negative_attention_mask = uncond_input.attention_mask.to(device)
+
+ if text_encoder.config.model_type == "clap":
+ negative_prompt_embeds = text_encoder.get_text_features(
+ uncond_input_ids,
+ attention_mask=negative_attention_mask,
+ )
+ # append the seq-len dim: (bs, hidden_size) -> (bs, seq_len, hidden_size)
+ negative_prompt_embeds = negative_prompt_embeds[:, None, :]
+ # make sure that we attend to this single hidden-state
+ negative_attention_mask = negative_attention_mask.new_ones((batch_size, 1))
+ else:
+ negative_prompt_embeds = text_encoder(
+ uncond_input_ids,
+ attention_mask=negative_attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+ negative_attention_mask_list.append(negative_attention_mask)
+
+ projection_output = self.projection_model(
+ hidden_states=negative_prompt_embeds_list[0],
+ hidden_states_1=negative_prompt_embeds_list[1],
+ attention_mask=negative_attention_mask_list[0],
+ attention_mask_1=negative_attention_mask_list[1],
+ )
+ negative_projected_prompt_embeds = projection_output.hidden_states
+ negative_projected_attention_mask = projection_output.attention_mask
+
+ negative_generated_prompt_embeds = self.generate_language_model(
+ negative_projected_prompt_embeds,
+ attention_mask=negative_projected_attention_mask,
+ max_new_tokens=max_new_tokens,
+ )
+
+ if do_classifier_free_guidance:
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ negative_attention_mask = (
+ negative_attention_mask.to(device=device)
+ if negative_attention_mask is not None
+ else torch.ones(negative_prompt_embeds.shape[:2], dtype=torch.long, device=device)
+ )
+ negative_generated_prompt_embeds = negative_generated_prompt_embeds.to(
+ dtype=self.language_model.dtype, device=device
+ )
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_waveforms_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_waveforms_per_prompt, seq_len, -1)
+
+ # duplicate unconditional attention mask for each generation per prompt
+ negative_attention_mask = negative_attention_mask.repeat(1, num_waveforms_per_prompt)
+ negative_attention_mask = negative_attention_mask.view(batch_size * num_waveforms_per_prompt, seq_len)
+
+ # duplicate unconditional generated embeddings for each generation per prompt
+ seq_len = negative_generated_prompt_embeds.shape[1]
+ negative_generated_prompt_embeds = negative_generated_prompt_embeds.repeat(1, num_waveforms_per_prompt, 1)
+ negative_generated_prompt_embeds = negative_generated_prompt_embeds.view(
+ batch_size * num_waveforms_per_prompt, seq_len, -1
+ )
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ attention_mask = torch.cat([negative_attention_mask, attention_mask])
+ generated_prompt_embeds = torch.cat([negative_generated_prompt_embeds, generated_prompt_embeds])
+
+ return prompt_embeds, attention_mask, generated_prompt_embeds
+
+ # Copied from diffusers.pipelines.audioldm.pipeline_audioldm.AudioLDMPipeline.mel_spectrogram_to_waveform
+ def mel_spectrogram_to_waveform(self, mel_spectrogram):
+ if mel_spectrogram.dim() == 4:
+ mel_spectrogram = mel_spectrogram.squeeze(1)
+
+ waveform = self.vocoder(mel_spectrogram)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ waveform = waveform.cpu().float()
+ return waveform
+
+ def score_waveforms(self, text, audio, num_waveforms_per_prompt, device, dtype):
+ if not is_librosa_available():
+ logger.info(
+ "Automatic scoring of the generated audio waveforms against the input prompt text requires the "
+ "`librosa` package to resample the generated waveforms. Returning the audios in the order they were "
+ "generated. To enable automatic scoring, install `librosa` with: `pip install librosa`."
+ )
+ return audio
+ inputs = self.tokenizer(text, return_tensors="pt", padding=True)
+ resampled_audio = librosa.resample(
+ audio.numpy(), orig_sr=self.vocoder.config.sampling_rate, target_sr=self.feature_extractor.sampling_rate
+ )
+ inputs["input_features"] = self.feature_extractor(
+ list(resampled_audio), return_tensors="pt", sampling_rate=self.feature_extractor.sampling_rate
+ ).input_features.type(dtype)
+ inputs = inputs.to(device)
+
+ # compute the audio-text similarity score using the CLAP model
+ logits_per_text = self.text_encoder(**inputs).logits_per_text
+ # sort by the highest matching generations per prompt
+ indices = torch.argsort(logits_per_text, dim=1, descending=True)[:, :num_waveforms_per_prompt]
+ audio = torch.index_select(audio, 0, indices.reshape(-1).cpu())
+ return audio
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ audio_length_in_s,
+ vocoder_upsample_factor,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ generated_prompt_embeds=None,
+ negative_generated_prompt_embeds=None,
+ attention_mask=None,
+ negative_attention_mask=None,
+ ):
+ min_audio_length_in_s = vocoder_upsample_factor * self.vae_scale_factor
+ if audio_length_in_s < min_audio_length_in_s:
+ raise ValueError(
+ f"`audio_length_in_s` has to be a positive value greater than or equal to {min_audio_length_in_s}, but "
+ f"is {audio_length_in_s}."
+ )
+
+ if self.vocoder.config.model_in_dim % self.vae_scale_factor != 0:
+ raise ValueError(
+ f"The number of frequency bins in the vocoder's log-mel spectrogram has to be divisible by the "
+ f"VAE scale factor, but got {self.vocoder.config.model_in_dim} bins and a scale factor of "
+ f"{self.vae_scale_factor}."
+ )
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and (prompt_embeds is None or generated_prompt_embeds is None):
+ raise ValueError(
+ "Provide either `prompt`, or `prompt_embeds` and `generated_prompt_embeds`. Cannot leave "
+ "`prompt` undefined without specifying both `prompt_embeds` and `generated_prompt_embeds`."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_embeds is not None and negative_generated_prompt_embeds is None:
+ raise ValueError(
+ "Cannot forward `negative_prompt_embeds` without `negative_generated_prompt_embeds`. Ensure that"
+ "both arguments are specified"
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+ if attention_mask is not None and attention_mask.shape != prompt_embeds.shape[:2]:
+ raise ValueError(
+ "`attention_mask should have the same batch size and sequence length as `prompt_embeds`, but got:"
+ f"`attention_mask: {attention_mask.shape} != `prompt_embeds` {prompt_embeds.shape}"
+ )
+
+ if generated_prompt_embeds is not None and negative_generated_prompt_embeds is not None:
+ if generated_prompt_embeds.shape != negative_generated_prompt_embeds.shape:
+ raise ValueError(
+ "`generated_prompt_embeds` and `negative_generated_prompt_embeds` must have the same shape when "
+ f"passed directly, but got: `generated_prompt_embeds` {generated_prompt_embeds.shape} != "
+ f"`negative_generated_prompt_embeds` {negative_generated_prompt_embeds.shape}."
+ )
+ if (
+ negative_attention_mask is not None
+ and negative_attention_mask.shape != negative_prompt_embeds.shape[:2]
+ ):
+ raise ValueError(
+ "`attention_mask should have the same batch size and sequence length as `prompt_embeds`, but got:"
+ f"`attention_mask: {negative_attention_mask.shape} != `prompt_embeds` {negative_prompt_embeds.shape}"
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents with width->self.vocoder.config.model_in_dim
+ def prepare_latents(self, batch_size, num_channels_latents, height, dtype, device, generator, latents=None):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ height // self.vae_scale_factor,
+ self.vocoder.config.model_in_dim // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ audio_length_in_s: Optional[float] = None,
+ num_inference_steps: int = 200,
+ guidance_scale: float = 3.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_waveforms_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ generated_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_generated_prompt_embeds: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.LongTensor] = None,
+ negative_attention_mask: Optional[torch.LongTensor] = None,
+ max_new_tokens: Optional[int] = None,
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ output_type: Optional[str] = "np",
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide audio generation. If not defined, you need to pass `prompt_embeds`.
+ audio_length_in_s (`int`, *optional*, defaults to 10.24):
+ The length of the generated audio sample in seconds.
+ num_inference_steps (`int`, *optional*, defaults to 200):
+ The number of denoising steps. More denoising steps usually lead to a higher quality audio at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 3.5):
+ A higher guidance scale value encourages the model to generate audio that is closely linked to the text
+ `prompt` at the expense of lower sound quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in audio generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_waveforms_per_prompt (`int`, *optional*, defaults to 1):
+ The number of waveforms to generate per prompt. If `num_waveforms_per_prompt > 1`, then automatic
+ scoring is performed between the generated outputs and the text prompt. This scoring ranks the
+ generated waveforms based on their cosine similarity with the text input in the joint text-audio
+ embedding space.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for spectrogram
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ generated_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings from the GPT2 langauge model. Can be used to easily tweak text inputs,
+ *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input
+ argument.
+ negative_generated_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings from the GPT2 language model. Can be used to easily tweak text
+ inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be computed from
+ `negative_prompt` input argument.
+ attention_mask (`torch.LongTensor`, *optional*):
+ Pre-computed attention mask to be applied to the `prompt_embeds`. If not provided, attention mask will
+ be computed from `prompt` input argument.
+ negative_attention_mask (`torch.LongTensor`, *optional*):
+ Pre-computed attention mask to be applied to the `negative_prompt_embeds`. If not provided, attention
+ mask will be computed from `negative_prompt` input argument.
+ max_new_tokens (`int`, *optional*, defaults to None):
+ Number of new tokens to generate with the GPT2 language model. If not provided, number of tokens will
+ be taken from the config of the model.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ output_type (`str`, *optional*, defaults to `"np"`):
+ The output format of the generated audio. Choose between `"np"` to return a NumPy `np.ndarray` or
+ `"pt"` to return a PyTorch `torch.Tensor` object. Set to `"latent"` to return the latent diffusion
+ model (LDM) output.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated audio.
+ """
+ # 0. Convert audio input length from seconds to spectrogram height
+ vocoder_upsample_factor = np.prod(self.vocoder.config.upsample_rates) / self.vocoder.config.sampling_rate
+
+ if audio_length_in_s is None:
+ audio_length_in_s = self.unet.config.sample_size * self.vae_scale_factor * vocoder_upsample_factor
+
+ height = int(audio_length_in_s / vocoder_upsample_factor)
+
+ original_waveform_length = int(audio_length_in_s * self.vocoder.config.sampling_rate)
+ if height % self.vae_scale_factor != 0:
+ height = int(np.ceil(height / self.vae_scale_factor)) * self.vae_scale_factor
+ logger.info(
+ f"Audio length in seconds {audio_length_in_s} is increased to {height * vocoder_upsample_factor} "
+ f"so that it can be handled by the model. It will be cut to {audio_length_in_s} after the "
+ f"denoising process."
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ audio_length_in_s,
+ vocoder_upsample_factor,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ generated_prompt_embeds,
+ negative_generated_prompt_embeds,
+ attention_mask,
+ negative_attention_mask,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, attention_mask, generated_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_waveforms_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ generated_prompt_embeds=generated_prompt_embeds,
+ negative_generated_prompt_embeds=negative_generated_prompt_embeds,
+ attention_mask=attention_mask,
+ negative_attention_mask=negative_attention_mask,
+ max_new_tokens=max_new_tokens,
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_waveforms_per_prompt,
+ num_channels_latents,
+ height,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=generated_prompt_embeds,
+ encoder_hidden_states_1=prompt_embeds,
+ encoder_attention_mask_1=attention_mask,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ self.maybe_free_model_hooks()
+
+ # 8. Post-processing
+ if not output_type == "latent":
+ latents = 1 / self.vae.config.scaling_factor * latents
+ mel_spectrogram = self.vae.decode(latents).sample
+ else:
+ return AudioPipelineOutput(audios=latents)
+
+ audio = self.mel_spectrogram_to_waveform(mel_spectrogram)
+
+ audio = audio[:, :original_waveform_length]
+
+ # 9. Automatic scoring
+ if num_waveforms_per_prompt > 1 and prompt is not None:
+ audio = self.score_waveforms(
+ text=prompt,
+ audio=audio,
+ num_waveforms_per_prompt=num_waveforms_per_prompt,
+ device=device,
+ dtype=prompt_embeds.dtype,
+ )
+
+ if output_type == "np":
+ audio = audio.numpy()
+
+ if not return_dict:
+ return (audio,)
+
+ return AudioPipelineOutput(audios=audio)
diff --git a/model/diffusers_c/pipelines/auto_pipeline.py b/model/diffusers_c/pipelines/auto_pipeline.py
new file mode 100644
index 0000000000000000000000000000000000000000..bf509f205cf95db42f9b03607eb8123f30149f11
--- /dev/null
+++ b/model/diffusers_c/pipelines/auto_pipeline.py
@@ -0,0 +1,986 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from collections import OrderedDict
+
+from huggingface_hub.utils import validate_hf_hub_args
+
+from ..configuration_utils import ConfigMixin
+from .controlnet import (
+ StableDiffusionControlNetImg2ImgPipeline,
+ StableDiffusionControlNetInpaintPipeline,
+ StableDiffusionControlNetPipeline,
+ StableDiffusionXLControlNetImg2ImgPipeline,
+ StableDiffusionXLControlNetInpaintPipeline,
+ StableDiffusionXLControlNetPipeline,
+)
+from .deepfloyd_if import IFImg2ImgPipeline, IFInpaintingPipeline, IFPipeline
+from .kandinsky import (
+ KandinskyCombinedPipeline,
+ KandinskyImg2ImgCombinedPipeline,
+ KandinskyImg2ImgPipeline,
+ KandinskyInpaintCombinedPipeline,
+ KandinskyInpaintPipeline,
+ KandinskyPipeline,
+)
+from .kandinsky2_2 import (
+ KandinskyV22CombinedPipeline,
+ KandinskyV22Img2ImgCombinedPipeline,
+ KandinskyV22Img2ImgPipeline,
+ KandinskyV22InpaintCombinedPipeline,
+ KandinskyV22InpaintPipeline,
+ KandinskyV22Pipeline,
+)
+from .kandinsky3 import Kandinsky3Img2ImgPipeline, Kandinsky3Pipeline
+from .latent_consistency_models import LatentConsistencyModelImg2ImgPipeline, LatentConsistencyModelPipeline
+from .pixart_alpha import PixArtAlphaPipeline
+from .stable_diffusion import (
+ StableDiffusionImg2ImgPipeline,
+ StableDiffusionInpaintPipeline,
+ StableDiffusionPipeline,
+)
+from .stable_diffusion_xl import (
+ StableDiffusionXLImg2ImgPipeline,
+ StableDiffusionXLInpaintPipeline,
+ StableDiffusionXLPipeline,
+)
+from .wuerstchen import WuerstchenCombinedPipeline, WuerstchenDecoderPipeline
+
+
+AUTO_TEXT2IMAGE_PIPELINES_MAPPING = OrderedDict(
+ [
+ ("stable-diffusion", StableDiffusionPipeline),
+ ("stable-diffusion-xl", StableDiffusionXLPipeline),
+ ("if", IFPipeline),
+ ("kandinsky", KandinskyCombinedPipeline),
+ ("kandinsky22", KandinskyV22CombinedPipeline),
+ ("kandinsky3", Kandinsky3Pipeline),
+ ("stable-diffusion-controlnet", StableDiffusionControlNetPipeline),
+ ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetPipeline),
+ ("wuerstchen", WuerstchenCombinedPipeline),
+ ("lcm", LatentConsistencyModelPipeline),
+ ("pixart", PixArtAlphaPipeline),
+ ]
+)
+
+AUTO_IMAGE2IMAGE_PIPELINES_MAPPING = OrderedDict(
+ [
+ ("stable-diffusion", StableDiffusionImg2ImgPipeline),
+ ("stable-diffusion-xl", StableDiffusionXLImg2ImgPipeline),
+ ("if", IFImg2ImgPipeline),
+ ("kandinsky", KandinskyImg2ImgCombinedPipeline),
+ ("kandinsky22", KandinskyV22Img2ImgCombinedPipeline),
+ ("kandinsky3", Kandinsky3Img2ImgPipeline),
+ ("stable-diffusion-controlnet", StableDiffusionControlNetImg2ImgPipeline),
+ ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetImg2ImgPipeline),
+ ("lcm", LatentConsistencyModelImg2ImgPipeline),
+ ]
+)
+
+AUTO_INPAINT_PIPELINES_MAPPING = OrderedDict(
+ [
+ ("stable-diffusion", StableDiffusionInpaintPipeline),
+ ("stable-diffusion-xl", StableDiffusionXLInpaintPipeline),
+ ("if", IFInpaintingPipeline),
+ ("kandinsky", KandinskyInpaintCombinedPipeline),
+ ("kandinsky22", KandinskyV22InpaintCombinedPipeline),
+ ("stable-diffusion-controlnet", StableDiffusionControlNetInpaintPipeline),
+ ("stable-diffusion-xl-controlnet", StableDiffusionXLControlNetInpaintPipeline),
+ ]
+)
+
+_AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict(
+ [
+ ("kandinsky", KandinskyPipeline),
+ ("kandinsky22", KandinskyV22Pipeline),
+ ("wuerstchen", WuerstchenDecoderPipeline),
+ ]
+)
+_AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING = OrderedDict(
+ [
+ ("kandinsky", KandinskyImg2ImgPipeline),
+ ("kandinsky22", KandinskyV22Img2ImgPipeline),
+ ]
+)
+_AUTO_INPAINT_DECODER_PIPELINES_MAPPING = OrderedDict(
+ [
+ ("kandinsky", KandinskyInpaintPipeline),
+ ("kandinsky22", KandinskyV22InpaintPipeline),
+ ]
+)
+
+SUPPORTED_TASKS_MAPPINGS = [
+ AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
+ AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
+ AUTO_INPAINT_PIPELINES_MAPPING,
+ _AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING,
+ _AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING,
+ _AUTO_INPAINT_DECODER_PIPELINES_MAPPING,
+]
+
+
+def _get_connected_pipeline(pipeline_cls):
+ # for now connected pipelines can only be loaded from decoder pipelines, such as kandinsky-community/kandinsky-2-2-decoder
+ if pipeline_cls in _AUTO_TEXT2IMAGE_DECODER_PIPELINES_MAPPING.values():
+ return _get_task_class(
+ AUTO_TEXT2IMAGE_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False
+ )
+ if pipeline_cls in _AUTO_IMAGE2IMAGE_DECODER_PIPELINES_MAPPING.values():
+ return _get_task_class(
+ AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False
+ )
+ if pipeline_cls in _AUTO_INPAINT_DECODER_PIPELINES_MAPPING.values():
+ return _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, pipeline_cls.__name__, throw_error_if_not_exist=False)
+
+
+def _get_task_class(mapping, pipeline_class_name, throw_error_if_not_exist: bool = True):
+ def get_model(pipeline_class_name):
+ for task_mapping in SUPPORTED_TASKS_MAPPINGS:
+ for model_name, pipeline in task_mapping.items():
+ if pipeline.__name__ == pipeline_class_name:
+ return model_name
+
+ model_name = get_model(pipeline_class_name)
+
+ if model_name is not None:
+ task_class = mapping.get(model_name, None)
+ if task_class is not None:
+ return task_class
+
+ if throw_error_if_not_exist:
+ raise ValueError(f"AutoPipeline can't find a pipeline linked to {pipeline_class_name} for {model_name}")
+
+
+class AutoPipelineForText2Image(ConfigMixin):
+ r"""
+
+ [`AutoPipelineForText2Image`] is a generic pipeline class that instantiates a text-to-image pipeline class. The
+ specific underlying pipeline class is automatically selected from either the
+ [`~AutoPipelineForText2Image.from_pretrained`] or [`~AutoPipelineForText2Image.from_pipe`] methods.
+
+ This class cannot be instantiated using `__init__()` (throws an error).
+
+ Class attributes:
+
+ - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
+ diffusion pipeline's components.
+
+ """
+
+ config_name = "model_index.json"
+
+ def __init__(self, *args, **kwargs):
+ raise EnvironmentError(
+ f"{self.__class__.__name__} is designed to be instantiated "
+ f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or "
+ f"`{self.__class__.__name__}.from_pipe(pipeline)` methods."
+ )
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(cls, pretrained_model_or_path, **kwargs):
+ r"""
+ Instantiates a text-to-image Pytorch diffusion pipeline from pretrained pipeline weight.
+
+ The from_pretrained() method takes care of returning the correct pipeline class instance by:
+ 1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its
+ config object
+ 2. Find the text-to-image pipeline linked to the pipeline class using pattern matching on pipeline class
+ name.
+
+ If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetPipeline`] object.
+
+ The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ If you get the error message below, you need to finetune the weights for your downstream task:
+
+ ```
+ Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+ - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+ ```
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
+ hosted on the Hub.
+ - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
+ saved using
+ [`~DiffusionPipeline.save_pretrained`].
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model with another dtype. If "auto" is passed, the
+ dtype is automatically derived from the model's weights.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ custom_revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
+ `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
+ custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+ device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+ A map that specifies where each submodule should go. It doesn’t need to be defined for each
+ parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
+ same device.
+
+ Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
+ more information about each option see [designing a device
+ map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+ max_memory (`Dict`, *optional*):
+ A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
+ each GPU and the available CPU RAM if unset.
+ offload_folder (`str` or `os.PathLike`, *optional*):
+ The path to offload weights if device_map contains the value `"disk"`.
+ offload_state_dict (`bool`, *optional*):
+ If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
+ the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
+ when there is some disk offload.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ use_safetensors (`bool`, *optional*, defaults to `None`):
+ If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+ safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+ weights. If set to `False`, safetensors weights are not loaded.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
+ class). The overwritten components are passed directly to the pipelines `__init__` method. See example
+ below for more information.
+ variant (`str`, *optional*):
+ Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
+ loading `from_flax`.
+
+
+
+ To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
+ `huggingface-cli login`.
+
+
+
+ Examples:
+
+ ```py
+ >>> from diffusers import AutoPipelineForText2Image
+
+ >>> pipeline = AutoPipelineForText2Image.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> image = pipeline(prompt).images[0]
+ ```
+ """
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ token = kwargs.pop("token", None)
+ local_files_only = kwargs.pop("local_files_only", False)
+ revision = kwargs.pop("revision", None)
+
+ load_config_kwargs = {
+ "cache_dir": cache_dir,
+ "force_download": force_download,
+ "resume_download": resume_download,
+ "proxies": proxies,
+ "token": token,
+ "local_files_only": local_files_only,
+ "revision": revision,
+ }
+
+ config = cls.load_config(pretrained_model_or_path, **load_config_kwargs)
+ orig_class_name = config["_class_name"]
+
+ if "controlnet" in kwargs:
+ orig_class_name = config["_class_name"].replace("Pipeline", "ControlNetPipeline")
+
+ text_2_image_cls = _get_task_class(AUTO_TEXT2IMAGE_PIPELINES_MAPPING, orig_class_name)
+
+ kwargs = {**load_config_kwargs, **kwargs}
+ return text_2_image_cls.from_pretrained(pretrained_model_or_path, **kwargs)
+
+ @classmethod
+ def from_pipe(cls, pipeline, **kwargs):
+ r"""
+ Instantiates a text-to-image Pytorch diffusion pipeline from another instantiated diffusion pipeline class.
+
+ The from_pipe() method takes care of returning the correct pipeline class instance by finding the text-to-image
+ pipeline linked to the pipeline class using pattern matching on pipeline class name.
+
+ All the modules the pipeline contains will be used to initialize the new pipeline without reallocating
+ additional memory.
+
+ The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ Parameters:
+ pipeline (`DiffusionPipeline`):
+ an instantiated `DiffusionPipeline` object
+
+ ```py
+ >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
+
+ >>> pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", requires_safety_checker=False
+ ... )
+
+ >>> pipe_t2i = AutoPipelineForText2Image.from_pipe(pipe_i2i)
+ >>> image = pipe_t2i(prompt).images[0]
+ ```
+ """
+
+ original_config = dict(pipeline.config)
+ original_cls_name = pipeline.__class__.__name__
+
+ # derive the pipeline class to instantiate
+ text_2_image_cls = _get_task_class(AUTO_TEXT2IMAGE_PIPELINES_MAPPING, original_cls_name)
+
+ if "controlnet" in kwargs:
+ if kwargs["controlnet"] is not None:
+ text_2_image_cls = _get_task_class(
+ AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
+ text_2_image_cls.__name__.replace("ControlNet", "").replace("Pipeline", "ControlNetPipeline"),
+ )
+ else:
+ text_2_image_cls = _get_task_class(
+ AUTO_TEXT2IMAGE_PIPELINES_MAPPING,
+ text_2_image_cls.__name__.replace("ControlNetPipeline", "Pipeline"),
+ )
+
+ # define expected module and optional kwargs given the pipeline signature
+ expected_modules, optional_kwargs = text_2_image_cls._get_signature_keys(text_2_image_cls)
+
+ pretrained_model_name_or_path = original_config.pop("_name_or_path", None)
+
+ # allow users pass modules in `kwargs` to override the original pipeline's components
+ passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
+ original_class_obj = {
+ k: pipeline.components[k]
+ for k, v in pipeline.components.items()
+ if k in expected_modules and k not in passed_class_obj
+ }
+
+ # allow users pass optional kwargs to override the original pipelines config attribute
+ passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
+ original_pipe_kwargs = {
+ k: original_config[k]
+ for k, v in original_config.items()
+ if k in optional_kwargs and k not in passed_pipe_kwargs
+ }
+
+ # config that were not expected by original pipeline is stored as private attribute
+ # we will pass them as optional arguments if they can be accepted by the pipeline
+ additional_pipe_kwargs = [
+ k[1:]
+ for k in original_config.keys()
+ if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs
+ ]
+ for k in additional_pipe_kwargs:
+ original_pipe_kwargs[k] = original_config.pop(f"_{k}")
+
+ text_2_image_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs}
+
+ # store unused config as private attribute
+ unused_original_config = {
+ f"{'' if k.startswith('_') else '_'}{k}": original_config[k]
+ for k, v in original_config.items()
+ if k not in text_2_image_kwargs
+ }
+
+ missing_modules = set(expected_modules) - set(pipeline._optional_components) - set(text_2_image_kwargs.keys())
+
+ if len(missing_modules) > 0:
+ raise ValueError(
+ f"Pipeline {text_2_image_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed"
+ )
+
+ model = text_2_image_cls(**text_2_image_kwargs)
+ model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+ model.register_to_config(**unused_original_config)
+
+ return model
+
+
+class AutoPipelineForImage2Image(ConfigMixin):
+ r"""
+
+ [`AutoPipelineForImage2Image`] is a generic pipeline class that instantiates an image-to-image pipeline class. The
+ specific underlying pipeline class is automatically selected from either the
+ [`~AutoPipelineForImage2Image.from_pretrained`] or [`~AutoPipelineForImage2Image.from_pipe`] methods.
+
+ This class cannot be instantiated using `__init__()` (throws an error).
+
+ Class attributes:
+
+ - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
+ diffusion pipeline's components.
+
+ """
+
+ config_name = "model_index.json"
+
+ def __init__(self, *args, **kwargs):
+ raise EnvironmentError(
+ f"{self.__class__.__name__} is designed to be instantiated "
+ f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or "
+ f"`{self.__class__.__name__}.from_pipe(pipeline)` methods."
+ )
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(cls, pretrained_model_or_path, **kwargs):
+ r"""
+ Instantiates a image-to-image Pytorch diffusion pipeline from pretrained pipeline weight.
+
+ The from_pretrained() method takes care of returning the correct pipeline class instance by:
+ 1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its
+ config object
+ 2. Find the image-to-image pipeline linked to the pipeline class using pattern matching on pipeline class
+ name.
+
+ If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetImg2ImgPipeline`]
+ object.
+
+ The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ If you get the error message below, you need to finetune the weights for your downstream task:
+
+ ```
+ Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+ - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+ ```
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
+ hosted on the Hub.
+ - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
+ saved using
+ [`~DiffusionPipeline.save_pretrained`].
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model with another dtype. If "auto" is passed, the
+ dtype is automatically derived from the model's weights.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ custom_revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
+ `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
+ custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+ device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+ A map that specifies where each submodule should go. It doesn’t need to be defined for each
+ parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
+ same device.
+
+ Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
+ more information about each option see [designing a device
+ map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+ max_memory (`Dict`, *optional*):
+ A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
+ each GPU and the available CPU RAM if unset.
+ offload_folder (`str` or `os.PathLike`, *optional*):
+ The path to offload weights if device_map contains the value `"disk"`.
+ offload_state_dict (`bool`, *optional*):
+ If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
+ the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
+ when there is some disk offload.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ use_safetensors (`bool`, *optional*, defaults to `None`):
+ If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+ safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+ weights. If set to `False`, safetensors weights are not loaded.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
+ class). The overwritten components are passed directly to the pipelines `__init__` method. See example
+ below for more information.
+ variant (`str`, *optional*):
+ Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
+ loading `from_flax`.
+
+
+
+ To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
+ `huggingface-cli login`.
+
+
+
+ Examples:
+
+ ```py
+ >>> from diffusers import AutoPipelineForImage2Image
+
+ >>> pipeline = AutoPipelineForImage2Image.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> image = pipeline(prompt, image).images[0]
+ ```
+ """
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ token = kwargs.pop("token", None)
+ local_files_only = kwargs.pop("local_files_only", False)
+ revision = kwargs.pop("revision", None)
+
+ load_config_kwargs = {
+ "cache_dir": cache_dir,
+ "force_download": force_download,
+ "resume_download": resume_download,
+ "proxies": proxies,
+ "token": token,
+ "local_files_only": local_files_only,
+ "revision": revision,
+ }
+
+ config = cls.load_config(pretrained_model_or_path, **load_config_kwargs)
+ orig_class_name = config["_class_name"]
+
+ if "controlnet" in kwargs:
+ orig_class_name = config["_class_name"].replace("Pipeline", "ControlNetPipeline")
+
+ image_2_image_cls = _get_task_class(AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, orig_class_name)
+
+ kwargs = {**load_config_kwargs, **kwargs}
+ return image_2_image_cls.from_pretrained(pretrained_model_or_path, **kwargs)
+
+ @classmethod
+ def from_pipe(cls, pipeline, **kwargs):
+ r"""
+ Instantiates a image-to-image Pytorch diffusion pipeline from another instantiated diffusion pipeline class.
+
+ The from_pipe() method takes care of returning the correct pipeline class instance by finding the
+ image-to-image pipeline linked to the pipeline class using pattern matching on pipeline class name.
+
+ All the modules the pipeline contains will be used to initialize the new pipeline without reallocating
+ additional memory.
+
+ The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ Parameters:
+ pipeline (`DiffusionPipeline`):
+ an instantiated `DiffusionPipeline` object
+
+ Examples:
+
+ ```py
+ >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForImage2Image
+
+ >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", requires_safety_checker=False
+ ... )
+
+ >>> pipe_i2i = AutoPipelineForImage2Image.from_pipe(pipe_t2i)
+ >>> image = pipe_i2i(prompt, image).images[0]
+ ```
+ """
+
+ original_config = dict(pipeline.config)
+ original_cls_name = pipeline.__class__.__name__
+
+ # derive the pipeline class to instantiate
+ image_2_image_cls = _get_task_class(AUTO_IMAGE2IMAGE_PIPELINES_MAPPING, original_cls_name)
+
+ if "controlnet" in kwargs:
+ if kwargs["controlnet"] is not None:
+ image_2_image_cls = _get_task_class(
+ AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
+ image_2_image_cls.__name__.replace("ControlNet", "").replace(
+ "Img2ImgPipeline", "ControlNetImg2ImgPipeline"
+ ),
+ )
+ else:
+ image_2_image_cls = _get_task_class(
+ AUTO_IMAGE2IMAGE_PIPELINES_MAPPING,
+ image_2_image_cls.__name__.replace("ControlNetImg2ImgPipeline", "Img2ImgPipeline"),
+ )
+
+ # define expected module and optional kwargs given the pipeline signature
+ expected_modules, optional_kwargs = image_2_image_cls._get_signature_keys(image_2_image_cls)
+
+ pretrained_model_name_or_path = original_config.pop("_name_or_path", None)
+
+ # allow users pass modules in `kwargs` to override the original pipeline's components
+ passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
+ original_class_obj = {
+ k: pipeline.components[k]
+ for k, v in pipeline.components.items()
+ if k in expected_modules and k not in passed_class_obj
+ }
+
+ # allow users pass optional kwargs to override the original pipelines config attribute
+ passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
+ original_pipe_kwargs = {
+ k: original_config[k]
+ for k, v in original_config.items()
+ if k in optional_kwargs and k not in passed_pipe_kwargs
+ }
+
+ # config attribute that were not expected by original pipeline is stored as its private attribute
+ # we will pass them as optional arguments if they can be accepted by the pipeline
+ additional_pipe_kwargs = [
+ k[1:]
+ for k in original_config.keys()
+ if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs
+ ]
+ for k in additional_pipe_kwargs:
+ original_pipe_kwargs[k] = original_config.pop(f"_{k}")
+
+ image_2_image_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs}
+
+ # store unused config as private attribute
+ unused_original_config = {
+ f"{'' if k.startswith('_') else '_'}{k}": original_config[k]
+ for k, v in original_config.items()
+ if k not in image_2_image_kwargs
+ }
+
+ missing_modules = set(expected_modules) - set(pipeline._optional_components) - set(image_2_image_kwargs.keys())
+
+ if len(missing_modules) > 0:
+ raise ValueError(
+ f"Pipeline {image_2_image_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed"
+ )
+
+ model = image_2_image_cls(**image_2_image_kwargs)
+ model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+ model.register_to_config(**unused_original_config)
+
+ return model
+
+
+class AutoPipelineForInpainting(ConfigMixin):
+ r"""
+
+ [`AutoPipelineForInpainting`] is a generic pipeline class that instantiates an inpainting pipeline class. The
+ specific underlying pipeline class is automatically selected from either the
+ [`~AutoPipelineForInpainting.from_pretrained`] or [`~AutoPipelineForInpainting.from_pipe`] methods.
+
+ This class cannot be instantiated using `__init__()` (throws an error).
+
+ Class attributes:
+
+ - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
+ diffusion pipeline's components.
+
+ """
+
+ config_name = "model_index.json"
+
+ def __init__(self, *args, **kwargs):
+ raise EnvironmentError(
+ f"{self.__class__.__name__} is designed to be instantiated "
+ f"using the `{self.__class__.__name__}.from_pretrained(pretrained_model_name_or_path)` or "
+ f"`{self.__class__.__name__}.from_pipe(pipeline)` methods."
+ )
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(cls, pretrained_model_or_path, **kwargs):
+ r"""
+ Instantiates a inpainting Pytorch diffusion pipeline from pretrained pipeline weight.
+
+ The from_pretrained() method takes care of returning the correct pipeline class instance by:
+ 1. Detect the pipeline class of the pretrained_model_or_path based on the _class_name property of its
+ config object
+ 2. Find the inpainting pipeline linked to the pipeline class using pattern matching on pipeline class name.
+
+ If a `controlnet` argument is passed, it will instantiate a [`StableDiffusionControlNetInpaintPipeline`]
+ object.
+
+ The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ If you get the error message below, you need to finetune the weights for your downstream task:
+
+ ```
+ Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+ - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+ ```
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
+ hosted on the Hub.
+ - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
+ saved using
+ [`~DiffusionPipeline.save_pretrained`].
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model with another dtype. If "auto" is passed, the
+ dtype is automatically derived from the model's weights.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ custom_revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
+ `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
+ custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+ device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+ A map that specifies where each submodule should go. It doesn’t need to be defined for each
+ parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
+ same device.
+
+ Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
+ more information about each option see [designing a device
+ map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+ max_memory (`Dict`, *optional*):
+ A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
+ each GPU and the available CPU RAM if unset.
+ offload_folder (`str` or `os.PathLike`, *optional*):
+ The path to offload weights if device_map contains the value `"disk"`.
+ offload_state_dict (`bool`, *optional*):
+ If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
+ the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
+ when there is some disk offload.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ use_safetensors (`bool`, *optional*, defaults to `None`):
+ If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+ safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+ weights. If set to `False`, safetensors weights are not loaded.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
+ class). The overwritten components are passed directly to the pipelines `__init__` method. See example
+ below for more information.
+ variant (`str`, *optional*):
+ Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
+ loading `from_flax`.
+
+
+
+ To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
+ `huggingface-cli login`.
+
+
+
+ Examples:
+
+ ```py
+ >>> from diffusers import AutoPipelineForInpainting
+
+ >>> pipeline = AutoPipelineForInpainting.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> image = pipeline(prompt, image=init_image, mask_image=mask_image).images[0]
+ ```
+ """
+ cache_dir = kwargs.pop("cache_dir", None)
+ force_download = kwargs.pop("force_download", False)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ token = kwargs.pop("token", None)
+ local_files_only = kwargs.pop("local_files_only", False)
+ revision = kwargs.pop("revision", None)
+
+ load_config_kwargs = {
+ "cache_dir": cache_dir,
+ "force_download": force_download,
+ "resume_download": resume_download,
+ "proxies": proxies,
+ "token": token,
+ "local_files_only": local_files_only,
+ "revision": revision,
+ }
+
+ config = cls.load_config(pretrained_model_or_path, **load_config_kwargs)
+ orig_class_name = config["_class_name"]
+
+ if "controlnet" in kwargs:
+ orig_class_name = config["_class_name"].replace("Pipeline", "ControlNetPipeline")
+
+ inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, orig_class_name)
+
+ kwargs = {**load_config_kwargs, **kwargs}
+ return inpainting_cls.from_pretrained(pretrained_model_or_path, **kwargs)
+
+ @classmethod
+ def from_pipe(cls, pipeline, **kwargs):
+ r"""
+ Instantiates a inpainting Pytorch diffusion pipeline from another instantiated diffusion pipeline class.
+
+ The from_pipe() method takes care of returning the correct pipeline class instance by finding the inpainting
+ pipeline linked to the pipeline class using pattern matching on pipeline class name.
+
+ All the modules the pipeline class contain will be used to initialize the new pipeline without reallocating
+ additional memory.
+
+ The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ Parameters:
+ pipeline (`DiffusionPipeline`):
+ an instantiated `DiffusionPipeline` object
+
+ Examples:
+
+ ```py
+ >>> from diffusers import AutoPipelineForText2Image, AutoPipelineForInpainting
+
+ >>> pipe_t2i = AutoPipelineForText2Image.from_pretrained(
+ ... "DeepFloyd/IF-I-XL-v1.0", requires_safety_checker=False
+ ... )
+
+ >>> pipe_inpaint = AutoPipelineForInpainting.from_pipe(pipe_t2i)
+ >>> image = pipe_inpaint(prompt, image=init_image, mask_image=mask_image).images[0]
+ ```
+ """
+ original_config = dict(pipeline.config)
+ original_cls_name = pipeline.__class__.__name__
+
+ # derive the pipeline class to instantiate
+ inpainting_cls = _get_task_class(AUTO_INPAINT_PIPELINES_MAPPING, original_cls_name)
+
+ if "controlnet" in kwargs:
+ if kwargs["controlnet"] is not None:
+ inpainting_cls = _get_task_class(
+ AUTO_INPAINT_PIPELINES_MAPPING,
+ inpainting_cls.__name__.replace("ControlNet", "").replace(
+ "InpaintPipeline", "ControlNetInpaintPipeline"
+ ),
+ )
+ else:
+ inpainting_cls = _get_task_class(
+ AUTO_INPAINT_PIPELINES_MAPPING,
+ inpainting_cls.__name__.replace("ControlNetInpaintPipeline", "InpaintPipeline"),
+ )
+
+ # define expected module and optional kwargs given the pipeline signature
+ expected_modules, optional_kwargs = inpainting_cls._get_signature_keys(inpainting_cls)
+
+ pretrained_model_name_or_path = original_config.pop("_name_or_path", None)
+
+ # allow users pass modules in `kwargs` to override the original pipeline's components
+ passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
+ original_class_obj = {
+ k: pipeline.components[k]
+ for k, v in pipeline.components.items()
+ if k in expected_modules and k not in passed_class_obj
+ }
+
+ # allow users pass optional kwargs to override the original pipelines config attribute
+ passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
+ original_pipe_kwargs = {
+ k: original_config[k]
+ for k, v in original_config.items()
+ if k in optional_kwargs and k not in passed_pipe_kwargs
+ }
+
+ # config that were not expected by original pipeline is stored as private attribute
+ # we will pass them as optional arguments if they can be accepted by the pipeline
+ additional_pipe_kwargs = [
+ k[1:]
+ for k in original_config.keys()
+ if k.startswith("_") and k[1:] in optional_kwargs and k[1:] not in passed_pipe_kwargs
+ ]
+ for k in additional_pipe_kwargs:
+ original_pipe_kwargs[k] = original_config.pop(f"_{k}")
+
+ inpainting_kwargs = {**passed_class_obj, **original_class_obj, **passed_pipe_kwargs, **original_pipe_kwargs}
+
+ # store unused config as private attribute
+ unused_original_config = {
+ f"{'' if k.startswith('_') else '_'}{k}": original_config[k]
+ for k, v in original_config.items()
+ if k not in inpainting_kwargs
+ }
+
+ missing_modules = set(expected_modules) - set(pipeline._optional_components) - set(inpainting_kwargs.keys())
+
+ if len(missing_modules) > 0:
+ raise ValueError(
+ f"Pipeline {inpainting_cls} expected {expected_modules}, but only {set(list(passed_class_obj.keys()) + list(original_class_obj.keys()))} were passed"
+ )
+
+ model = inpainting_cls(**inpainting_kwargs)
+ model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+ model.register_to_config(**unused_original_config)
+
+ return model
diff --git a/model/diffusers_c/pipelines/blip_diffusion/__init__.py b/model/diffusers_c/pipelines/blip_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..af6c879d5ce88aa8edec0691e987444ff1d3dfec
--- /dev/null
+++ b/model/diffusers_c/pipelines/blip_diffusion/__init__.py
@@ -0,0 +1,20 @@
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL
+from PIL import Image
+
+from ...utils import OptionalDependencyNotAvailable, is_torch_available, is_transformers_available
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import ShapEPipeline
+else:
+ from .blip_image_processing import BlipImageProcessor
+ from .modeling_blip2 import Blip2QFormerModel
+ from .modeling_ctx_clip import ContextCLIPTextModel
+ from .pipeline_blip_diffusion import BlipDiffusionPipeline
diff --git a/model/diffusers_c/pipelines/blip_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5cf8598be2ccde42bf39dade39db3305aa5cdc88
Binary files /dev/null and b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/blip_diffusion/__pycache__/blip_image_processing.cpython-39.pyc b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/blip_image_processing.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b114393b3060df5cf3e7e203edecedba36619a28
Binary files /dev/null and b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/blip_image_processing.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/blip_diffusion/__pycache__/modeling_blip2.cpython-39.pyc b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/modeling_blip2.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..84cde42db04e63067fcdb2f6b416b8eb25cea941
Binary files /dev/null and b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/modeling_blip2.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/blip_diffusion/__pycache__/modeling_ctx_clip.cpython-39.pyc b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/modeling_ctx_clip.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5e64c0e0da24aad1fcff29b65d521ab7ca7ff20b
Binary files /dev/null and b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/modeling_ctx_clip.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/blip_diffusion/__pycache__/pipeline_blip_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/pipeline_blip_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d884b282cbde7871b7f0c3af0fdbe6bc59758324
Binary files /dev/null and b/model/diffusers_c/pipelines/blip_diffusion/__pycache__/pipeline_blip_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/blip_diffusion/blip_image_processing.py b/model/diffusers_c/pipelines/blip_diffusion/blip_image_processing.py
new file mode 100644
index 0000000000000000000000000000000000000000..9e32098321fc4825741a91f44be611e1dd7bf791
--- /dev/null
+++ b/model/diffusers_c/pipelines/blip_diffusion/blip_image_processing.py
@@ -0,0 +1,317 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Image processor class for BLIP."""
+
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature, get_size_dict
+from transformers.image_transforms import convert_to_rgb, resize, to_channel_dimension_format
+from transformers.image_utils import (
+ OPENAI_CLIP_MEAN,
+ OPENAI_CLIP_STD,
+ ChannelDimension,
+ ImageInput,
+ PILImageResampling,
+ infer_channel_dimension_format,
+ is_scaled_image,
+ make_list_of_images,
+ to_numpy_array,
+ valid_images,
+)
+from transformers.utils import TensorType, is_vision_available, logging
+
+from diffusers.utils import numpy_to_pil
+
+
+if is_vision_available():
+ import PIL.Image
+
+
+logger = logging.get_logger(__name__)
+
+
+# We needed some extra functions on top of the ones in transformers.image_processing_utils.BaseImageProcessor, namely center crop
+# Copy-pasted from transformers.models.blip.image_processing_blip.BlipImageProcessor
+class BlipImageProcessor(BaseImageProcessor):
+ r"""
+ Constructs a BLIP image processor.
+
+ Args:
+ do_resize (`bool`, *optional*, defaults to `True`):
+ Whether to resize the image's (height, width) dimensions to the specified `size`. Can be overridden by the
+ `do_resize` parameter in the `preprocess` method.
+ size (`dict`, *optional*, defaults to `{"height": 384, "width": 384}`):
+ Size of the output image after resizing. Can be overridden by the `size` parameter in the `preprocess`
+ method.
+ resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+ Resampling filter to use if resizing the image. Only has an effect if `do_resize` is set to `True`. Can be
+ overridden by the `resample` parameter in the `preprocess` method.
+ do_rescale (`bool`, *optional*, defaults to `True`):
+ Wwhether to rescale the image by the specified scale `rescale_factor`. Can be overridden by the
+ `do_rescale` parameter in the `preprocess` method.
+ rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+ Scale factor to use if rescaling the image. Only has an effect if `do_rescale` is set to `True`. Can be
+ overridden by the `rescale_factor` parameter in the `preprocess` method.
+ do_normalize (`bool`, *optional*, defaults to `True`):
+ Whether to normalize the image. Can be overridden by the `do_normalize` parameter in the `preprocess`
+ method. Can be overridden by the `do_normalize` parameter in the `preprocess` method.
+ image_mean (`float` or `List[float]`, *optional*, defaults to `IMAGENET_STANDARD_MEAN`):
+ Mean to use if normalizing the image. This is a float or list of floats the length of the number of
+ channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method. Can be
+ overridden by the `image_mean` parameter in the `preprocess` method.
+ image_std (`float` or `List[float]`, *optional*, defaults to `IMAGENET_STANDARD_STD`):
+ Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
+ number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
+ Can be overridden by the `image_std` parameter in the `preprocess` method.
+ do_convert_rgb (`bool`, *optional*, defaults to `True`):
+ Whether to convert the image to RGB.
+ """
+
+ model_input_names = ["pixel_values"]
+
+ def __init__(
+ self,
+ do_resize: bool = True,
+ size: Dict[str, int] = None,
+ resample: PILImageResampling = PILImageResampling.BICUBIC,
+ do_rescale: bool = True,
+ rescale_factor: Union[int, float] = 1 / 255,
+ do_normalize: bool = True,
+ image_mean: Optional[Union[float, List[float]]] = None,
+ image_std: Optional[Union[float, List[float]]] = None,
+ do_convert_rgb: bool = True,
+ do_center_crop: bool = True,
+ **kwargs,
+ ) -> None:
+ super().__init__(**kwargs)
+ size = size if size is not None else {"height": 224, "width": 224}
+ size = get_size_dict(size, default_to_square=True)
+
+ self.do_resize = do_resize
+ self.size = size
+ self.resample = resample
+ self.do_rescale = do_rescale
+ self.rescale_factor = rescale_factor
+ self.do_normalize = do_normalize
+ self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+ self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
+ self.do_convert_rgb = do_convert_rgb
+ self.do_center_crop = do_center_crop
+
+ # Copy-pasted from transformers.models.vit.image_processing_vit.ViTImageProcessor.resize with PILImageResampling.BILINEAR->PILImageResampling.BICUBIC
+ def resize(
+ self,
+ image: np.ndarray,
+ size: Dict[str, int],
+ resample: PILImageResampling = PILImageResampling.BICUBIC,
+ data_format: Optional[Union[str, ChannelDimension]] = None,
+ input_data_format: Optional[Union[str, ChannelDimension]] = None,
+ **kwargs,
+ ) -> np.ndarray:
+ """
+ Resize an image to `(size["height"], size["width"])`.
+
+ Args:
+ image (`np.ndarray`):
+ Image to resize.
+ size (`Dict[str, int]`):
+ Dictionary in the format `{"height": int, "width": int}` specifying the size of the output image.
+ resample (`PILImageResampling`, *optional*, defaults to `PILImageResampling.BICUBIC`):
+ `PILImageResampling` filter to use when resizing the image e.g. `PILImageResampling.BICUBIC`.
+ data_format (`ChannelDimension` or `str`, *optional*):
+ The channel dimension format for the output image. If unset, the channel dimension format of the input
+ image is used. Can be one of:
+ - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+ - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+ - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+ input_data_format (`ChannelDimension` or `str`, *optional*):
+ The channel dimension format for the input image. If unset, the channel dimension format is inferred
+ from the input image. Can be one of:
+ - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+ - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+ - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+
+ Returns:
+ `np.ndarray`: The resized image.
+ """
+ size = get_size_dict(size)
+ if "height" not in size or "width" not in size:
+ raise ValueError(f"The `size` dictionary must contain the keys `height` and `width`. Got {size.keys()}")
+ output_size = (size["height"], size["width"])
+ return resize(
+ image,
+ size=output_size,
+ resample=resample,
+ data_format=data_format,
+ input_data_format=input_data_format,
+ **kwargs,
+ )
+
+ def preprocess(
+ self,
+ images: ImageInput,
+ do_resize: Optional[bool] = None,
+ size: Optional[Dict[str, int]] = None,
+ resample: PILImageResampling = None,
+ do_rescale: Optional[bool] = None,
+ do_center_crop: Optional[bool] = None,
+ rescale_factor: Optional[float] = None,
+ do_normalize: Optional[bool] = None,
+ image_mean: Optional[Union[float, List[float]]] = None,
+ image_std: Optional[Union[float, List[float]]] = None,
+ return_tensors: Optional[Union[str, TensorType]] = None,
+ do_convert_rgb: bool = None,
+ data_format: ChannelDimension = ChannelDimension.FIRST,
+ input_data_format: Optional[Union[str, ChannelDimension]] = None,
+ **kwargs,
+ ) -> PIL.Image.Image:
+ """
+ Preprocess an image or batch of images.
+
+ Args:
+ images (`ImageInput`):
+ Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
+ passing in images with pixel values between 0 and 1, set `do_rescale=False`.
+ do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+ Whether to resize the image.
+ size (`Dict[str, int]`, *optional*, defaults to `self.size`):
+ Controls the size of the image after `resize`. The shortest edge of the image is resized to
+ `size["shortest_edge"]` whilst preserving the aspect ratio. If the longest edge of this resized image
+ is > `int(size["shortest_edge"] * (1333 / 800))`, then the image is resized again to make the longest
+ edge equal to `int(size["shortest_edge"] * (1333 / 800))`.
+ resample (`PILImageResampling`, *optional*, defaults to `self.resample`):
+ Resampling filter to use if resizing the image. Only has an effect if `do_resize` is set to `True`.
+ do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+ Whether to rescale the image values between [0 - 1].
+ rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+ Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+ do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+ Whether to normalize the image.
+ image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+ Image mean to normalize the image by if `do_normalize` is set to `True`.
+ image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+ Image standard deviation to normalize the image by if `do_normalize` is set to `True`.
+ do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+ Whether to convert the image to RGB.
+ return_tensors (`str` or `TensorType`, *optional*):
+ The type of tensors to return. Can be one of:
+ - Unset: Return a list of `np.ndarray`.
+ - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+ - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+ - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+ - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+ data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+ The channel dimension format for the output image. Can be one of:
+ - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+ - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+ - Unset: Use the channel dimension format of the input image.
+ input_data_format (`ChannelDimension` or `str`, *optional*):
+ The channel dimension format for the input image. If unset, the channel dimension format is inferred
+ from the input image. Can be one of:
+ - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+ - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+ - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+ """
+ do_resize = do_resize if do_resize is not None else self.do_resize
+ resample = resample if resample is not None else self.resample
+ do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+ rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+ do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+ image_mean = image_mean if image_mean is not None else self.image_mean
+ image_std = image_std if image_std is not None else self.image_std
+ do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+ do_center_crop = do_center_crop if do_center_crop is not None else self.do_center_crop
+
+ size = size if size is not None else self.size
+ size = get_size_dict(size, default_to_square=False)
+ images = make_list_of_images(images)
+
+ if not valid_images(images):
+ raise ValueError(
+ "Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
+ "torch.Tensor, tf.Tensor or jax.ndarray."
+ )
+
+ if do_resize and size is None or resample is None:
+ raise ValueError("Size and resample must be specified if do_resize is True.")
+
+ if do_rescale and rescale_factor is None:
+ raise ValueError("Rescale factor must be specified if do_rescale is True.")
+
+ if do_normalize and (image_mean is None or image_std is None):
+ raise ValueError("Image mean and std must be specified if do_normalize is True.")
+
+ # PIL RGBA images are converted to RGB
+ if do_convert_rgb:
+ images = [convert_to_rgb(image) for image in images]
+
+ # All transformations expect numpy arrays.
+ images = [to_numpy_array(image) for image in images]
+
+ if is_scaled_image(images[0]) and do_rescale:
+ logger.warning_once(
+ "It looks like you are trying to rescale already rescaled images. If the input"
+ " images have pixel values between 0 and 1, set `do_rescale=False` to avoid rescaling them again."
+ )
+ if input_data_format is None:
+ # We assume that all images have the same channel dimension format.
+ input_data_format = infer_channel_dimension_format(images[0])
+
+ if do_resize:
+ images = [
+ self.resize(image=image, size=size, resample=resample, input_data_format=input_data_format)
+ for image in images
+ ]
+
+ if do_rescale:
+ images = [
+ self.rescale(image=image, scale=rescale_factor, input_data_format=input_data_format)
+ for image in images
+ ]
+ if do_normalize:
+ images = [
+ self.normalize(image=image, mean=image_mean, std=image_std, input_data_format=input_data_format)
+ for image in images
+ ]
+ if do_center_crop:
+ images = [self.center_crop(image, size, input_data_format=input_data_format) for image in images]
+
+ images = [
+ to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format) for image in images
+ ]
+
+ encoded_outputs = BatchFeature(data={"pixel_values": images}, tensor_type=return_tensors)
+ return encoded_outputs
+
+ # Follows diffusers.VaeImageProcessor.postprocess
+ def postprocess(self, sample: torch.FloatTensor, output_type: str = "pil"):
+ if output_type not in ["pt", "np", "pil"]:
+ raise ValueError(
+ f"output_type={output_type} is not supported. Make sure to choose one of ['pt', 'np', or 'pil']"
+ )
+
+ # Equivalent to diffusers.VaeImageProcessor.denormalize
+ sample = (sample / 2 + 0.5).clamp(0, 1)
+ if output_type == "pt":
+ return sample
+
+ # Equivalent to diffusers.VaeImageProcessor.pt_to_numpy
+ sample = sample.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "np":
+ return sample
+ # Output_type must be 'pil'
+ sample = numpy_to_pil(sample)
+ return sample
diff --git a/model/diffusers_c/pipelines/blip_diffusion/modeling_blip2.py b/model/diffusers_c/pipelines/blip_diffusion/modeling_blip2.py
new file mode 100644
index 0000000000000000000000000000000000000000..c8869ad9db23777a242a44ec8efeae16d0175c0e
--- /dev/null
+++ b/model/diffusers_c/pipelines/blip_diffusion/modeling_blip2.py
@@ -0,0 +1,642 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from transformers import BertTokenizer
+from transformers.activations import QuickGELUActivation as QuickGELU
+from transformers.modeling_outputs import (
+ BaseModelOutputWithPastAndCrossAttentions,
+ BaseModelOutputWithPooling,
+ BaseModelOutputWithPoolingAndCrossAttentions,
+)
+from transformers.models.blip_2.configuration_blip_2 import Blip2Config, Blip2VisionConfig
+from transformers.models.blip_2.modeling_blip_2 import (
+ Blip2Encoder,
+ Blip2PreTrainedModel,
+ Blip2QFormerAttention,
+ Blip2QFormerIntermediate,
+ Blip2QFormerOutput,
+)
+from transformers.pytorch_utils import apply_chunking_to_forward
+from transformers.utils import (
+ logging,
+ replace_return_docstrings,
+)
+
+
+logger = logging.get_logger(__name__)
+
+
+# There is an implementation of Blip2 in `transformers` : https://github.com/huggingface/transformers/blob/main/src/transformers/models/blip_2/modeling_blip_2.py.
+# But it doesn't support getting multimodal embeddings. So, this module can be
+# replaced with a future `transformers` version supports that.
+class Blip2TextEmbeddings(nn.Module):
+ """Construct the embeddings from word and position embeddings."""
+
+ def __init__(self, config):
+ super().__init__()
+ self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+ self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+
+ # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+ # any TensorFlow checkpoint file
+ self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+ self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+ # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+ self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+ self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+
+ self.config = config
+
+ def forward(
+ self,
+ input_ids=None,
+ position_ids=None,
+ query_embeds=None,
+ past_key_values_length=0,
+ ):
+ if input_ids is not None:
+ seq_length = input_ids.size()[1]
+ else:
+ seq_length = 0
+
+ if position_ids is None:
+ position_ids = self.position_ids[:, past_key_values_length : seq_length + past_key_values_length].clone()
+
+ if input_ids is not None:
+ embeddings = self.word_embeddings(input_ids)
+ if self.position_embedding_type == "absolute":
+ position_embeddings = self.position_embeddings(position_ids)
+ embeddings = embeddings + position_embeddings
+
+ if query_embeds is not None:
+ batch_size = embeddings.shape[0]
+ # repeat the query embeddings for batch size
+ query_embeds = query_embeds.repeat(batch_size, 1, 1)
+ embeddings = torch.cat((query_embeds, embeddings), dim=1)
+ else:
+ embeddings = query_embeds
+ embeddings = embeddings.to(query_embeds.dtype)
+ embeddings = self.LayerNorm(embeddings)
+ embeddings = self.dropout(embeddings)
+ return embeddings
+
+
+# Copy-pasted from transformers.models.blip.modeling_blip.BlipVisionEmbeddings with Blip->Blip2
+class Blip2VisionEmbeddings(nn.Module):
+ def __init__(self, config: Blip2VisionConfig):
+ super().__init__()
+ self.config = config
+ self.embed_dim = config.hidden_size
+ self.image_size = config.image_size
+ self.patch_size = config.patch_size
+
+ self.class_embedding = nn.Parameter(torch.randn(1, 1, self.embed_dim))
+
+ self.patch_embedding = nn.Conv2d(
+ in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size, bias=False
+ )
+
+ self.num_patches = (self.image_size // self.patch_size) ** 2
+ self.num_positions = self.num_patches + 1
+
+ self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))
+
+ def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+ batch_size = pixel_values.shape[0]
+ target_dtype = self.patch_embedding.weight.dtype
+ patch_embeds = self.patch_embedding(pixel_values.to(dtype=target_dtype)) # shape = [*, width, grid, grid]
+ patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+ class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
+ embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+ embeddings = embeddings + self.position_embedding[:, : embeddings.size(1), :].to(target_dtype)
+ return embeddings
+
+
+# The Qformer encoder, which takes the visual embeddings, and the text input, to get multimodal embeddings
+class Blip2QFormerEncoder(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ self.config = config
+ self.layer = nn.ModuleList(
+ [Blip2QFormerLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+ )
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states,
+ attention_mask=None,
+ head_mask=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ past_key_values=None,
+ use_cache=None,
+ output_attentions=False,
+ output_hidden_states=False,
+ return_dict=True,
+ query_length=0,
+ ):
+ all_hidden_states = () if output_hidden_states else None
+ all_self_attentions = () if output_attentions else None
+ all_cross_attentions = () if output_attentions else None
+
+ next_decoder_cache = () if use_cache else None
+
+ for i in range(self.config.num_hidden_layers):
+ layer_module = self.layer[i]
+ if output_hidden_states:
+ all_hidden_states = all_hidden_states + (hidden_states,)
+
+ layer_head_mask = head_mask[i] if head_mask is not None else None
+ past_key_value = past_key_values[i] if past_key_values is not None else None
+
+ if getattr(self.config, "gradient_checkpointing", False) and self.training:
+ if use_cache:
+ logger.warning(
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+ )
+ use_cache = False
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs, past_key_value, output_attentions, query_length)
+
+ return custom_forward
+
+ layer_outputs = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(layer_module),
+ hidden_states,
+ attention_mask,
+ layer_head_mask,
+ encoder_hidden_states,
+ encoder_attention_mask,
+ )
+ else:
+ layer_outputs = layer_module(
+ hidden_states,
+ attention_mask,
+ layer_head_mask,
+ encoder_hidden_states,
+ encoder_attention_mask,
+ past_key_value,
+ output_attentions,
+ query_length,
+ )
+
+ hidden_states = layer_outputs[0]
+ if use_cache:
+ next_decoder_cache += (layer_outputs[-1],)
+ if output_attentions:
+ all_self_attentions = all_self_attentions + (layer_outputs[1],)
+ if layer_module.has_cross_attention:
+ all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+ if output_hidden_states:
+ all_hidden_states = all_hidden_states + (hidden_states,)
+
+ if not return_dict:
+ return tuple(
+ v
+ for v in [
+ hidden_states,
+ next_decoder_cache,
+ all_hidden_states,
+ all_self_attentions,
+ all_cross_attentions,
+ ]
+ if v is not None
+ )
+ return BaseModelOutputWithPastAndCrossAttentions(
+ last_hidden_state=hidden_states,
+ past_key_values=next_decoder_cache,
+ hidden_states=all_hidden_states,
+ attentions=all_self_attentions,
+ cross_attentions=all_cross_attentions,
+ )
+
+
+# The layers making up the Qformer encoder
+class Blip2QFormerLayer(nn.Module):
+ def __init__(self, config, layer_idx):
+ super().__init__()
+ self.chunk_size_feed_forward = config.chunk_size_feed_forward
+ self.seq_len_dim = 1
+ self.attention = Blip2QFormerAttention(config)
+
+ self.layer_idx = layer_idx
+
+ if layer_idx % config.cross_attention_frequency == 0:
+ self.crossattention = Blip2QFormerAttention(config, is_cross_attention=True)
+ self.has_cross_attention = True
+ else:
+ self.has_cross_attention = False
+
+ self.intermediate = Blip2QFormerIntermediate(config)
+ self.intermediate_query = Blip2QFormerIntermediate(config)
+ self.output_query = Blip2QFormerOutput(config)
+ self.output = Blip2QFormerOutput(config)
+
+ def forward(
+ self,
+ hidden_states,
+ attention_mask=None,
+ head_mask=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ past_key_value=None,
+ output_attentions=False,
+ query_length=0,
+ ):
+ # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+ self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+ self_attention_outputs = self.attention(
+ hidden_states,
+ attention_mask,
+ head_mask,
+ output_attentions=output_attentions,
+ past_key_value=self_attn_past_key_value,
+ )
+ attention_output = self_attention_outputs[0]
+ outputs = self_attention_outputs[1:-1]
+
+ present_key_value = self_attention_outputs[-1]
+
+ if query_length > 0:
+ query_attention_output = attention_output[:, :query_length, :]
+
+ if self.has_cross_attention:
+ if encoder_hidden_states is None:
+ raise ValueError("encoder_hidden_states must be given for cross-attention layers")
+ cross_attention_outputs = self.crossattention(
+ query_attention_output,
+ attention_mask,
+ head_mask,
+ encoder_hidden_states,
+ encoder_attention_mask,
+ output_attentions=output_attentions,
+ )
+ query_attention_output = cross_attention_outputs[0]
+ # add cross attentions if we output attention weights
+ outputs = outputs + cross_attention_outputs[1:-1]
+
+ layer_output = apply_chunking_to_forward(
+ self.feed_forward_chunk_query,
+ self.chunk_size_feed_forward,
+ self.seq_len_dim,
+ query_attention_output,
+ )
+
+ if attention_output.shape[1] > query_length:
+ layer_output_text = apply_chunking_to_forward(
+ self.feed_forward_chunk,
+ self.chunk_size_feed_forward,
+ self.seq_len_dim,
+ attention_output[:, query_length:, :],
+ )
+ layer_output = torch.cat([layer_output, layer_output_text], dim=1)
+ else:
+ layer_output = apply_chunking_to_forward(
+ self.feed_forward_chunk,
+ self.chunk_size_feed_forward,
+ self.seq_len_dim,
+ attention_output,
+ )
+ outputs = (layer_output,) + outputs
+
+ outputs = outputs + (present_key_value,)
+
+ return outputs
+
+ def feed_forward_chunk(self, attention_output):
+ intermediate_output = self.intermediate(attention_output)
+ layer_output = self.output(intermediate_output, attention_output)
+ return layer_output
+
+ def feed_forward_chunk_query(self, attention_output):
+ intermediate_output = self.intermediate_query(attention_output)
+ layer_output = self.output_query(intermediate_output, attention_output)
+ return layer_output
+
+
+# ProjLayer used to project the multimodal Blip2 embeddings to be used in the text encoder
+class ProjLayer(nn.Module):
+ def __init__(self, in_dim, out_dim, hidden_dim, drop_p=0.1, eps=1e-12):
+ super().__init__()
+
+ # Dense1 -> Act -> Dense2 -> Drop -> Res -> Norm
+ self.dense1 = nn.Linear(in_dim, hidden_dim)
+ self.act_fn = QuickGELU()
+ self.dense2 = nn.Linear(hidden_dim, out_dim)
+ self.dropout = nn.Dropout(drop_p)
+
+ self.LayerNorm = nn.LayerNorm(out_dim, eps=eps)
+
+ def forward(self, x):
+ x_in = x
+
+ x = self.LayerNorm(x)
+ x = self.dropout(self.dense2(self.act_fn(self.dense1(x)))) + x_in
+
+ return x
+
+
+# Copy-pasted from transformers.models.blip.modeling_blip.BlipVisionModel with Blip->Blip2, BLIP->BLIP_2
+class Blip2VisionModel(Blip2PreTrainedModel):
+ main_input_name = "pixel_values"
+ config_class = Blip2VisionConfig
+
+ def __init__(self, config: Blip2VisionConfig):
+ super().__init__(config)
+ self.config = config
+ embed_dim = config.hidden_size
+ self.embeddings = Blip2VisionEmbeddings(config)
+ self.pre_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+ self.encoder = Blip2Encoder(config)
+ self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+ self.post_init()
+
+ @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=Blip2VisionConfig)
+ def forward(
+ self,
+ pixel_values: Optional[torch.FloatTensor] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ ) -> Union[Tuple, BaseModelOutputWithPooling]:
+ r"""
+ Returns:
+
+ """
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+ output_hidden_states = (
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+ )
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+ if pixel_values is None:
+ raise ValueError("You have to specify pixel_values")
+
+ hidden_states = self.embeddings(pixel_values)
+ hidden_states = self.pre_layernorm(hidden_states)
+ encoder_outputs = self.encoder(
+ inputs_embeds=hidden_states,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+ last_hidden_state = encoder_outputs[0]
+ last_hidden_state = self.post_layernorm(last_hidden_state)
+
+ pooled_output = last_hidden_state[:, 0, :]
+ pooled_output = self.post_layernorm(pooled_output)
+
+ if not return_dict:
+ return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+ return BaseModelOutputWithPooling(
+ last_hidden_state=last_hidden_state,
+ pooler_output=pooled_output,
+ hidden_states=encoder_outputs.hidden_states,
+ attentions=encoder_outputs.attentions,
+ )
+
+ def get_input_embeddings(self):
+ return self.embeddings
+
+
+# Qformer model, used to get multimodal embeddings from the text and image inputs
+class Blip2QFormerModel(Blip2PreTrainedModel):
+ """
+ Querying Transformer (Q-Former), used in BLIP-2.
+ """
+
+ def __init__(self, config: Blip2Config):
+ super().__init__(config)
+ self.config = config
+ self.embeddings = Blip2TextEmbeddings(config.qformer_config)
+ self.visual_encoder = Blip2VisionModel(config.vision_config)
+ self.query_tokens = nn.Parameter(torch.zeros(1, config.num_query_tokens, config.qformer_config.hidden_size))
+ if not hasattr(config, "tokenizer") or config.tokenizer is None:
+ self.tokenizer = BertTokenizer.from_pretrained("bert-base-uncased", truncation_side="right")
+ else:
+ self.tokenizer = BertTokenizer.from_pretrained(config.tokenizer, truncation_side="right")
+ self.tokenizer.add_special_tokens({"bos_token": "[DEC]"})
+ self.proj_layer = ProjLayer(
+ in_dim=config.qformer_config.hidden_size,
+ out_dim=config.qformer_config.hidden_size,
+ hidden_dim=config.qformer_config.hidden_size * 4,
+ drop_p=0.1,
+ eps=1e-12,
+ )
+
+ self.encoder = Blip2QFormerEncoder(config.qformer_config)
+
+ self.post_init()
+
+ def get_input_embeddings(self):
+ return self.embeddings.word_embeddings
+
+ def set_input_embeddings(self, value):
+ self.embeddings.word_embeddings = value
+
+ def _prune_heads(self, heads_to_prune):
+ """
+ Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+ class PreTrainedModel
+ """
+ for layer, heads in heads_to_prune.items():
+ self.encoder.layer[layer].attention.prune_heads(heads)
+
+ def get_extended_attention_mask(
+ self,
+ attention_mask: torch.Tensor,
+ input_shape: Tuple[int],
+ device: torch.device,
+ has_query: bool = False,
+ ) -> torch.Tensor:
+ """
+ Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+ Arguments:
+ attention_mask (`torch.Tensor`):
+ Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+ input_shape (`Tuple[int]`):
+ The shape of the input to the model.
+ device (`torch.device`):
+ The device of the input to the model.
+
+ Returns:
+ `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
+ """
+ # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+ # ourselves in which case we just need to make it broadcastable to all heads.
+ if attention_mask.dim() == 3:
+ extended_attention_mask = attention_mask[:, None, :, :]
+ elif attention_mask.dim() == 2:
+ # Provided a padding mask of dimensions [batch_size, seq_length]
+ # - the model is an encoder, so make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+ extended_attention_mask = attention_mask[:, None, None, :]
+ else:
+ raise ValueError(
+ "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+ input_shape, attention_mask.shape
+ )
+ )
+
+ # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+ # masked positions, this operation will create a tensor which is 0.0 for
+ # positions we want to attend and -10000.0 for masked positions.
+ # Since we are adding it to the raw scores before the softmax, this is
+ # effectively the same as removing these entirely.
+ extended_attention_mask = extended_attention_mask.to(dtype=self.dtype) # fp16 compatibility
+ extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+ return extended_attention_mask
+
+ def forward(
+ self,
+ text_input=None,
+ image_input=None,
+ head_mask=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ past_key_values=None,
+ use_cache=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ ):
+ r"""
+ encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, `optional`):
+ Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+ the model is configured as a decoder.
+ encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, `optional`):
+ Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+ the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+ - 1 for tokens that are **not masked**,
+ - 0 for tokens that are **masked**.
+ past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of:
+ shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and
+ value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are
+ used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key
+ value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape
+ `(batch_size, sequence_length)`.
+ use_cache (`bool`, `optional`):
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+ `past_key_values`).
+ """
+
+ text = self.tokenizer(text_input, return_tensors="pt", padding=True)
+ text = text.to(self.device)
+ input_ids = text.input_ids
+ batch_size = input_ids.shape[0]
+ query_atts = torch.ones((batch_size, self.query_tokens.size()[1]), dtype=torch.long).to(self.device)
+ attention_mask = torch.cat([query_atts, text.attention_mask], dim=1)
+
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+ output_hidden_states = (
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+ )
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+ # past_key_values_length
+ past_key_values_length = (
+ past_key_values[0][0].shape[2] - self.config.query_length if past_key_values is not None else 0
+ )
+
+ query_length = self.query_tokens.shape[1]
+
+ embedding_output = self.embeddings(
+ input_ids=input_ids,
+ query_embeds=self.query_tokens,
+ past_key_values_length=past_key_values_length,
+ )
+
+ # embedding_output = self.layernorm(query_embeds)
+ # embedding_output = self.dropout(embedding_output)
+
+ input_shape = embedding_output.size()[:-1]
+ batch_size, seq_length = input_shape
+ device = embedding_output.device
+
+ image_embeds_frozen = self.visual_encoder(image_input).last_hidden_state
+ # image_embeds_frozen = torch.ones_like(image_embeds_frozen)
+ encoder_hidden_states = image_embeds_frozen
+
+ if attention_mask is None:
+ attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
+ # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+ # ourselves in which case we just need to make it broadcastable to all heads.
+ extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+ # If a 2D or 3D attention mask is provided for the cross-attention
+ # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+ if encoder_hidden_states is not None:
+ if isinstance(encoder_hidden_states, list):
+ encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size()
+ else:
+ encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+ encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+
+ if isinstance(encoder_attention_mask, list):
+ encoder_extended_attention_mask = [self.invert_attention_mask(mask) for mask in encoder_attention_mask]
+ elif encoder_attention_mask is None:
+ encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+ encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+ else:
+ encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+ else:
+ encoder_extended_attention_mask = None
+
+ # Prepare head mask if needed
+ # 1.0 in head_mask indicate we keep the head
+ # attention_probs has shape bsz x n_heads x N x N
+ # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+ # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+ head_mask = self.get_head_mask(head_mask, self.config.qformer_config.num_hidden_layers)
+
+ encoder_outputs = self.encoder(
+ embedding_output,
+ attention_mask=extended_attention_mask,
+ head_mask=head_mask,
+ encoder_hidden_states=encoder_hidden_states,
+ encoder_attention_mask=encoder_extended_attention_mask,
+ past_key_values=past_key_values,
+ use_cache=use_cache,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ query_length=query_length,
+ )
+ sequence_output = encoder_outputs[0]
+ pooled_output = sequence_output[:, 0, :]
+
+ if not return_dict:
+ return self.proj_layer(sequence_output[:, :query_length, :])
+
+ return BaseModelOutputWithPoolingAndCrossAttentions(
+ last_hidden_state=sequence_output,
+ pooler_output=pooled_output,
+ past_key_values=encoder_outputs.past_key_values,
+ hidden_states=encoder_outputs.hidden_states,
+ attentions=encoder_outputs.attentions,
+ cross_attentions=encoder_outputs.cross_attentions,
+ )
diff --git a/model/diffusers_c/pipelines/blip_diffusion/modeling_ctx_clip.py b/model/diffusers_c/pipelines/blip_diffusion/modeling_ctx_clip.py
new file mode 100644
index 0000000000000000000000000000000000000000..c6772fc8880708154dd802cda78c5e0664d2426d
--- /dev/null
+++ b/model/diffusers_c/pipelines/blip_diffusion/modeling_ctx_clip.py
@@ -0,0 +1,223 @@
+# Copyright 2024 Salesforce.com, inc.
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional, Tuple, Union
+
+import torch
+from torch import nn
+from transformers import CLIPPreTrainedModel
+from transformers.modeling_outputs import BaseModelOutputWithPooling
+from transformers.models.clip.configuration_clip import CLIPTextConfig
+from transformers.models.clip.modeling_clip import CLIPEncoder
+
+
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+ """
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+ """
+ bsz, src_len = mask.size()
+ tgt_len = tgt_len if tgt_len is not None else src_len
+
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+ inverted_mask = 1.0 - expanded_mask
+
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+# This is a modified version of the CLIPTextModel from transformers.models.clip.modeling_clip
+# Which allows for an extra input of "context embeddings", which are the query embeddings used in Qformer
+# They pass through the clip model, along with the text embeddings, and interact with them using self attention
+class ContextCLIPTextModel(CLIPPreTrainedModel):
+ config_class = CLIPTextConfig
+
+ _no_split_modules = ["CLIPEncoderLayer"]
+
+ def __init__(self, config: CLIPTextConfig):
+ super().__init__(config)
+ self.text_model = ContextCLIPTextTransformer(config)
+ # Initialize weights and apply final processing
+ self.post_init()
+
+ def forward(
+ self,
+ ctx_embeddings: torch.Tensor = None,
+ ctx_begin_pos: list = None,
+ input_ids: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.Tensor] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ ) -> Union[Tuple, BaseModelOutputWithPooling]:
+ return self.text_model(
+ ctx_embeddings=ctx_embeddings,
+ ctx_begin_pos=ctx_begin_pos,
+ input_ids=input_ids,
+ attention_mask=attention_mask,
+ position_ids=position_ids,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+
+class ContextCLIPTextTransformer(nn.Module):
+ def __init__(self, config: CLIPTextConfig):
+ super().__init__()
+ self.config = config
+ embed_dim = config.hidden_size
+ self.embeddings = ContextCLIPTextEmbeddings(config)
+ self.encoder = CLIPEncoder(config)
+ self.final_layer_norm = nn.LayerNorm(embed_dim)
+
+ def forward(
+ self,
+ ctx_embeddings: torch.Tensor,
+ ctx_begin_pos: list,
+ input_ids: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.Tensor] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ ) -> Union[Tuple, BaseModelOutputWithPooling]:
+ r"""
+ Returns:
+
+ """
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+ output_hidden_states = (
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+ )
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+ if input_ids is None:
+ raise ValueError("You have to specify either input_ids")
+
+ input_shape = input_ids.size()
+ input_ids = input_ids.view(-1, input_shape[-1])
+
+ hidden_states = self.embeddings(
+ input_ids=input_ids,
+ position_ids=position_ids,
+ ctx_embeddings=ctx_embeddings,
+ ctx_begin_pos=ctx_begin_pos,
+ )
+
+ bsz, seq_len = input_shape
+ if ctx_embeddings is not None:
+ seq_len += ctx_embeddings.size(1)
+ # CLIP's text model uses causal mask, prepare it here.
+ # https://github.com/openai/CLIP/blob/cfcffb90e69f37bf2ff1e988237a0fbe41f33c04/clip/model.py#L324
+ causal_attention_mask = self._build_causal_attention_mask(bsz, seq_len, hidden_states.dtype).to(
+ hidden_states.device
+ )
+ # expand attention_mask
+ if attention_mask is not None:
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+ attention_mask = _expand_mask(attention_mask, hidden_states.dtype)
+
+ encoder_outputs = self.encoder(
+ inputs_embeds=hidden_states,
+ attention_mask=attention_mask,
+ causal_attention_mask=causal_attention_mask,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ last_hidden_state = encoder_outputs[0]
+ last_hidden_state = self.final_layer_norm(last_hidden_state)
+
+ # text_embeds.shape = [batch_size, sequence_length, transformer.width]
+ # take features from the eot embedding (eot_token is the highest number in each sequence)
+ # casting to torch.int for onnx compatibility: argmax doesn't support int64 inputs with opset 14
+ pooled_output = last_hidden_state[
+ torch.arange(last_hidden_state.shape[0], device=input_ids.device),
+ input_ids.to(torch.int).argmax(dim=-1),
+ ]
+
+ if not return_dict:
+ return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+ return BaseModelOutputWithPooling(
+ last_hidden_state=last_hidden_state,
+ pooler_output=pooled_output,
+ hidden_states=encoder_outputs.hidden_states,
+ attentions=encoder_outputs.attentions,
+ )
+
+ def _build_causal_attention_mask(self, bsz, seq_len, dtype):
+ # lazily create causal attention mask, with full attention between the vision tokens
+ # pytorch uses additive attention mask; fill with -inf
+ mask = torch.empty(bsz, seq_len, seq_len, dtype=dtype)
+ mask.fill_(torch.tensor(torch.finfo(dtype).min))
+ mask.triu_(1) # zero out the lower diagonal
+ mask = mask.unsqueeze(1) # expand mask
+ return mask
+
+
+class ContextCLIPTextEmbeddings(nn.Module):
+ def __init__(self, config: CLIPTextConfig):
+ super().__init__()
+ embed_dim = config.hidden_size
+
+ self.token_embedding = nn.Embedding(config.vocab_size, embed_dim)
+ self.position_embedding = nn.Embedding(config.max_position_embeddings, embed_dim)
+
+ # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+ self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+
+ def forward(
+ self,
+ ctx_embeddings: torch.Tensor,
+ ctx_begin_pos: list,
+ input_ids: Optional[torch.LongTensor] = None,
+ position_ids: Optional[torch.LongTensor] = None,
+ inputs_embeds: Optional[torch.FloatTensor] = None,
+ ) -> torch.Tensor:
+ if ctx_embeddings is None:
+ ctx_len = 0
+ else:
+ ctx_len = ctx_embeddings.shape[1]
+
+ seq_length = (input_ids.shape[-1] if input_ids is not None else inputs_embeds.shape[-2]) + ctx_len
+
+ if position_ids is None:
+ position_ids = self.position_ids[:, :seq_length]
+
+ if inputs_embeds is None:
+ inputs_embeds = self.token_embedding(input_ids)
+
+ # for each input embeddings, add the ctx embeddings at the correct position
+ input_embeds_ctx = []
+ bsz = inputs_embeds.shape[0]
+
+ if ctx_embeddings is not None:
+ for i in range(bsz):
+ cbp = ctx_begin_pos[i]
+
+ prefix = inputs_embeds[i, :cbp]
+ # remove the special token embedding
+ suffix = inputs_embeds[i, cbp:]
+
+ input_embeds_ctx.append(torch.cat([prefix, ctx_embeddings[i], suffix], dim=0))
+
+ inputs_embeds = torch.stack(input_embeds_ctx, dim=0)
+
+ position_embeddings = self.position_embedding(position_ids)
+ embeddings = inputs_embeds + position_embeddings
+
+ return embeddings
diff --git a/model/diffusers_c/pipelines/blip_diffusion/pipeline_blip_diffusion.py b/model/diffusers_c/pipelines/blip_diffusion/pipeline_blip_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..ba43b2e53d2d21e48fd5708321c6f422ba47538d
--- /dev/null
+++ b/model/diffusers_c/pipelines/blip_diffusion/pipeline_blip_diffusion.py
@@ -0,0 +1,348 @@
+# Copyright 2024 Salesforce.com, inc.
+# Copyright 2024 The HuggingFace Team. All rights reserved.#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPTokenizer
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...schedulers import PNDMScheduler
+from ...utils import (
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .blip_image_processing import BlipImageProcessor
+from .modeling_blip2 import Blip2QFormerModel
+from .modeling_ctx_clip import ContextCLIPTextModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers.pipelines import BlipDiffusionPipeline
+ >>> from diffusers.utils import load_image
+ >>> import torch
+
+ >>> blip_diffusion_pipe = BlipDiffusionPipeline.from_pretrained(
+ ... "Salesforce/blipdiffusion", torch_dtype=torch.float16
+ ... ).to("cuda")
+
+
+ >>> cond_subject = "dog"
+ >>> tgt_subject = "dog"
+ >>> text_prompt_input = "swimming underwater"
+
+ >>> cond_image = load_image(
+ ... "https://huggingface.co/datasets/ayushtues/blipdiffusion_images/resolve/main/dog.jpg"
+ ... )
+ >>> guidance_scale = 7.5
+ >>> num_inference_steps = 25
+ >>> negative_prompt = "over-exposure, under-exposure, saturated, duplicate, out of frame, lowres, cropped, worst quality, low quality, jpeg artifacts, morbid, mutilated, out of frame, ugly, bad anatomy, bad proportions, deformed, blurry, duplicate"
+
+
+ >>> output = blip_diffusion_pipe(
+ ... text_prompt_input,
+ ... cond_image,
+ ... cond_subject,
+ ... tgt_subject,
+ ... guidance_scale=guidance_scale,
+ ... num_inference_steps=num_inference_steps,
+ ... neg_prompt=negative_prompt,
+ ... height=512,
+ ... width=512,
+ ... ).images
+ >>> output[0].save("image.png")
+ ```
+"""
+
+
+class BlipDiffusionPipeline(DiffusionPipeline):
+ """
+ Pipeline for Zero-Shot Subject Driven Generation using Blip Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ tokenizer ([`CLIPTokenizer`]):
+ Tokenizer for the text encoder
+ text_encoder ([`ContextCLIPTextModel`]):
+ Text encoder to encode the text prompt
+ vae ([`AutoencoderKL`]):
+ VAE model to map the latents to the image
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ scheduler ([`PNDMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ qformer ([`Blip2QFormerModel`]):
+ QFormer model to get multi-modal embeddings from the text and image.
+ image_processor ([`BlipImageProcessor`]):
+ Image Processor to preprocess and postprocess the image.
+ ctx_begin_pos (int, `optional`, defaults to 2):
+ Position of the context token in the text encoder.
+ """
+
+ model_cpu_offload_seq = "qformer->text_encoder->unet->vae"
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ text_encoder: ContextCLIPTextModel,
+ vae: AutoencoderKL,
+ unet: UNet2DConditionModel,
+ scheduler: PNDMScheduler,
+ qformer: Blip2QFormerModel,
+ image_processor: BlipImageProcessor,
+ ctx_begin_pos: int = 2,
+ mean: List[float] = None,
+ std: List[float] = None,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ vae=vae,
+ unet=unet,
+ scheduler=scheduler,
+ qformer=qformer,
+ image_processor=image_processor,
+ )
+ self.register_to_config(ctx_begin_pos=ctx_begin_pos, mean=mean, std=std)
+
+ def get_query_embeddings(self, input_image, src_subject):
+ return self.qformer(image_input=input_image, text_input=src_subject, return_dict=False)
+
+ # from the original Blip Diffusion code, speciefies the target subject and augments the prompt by repeating it
+ def _build_prompt(self, prompts, tgt_subjects, prompt_strength=1.0, prompt_reps=20):
+ rv = []
+ for prompt, tgt_subject in zip(prompts, tgt_subjects):
+ prompt = f"a {tgt_subject} {prompt.strip()}"
+ # a trick to amplify the prompt
+ rv.append(", ".join([prompt] * int(prompt_strength * prompt_reps)))
+
+ return rv
+
+ # Copied from diffusers.pipelines.consistency_models.pipeline_consistency_models.ConsistencyModelPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels, height, width)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device=device, dtype=dtype)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def encode_prompt(self, query_embeds, prompt, device=None):
+ device = device or self._execution_device
+
+ # embeddings for prompt, with query_embeds as context
+ max_len = self.text_encoder.text_model.config.max_position_embeddings
+ max_len -= self.qformer.config.num_query_tokens
+
+ tokenized_prompt = self.tokenizer(
+ prompt,
+ padding="max_length",
+ truncation=True,
+ max_length=max_len,
+ return_tensors="pt",
+ ).to(device)
+
+ batch_size = query_embeds.shape[0]
+ ctx_begin_pos = [self.config.ctx_begin_pos] * batch_size
+
+ text_embeddings = self.text_encoder(
+ input_ids=tokenized_prompt.input_ids,
+ ctx_embeddings=query_embeds,
+ ctx_begin_pos=ctx_begin_pos,
+ )[0]
+
+ return text_embeddings
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: List[str],
+ reference_image: PIL.Image.Image,
+ source_subject_category: List[str],
+ target_subject_category: List[str],
+ latents: Optional[torch.FloatTensor] = None,
+ guidance_scale: float = 7.5,
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 50,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ neg_prompt: Optional[str] = "",
+ prompt_strength: float = 1.0,
+ prompt_reps: int = 20,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`List[str]`):
+ The prompt or prompts to guide the image generation.
+ reference_image (`PIL.Image.Image`):
+ The reference image to condition the generation on.
+ source_subject_category (`List[str]`):
+ The source subject category.
+ target_subject_category (`List[str]`):
+ The target subject category.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by random sampling.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ height (`int`, *optional*, defaults to 512):
+ The height of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ neg_prompt (`str`, *optional*, defaults to ""):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_strength (`float`, *optional*, defaults to 1.0):
+ The strength of the prompt. Specifies the number of times the prompt is repeated along with prompt_reps
+ to amplify the prompt.
+ prompt_reps (`int`, *optional*, defaults to 20):
+ The number of times the prompt is repeated along with prompt_strength to amplify the prompt.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ device = self._execution_device
+
+ reference_image = self.image_processor.preprocess(
+ reference_image, image_mean=self.config.mean, image_std=self.config.std, return_tensors="pt"
+ )["pixel_values"]
+ reference_image = reference_image.to(device)
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ if isinstance(source_subject_category, str):
+ source_subject_category = [source_subject_category]
+ if isinstance(target_subject_category, str):
+ target_subject_category = [target_subject_category]
+
+ batch_size = len(prompt)
+
+ prompt = self._build_prompt(
+ prompts=prompt,
+ tgt_subjects=target_subject_category,
+ prompt_strength=prompt_strength,
+ prompt_reps=prompt_reps,
+ )
+ query_embeds = self.get_query_embeddings(reference_image, source_subject_category)
+ text_embeddings = self.encode_prompt(query_embeds, prompt, device)
+ do_classifier_free_guidance = guidance_scale > 1.0
+ if do_classifier_free_guidance:
+ max_length = self.text_encoder.text_model.config.max_position_embeddings
+
+ uncond_input = self.tokenizer(
+ [neg_prompt] * batch_size,
+ padding="max_length",
+ max_length=max_length,
+ return_tensors="pt",
+ )
+ uncond_embeddings = self.text_encoder(
+ input_ids=uncond_input.input_ids.to(device),
+ ctx_embeddings=None,
+ )[0]
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+
+ scale_down_factor = 2 ** (len(self.unet.config.block_out_channels) - 1)
+ latents = self.prepare_latents(
+ batch_size=batch_size,
+ num_channels=self.unet.config.in_channels,
+ height=height // scale_down_factor,
+ width=width // scale_down_factor,
+ generator=generator,
+ latents=latents,
+ dtype=self.unet.dtype,
+ device=device,
+ )
+ # set timesteps
+ extra_set_kwargs = {}
+ self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)
+
+ for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ noise_pred = self.unet(
+ latent_model_input,
+ timestep=t,
+ encoder_hidden_states=text_embeddings,
+ down_block_additional_residuals=None,
+ mid_block_additional_residual=None,
+ )["sample"]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ )["prev_sample"]
+
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/brushnet/__init__.py b/model/diffusers_c/pipelines/brushnet/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..f099981e42c12c50a46f639a8901759ef8f29bcd
--- /dev/null
+++ b/model/diffusers_c/pipelines/brushnet/__init__.py
@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_flax_available,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_brushnet"] = ["StableDiffusionBrushNetPipeline"]
+ _import_structure["pipeline_brushnet_powerpaint"] = ["StableDiffusionBrushNetPowerPaintPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_brushnet import StableDiffusionBrushNetPipeline
+ from .pipeline_brushnet_powerpaint import StableDiffusionBrushNetPowerPaintPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/brushnet/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/brushnet/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..587f8bd13288a9da0b137ccc764ac08c5d8d01d0
Binary files /dev/null and b/model/diffusers_c/pipelines/brushnet/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/brushnet/__pycache__/pipeline_brushnet.cpython-39.pyc b/model/diffusers_c/pipelines/brushnet/__pycache__/pipeline_brushnet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..51bbf8b0c6c82a17149d657b61d835d9e2438411
Binary files /dev/null and b/model/diffusers_c/pipelines/brushnet/__pycache__/pipeline_brushnet.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/brushnet/__pycache__/pipeline_brushnet_powerpaint.cpython-39.pyc b/model/diffusers_c/pipelines/brushnet/__pycache__/pipeline_brushnet_powerpaint.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3dbb00e269050e91118753ca9e1200da6d696189
Binary files /dev/null and b/model/diffusers_c/pipelines/brushnet/__pycache__/pipeline_brushnet_powerpaint.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/brushnet/pipeline_brushnet.py b/model/diffusers_c/pipelines/brushnet/pipeline_brushnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..ffc1744e2f99c15fe6225c492ce9e088458b1190
--- /dev/null
+++ b/model/diffusers_c/pipelines/brushnet/pipeline_brushnet.py
@@ -0,0 +1,1274 @@
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, BrushNetModel, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import StableDiffusionBrushNetPipeline, BrushNetModel, UniPCMultistepScheduler
+ from diffusers.utils import load_image
+ import torch
+ import cv2
+ import numpy as np
+ from PIL import Image
+
+ base_model_path = "runwayml/stable-diffusion-v1-5"
+ brushnet_path = "ckpt_path"
+
+ brushnet = BrushNetModel.from_pretrained(brushnet_path, torch_dtype=torch.float16)
+ pipe = StableDiffusionBrushNetPipeline.from_pretrained(
+ base_model_path, brushnet=brushnet, torch_dtype=torch.float16, low_cpu_mem_usage=False
+ )
+
+ # speed up diffusion process with faster scheduler and memory optimization
+ pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+ # remove following line if xformers is not installed or when using Torch 2.0.
+ # pipe.enable_xformers_memory_efficient_attention()
+ # memory optimization.
+ pipe.enable_model_cpu_offload()
+
+ image_path="examples/brushnet/src/test_image.jpg"
+ mask_path="examples/brushnet/src/test_mask.jpg"
+ caption="A cake on the table."
+
+ init_image = cv2.imread(image_path)
+ mask_image = 1.*(cv2.imread(mask_path).sum(-1)>255)[:,:,np.newaxis]
+ init_image = init_image * (1-mask_image)
+
+ init_image = Image.fromarray(init_image.astype(np.uint8)).convert("RGB")
+ mask_image = Image.fromarray(mask_image.astype(np.uint8).repeat(3,-1)*255).convert("RGB")
+
+ generator = torch.Generator("cuda").manual_seed(1234)
+
+ image = pipe(
+ caption,
+ init_image,
+ mask_image,
+ num_inference_steps=50,
+ generator=generator,
+ paintingnet_conditioning_scale=1.0
+ ).images[0]
+ image.save("output.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionBrushNetPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with BrushNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ brushnet ([`BrushNetModel`]`):
+ Provides additional conditioning to the `unet` during the denoising process.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ brushnet: BrushNetModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ brushnet=brushnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ mask,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ brushnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.brushnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.brushnet, BrushNetModel)
+ or is_compiled
+ and isinstance(self.brushnet._orig_mod, BrushNetModel)
+ ):
+ self.check_image(image, mask, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `brushnet_conditioning_scale`
+ if (
+ isinstance(self.brushnet, BrushNetModel)
+ or is_compiled
+ and isinstance(self.brushnet._orig_mod, BrushNetModel)
+ ):
+ if not isinstance(brushnet_conditioning_scale, float):
+ raise TypeError("For single brushnet: `brushnet_conditioning_scale` must be type `float`.")
+ else:
+ assert False
+
+ if not isinstance(control_guidance_start, (tuple, list)):
+ control_guidance_start = [control_guidance_start]
+
+ if not isinstance(control_guidance_end, (tuple, list)):
+ control_guidance_end = [control_guidance_end]
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def check_image(self, image, mask, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ mask_is_pil = isinstance(mask, PIL.Image.Image)
+ mask_is_tensor = isinstance(mask, torch.Tensor)
+ mask_is_np = isinstance(mask, np.ndarray)
+ mask_is_pil_list = isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image)
+ mask_is_tensor_list = isinstance(mask, list) and isinstance(mask[0], torch.Tensor)
+ mask_is_np_list = isinstance(mask, list) and isinstance(mask[0], np.ndarray)
+
+ if (
+ not mask_is_pil
+ and not mask_is_tensor
+ and not mask_is_np
+ and not mask_is_pil_list
+ and not mask_is_tensor_list
+ and not mask_is_np_list
+ ):
+ raise TypeError(
+ f"mask must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(mask)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ def prepare_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ noise = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = noise * self.scheduler.init_noise_sigma
+ return latents, noise
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ mask: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ brushnet_conditioning_scale: Union[float, List[float]] = 1.0,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The BrushNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to BrushNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple BrushNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single BrushNet. When `prompt` is a list, and if a list of images is passed for a single BrushNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple BrushNets,
+ where a list of image lists can be passed to batch for each prompt and each BrushNet.
+ mask (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The BrushNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to BrushNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple BrushNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single BrushNet. When `prompt` is a list, and if a list of images is passed for a single BrushNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple BrushNets,
+ where a list of image lists can be passed to batch for each prompt and each BrushNet.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ brushnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the BrushNet are multiplied by `brushnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple BrushNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ The BrushNet encoder tries to recognize the content of the input image even if you remove all
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the BrushNet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the BrushNet stops applying.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ brushnet = self.brushnet._orig_mod if is_compiled_module(self.brushnet) else self.brushnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ control_guidance_start, control_guidance_end = (
+ [control_guidance_start],
+ [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ image,
+ mask,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ brushnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ global_pool_conditions = (
+ brushnet.config.global_pool_conditions
+ if isinstance(brushnet, BrushNetModel)
+ else brushnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image
+ if isinstance(brushnet, BrushNetModel):
+ image = self.prepare_image(
+ image=image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=brushnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ original_mask = self.prepare_image(
+ image=mask,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=brushnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ original_mask = (original_mask.sum(1)[:, None, :, :] < 0).to(image.dtype)
+ height, width = image.shape[-2:]
+ else:
+ assert False
+
+ # 5. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents, noise = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6.1 prepare condition latents
+ conditioning_latents = self.vae.encode(image).latent_dist.sample() * self.vae.config.scaling_factor
+ mask = torch.nn.functional.interpolate(
+ original_mask, size=(conditioning_latents.shape[-2], conditioning_latents.shape[-1])
+ )
+ conditioning_latents = torch.concat([conditioning_latents, mask], 1)
+
+ # 6.5 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 7.2 Create tensor stating which brushnets to keep
+ brushnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ brushnet_keep.append(keeps[0] if isinstance(brushnet, BrushNetModel) else keeps)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ is_unet_compiled = is_compiled_module(self.unet)
+ is_brushnet_compiled = is_compiled_module(self.brushnet)
+ is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Relevant thread:
+ # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+ if (is_unet_compiled and is_brushnet_compiled) and is_torch_higher_equal_2_1:
+ torch._inductor.cudagraph_mark_step_begin()
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # brushnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer BrushNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ brushnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ brushnet_prompt_embeds = prompt_embeds
+
+ if isinstance(brushnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(brushnet_conditioning_scale, brushnet_keep[i])]
+ else:
+ brushnet_cond_scale = brushnet_conditioning_scale
+ if isinstance(brushnet_cond_scale, list):
+ brushnet_cond_scale = brushnet_cond_scale[0]
+ cond_scale = brushnet_cond_scale * brushnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample, up_block_res_samples = self.brushnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=brushnet_prompt_embeds,
+ brushnet_cond=conditioning_latents,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered BrushNet only for the conditional batch.
+ # To apply the output of BrushNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+ up_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in up_block_res_samples]
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_add_samples=down_block_res_samples,
+ mid_block_add_sample=mid_block_res_sample,
+ up_block_add_samples=up_block_res_samples,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and brushnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.brushnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/brushnet/pipeline_brushnet_powerpaint.py b/model/diffusers_c/pipelines/brushnet/pipeline_brushnet_powerpaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..2f03fe1826f91f1a7c547ed4e464308d42cfd7a3
--- /dev/null
+++ b/model/diffusers_c/pipelines/brushnet/pipeline_brushnet_powerpaint.py
@@ -0,0 +1,1310 @@
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, BrushNetModel, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import StableDiffusionBrushNetPipeline, BrushNetModel, UniPCMultistepScheduler
+ from diffusers.utils import load_image
+ import torch
+ import cv2
+ import numpy as np
+ from PIL import Image
+
+ base_model_path = "runwayml/stable-diffusion-v1-5"
+ brushnet_path = "ckpt_path"
+
+ brushnet = BrushNetModel.from_pretrained(brushnet_path, torch_dtype=torch.float16)
+ pipe = StableDiffusionBrushNetPipeline.from_pretrained(
+ base_model_path, brushnet=brushnet, torch_dtype=torch.float16, low_cpu_mem_usage=False
+ )
+
+ # speed up diffusion process with faster scheduler and memory optimization
+ pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+ # remove following line if xformers is not installed or when using Torch 2.0.
+ # pipe.enable_xformers_memory_efficient_attention()
+ # memory optimization.
+ pipe.enable_model_cpu_offload()
+
+ image_path="examples/brushnet/src/test_image.jpg"
+ mask_path="examples/brushnet/src/test_mask.jpg"
+ caption="A cake on the table."
+
+ init_image = cv2.imread(image_path)
+ mask_image = 1.*(cv2.imread(mask_path).sum(-1)>255)[:,:,np.newaxis]
+ init_image = init_image * (1-mask_image)
+
+ init_image = Image.fromarray(init_image.astype(np.uint8)).convert("RGB")
+ mask_image = Image.fromarray(mask_image.astype(np.uint8).repeat(3,-1)*255).convert("RGB")
+
+ generator = torch.Generator("cuda").manual_seed(1234)
+
+ image = pipe(
+ caption,
+ init_image,
+ mask_image,
+ num_inference_steps=50,
+ generator=generator,
+ paintingnet_conditioning_scale=1.0
+ ).images[0]
+ image.save("output.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionBrushNetPowerPaintPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with BrushNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ brushnet ([`BrushNetModel`]`):
+ Provides additional conditioning to the `unet` during the denoising process.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ brushnet: BrushNetModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ brushnet=brushnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ print("1 ", prompt, negative_prompt)
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+ print("2 ", prompt, negative_prompt)
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+ print("3 ", prompt, negative_prompt)
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ print("4 ", prompt, negative_prompt)
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ print("5 ", prompt, negative_prompt)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ print(prompt, text_input_ids)
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ print("neg: ", uncond_input.input_ids)
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ mask,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ brushnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.brushnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.brushnet, BrushNetModel)
+ or is_compiled
+ and isinstance(self.brushnet._orig_mod, BrushNetModel)
+ ):
+ self.check_image(image, mask, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `brushnet_conditioning_scale`
+ if (
+ isinstance(self.brushnet, BrushNetModel)
+ or is_compiled
+ and isinstance(self.brushnet._orig_mod, BrushNetModel)
+ ):
+ if not isinstance(brushnet_conditioning_scale, float):
+ raise TypeError("For single brushnet: `brushnet_conditioning_scale` must be type `float`.")
+ else:
+ assert False
+
+ if not isinstance(control_guidance_start, (tuple, list)):
+ control_guidance_start = [control_guidance_start]
+
+ if not isinstance(control_guidance_end, (tuple, list)):
+ control_guidance_end = [control_guidance_end]
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def check_image(self, image, mask, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ mask_is_pil = isinstance(mask, PIL.Image.Image)
+ mask_is_tensor = isinstance(mask, torch.Tensor)
+ mask_is_np = isinstance(mask, np.ndarray)
+ mask_is_pil_list = isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image)
+ mask_is_tensor_list = isinstance(mask, list) and isinstance(mask[0], torch.Tensor)
+ mask_is_np_list = isinstance(mask, list) and isinstance(mask[0], np.ndarray)
+
+ if (
+ not mask_is_pil
+ and not mask_is_tensor
+ and not mask_is_np
+ and not mask_is_pil_list
+ and not mask_is_tensor_list
+ and not mask_is_np_list
+ ):
+ raise TypeError(
+ f"mask must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(mask)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ def prepare_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ noise = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = noise * self.scheduler.init_noise_sigma
+ return latents, noise
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ mask: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ brushnet_conditioning_scale: Union[float, List[float]] = 1.0,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ con_prompt=None,
+ con_negprompt=None,
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The BrushNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to BrushNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple BrushNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single BrushNet. When `prompt` is a list, and if a list of images is passed for a single BrushNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple BrushNets,
+ where a list of image lists can be passed to batch for each prompt and each BrushNet.
+ mask (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The BrushNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to BrushNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple BrushNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single BrushNet. When `prompt` is a list, and if a list of images is passed for a single BrushNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple BrushNets,
+ where a list of image lists can be passed to batch for each prompt and each BrushNet.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ brushnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the BrushNet are multiplied by `brushnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple BrushNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ The BrushNet encoder tries to recognize the content of the input image even if you remove all
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the BrushNet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the BrushNet stops applying.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ brushnet = self.brushnet._orig_mod if is_compiled_module(self.brushnet) else self.brushnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ control_guidance_start, control_guidance_end = (
+ [control_guidance_start],
+ [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ image,
+ mask,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ brushnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ global_pool_conditions = (
+ brushnet.config.global_pool_conditions
+ if isinstance(brushnet, BrushNetModel)
+ else brushnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ print(con_prompt, con_negprompt)
+ con_negative_prompt_embeds = None
+ con_prompt_embeds = None
+ con_prompt_embeds, con_negative_prompt_embeds = self.encode_prompt(
+ con_prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ con_negprompt,
+ prompt_embeds=con_prompt_embeds,
+ negative_prompt_embeds=con_negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ con_prompt_embeds = torch.cat([con_negative_prompt_embeds, con_prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image
+ if isinstance(brushnet, BrushNetModel):
+ image = self.prepare_image(
+ image=image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=brushnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ original_mask = self.prepare_image(
+ image=mask,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=brushnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ original_mask = (original_mask.sum(1)[:, None, :, :] < 0).to(image.dtype)
+ height, width = image.shape[-2:]
+ else:
+ assert False
+
+ # 5. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents, noise = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6.1 prepare condition latents
+ from torchvision import transforms
+ # mask_i = transforms.ToPILImage()(image[0:1,:,:,:].squeeze(0))
+ # mask_i.save('_mask.png')
+
+ print(image)
+ mask_i = transforms.ToPILImage()(image[0:1, :, :, :].squeeze(0) / 2 + 0.5)
+ mask_i.save(str(timesteps[0]) + "_CC.png")
+ conditioning_latents = self.vae.encode(image).latent_dist.sample() * self.vae.config.scaling_factor
+ mask = torch.nn.functional.interpolate(
+ original_mask, size=(conditioning_latents.shape[-2], conditioning_latents.shape[-1])
+ )
+ conditioning_latents = torch.concat([conditioning_latents, mask], 1)
+ # image = self.vae.decode(conditioning_latents[:1,:4,:,:] / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]
+ # from torchvision import transforms
+ # mask_i = transforms.ToPILImage()(image[0:1,:,:,:].squeeze(0)/2+0.5)
+ # mask_i.save(str(timesteps[0]) +'_C.png')
+
+ # 6.5 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 7.2 Create tensor stating which brushnets to keep
+ brushnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ brushnet_keep.append(keeps[0] if isinstance(brushnet, BrushNetModel) else keeps)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ is_unet_compiled = is_compiled_module(self.unet)
+ is_brushnet_compiled = is_compiled_module(self.brushnet)
+ is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Relevant thread:
+ # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+ if (is_unet_compiled and is_brushnet_compiled) and is_torch_higher_equal_2_1:
+ torch._inductor.cudagraph_mark_step_begin()
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # brushnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer BrushNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ brushnet_prompt_embeds = con_prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ brushnet_prompt_embeds = con_prompt_embeds
+
+ if isinstance(brushnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(brushnet_conditioning_scale, brushnet_keep[i])]
+ else:
+ brushnet_cond_scale = brushnet_conditioning_scale
+ if isinstance(brushnet_cond_scale, list):
+ brushnet_cond_scale = brushnet_cond_scale[0]
+ cond_scale = brushnet_cond_scale * brushnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample, up_block_res_samples = self.brushnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=brushnet_prompt_embeds,
+ brushnet_cond=conditioning_latents,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered BrushNet only for the conditional batch.
+ # To apply the output of BrushNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+ up_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in up_block_res_samples]
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_add_samples=down_block_res_samples,
+ mid_block_add_sample=mid_block_res_sample,
+ up_block_add_samples=up_block_res_samples,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and brushnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.brushnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/consistency_models/__init__.py b/model/diffusers_c/pipelines/consistency_models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..162d91c010acf95aa2daf87c51ab1e0c68361fd5
--- /dev/null
+++ b/model/diffusers_c/pipelines/consistency_models/__init__.py
@@ -0,0 +1,24 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ _LazyModule,
+)
+
+
+_import_structure = {
+ "pipeline_consistency_models": ["ConsistencyModelPipeline"],
+}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_consistency_models import ConsistencyModelPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/consistency_models/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/consistency_models/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..82d47baa6ed5cb1e86268019abb4d23ee5562725
Binary files /dev/null and b/model/diffusers_c/pipelines/consistency_models/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/consistency_models/__pycache__/pipeline_consistency_models.cpython-39.pyc b/model/diffusers_c/pipelines/consistency_models/__pycache__/pipeline_consistency_models.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..04677c06cf5d3d82e0ceea7fde4aa2257f8f1ac0
Binary files /dev/null and b/model/diffusers_c/pipelines/consistency_models/__pycache__/pipeline_consistency_models.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/consistency_models/pipeline_consistency_models.py b/model/diffusers_c/pipelines/consistency_models/pipeline_consistency_models.py
new file mode 100644
index 0000000000000000000000000000000000000000..befac79c63f3f621c7a475099f0e0119b4adffd7
--- /dev/null
+++ b/model/diffusers_c/pipelines/consistency_models/pipeline_consistency_models.py
@@ -0,0 +1,275 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, List, Optional, Union
+
+import torch
+
+from ...models import UNet2DModel
+from ...schedulers import CMStochasticIterativeScheduler
+from ...utils import (
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+
+ >>> from diffusers import ConsistencyModelPipeline
+
+ >>> device = "cuda"
+ >>> # Load the cd_imagenet64_l2 checkpoint.
+ >>> model_id_or_path = "openai/diffusers-cd_imagenet64_l2"
+ >>> pipe = ConsistencyModelPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
+ >>> pipe.to(device)
+
+ >>> # Onestep Sampling
+ >>> image = pipe(num_inference_steps=1).images[0]
+ >>> image.save("cd_imagenet64_l2_onestep_sample.png")
+
+ >>> # Onestep sampling, class-conditional image generation
+ >>> # ImageNet-64 class label 145 corresponds to king penguins
+ >>> image = pipe(num_inference_steps=1, class_labels=145).images[0]
+ >>> image.save("cd_imagenet64_l2_onestep_sample_penguin.png")
+
+ >>> # Multistep sampling, class-conditional image generation
+ >>> # Timesteps can be explicitly specified; the particular timesteps below are from the original Github repo:
+ >>> # https://github.com/openai/consistency_models/blob/main/scripts/launch.sh#L77
+ >>> image = pipe(num_inference_steps=None, timesteps=[22, 0], class_labels=145).images[0]
+ >>> image.save("cd_imagenet64_l2_multistep_sample_penguin.png")
+ ```
+"""
+
+
+class ConsistencyModelPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for unconditional or class-conditional image generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Currently only
+ compatible with [`CMStochasticIterativeScheduler`].
+ """
+
+ model_cpu_offload_seq = "unet"
+
+ def __init__(self, unet: UNet2DModel, scheduler: CMStochasticIterativeScheduler) -> None:
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ )
+
+ self.safety_checker = None
+
+ def prepare_latents(self, batch_size, num_channels, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels, height, width)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device=device, dtype=dtype)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Follows diffusers.VaeImageProcessor.postprocess
+ def postprocess_image(self, sample: torch.FloatTensor, output_type: str = "pil"):
+ if output_type not in ["pt", "np", "pil"]:
+ raise ValueError(
+ f"output_type={output_type} is not supported. Make sure to choose one of ['pt', 'np', or 'pil']"
+ )
+
+ # Equivalent to diffusers.VaeImageProcessor.denormalize
+ sample = (sample / 2 + 0.5).clamp(0, 1)
+ if output_type == "pt":
+ return sample
+
+ # Equivalent to diffusers.VaeImageProcessor.pt_to_numpy
+ sample = sample.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "np":
+ return sample
+
+ # Output_type must be 'pil'
+ sample = self.numpy_to_pil(sample)
+ return sample
+
+ def prepare_class_labels(self, batch_size, device, class_labels=None):
+ if self.unet.config.num_class_embeds is not None:
+ if isinstance(class_labels, list):
+ class_labels = torch.tensor(class_labels, dtype=torch.int)
+ elif isinstance(class_labels, int):
+ assert batch_size == 1, "Batch size must be 1 if classes is an int"
+ class_labels = torch.tensor([class_labels], dtype=torch.int)
+ elif class_labels is None:
+ # Randomly generate batch_size class labels
+ # TODO: should use generator here? int analogue of randn_tensor is not exposed in ...utils
+ class_labels = torch.randint(0, self.unet.config.num_class_embeds, size=(batch_size,))
+ class_labels = class_labels.to(device)
+ else:
+ class_labels = None
+ return class_labels
+
+ def check_inputs(self, num_inference_steps, timesteps, latents, batch_size, img_size, callback_steps):
+ if num_inference_steps is None and timesteps is None:
+ raise ValueError("Exactly one of `num_inference_steps` or `timesteps` must be supplied.")
+
+ if num_inference_steps is not None and timesteps is not None:
+ logger.warning(
+ f"Both `num_inference_steps`: {num_inference_steps} and `timesteps`: {timesteps} are supplied;"
+ " `timesteps` will be used over `num_inference_steps`."
+ )
+
+ if latents is not None:
+ expected_shape = (batch_size, 3, img_size, img_size)
+ if latents.shape != expected_shape:
+ raise ValueError(f"The shape of latents is {latents.shape} but is expected to be {expected_shape}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ batch_size: int = 1,
+ class_labels: Optional[Union[torch.Tensor, List[int], int]] = None,
+ num_inference_steps: int = 1,
+ timesteps: List[int] = None,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ Args:
+ batch_size (`int`, *optional*, defaults to 1):
+ The number of images to generate.
+ class_labels (`torch.Tensor` or `List[int]` or `int`, *optional*):
+ Optional class labels for conditioning class-conditional consistency models. Not used if the model is
+ not class-conditional.
+ num_inference_steps (`int`, *optional*, defaults to 1):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+ # 0. Prepare call parameters
+ img_size = self.unet.config.sample_size
+ device = self._execution_device
+
+ # 1. Check inputs
+ self.check_inputs(num_inference_steps, timesteps, latents, batch_size, img_size, callback_steps)
+
+ # 2. Prepare image latents
+ # Sample image latents x_0 ~ N(0, sigma_0^2 * I)
+ sample = self.prepare_latents(
+ batch_size=batch_size,
+ num_channels=self.unet.config.in_channels,
+ height=img_size,
+ width=img_size,
+ dtype=self.unet.dtype,
+ device=device,
+ generator=generator,
+ latents=latents,
+ )
+
+ # 3. Handle class_labels for class-conditional models
+ class_labels = self.prepare_class_labels(batch_size, device, class_labels=class_labels)
+
+ # 4. Prepare timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Denoising loop
+ # Multistep sampling: implements Algorithm 1 in the paper
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ scaled_sample = self.scheduler.scale_model_input(sample, t)
+ model_output = self.unet(scaled_sample, t, class_labels=class_labels, return_dict=False)[0]
+
+ sample = self.scheduler.step(model_output, t, sample, generator=generator)[0]
+
+ # call the callback, if provided
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, sample)
+
+ # 6. Post-process image sample
+ image = self.postprocess_image(sample, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/controlnet/__init__.py b/model/diffusers_c/pipelines/controlnet/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..3b832c0170641b1ab895dabd6deb523ca486c089
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/__init__.py
@@ -0,0 +1,80 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_flax_available,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["multicontrolnet"] = ["MultiControlNetModel"]
+ _import_structure["pipeline_controlnet"] = ["StableDiffusionControlNetPipeline"]
+ _import_structure["pipeline_controlnet_blip_diffusion"] = ["BlipDiffusionControlNetPipeline"]
+ _import_structure["pipeline_controlnet_img2img"] = ["StableDiffusionControlNetImg2ImgPipeline"]
+ _import_structure["pipeline_controlnet_inpaint"] = ["StableDiffusionControlNetInpaintPipeline"]
+ _import_structure["pipeline_controlnet_inpaint_sd_xl"] = ["StableDiffusionXLControlNetInpaintPipeline"]
+ _import_structure["pipeline_controlnet_sd_xl"] = ["StableDiffusionXLControlNetPipeline"]
+ _import_structure["pipeline_controlnet_sd_xl_img2img"] = ["StableDiffusionXLControlNetImg2ImgPipeline"]
+try:
+ if not (is_transformers_available() and is_flax_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_flax_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_flax_and_transformers_objects))
+else:
+ _import_structure["pipeline_flax_controlnet"] = ["FlaxStableDiffusionControlNetPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .multicontrolnet import MultiControlNetModel
+ from .pipeline_controlnet import StableDiffusionControlNetPipeline
+ from .pipeline_controlnet_blip_diffusion import BlipDiffusionControlNetPipeline
+ from .pipeline_controlnet_img2img import StableDiffusionControlNetImg2ImgPipeline
+ from .pipeline_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline
+ from .pipeline_controlnet_inpaint_sd_xl import StableDiffusionXLControlNetInpaintPipeline
+ from .pipeline_controlnet_sd_xl import StableDiffusionXLControlNetPipeline
+ from .pipeline_controlnet_sd_xl_img2img import StableDiffusionXLControlNetImg2ImgPipeline
+
+ try:
+ if not (is_transformers_available() and is_flax_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_flax_and_transformers_objects import * # noqa F403
+ else:
+ from .pipeline_flax_controlnet import FlaxStableDiffusionControlNetPipeline
+
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e22a3d82fcc66e1b20623d0eff1d8021b7ec1f08
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/multicontrolnet.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/multicontrolnet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..84ca4b33fd2e37dd24f9e3e54fc715a43cc9a10a
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/multicontrolnet.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d19a47d339a1bb537ed97a4a2ef137e22e4711cf
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_blip_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_blip_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..dcb55c57766a145be6aaad94d626b678fa2b4baa
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_blip_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6f327e9d7eb0f905db6026db0a742724b9a41337
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_inpaint.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_inpaint.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6cf97ceb9dcd06d902599b6cf2fa0800b25c2b50
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_inpaint.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_inpaint_sd_xl.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_inpaint_sd_xl.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..813fb928d3a38e81d170a7de06fa552d9ad6fb2e
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_inpaint_sd_xl.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_sd_xl.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_sd_xl.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..90d7dec98193b981e8196c097ab37ef3f09197ab
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_sd_xl.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_sd_xl_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_sd_xl_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c1a499ad2d3d7ab360b6549a76b22a77c0cf1ccc
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet/__pycache__/pipeline_controlnet_sd_xl_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/controlnet/multicontrolnet.py b/model/diffusers_c/pipelines/controlnet/multicontrolnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..7d284f2d26d3772414767f1d8a7815306cad4388
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/multicontrolnet.py
@@ -0,0 +1,187 @@
+import os
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+
+from ...models.controlnet import ControlNetModel, ControlNetOutput
+from ...models.modeling_utils import ModelMixin
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class MultiControlNetModel(ModelMixin):
+ r"""
+ Multiple `ControlNetModel` wrapper class for Multi-ControlNet
+
+ This module is a wrapper for multiple instances of the `ControlNetModel`. The `forward()` API is designed to be
+ compatible with `ControlNetModel`.
+
+ Args:
+ controlnets (`List[ControlNetModel]`):
+ Provides additional conditioning to the unet during the denoising process. You must set multiple
+ `ControlNetModel` as a list.
+ """
+
+ def __init__(self, controlnets: Union[List[ControlNetModel], Tuple[ControlNetModel]]):
+ super().__init__()
+ self.nets = nn.ModuleList(controlnets)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ controlnet_cond: List[torch.tensor],
+ conditioning_scale: List[float],
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guess_mode: bool = False,
+ return_dict: bool = True,
+ ) -> Union[ControlNetOutput, Tuple]:
+ for i, (image, scale, controlnet) in enumerate(zip(controlnet_cond, conditioning_scale, self.nets)):
+ down_samples, mid_sample = controlnet(
+ sample=sample,
+ timestep=timestep,
+ encoder_hidden_states=encoder_hidden_states,
+ controlnet_cond=image,
+ conditioning_scale=scale,
+ class_labels=class_labels,
+ timestep_cond=timestep_cond,
+ attention_mask=attention_mask,
+ added_cond_kwargs=added_cond_kwargs,
+ cross_attention_kwargs=cross_attention_kwargs,
+ guess_mode=guess_mode,
+ return_dict=return_dict,
+ )
+
+ # merge samples
+ if i == 0:
+ down_block_res_samples, mid_block_res_sample = down_samples, mid_sample
+ else:
+ down_block_res_samples = [
+ samples_prev + samples_curr
+ for samples_prev, samples_curr in zip(down_block_res_samples, down_samples)
+ ]
+ mid_block_res_sample += mid_sample
+
+ return down_block_res_samples, mid_block_res_sample
+
+ def save_pretrained(
+ self,
+ save_directory: Union[str, os.PathLike],
+ is_main_process: bool = True,
+ save_function: Callable = None,
+ safe_serialization: bool = True,
+ variant: Optional[str] = None,
+ ):
+ """
+ Save a model and its configuration file to a directory, so that it can be re-loaded using the
+ `[`~pipelines.controlnet.MultiControlNetModel.from_pretrained`]` class method.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to which to save. Will be created if it doesn't exist.
+ is_main_process (`bool`, *optional*, defaults to `True`):
+ Whether the process calling this is the main process or not. Useful when in distributed training like
+ TPUs and need to call this function on all processes. In this case, set `is_main_process=True` only on
+ the main process to avoid race conditions.
+ save_function (`Callable`):
+ The function to use to save the state dictionary. Useful on distributed training like TPUs when one
+ need to replace `torch.save` by another method. Can be configured with the environment variable
+ `DIFFUSERS_SAVE_MODE`.
+ safe_serialization (`bool`, *optional*, defaults to `True`):
+ Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
+ variant (`str`, *optional*):
+ If specified, weights are saved in the format pytorch_model..bin.
+ """
+ idx = 0
+ model_path_to_save = save_directory
+ for controlnet in self.nets:
+ controlnet.save_pretrained(
+ model_path_to_save,
+ is_main_process=is_main_process,
+ save_function=save_function,
+ safe_serialization=safe_serialization,
+ variant=variant,
+ )
+
+ idx += 1
+ model_path_to_save = model_path_to_save + f"_{idx}"
+
+ @classmethod
+ def from_pretrained(cls, pretrained_model_path: Optional[Union[str, os.PathLike]], **kwargs):
+ r"""
+ Instantiate a pretrained MultiControlNet model from multiple pre-trained controlnet models.
+
+ The model is set in evaluation mode by default using `model.eval()` (Dropout modules are deactivated). To train
+ the model, you should first set it back in training mode with `model.train()`.
+
+ The warning *Weights from XXX not initialized from pretrained model* means that the weights of XXX do not come
+ pretrained with the rest of the model. It is up to you to train those weights with a downstream fine-tuning
+ task.
+
+ The warning *Weights from XXX not used in YYY* means that the layer XXX is not used by YYY, therefore those
+ weights are discarded.
+
+ Parameters:
+ pretrained_model_path (`os.PathLike`):
+ A path to a *directory* containing model weights saved using
+ [`~diffusers.pipelines.controlnet.MultiControlNetModel.save_pretrained`], e.g.,
+ `./my_model_directory/controlnet`.
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model under this dtype. If `"auto"` is passed the dtype
+ will be automatically derived from the model's weights.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+ A map that specifies where each submodule should go. It doesn't need to be refined to each
+ parameter/buffer name, once a given module name is inside, every submodule of it will be sent to the
+ same device.
+
+ To have Accelerate compute the most optimized `device_map` automatically, set `device_map="auto"`. For
+ more information about each option see [designing a device
+ map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+ max_memory (`Dict`, *optional*):
+ A dictionary device identifier to maximum memory. Will default to the maximum memory available for each
+ GPU and the available CPU RAM if unset.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading by not initializing the weights and only loading the pre-trained weights. This
+ also tries to not use more than 1x model size in CPU memory (including peak memory) while loading the
+ model. This is only supported when torch version >= 1.9.0. If you are using an older version of torch,
+ setting this argument to `True` will raise an error.
+ variant (`str`, *optional*):
+ If specified load weights from `variant` filename, *e.g.* pytorch_model..bin. `variant` is
+ ignored when using `from_flax`.
+ use_safetensors (`bool`, *optional*, defaults to `None`):
+ If set to `None`, the `safetensors` weights will be downloaded if they're available **and** if the
+ `safetensors` library is installed. If set to `True`, the model will be forcibly loaded from
+ `safetensors` weights. If set to `False`, loading will *not* use `safetensors`.
+ """
+ idx = 0
+ controlnets = []
+
+ # load controlnet and append to list until no controlnet directory exists anymore
+ # first controlnet has to be saved under `./mydirectory/controlnet` to be compliant with `DiffusionPipeline.from_prertained`
+ # second, third, ... controlnets have to be saved under `./mydirectory/controlnet_1`, `./mydirectory/controlnet_2`, ...
+ model_path_to_load = pretrained_model_path
+ while os.path.isdir(model_path_to_load):
+ controlnet = ControlNetModel.from_pretrained(model_path_to_load, **kwargs)
+ controlnets.append(controlnet)
+
+ idx += 1
+ model_path_to_load = pretrained_model_path + f"_{idx}"
+
+ logger.info(f"{len(controlnets)} controlnets loaded from {pretrained_model_path}.")
+
+ if len(controlnets) == 0:
+ raise ValueError(
+ f"No ControlNets found under {os.path.dirname(pretrained_model_path)}. Expected at least {pretrained_model_path + '_0'}."
+ )
+
+ return cls(controlnets)
diff --git a/model/diffusers_c/pipelines/controlnet/pipeline_controlnet.py b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..8f31dfc2678a1aca7f538c9c10f7406822a60b19
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet.py
@@ -0,0 +1,1318 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ControlNetModel, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from .multicontrolnet import MultiControlNetModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> # !pip install opencv-python transformers accelerate
+ >>> from diffusers import StableDiffusionControlNetPipeline, ControlNetModel, UniPCMultistepScheduler
+ >>> from diffusers.utils import load_image
+ >>> import numpy as np
+ >>> import torch
+
+ >>> import cv2
+ >>> from PIL import Image
+
+ >>> # download an image
+ >>> image = load_image(
+ ... "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png"
+ ... )
+ >>> image = np.array(image)
+
+ >>> # get canny image
+ >>> image = cv2.Canny(image, 100, 200)
+ >>> image = image[:, :, None]
+ >>> image = np.concatenate([image, image, image], axis=2)
+ >>> canny_image = Image.fromarray(image)
+
+ >>> # load control net and stable diffusion v1-5
+ >>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
+ >>> pipe = StableDiffusionControlNetPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
+ ... )
+
+ >>> # speed up diffusion process with faster scheduler and memory optimization
+ >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+ >>> # remove following line if xformers is not installed
+ >>> pipe.enable_xformers_memory_efficient_attention()
+
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> # generate image
+ >>> generator = torch.manual_seed(0)
+ >>> image = pipe(
+ ... "futuristic-looking woman", num_inference_steps=20, generator=generator, image=canny_image
+ ... ).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionControlNetPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with ControlNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+ Provides additional conditioning to the `unet` during the denoising process. If you set multiple
+ ControlNets as a list, the outputs from each ControlNet are added together to create one combined
+ additional conditioning.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ if isinstance(controlnet, (list, tuple)):
+ controlnet = MultiControlNetModel(controlnet)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.control_image_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ controlnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ self.check_image(image, prompt, prompt_embeds)
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if not isinstance(image, list):
+ raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+ # When `image` is a nested list:
+ # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+ elif any(isinstance(i, list) for i in image):
+ transposed_image = [list(t) for t in zip(*image)]
+ if len(transposed_image) != len(self.controlnet.nets):
+ raise ValueError(
+ f"For multiple controlnets: if you pass`image` as a list of list, each sublist must have the same length as the number of controlnets, but the sublists in `image` got {len(transposed_image)} images and {len(self.controlnet.nets)} ControlNets."
+ )
+ for image_ in transposed_image:
+ self.check_image(image_, prompt, prompt_embeds)
+ elif len(image) != len(self.controlnet.nets):
+ raise ValueError(
+ f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+ )
+
+ for image_ in image:
+ self.check_image(image_, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `controlnet_conditioning_scale`
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ if not isinstance(controlnet_conditioning_scale, float):
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if isinstance(controlnet_conditioning_scale, list):
+ if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+ raise ValueError(
+ "A single batch of varying conditioning scale settings (e.g. [[1.0, 0.5], [0.2, 0.8]]) is not supported at the moment. "
+ "The conditioning scale must be fixed across the batch."
+ )
+ elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+ self.controlnet.nets
+ ):
+ raise ValueError(
+ "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+ " the same length as the number of controlnets"
+ )
+ else:
+ assert False
+
+ if not isinstance(control_guidance_start, (tuple, list)):
+ control_guidance_start = [control_guidance_start]
+
+ if not isinstance(control_guidance_end, (tuple, list)):
+ control_guidance_end = [control_guidance_end]
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if len(control_guidance_start) != len(self.controlnet.nets):
+ raise ValueError(
+ f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def check_image(self, image, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ def prepare_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single ControlNet. When `prompt` is a list, and if a list of images is passed for a single ControlNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple ControlNets,
+ where a list of image lists can be passed to batch for each prompt and each ControlNet.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ The ControlNet encoder tries to recognize the content of the input image even if you remove all
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the ControlNet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the ControlNet stops applying.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+ control_guidance_start, control_guidance_end = (
+ mult * [control_guidance_start],
+ mult * [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ image,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ controlnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+ controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+ global_pool_conditions = (
+ controlnet.config.global_pool_conditions
+ if isinstance(controlnet, ControlNetModel)
+ else controlnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image
+ if isinstance(controlnet, ControlNetModel):
+ image = self.prepare_image(
+ image=image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ height, width = image.shape[-2:]
+ elif isinstance(controlnet, MultiControlNetModel):
+ images = []
+
+ # Nested lists as ControlNet condition
+ if isinstance(image[0], list):
+ # Transpose the nested image list
+ image = [list(t) for t in zip(*image)]
+
+ for image_ in image:
+ image_ = self.prepare_image(
+ image=image_,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+
+ images.append(image_)
+
+ image = images
+ height, width = image[0].shape[-2:]
+ else:
+ assert False
+
+ # 5. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6.5 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 7.2 Create tensor stating which controlnets to keep
+ controlnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ is_unet_compiled = is_compiled_module(self.unet)
+ is_controlnet_compiled = is_compiled_module(self.controlnet)
+ is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Relevant thread:
+ # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+ if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
+ torch._inductor.cudagraph_mark_step_begin()
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # controlnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer ControlNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ controlnet_prompt_embeds = prompt_embeds
+
+ if isinstance(controlnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+ else:
+ controlnet_cond_scale = controlnet_conditioning_scale
+ if isinstance(controlnet_cond_scale, list):
+ controlnet_cond_scale = controlnet_cond_scale[0]
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=controlnet_prompt_embeds,
+ controlnet_cond=image,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered ControlNet only for the conditional batch.
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and controlnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.controlnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_blip_diffusion.py b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_blip_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..b983a3f8d4fc981d6ac3b55d91fd44ab816b35c7
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_blip_diffusion.py
@@ -0,0 +1,413 @@
+# Copyright 2024 Salesforce.com, inc.
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPTokenizer
+
+from ...models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
+from ...schedulers import PNDMScheduler
+from ...utils import (
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..blip_diffusion.blip_image_processing import BlipImageProcessor
+from ..blip_diffusion.modeling_blip2 import Blip2QFormerModel
+from ..blip_diffusion.modeling_ctx_clip import ContextCLIPTextModel
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers.pipelines import BlipDiffusionControlNetPipeline
+ >>> from diffusers.utils import load_image
+ >>> from controlnet_aux import CannyDetector
+ >>> import torch
+
+ >>> blip_diffusion_pipe = BlipDiffusionControlNetPipeline.from_pretrained(
+ ... "Salesforce/blipdiffusion-controlnet", torch_dtype=torch.float16
+ ... ).to("cuda")
+
+ >>> style_subject = "flower"
+ >>> tgt_subject = "teapot"
+ >>> text_prompt = "on a marble table"
+
+ >>> cldm_cond_image = load_image(
+ ... "https://huggingface.co/datasets/ayushtues/blipdiffusion_images/resolve/main/kettle.jpg"
+ ... ).resize((512, 512))
+ >>> canny = CannyDetector()
+ >>> cldm_cond_image = canny(cldm_cond_image, 30, 70, output_type="pil")
+ >>> style_image = load_image(
+ ... "https://huggingface.co/datasets/ayushtues/blipdiffusion_images/resolve/main/flower.jpg"
+ ... )
+ >>> guidance_scale = 7.5
+ >>> num_inference_steps = 50
+ >>> negative_prompt = "over-exposure, under-exposure, saturated, duplicate, out of frame, lowres, cropped, worst quality, low quality, jpeg artifacts, morbid, mutilated, out of frame, ugly, bad anatomy, bad proportions, deformed, blurry, duplicate"
+
+
+ >>> output = blip_diffusion_pipe(
+ ... text_prompt,
+ ... style_image,
+ ... cldm_cond_image,
+ ... style_subject,
+ ... tgt_subject,
+ ... guidance_scale=guidance_scale,
+ ... num_inference_steps=num_inference_steps,
+ ... neg_prompt=negative_prompt,
+ ... height=512,
+ ... width=512,
+ ... ).images
+ >>> output[0].save("image.png")
+ ```
+"""
+
+
+class BlipDiffusionControlNetPipeline(DiffusionPipeline):
+ """
+ Pipeline for Canny Edge based Controlled subject-driven generation using Blip Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ tokenizer ([`CLIPTokenizer`]):
+ Tokenizer for the text encoder
+ text_encoder ([`ContextCLIPTextModel`]):
+ Text encoder to encode the text prompt
+ vae ([`AutoencoderKL`]):
+ VAE model to map the latents to the image
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ scheduler ([`PNDMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ qformer ([`Blip2QFormerModel`]):
+ QFormer model to get multi-modal embeddings from the text and image.
+ controlnet ([`ControlNetModel`]):
+ ControlNet model to get the conditioning image embedding.
+ image_processor ([`BlipImageProcessor`]):
+ Image Processor to preprocess and postprocess the image.
+ ctx_begin_pos (int, `optional`, defaults to 2):
+ Position of the context token in the text encoder.
+ """
+
+ model_cpu_offload_seq = "qformer->text_encoder->unet->vae"
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ text_encoder: ContextCLIPTextModel,
+ vae: AutoencoderKL,
+ unet: UNet2DConditionModel,
+ scheduler: PNDMScheduler,
+ qformer: Blip2QFormerModel,
+ controlnet: ControlNetModel,
+ image_processor: BlipImageProcessor,
+ ctx_begin_pos: int = 2,
+ mean: List[float] = None,
+ std: List[float] = None,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ vae=vae,
+ unet=unet,
+ scheduler=scheduler,
+ qformer=qformer,
+ controlnet=controlnet,
+ image_processor=image_processor,
+ )
+ self.register_to_config(ctx_begin_pos=ctx_begin_pos, mean=mean, std=std)
+
+ def get_query_embeddings(self, input_image, src_subject):
+ return self.qformer(image_input=input_image, text_input=src_subject, return_dict=False)
+
+ # from the original Blip Diffusion code, speciefies the target subject and augments the prompt by repeating it
+ def _build_prompt(self, prompts, tgt_subjects, prompt_strength=1.0, prompt_reps=20):
+ rv = []
+ for prompt, tgt_subject in zip(prompts, tgt_subjects):
+ prompt = f"a {tgt_subject} {prompt.strip()}"
+ # a trick to amplify the prompt
+ rv.append(", ".join([prompt] * int(prompt_strength * prompt_reps)))
+
+ return rv
+
+ # Copied from diffusers.pipelines.consistency_models.pipeline_consistency_models.ConsistencyModelPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels, height, width)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device=device, dtype=dtype)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def encode_prompt(self, query_embeds, prompt, device=None):
+ device = device or self._execution_device
+
+ # embeddings for prompt, with query_embeds as context
+ max_len = self.text_encoder.text_model.config.max_position_embeddings
+ max_len -= self.qformer.config.num_query_tokens
+
+ tokenized_prompt = self.tokenizer(
+ prompt,
+ padding="max_length",
+ truncation=True,
+ max_length=max_len,
+ return_tensors="pt",
+ ).to(device)
+
+ batch_size = query_embeds.shape[0]
+ ctx_begin_pos = [self.config.ctx_begin_pos] * batch_size
+
+ text_embeddings = self.text_encoder(
+ input_ids=tokenized_prompt.input_ids,
+ ctx_embeddings=query_embeds,
+ ctx_begin_pos=ctx_begin_pos,
+ )[0]
+
+ return text_embeddings
+
+ # Adapted from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
+ def prepare_control_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ ):
+ image = self.image_processor.preprocess(
+ image,
+ size={"width": width, "height": height},
+ do_rescale=True,
+ do_center_crop=False,
+ do_normalize=False,
+ return_tensors="pt",
+ )["pixel_values"].to(device)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: List[str],
+ reference_image: PIL.Image.Image,
+ condtioning_image: PIL.Image.Image,
+ source_subject_category: List[str],
+ target_subject_category: List[str],
+ latents: Optional[torch.FloatTensor] = None,
+ guidance_scale: float = 7.5,
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 50,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ neg_prompt: Optional[str] = "",
+ prompt_strength: float = 1.0,
+ prompt_reps: int = 20,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`List[str]`):
+ The prompt or prompts to guide the image generation.
+ reference_image (`PIL.Image.Image`):
+ The reference image to condition the generation on.
+ condtioning_image (`PIL.Image.Image`):
+ The conditioning canny edge image to condition the generation on.
+ source_subject_category (`List[str]`):
+ The source subject category.
+ target_subject_category (`List[str]`):
+ The target subject category.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by random sampling.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ height (`int`, *optional*, defaults to 512):
+ The height of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width of the generated image.
+ seed (`int`, *optional*, defaults to 42):
+ The seed to use for random generation.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ neg_prompt (`str`, *optional*, defaults to ""):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_strength (`float`, *optional*, defaults to 1.0):
+ The strength of the prompt. Specifies the number of times the prompt is repeated along with prompt_reps
+ to amplify the prompt.
+ prompt_reps (`int`, *optional*, defaults to 20):
+ The number of times the prompt is repeated along with prompt_strength to amplify the prompt.
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ device = self._execution_device
+
+ reference_image = self.image_processor.preprocess(
+ reference_image, image_mean=self.config.mean, image_std=self.config.std, return_tensors="pt"
+ )["pixel_values"]
+ reference_image = reference_image.to(device)
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ if isinstance(source_subject_category, str):
+ source_subject_category = [source_subject_category]
+ if isinstance(target_subject_category, str):
+ target_subject_category = [target_subject_category]
+
+ batch_size = len(prompt)
+
+ prompt = self._build_prompt(
+ prompts=prompt,
+ tgt_subjects=target_subject_category,
+ prompt_strength=prompt_strength,
+ prompt_reps=prompt_reps,
+ )
+ query_embeds = self.get_query_embeddings(reference_image, source_subject_category)
+ text_embeddings = self.encode_prompt(query_embeds, prompt, device)
+ # 3. unconditional embedding
+ do_classifier_free_guidance = guidance_scale > 1.0
+ if do_classifier_free_guidance:
+ max_length = self.text_encoder.text_model.config.max_position_embeddings
+
+ uncond_input = self.tokenizer(
+ [neg_prompt] * batch_size,
+ padding="max_length",
+ max_length=max_length,
+ return_tensors="pt",
+ )
+ uncond_embeddings = self.text_encoder(
+ input_ids=uncond_input.input_ids.to(device),
+ ctx_embeddings=None,
+ )[0]
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+ scale_down_factor = 2 ** (len(self.unet.config.block_out_channels) - 1)
+ latents = self.prepare_latents(
+ batch_size=batch_size,
+ num_channels=self.unet.config.in_channels,
+ height=height // scale_down_factor,
+ width=width // scale_down_factor,
+ generator=generator,
+ latents=latents,
+ dtype=self.unet.dtype,
+ device=device,
+ )
+ # set timesteps
+ extra_set_kwargs = {}
+ self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)
+
+ cond_image = self.prepare_control_image(
+ image=condtioning_image,
+ width=width,
+ height=height,
+ batch_size=batch_size,
+ num_images_per_prompt=1,
+ device=device,
+ dtype=self.controlnet.dtype,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ )
+
+ for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=text_embeddings,
+ controlnet_cond=cond_image,
+ return_dict=False,
+ )
+
+ noise_pred = self.unet(
+ latent_model_input,
+ timestep=t,
+ encoder_hidden_states=text_embeddings,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ )["sample"]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ )["prev_sample"]
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_img2img.py b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..9d2c76fd748363229eed17dafc229b925eddaf2f
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_img2img.py
@@ -0,0 +1,1310 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ControlNetModel, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from .multicontrolnet import MultiControlNetModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> # !pip install opencv-python transformers accelerate
+ >>> from diffusers import StableDiffusionControlNetImg2ImgPipeline, ControlNetModel, UniPCMultistepScheduler
+ >>> from diffusers.utils import load_image
+ >>> import numpy as np
+ >>> import torch
+
+ >>> import cv2
+ >>> from PIL import Image
+
+ >>> # download an image
+ >>> image = load_image(
+ ... "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/input_image_vermeer.png"
+ ... )
+ >>> np_image = np.array(image)
+
+ >>> # get canny image
+ >>> np_image = cv2.Canny(np_image, 100, 200)
+ >>> np_image = np_image[:, :, None]
+ >>> np_image = np.concatenate([np_image, np_image, np_image], axis=2)
+ >>> canny_image = Image.fromarray(np_image)
+
+ >>> # load control net and stable diffusion v1-5
+ >>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
+ >>> pipe = StableDiffusionControlNetImg2ImgPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
+ ... )
+
+ >>> # speed up diffusion process with faster scheduler and memory optimization
+ >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> # generate image
+ >>> generator = torch.manual_seed(0)
+ >>> image = pipe(
+ ... "futuristic-looking woman",
+ ... num_inference_steps=20,
+ ... generator=generator,
+ ... image=image,
+ ... control_image=canny_image,
+ ... ).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+def prepare_image(image):
+ if isinstance(image, torch.Tensor):
+ # Batch single image
+ if image.ndim == 3:
+ image = image.unsqueeze(0)
+
+ image = image.to(dtype=torch.float32)
+ else:
+ # preprocess image
+ if isinstance(image, (PIL.Image.Image, np.ndarray)):
+ image = [image]
+
+ if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+ image = [np.array(i.convert("RGB"))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+ image = np.concatenate([i[None, :] for i in image], axis=0)
+
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ return image
+
+
+class StableDiffusionControlNetImg2ImgPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for image-to-image generation using Stable Diffusion with ControlNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+ Provides additional conditioning to the `unet` during the denoising process. If you set multiple
+ ControlNets as a list, the outputs from each ControlNet are added together to create one combined
+ additional conditioning.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ if isinstance(controlnet, (list, tuple)):
+ controlnet = MultiControlNetModel(controlnet)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.control_image_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ controlnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # `prompt` needs more sophisticated handling when there are multiple
+ # conditionings.
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if isinstance(prompt, list):
+ logger.warning(
+ f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+ " prompts. The conditionings will be fixed across the prompts."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ self.check_image(image, prompt, prompt_embeds)
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if not isinstance(image, list):
+ raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+ # When `image` is a nested list:
+ # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+ elif any(isinstance(i, list) for i in image):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif len(image) != len(self.controlnet.nets):
+ raise ValueError(
+ f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+ )
+
+ for image_ in image:
+ self.check_image(image_, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `controlnet_conditioning_scale`
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ if not isinstance(controlnet_conditioning_scale, float):
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if isinstance(controlnet_conditioning_scale, list):
+ if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+ self.controlnet.nets
+ ):
+ raise ValueError(
+ "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+ " the same length as the number of controlnets"
+ )
+ else:
+ assert False
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if len(control_guidance_start) != len(self.controlnet.nets):
+ raise ValueError(
+ f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
+ def check_image(self, image, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
+ def prepare_control_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.prepare_latents
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+
+ return latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ control_image: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ strength: float = 0.8,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ controlnet_conditioning_scale: Union[float, List[float]] = 0.8,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The initial image to be used as the starting point for the image generation process. Can also accept
+ image latents as `image`, and if passing latents directly they are not encoded again.
+ control_image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single ControlNet.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ The ControlNet encoder tries to recognize the content of the input image even if you remove all
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the ControlNet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the ControlNet stops applying.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+ control_guidance_start, control_guidance_end = (
+ mult * [control_guidance_start],
+ mult * [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ control_image,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ controlnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+ controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+ global_pool_conditions = (
+ controlnet.config.global_pool_conditions
+ if isinstance(controlnet, ControlNetModel)
+ else controlnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image
+ image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+
+ # 5. Prepare controlnet_conditioning_image
+ if isinstance(controlnet, ControlNetModel):
+ control_image = self.prepare_control_image(
+ image=control_image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ elif isinstance(controlnet, MultiControlNetModel):
+ control_images = []
+
+ for control_image_ in control_image:
+ control_image_ = self.prepare_control_image(
+ image=control_image_,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+
+ control_images.append(control_image_)
+
+ control_image = control_images
+ else:
+ assert False
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ latents = self.prepare_latents(
+ image,
+ latent_timestep,
+ batch_size,
+ num_images_per_prompt,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 7.2 Create tensor stating which controlnets to keep
+ controlnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # controlnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer ControlNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ controlnet_prompt_embeds = prompt_embeds
+
+ if isinstance(controlnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+ else:
+ controlnet_cond_scale = controlnet_conditioning_scale
+ if isinstance(controlnet_cond_scale, list):
+ controlnet_cond_scale = controlnet_cond_scale[0]
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=controlnet_prompt_embeds,
+ controlnet_cond=control_image,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered ControlNet only for the conditional batch.
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and controlnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.controlnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_inpaint.py b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_inpaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..c4f1bff5efcdd5d3c62d31758e017402ed8db06e
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_inpaint.py
@@ -0,0 +1,1620 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This model implementation is heavily inspired by https://github.com/haofanwang/ControlNet-for-Diffusers/
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ControlNetModel, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from .multicontrolnet import MultiControlNetModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> # !pip install transformers accelerate
+ >>> from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel, DDIMScheduler
+ >>> from diffusers.utils import load_image
+ >>> import numpy as np
+ >>> import torch
+
+ >>> init_image = load_image(
+ ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy.png"
+ ... )
+ >>> init_image = init_image.resize((512, 512))
+
+ >>> generator = torch.Generator(device="cpu").manual_seed(1)
+
+ >>> mask_image = load_image(
+ ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy_mask.png"
+ ... )
+ >>> mask_image = mask_image.resize((512, 512))
+
+
+ >>> def make_canny_condition(image):
+ ... image = np.array(image)
+ ... image = cv2.Canny(image, 100, 200)
+ ... image = image[:, :, None]
+ ... image = np.concatenate([image, image, image], axis=2)
+ ... image = Image.fromarray(image)
+ ... return image
+
+
+ >>> control_image = make_canny_condition(init_image)
+
+ >>> controlnet = ControlNetModel.from_pretrained(
+ ... "lllyasviel/control_v11p_sd15_inpaint", torch_dtype=torch.float16
+ ... )
+ >>> pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
+ ... )
+
+ >>> pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> # generate image
+ >>> image = pipe(
+ ... "a handsome man with ray-ban sunglasses",
+ ... num_inference_steps=20,
+ ... generator=generator,
+ ... eta=1.0,
+ ... image=init_image,
+ ... mask_image=mask_image,
+ ... control_image=control_image,
+ ... ).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.prepare_mask_and_masked_image
+def prepare_mask_and_masked_image(image, mask, height, width, return_image=False):
+ """
+ Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
+ converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
+ ``image`` and ``1`` for the ``mask``.
+
+ The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+ binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+ Args:
+ image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+ ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+ mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+ ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+
+
+ Raises:
+ ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+ should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+ TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+ (ot the other way around).
+
+ Returns:
+ tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
+ dimensions: ``batch x channels x height x width``.
+ """
+ deprecation_message = "The prepare_mask_and_masked_image method is deprecated and will be removed in a future version. Please use VaeImageProcessor.preprocess instead"
+ deprecate(
+ "prepare_mask_and_masked_image",
+ "0.30.0",
+ deprecation_message,
+ )
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ if mask is None:
+ raise ValueError("`mask_image` input cannot be undefined.")
+
+ if isinstance(image, torch.Tensor):
+ if not isinstance(mask, torch.Tensor):
+ raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
+
+ # Batch single image
+ if image.ndim == 3:
+ assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
+ image = image.unsqueeze(0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Single batched mask, no channel dim or single mask not batched but channel dim
+ if mask.shape[0] == 1:
+ mask = mask.unsqueeze(0)
+
+ # Batched masks no channel dim
+ else:
+ mask = mask.unsqueeze(1)
+
+ assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+ assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+ assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+
+ # Check image is in [-1, 1]
+ if image.min() < -1 or image.max() > 1:
+ raise ValueError("Image should be in [-1, 1] range")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("Mask should be in [0, 1] range")
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ # Image as float32
+ image = image.to(dtype=torch.float32)
+ elif isinstance(mask, torch.Tensor):
+ raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+ else:
+ # preprocess image
+ if isinstance(image, (PIL.Image.Image, np.ndarray)):
+ image = [image]
+ if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+ # resize all images w.r.t passed height an width
+ image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
+ image = [np.array(i.convert("RGB"))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+ image = np.concatenate([i[None, :] for i in image], axis=0)
+
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ # preprocess mask
+ if isinstance(mask, (PIL.Image.Image, np.ndarray)):
+ mask = [mask]
+
+ if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
+ mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
+ mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+ mask = mask.astype(np.float32) / 255.0
+ elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+ mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+ mask = torch.from_numpy(mask)
+
+ masked_image = image * (mask < 0.5)
+
+ # n.b. ensure backwards compatibility as old function does not return image
+ if return_image:
+ return mask, masked_image, image
+
+ return mask, masked_image
+
+
+class StableDiffusionControlNetInpaintPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for image inpainting using Stable Diffusion with ControlNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+
+
+ This pipeline can be used with checkpoints that have been specifically fine-tuned for inpainting
+ ([runwayml/stable-diffusion-inpainting](https://huggingface.co/runwayml/stable-diffusion-inpainting)) as well as
+ default text-to-image Stable Diffusion checkpoints
+ ([runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5)). Default text-to-image
+ Stable Diffusion checkpoints might be preferable for ControlNets that have been fine-tuned on those, such as
+ [lllyasviel/control_v11p_sd15_inpaint](https://huggingface.co/lllyasviel/control_v11p_sd15_inpaint).
+
+
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+ Provides additional conditioning to the `unet` during the denoising process. If you set multiple
+ ControlNets as a list, the outputs from each ControlNet are added together to create one combined
+ additional conditioning.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ if isinstance(controlnet, (list, tuple)):
+ controlnet = MultiControlNetModel(controlnet)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.mask_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
+ )
+ self.control_image_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ mask_image,
+ height,
+ width,
+ callback_steps,
+ output_type,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ controlnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ padding_mask_crop=None,
+ ):
+ if height is not None and height % 8 != 0 or width is not None and width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if padding_mask_crop is not None:
+ if not isinstance(image, PIL.Image.Image):
+ raise ValueError(
+ f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+ )
+ if not isinstance(mask_image, PIL.Image.Image):
+ raise ValueError(
+ f"The mask image should be a PIL image when inpainting mask crop, but is of type"
+ f" {type(mask_image)}."
+ )
+ if output_type != "pil":
+ raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+
+ # `prompt` needs more sophisticated handling when there are multiple
+ # conditionings.
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if isinstance(prompt, list):
+ logger.warning(
+ f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+ " prompts. The conditionings will be fixed across the prompts."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ self.check_image(image, prompt, prompt_embeds)
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if not isinstance(image, list):
+ raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+ # When `image` is a nested list:
+ # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+ elif any(isinstance(i, list) for i in image):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif len(image) != len(self.controlnet.nets):
+ raise ValueError(
+ f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+ )
+
+ for image_ in image:
+ self.check_image(image_, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `controlnet_conditioning_scale`
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ if not isinstance(controlnet_conditioning_scale, float):
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if isinstance(controlnet_conditioning_scale, list):
+ if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+ self.controlnet.nets
+ ):
+ raise ValueError(
+ "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+ " the same length as the number of controlnets"
+ )
+ else:
+ assert False
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if len(control_guidance_start) != len(self.controlnet.nets):
+ raise ValueError(
+ f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
+ def check_image(self, image, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ def prepare_control_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ crops_coords,
+ resize_mode,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.control_image_processor.preprocess(
+ image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
+ ).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_latents
+ def prepare_latents(
+ self,
+ batch_size,
+ num_channels_latents,
+ height,
+ width,
+ dtype,
+ device,
+ generator,
+ latents=None,
+ image=None,
+ timestep=None,
+ is_strength_max=True,
+ return_noise=False,
+ return_image_latents=False,
+ ):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if (image is None or timestep is None) and not is_strength_max:
+ raise ValueError(
+ "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
+ "However, either the image or the noise timestep has not been provided."
+ )
+
+ if return_image_latents or (latents is None and not is_strength_max):
+ image = image.to(device=device, dtype=dtype)
+
+ if image.shape[1] == 4:
+ image_latents = image
+ else:
+ image_latents = self._encode_vae_image(image=image, generator=generator)
+ image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
+
+ if latents is None:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ # if strength is 1. then initialise the latents to noise, else initial to image + noise
+ latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
+ # if pure noise then scale the initial latents by the Scheduler's init sigma
+ latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+ else:
+ noise = latents.to(device)
+ latents = noise * self.scheduler.init_noise_sigma
+
+ outputs = (latents,)
+
+ if return_noise:
+ outputs += (noise,)
+
+ if return_image_latents:
+ outputs += (image_latents,)
+
+ return outputs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_mask_latents
+ def prepare_mask_latents(
+ self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+ ):
+ # resize the mask to latents shape as we concatenate the mask to the latents
+ # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+ # and half precision
+ mask = torch.nn.functional.interpolate(
+ mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+ )
+ mask = mask.to(device=device, dtype=dtype)
+
+ masked_image = masked_image.to(device=device, dtype=dtype)
+
+ if masked_image.shape[1] == 4:
+ masked_image_latents = masked_image
+ else:
+ masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+
+ # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+ if mask.shape[0] < batch_size:
+ if not batch_size % mask.shape[0] == 0:
+ raise ValueError(
+ "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+ f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+ " of masks that you pass is divisible by the total requested batch size."
+ )
+ mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+ if masked_image_latents.shape[0] < batch_size:
+ if not batch_size % masked_image_latents.shape[0] == 0:
+ raise ValueError(
+ "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+ f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+ " Make sure the number of images that you pass is divisible by the total requested batch size."
+ )
+ masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)
+
+ mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+ masked_image_latents = (
+ torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+ )
+
+ # aligning device to prevent device errors when concating it with the latent model input
+ masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+ return mask, masked_image_latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline._encode_vae_image
+ def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+ if isinstance(generator, list):
+ image_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(image.shape[0])
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ image_latents = self.vae.config.scaling_factor * image_latents
+
+ return image_latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ mask_image: PipelineImageInput = None,
+ control_image: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ padding_mask_crop: Optional[int] = None,
+ strength: float = 1.0,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ controlnet_conditioning_scale: Union[float, List[float]] = 0.5,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`,
+ `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, NumPy array or tensor representing an image batch to be used as the starting point. For both
+ NumPy array and PyTorch tensor, the expected value range is between `[0, 1]`. If it's a tensor or a
+ list or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a NumPy array or
+ a list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)`. It can also accept image
+ latents as `image`, but if passing latents directly it is not encoded again.
+ mask_image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`,
+ `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, NumPy array or tensor representing an image batch to mask `image`. White pixels in the mask
+ are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a
+ single channel (luminance) before use. If it's a NumPy array or PyTorch tensor, it should contain one
+ color channel (L) instead of 3, so the expected shape for PyTorch tensor would be `(B, 1, H, W)`, `(B,
+ H, W)`, `(1, H, W)`, `(H, W)`. And for NumPy array, it would be for `(B, H, W, 1)`, `(B, H, W)`, `(H,
+ W, 1)`, or `(H, W)`.
+ control_image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`,
+ `List[List[torch.FloatTensor]]`, or `List[List[PIL.Image.Image]]`):
+ The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single ControlNet.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ padding_mask_crop (`int`, *optional*, defaults to `None`):
+ The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to image and mask_image. If
+ `padding_mask_crop` is not `None`, it will first find a rectangular region with the same aspect ration of the image and
+ contains all masked area, and then expand that area based on `padding_mask_crop`. The image and mask_image will then be cropped based on
+ the expanded area before resizing to the original image size for inpainting. This is useful when the masked area is small while the image is large
+ and contain information inreleant for inpainging, such as background.
+ strength (`float`, *optional*, defaults to 1.0):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 0.5):
+ The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ The ControlNet encoder tries to recognize the content of the input image even if you remove all
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the ControlNet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the ControlNet stops applying.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+ control_guidance_start, control_guidance_end = (
+ mult * [control_guidance_start],
+ mult * [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ control_image,
+ mask_image,
+ height,
+ width,
+ callback_steps,
+ output_type,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ controlnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ padding_mask_crop,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if padding_mask_crop is not None:
+ height, width = self.image_processor.get_default_height_width(image, height, width)
+ crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop)
+ resize_mode = "fill"
+ else:
+ crops_coords = None
+ resize_mode = "default"
+
+ device = self._execution_device
+
+ if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+ controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+ global_pool_conditions = (
+ controlnet.config.global_pool_conditions
+ if isinstance(controlnet, ControlNetModel)
+ else controlnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image
+ if isinstance(controlnet, ControlNetModel):
+ control_image = self.prepare_control_image(
+ image=control_image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ crops_coords=crops_coords,
+ resize_mode=resize_mode,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ elif isinstance(controlnet, MultiControlNetModel):
+ control_images = []
+
+ for control_image_ in control_image:
+ control_image_ = self.prepare_control_image(
+ image=control_image_,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ crops_coords=crops_coords,
+ resize_mode=resize_mode,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+
+ control_images.append(control_image_)
+
+ control_image = control_images
+ else:
+ assert False
+
+ # 4.1 Preprocess mask and image - resizes image and mask w.r.t height and width
+ original_image = image
+ init_image = self.image_processor.preprocess(
+ image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
+ )
+ init_image = init_image.to(dtype=torch.float32)
+
+ mask = self.mask_processor.preprocess(
+ mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
+ )
+
+ masked_image = init_image * (mask < 0.5)
+ _, _, height, width = init_image.shape
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(
+ num_inference_steps=num_inference_steps, strength=strength, device=device
+ )
+ # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+ is_strength_max = strength == 1.0
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ num_channels_unet = self.unet.config.in_channels
+ return_image_latents = num_channels_unet == 4
+ latents_outputs = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ image=init_image,
+ timestep=latent_timestep,
+ is_strength_max=is_strength_max,
+ return_noise=True,
+ return_image_latents=return_image_latents,
+ )
+
+ if return_image_latents:
+ latents, noise, image_latents = latents_outputs
+ else:
+ latents, noise = latents_outputs
+
+ # 7. Prepare mask latent variables
+ mask, masked_image_latents = self.prepare_mask_latents(
+ mask,
+ masked_image,
+ batch_size * num_images_per_prompt,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ self.do_classifier_free_guidance,
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 7.2 Create tensor stating which controlnets to keep
+ controlnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # controlnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer ControlNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ controlnet_prompt_embeds = prompt_embeds
+
+ if isinstance(controlnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+ else:
+ controlnet_cond_scale = controlnet_conditioning_scale
+ if isinstance(controlnet_cond_scale, list):
+ controlnet_cond_scale = controlnet_cond_scale[0]
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=controlnet_prompt_embeds,
+ controlnet_cond=control_image,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered ControlNet only for the conditional batch.
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+ # predict the noise residual
+ if num_channels_unet == 9:
+ latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if num_channels_unet == 4:
+ init_latents_proper = image_latents
+ if self.do_classifier_free_guidance:
+ init_mask, _ = mask.chunk(2)
+ else:
+ init_mask = mask
+
+ if i < len(timesteps) - 1:
+ noise_timestep = timesteps[i + 1]
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_proper, noise, torch.tensor([noise_timestep])
+ )
+
+ latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and controlnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.controlnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ if padding_mask_crop is not None:
+ image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image]
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py
new file mode 100644
index 0000000000000000000000000000000000000000..52ffe5a3f356a4845962d61a3b9cf108a7918e86
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_inpaint_sd_xl.py
@@ -0,0 +1,1818 @@
+# Copyright 2024 Harutatsu Akiyama, Jinbin Bai, and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import (
+ FromSingleFileMixin,
+ IPAdapterMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, ControlNetModel, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ is_invisible_watermark_available,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+from .multicontrolnet import MultiControlNetModel
+
+
+if is_invisible_watermark_available():
+ from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> # !pip install transformers accelerate
+ >>> from diffusers import StableDiffusionXLControlNetInpaintPipeline, ControlNetModel, DDIMScheduler
+ >>> from diffusers.utils import load_image
+ >>> import numpy as np
+ >>> import torch
+
+ >>> init_image = load_image(
+ ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy.png"
+ ... )
+ >>> init_image = init_image.resize((1024, 1024))
+
+ >>> generator = torch.Generator(device="cpu").manual_seed(1)
+
+ >>> mask_image = load_image(
+ ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy_mask.png"
+ ... )
+ >>> mask_image = mask_image.resize((1024, 1024))
+
+
+ >>> def make_canny_condition(image):
+ ... image = np.array(image)
+ ... image = cv2.Canny(image, 100, 200)
+ ... image = image[:, :, None]
+ ... image = np.concatenate([image, image, image], axis=2)
+ ... image = Image.fromarray(image)
+ ... return image
+
+
+ >>> control_image = make_canny_condition(init_image)
+
+ >>> controlnet = ControlNetModel.from_pretrained(
+ ... "diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16
+ ... )
+ >>> pipe = StableDiffusionXLControlNetInpaintPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, torch_dtype=torch.float16
+ ... )
+
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> # generate image
+ >>> image = pipe(
+ ... "a handsome man with ray-ban sunglasses",
+ ... num_inference_steps=20,
+ ... generator=generator,
+ ... eta=1.0,
+ ... image=init_image,
+ ... mask_image=mask_image,
+ ... control_image=control_image,
+ ... ).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+class StableDiffusionXLControlNetInpaintPipeline(
+ DiffusionPipeline, StableDiffusionMixin, StableDiffusionXLLoraLoaderMixin, FromSingleFileMixin, IPAdapterMixin
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion XL.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+ specifically the
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+ variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ tokenizer_2 (`CLIPTokenizer`):
+ Second Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
+ _optional_components = ["tokenizer", "tokenizer_2", "text_encoder", "text_encoder_2"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ controlnet: ControlNetModel,
+ scheduler: KarrasDiffusionSchedulers,
+ requires_aesthetics_score: bool = False,
+ force_zeros_for_empty_prompt: bool = True,
+ add_watermarker: Optional[bool] = None,
+ feature_extractor: Optional[CLIPImageProcessor] = None,
+ image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ ):
+ super().__init__()
+
+ if isinstance(controlnet, (list, tuple)):
+ controlnet = MultiControlNetModel(controlnet)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.mask_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
+ )
+ self.control_image_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+ )
+
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+ if add_watermarker:
+ self.watermark = StableDiffusionXLWatermarker()
+ else:
+ self.watermark = None
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_image(self, image, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ image,
+ mask_image,
+ strength,
+ num_inference_steps,
+ callback_steps,
+ output_type,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ pooled_prompt_embeds=None,
+ negative_pooled_prompt_embeds=None,
+ controlnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ padding_mask_crop=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+ if num_inference_steps is None:
+ raise ValueError("`num_inference_steps` cannot be None.")
+ elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0:
+ raise ValueError(
+ f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type"
+ f" {type(num_inference_steps)}."
+ )
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if padding_mask_crop is not None:
+ if not isinstance(image, PIL.Image.Image):
+ raise ValueError(
+ f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+ )
+ if not isinstance(mask_image, PIL.Image.Image):
+ raise ValueError(
+ f"The mask image should be a PIL image when inpainting mask crop, but is of type"
+ f" {type(mask_image)}."
+ )
+ if output_type != "pil":
+ raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+ )
+
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+ )
+
+ # `prompt` needs more sophisticated handling when there are multiple
+ # conditionings.
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if isinstance(prompt, list):
+ logger.warning(
+ f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+ " prompts. The conditionings will be fixed across the prompts."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ self.check_image(image, prompt, prompt_embeds)
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if not isinstance(image, list):
+ raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+ # When `image` is a nested list:
+ # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+ elif any(isinstance(i, list) for i in image):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif len(image) != len(self.controlnet.nets):
+ raise ValueError(
+ f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+ )
+
+ for image_ in image:
+ self.check_image(image_, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `controlnet_conditioning_scale`
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ if not isinstance(controlnet_conditioning_scale, float):
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if isinstance(controlnet_conditioning_scale, list):
+ if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+ self.controlnet.nets
+ ):
+ raise ValueError(
+ "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+ " the same length as the number of controlnets"
+ )
+ else:
+ assert False
+
+ if not isinstance(control_guidance_start, (tuple, list)):
+ control_guidance_start = [control_guidance_start]
+
+ if not isinstance(control_guidance_end, (tuple, list)):
+ control_guidance_end = [control_guidance_end]
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if len(control_guidance_start) != len(self.controlnet.nets):
+ raise ValueError(
+ f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def prepare_control_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ crops_coords,
+ resize_mode,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.control_image_processor.preprocess(
+ image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
+ ).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ def prepare_latents(
+ self,
+ batch_size,
+ num_channels_latents,
+ height,
+ width,
+ dtype,
+ device,
+ generator,
+ latents=None,
+ image=None,
+ timestep=None,
+ is_strength_max=True,
+ add_noise=True,
+ return_noise=False,
+ return_image_latents=False,
+ ):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if (image is None or timestep is None) and not is_strength_max:
+ raise ValueError(
+ "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
+ "However, either the image or the noise timestep has not been provided."
+ )
+
+ if return_image_latents or (latents is None and not is_strength_max):
+ image = image.to(device=device, dtype=dtype)
+
+ if image.shape[1] == 4:
+ image_latents = image
+ else:
+ image_latents = self._encode_vae_image(image=image, generator=generator)
+ image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
+
+ if latents is None and add_noise:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ # if strength is 1. then initialise the latents to noise, else initial to image + noise
+ latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
+ # if pure noise then scale the initial latents by the Scheduler's init sigma
+ latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+ elif add_noise:
+ noise = latents.to(device)
+ latents = noise * self.scheduler.init_noise_sigma
+ else:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ latents = image_latents.to(device)
+
+ outputs = (latents,)
+
+ if return_noise:
+ outputs += (noise,)
+
+ if return_image_latents:
+ outputs += (image_latents,)
+
+ return outputs
+
+ def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+ dtype = image.dtype
+ if self.vae.config.force_upcast:
+ image = image.float()
+ self.vae.to(dtype=torch.float32)
+
+ if isinstance(generator, list):
+ image_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(image.shape[0])
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ if self.vae.config.force_upcast:
+ self.vae.to(dtype)
+
+ image_latents = image_latents.to(dtype)
+ image_latents = self.vae.config.scaling_factor * image_latents
+
+ return image_latents
+
+ def prepare_mask_latents(
+ self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+ ):
+ # resize the mask to latents shape as we concatenate the mask to the latents
+ # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+ # and half precision
+ mask = torch.nn.functional.interpolate(
+ mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+ )
+ mask = mask.to(device=device, dtype=dtype)
+
+ # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+ if mask.shape[0] < batch_size:
+ if not batch_size % mask.shape[0] == 0:
+ raise ValueError(
+ "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+ f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+ " of masks that you pass is divisible by the total requested batch size."
+ )
+ mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+
+ mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+
+ masked_image_latents = None
+ if masked_image is not None:
+ masked_image = masked_image.to(device=device, dtype=dtype)
+ masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+ if masked_image_latents.shape[0] < batch_size:
+ if not batch_size % masked_image_latents.shape[0] == 0:
+ raise ValueError(
+ "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+ f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+ " Make sure the number of images that you pass is divisible by the total requested batch size."
+ )
+ masked_image_latents = masked_image_latents.repeat(
+ batch_size // masked_image_latents.shape[0], 1, 1, 1
+ )
+
+ masked_image_latents = (
+ torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+ )
+
+ # aligning device to prevent device errors when concating it with the latent model input
+ masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+
+ return mask, masked_image_latents
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
+ # get the original timestep using init_timestep
+ if denoising_start is None:
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+ t_start = max(num_inference_steps - init_timestep, 0)
+ else:
+ t_start = 0
+
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+ # Strength is irrelevant if we directly request a timestep to start at;
+ # that is, strength is determined by the denoising_start instead.
+ if denoising_start is not None:
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (denoising_start * self.scheduler.config.num_train_timesteps)
+ )
+ )
+
+ num_inference_steps = (timesteps < discrete_timestep_cutoff).sum().item()
+ if self.scheduler.order == 2 and num_inference_steps % 2 == 0:
+ # if the scheduler is a 2nd order scheduler we might have to do +1
+ # because `num_inference_steps` might be even given that every timestep
+ # (except the highest one) is duplicated. If `num_inference_steps` is even it would
+ # mean that we cut the timesteps in the middle of the denoising step
+ # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1
+ # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
+ num_inference_steps = num_inference_steps + 1
+
+ # because t_n+1 >= t_n, we slice the timesteps starting from the end
+ timesteps = timesteps[-num_inference_steps:]
+ return timesteps, num_inference_steps
+
+ return timesteps, num_inference_steps - t_start
+
+ def _get_add_time_ids(
+ self,
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ aesthetic_score,
+ negative_aesthetic_score,
+ dtype,
+ text_encoder_projection_dim=None,
+ ):
+ if self.config.requires_aesthetics_score:
+ add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,))
+ add_neg_time_ids = list(original_size + crops_coords_top_left + (negative_aesthetic_score,))
+ else:
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+ add_neg_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if (
+ expected_add_embed_dim > passed_add_embed_dim
+ and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim
+ ):
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model."
+ )
+ elif (
+ expected_add_embed_dim < passed_add_embed_dim
+ and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim
+ ):
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model."
+ )
+ elif expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype)
+
+ return add_time_ids, add_neg_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ image: PipelineImageInput = None,
+ mask_image: PipelineImageInput = None,
+ control_image: Union[
+ PipelineImageInput,
+ List[PipelineImageInput],
+ ] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ padding_mask_crop: Optional[int] = None,
+ strength: float = 0.9999,
+ num_inference_steps: int = 50,
+ denoising_start: Optional[float] = None,
+ denoising_end: Optional[float] = None,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ guidance_rescale: float = 0.0,
+ original_size: Tuple[int, int] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Tuple[int, int] = None,
+ aesthetic_score: float = 6.0,
+ negative_aesthetic_score: float = 2.5,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
+ be masked out with `mask_image` and repainted according to `prompt`.
+ mask_image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+ repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
+ to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
+ instead of 3, so the expected shape would be `(B, H, W, 1)`.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image.
+ padding_mask_crop (`int`, *optional*, defaults to `None`):
+ The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to image and mask_image. If
+ `padding_mask_crop` is not `None`, it will first find a rectangular region with the same aspect ration of the image and
+ contains all masked area, and then expand that area based on `padding_mask_crop`. The image and mask_image will then be cropped based on
+ the expanded area before resizing to the original image size for inpainting. This is useful when the masked area is small while the image is large
+ and contain information inreleant for inpainging, such as background.
+ strength (`float`, *optional*, defaults to 0.9999):
+ Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be
+ between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the
+ `strength`. The number of denoising steps depends on the amount of noise initially added. When
+ `strength` is 1, added noise will be maximum and the denoising process will run for the full number of
+ iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked
+ portion of the reference `image`. Note that in the case of `denoising_start` being declared as an
+ integer, the value of `strength` will be ignored.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ denoising_start (`float`, *optional*):
+ When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be
+ bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and
+ it is assumed that the passed `image` is a partly denoised image. Note that when this is specified,
+ strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline
+ is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
+ denoising_end (`float`, *optional*):
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
+ still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be
+ denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the
+ final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline
+ forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ aesthetic_score (`float`, *optional*, defaults to 6.0):
+ Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
+ simulate an aesthetic score of the generated image by influencing the negative text condition.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+ `tuple. `tuple. When returning a tuple, the first element is a list with the generated images.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+ control_guidance_start, control_guidance_end = (
+ mult * [control_guidance_start],
+ mult * [control_guidance_end],
+ )
+
+ # # 0.0 Default height and width to unet
+ # height = height or self.unet.config.sample_size * self.vae_scale_factor
+ # width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 0.1 align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+ control_guidance_start, control_guidance_end = (
+ mult * [control_guidance_start],
+ mult * [control_guidance_end],
+ )
+
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ control_image,
+ mask_image,
+ strength,
+ num_inference_steps,
+ callback_steps,
+ output_type,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ controlnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ padding_mask_crop,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+ controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ prompt_2=prompt_2,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ negative_prompt_2=negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 3.1 Encode ip_adapter_image
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. set timesteps
+ def denoising_value_valid(dnv):
+ return isinstance(dnv, float) and 0 < dnv < 1
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(
+ num_inference_steps,
+ strength,
+ device,
+ denoising_start=denoising_start if denoising_value_valid(denoising_start) else None,
+ )
+ # check that number of inference steps is not < 1 - as this doesn't make sense
+ if num_inference_steps < 1:
+ raise ValueError(
+ f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+ f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+ )
+ # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+ is_strength_max = strength == 1.0
+ self._num_timesteps = len(timesteps)
+
+ # 5. Preprocess mask and image - resizes image and mask w.r.t height and width
+ # 5.1 Prepare init image
+ if padding_mask_crop is not None:
+ height, width = self.image_processor.get_default_height_width(image, height, width)
+ crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop)
+ resize_mode = "fill"
+ else:
+ crops_coords = None
+ resize_mode = "default"
+
+ original_image = image
+ init_image = self.image_processor.preprocess(
+ image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
+ )
+ init_image = init_image.to(dtype=torch.float32)
+
+ # 5.2 Prepare control images
+ if isinstance(controlnet, ControlNetModel):
+ control_image = self.prepare_control_image(
+ image=control_image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ crops_coords=crops_coords,
+ resize_mode=resize_mode,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ elif isinstance(controlnet, MultiControlNetModel):
+ control_images = []
+
+ for control_image_ in control_image:
+ control_image_ = self.prepare_control_image(
+ image=control_image_,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ crops_coords=crops_coords,
+ resize_mode=resize_mode,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+
+ control_images.append(control_image_)
+
+ control_image = control_images
+ else:
+ raise ValueError(f"{controlnet.__class__} is not supported.")
+
+ # 5.3 Prepare mask
+ mask = self.mask_processor.preprocess(
+ mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
+ )
+
+ masked_image = init_image * (mask < 0.5)
+ _, _, height, width = init_image.shape
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ num_channels_unet = self.unet.config.in_channels
+ return_image_latents = num_channels_unet == 4
+
+ add_noise = True if denoising_start is None else False
+ latents_outputs = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ image=init_image,
+ timestep=latent_timestep,
+ is_strength_max=is_strength_max,
+ add_noise=add_noise,
+ return_noise=True,
+ return_image_latents=return_image_latents,
+ )
+
+ if return_image_latents:
+ latents, noise, image_latents = latents_outputs
+ else:
+ latents, noise = latents_outputs
+
+ # 7. Prepare mask latent variables
+ mask, masked_image_latents = self.prepare_mask_latents(
+ mask,
+ masked_image,
+ batch_size * num_images_per_prompt,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ self.do_classifier_free_guidance,
+ )
+
+ # 8. Check that sizes of mask, masked image and latents match
+ if num_channels_unet == 9:
+ # default case for runwayml/stable-diffusion-inpainting
+ num_channels_mask = mask.shape[1]
+ num_channels_masked_image = masked_image_latents.shape[1]
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+ elif num_channels_unet != 4:
+ raise ValueError(
+ f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
+ )
+ # 8.1 Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 8.2 Create tensor stating which controlnets to keep
+ controlnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ if isinstance(self.controlnet, MultiControlNetModel):
+ controlnet_keep.append(keeps)
+ else:
+ controlnet_keep.append(keeps[0])
+
+ # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ height, width = latents.shape[-2:]
+ height = height * self.vae_scale_factor
+ width = width * self.vae_scale_factor
+
+ original_size = original_size or (height, width)
+ target_size = target_size or (height, width)
+
+ # 10. Prepare added time ids & embeddings
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids, add_neg_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ aesthetic_score,
+ negative_aesthetic_score,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+ add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device)
+
+ # 11. Denoising loop
+ num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+ if (
+ denoising_end is not None
+ and denoising_start is not None
+ and denoising_value_valid(denoising_end)
+ and denoising_value_valid(denoising_start)
+ and denoising_start >= denoising_end
+ ):
+ raise ValueError(
+ f"`denoising_start`: {denoising_start} cannot be larger than or equal to `denoising_end`: "
+ + f" {denoising_end} when using type float."
+ )
+ elif denoising_end is not None and denoising_value_valid(denoising_end):
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (denoising_end * self.scheduler.config.num_train_timesteps)
+ )
+ )
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+ timesteps = timesteps[:num_inference_steps]
+
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+
+ # controlnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer ControlNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ controlnet_added_cond_kwargs = {
+ "text_embeds": add_text_embeds.chunk(2)[1],
+ "time_ids": add_time_ids.chunk(2)[1],
+ }
+ else:
+ control_model_input = latent_model_input
+ controlnet_prompt_embeds = prompt_embeds
+ controlnet_added_cond_kwargs = added_cond_kwargs
+
+ if isinstance(controlnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+ else:
+ controlnet_cond_scale = controlnet_conditioning_scale
+ if isinstance(controlnet_cond_scale, list):
+ controlnet_cond_scale = controlnet_cond_scale[0]
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+ # # Resize control_image to match the size of the input to the controlnet
+ # if control_image.shape[-2:] != control_model_input.shape[-2:]:
+ # control_image = F.interpolate(control_image, size=control_model_input.shape[-2:], mode="bilinear", align_corners=False)
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=controlnet_prompt_embeds,
+ controlnet_cond=control_image,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ added_cond_kwargs=controlnet_added_cond_kwargs,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered ControlNet only for the conditional batch.
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+ if ip_adapter_image is not None:
+ added_cond_kwargs["image_embeds"] = image_embeds
+
+ if num_channels_unet == 9:
+ latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if self.do_classifier_free_guidance and guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if num_channels_unet == 4:
+ init_latents_proper = image_latents
+ if self.do_classifier_free_guidance:
+ init_mask, _ = mask.chunk(2)
+ else:
+ init_mask = mask
+
+ if i < len(timesteps) - 1:
+ noise_timestep = timesteps[i + 1]
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_proper, noise, torch.tensor([noise_timestep])
+ )
+
+ latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ if self.vae.dtype == torch.float16 and self.vae.config.force_upcast:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ # If we do sequential model offloading, let's offload unet and controlnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.controlnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ return StableDiffusionXLPipelineOutput(images=latents)
+
+ # apply watermark if available
+ if self.watermark is not None:
+ image = self.watermark.apply_watermark(image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ if padding_mask_crop is not None:
+ image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image]
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_sd_xl.py b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_sd_xl.py
new file mode 100644
index 0000000000000000000000000000000000000000..eca81083be7be729ec4205c831373eddadc8f163
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_sd_xl.py
@@ -0,0 +1,1499 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from diffusers.utils.import_utils import is_invisible_watermark_available
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import (
+ FromSingleFileMixin,
+ IPAdapterMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, ControlNetModel, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+ from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
+
+from .multicontrolnet import MultiControlNetModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> # !pip install opencv-python transformers accelerate
+ >>> from diffusers import StableDiffusionXLControlNetPipeline, ControlNetModel, AutoencoderKL
+ >>> from diffusers.utils import load_image
+ >>> import numpy as np
+ >>> import torch
+
+ >>> import cv2
+ >>> from PIL import Image
+
+ >>> prompt = "aerial view, a futuristic research complex in a bright foggy jungle, hard lighting"
+ >>> negative_prompt = "low quality, bad quality, sketches"
+
+ >>> # download an image
+ >>> image = load_image(
+ ... "https://hf.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd_controlnet/hf-logo.png"
+ ... )
+
+ >>> # initialize the models and pipeline
+ >>> controlnet_conditioning_scale = 0.5 # recommended for good generalization
+ >>> controlnet = ControlNetModel.from_pretrained(
+ ... "diffusers/controlnet-canny-sdxl-1.0", torch_dtype=torch.float16
+ ... )
+ >>> vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
+ >>> pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16
+ ... )
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> # get canny image
+ >>> image = np.array(image)
+ >>> image = cv2.Canny(image, 100, 200)
+ >>> image = image[:, :, None]
+ >>> image = np.concatenate([image, image, image], axis=2)
+ >>> canny_image = Image.fromarray(image)
+
+ >>> # generate image
+ >>> image = pipe(
+ ... prompt, controlnet_conditioning_scale=controlnet_conditioning_scale, image=canny_image
+ ... ).images[0]
+ ```
+"""
+
+
+class StableDiffusionXLControlNetPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]):
+ Second frozen text-encoder
+ ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ tokenizer_2 ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+ Provides additional conditioning to the `unet` during the denoising process. If you set multiple
+ ControlNets as a list, the outputs from each ControlNet are added together to create one combined
+ additional conditioning.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+ Whether the negative prompt embeddings should always be set to 0. Also see the config of
+ `stabilityai/stable-diffusion-xl-base-1-0`.
+ add_watermarker (`bool`, *optional*):
+ Whether to use the [invisible_watermark](https://github.com/ShieldMnt/invisible-watermark/) library to
+ watermark output images. If not defined, it defaults to `True` if the package is installed; otherwise no
+ watermarker is used.
+ """
+
+ # leave controlnet out on purpose because it iterates with unet
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+ _optional_components = [
+ "tokenizer",
+ "tokenizer_2",
+ "text_encoder",
+ "text_encoder_2",
+ "feature_extractor",
+ "image_encoder",
+ ]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
+ scheduler: KarrasDiffusionSchedulers,
+ force_zeros_for_empty_prompt: bool = True,
+ add_watermarker: Optional[bool] = None,
+ feature_extractor: CLIPImageProcessor = None,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ ):
+ super().__init__()
+
+ if isinstance(controlnet, (list, tuple)):
+ controlnet = MultiControlNetModel(controlnet)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.control_image_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+ )
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+ if add_watermarker:
+ self.watermark = StableDiffusionXLWatermarker()
+ else:
+ self.watermark = None
+
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ image,
+ callback_steps,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ pooled_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ negative_pooled_prompt_embeds=None,
+ controlnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+ )
+
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+ )
+
+ # `prompt` needs more sophisticated handling when there are multiple
+ # conditionings.
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if isinstance(prompt, list):
+ logger.warning(
+ f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+ " prompts. The conditionings will be fixed across the prompts."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ self.check_image(image, prompt, prompt_embeds)
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if not isinstance(image, list):
+ raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+ # When `image` is a nested list:
+ # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+ elif any(isinstance(i, list) for i in image):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif len(image) != len(self.controlnet.nets):
+ raise ValueError(
+ f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+ )
+
+ for image_ in image:
+ self.check_image(image_, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `controlnet_conditioning_scale`
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ if not isinstance(controlnet_conditioning_scale, float):
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if isinstance(controlnet_conditioning_scale, list):
+ if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+ self.controlnet.nets
+ ):
+ raise ValueError(
+ "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+ " the same length as the number of controlnets"
+ )
+ else:
+ assert False
+
+ if not isinstance(control_guidance_start, (tuple, list)):
+ control_guidance_start = [control_guidance_start]
+
+ if not isinstance(control_guidance_end, (tuple, list)):
+ control_guidance_end = [control_guidance_end]
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if len(control_guidance_start) != len(self.controlnet.nets):
+ raise ValueError(
+ f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
+ def check_image(self, image, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
+ def prepare_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
+ def _get_add_time_ids(
+ self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+ ):
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ return add_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def denoising_end(self):
+ return self._denoising_end
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ image: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ denoising_end: Optional[float] = None,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ original_size: Tuple[int, int] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Tuple[int, int] = None,
+ negative_original_size: Optional[Tuple[int, int]] = None,
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+ negative_target_size: Optional[Tuple[int, int]] = None,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The ControlNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single ControlNet.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image. Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image. Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ denoising_end (`float`, *optional*):
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
+ still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+ scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+ "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+ guidance_scale (`float`, *optional*, defaults to 5.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2`
+ and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, pooled text embeddings are generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt
+ weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input
+ argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ The ControlNet encoder tries to recognize the content of the input image even if you remove all
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the ControlNet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the ControlNet stops applying.
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a target image resolution. It should be as same
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned containing the output images.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+ control_guidance_start, control_guidance_end = (
+ mult * [control_guidance_start],
+ mult * [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ image,
+ callback_steps,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ negative_pooled_prompt_embeds,
+ controlnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+ self._denoising_end = denoising_end
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+ controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+ global_pool_conditions = (
+ controlnet.config.global_pool_conditions
+ if isinstance(controlnet, ControlNetModel)
+ else controlnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3.1 Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt,
+ prompt_2,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 3.2 Encode ip_adapter_image
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image
+ if isinstance(controlnet, ControlNetModel):
+ image = self.prepare_image(
+ image=image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ height, width = image.shape[-2:]
+ elif isinstance(controlnet, MultiControlNetModel):
+ images = []
+
+ for image_ in image:
+ image_ = self.prepare_image(
+ image=image_,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+
+ images.append(image_)
+
+ image = images
+ height, width = image[0].shape[-2:]
+ else:
+ assert False
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6.5 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Create tensor stating which controlnets to keep
+ controlnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
+
+ # 7.2 Prepare added time ids & embeddings
+ if isinstance(image, list):
+ original_size = original_size or image[0].shape[-2:]
+ else:
+ original_size = original_size or image.shape[-2:]
+ target_size = target_size or (height, width)
+
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+
+ if negative_original_size is not None and negative_target_size is not None:
+ negative_add_time_ids = self._get_add_time_ids(
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ else:
+ negative_add_time_ids = add_time_ids
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+
+ # 8.1 Apply denoising_end
+ if (
+ self.denoising_end is not None
+ and isinstance(self.denoising_end, float)
+ and self.denoising_end > 0
+ and self.denoising_end < 1
+ ):
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+ )
+ )
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+ timesteps = timesteps[:num_inference_steps]
+
+ is_unet_compiled = is_compiled_module(self.unet)
+ is_controlnet_compiled = is_compiled_module(self.controlnet)
+ is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Relevant thread:
+ # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+ if (is_unet_compiled and is_controlnet_compiled) and is_torch_higher_equal_2_1:
+ torch._inductor.cudagraph_mark_step_begin()
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+
+ # controlnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer ControlNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ controlnet_added_cond_kwargs = {
+ "text_embeds": add_text_embeds.chunk(2)[1],
+ "time_ids": add_time_ids.chunk(2)[1],
+ }
+ else:
+ control_model_input = latent_model_input
+ controlnet_prompt_embeds = prompt_embeds
+ controlnet_added_cond_kwargs = added_cond_kwargs
+
+ if isinstance(controlnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+ else:
+ controlnet_cond_scale = controlnet_conditioning_scale
+ if isinstance(controlnet_cond_scale, list):
+ controlnet_cond_scale = controlnet_cond_scale[0]
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=controlnet_prompt_embeds,
+ controlnet_cond=image,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ added_cond_kwargs=controlnet_added_cond_kwargs,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered ControlNet only for the conditional batch.
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ added_cond_kwargs["image_embeds"] = image_embeds
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ # unscale/denormalize the latents
+ # denormalize with the mean and std if available and not None
+ has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+ has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+ if has_latents_mean and has_latents_std:
+ latents_mean = (
+ torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents_std = (
+ torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+ else:
+ latents = latents / self.vae.config.scaling_factor
+
+ image = self.vae.decode(latents, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ image = latents
+
+ if not output_type == "latent":
+ # apply watermark if available
+ if self.watermark is not None:
+ image = self.watermark.apply_watermark(image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..86a0e2c570d8eadfedc05b54dae55f492e5e0e98
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/pipeline_controlnet_sd_xl_img2img.py
@@ -0,0 +1,1626 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from diffusers.utils.import_utils import is_invisible_watermark_available
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import (
+ IPAdapterMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, ControlNetModel, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+ from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
+
+from .multicontrolnet import MultiControlNetModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> # pip install accelerate transformers safetensors diffusers
+
+ >>> import torch
+ >>> import numpy as np
+ >>> from PIL import Image
+
+ >>> from transformers import DPTFeatureExtractor, DPTForDepthEstimation
+ >>> from diffusers import ControlNetModel, StableDiffusionXLControlNetImg2ImgPipeline, AutoencoderKL
+ >>> from diffusers.utils import load_image
+
+
+ >>> depth_estimator = DPTForDepthEstimation.from_pretrained("Intel/dpt-hybrid-midas").to("cuda")
+ >>> feature_extractor = DPTFeatureExtractor.from_pretrained("Intel/dpt-hybrid-midas")
+ >>> controlnet = ControlNetModel.from_pretrained(
+ ... "diffusers/controlnet-depth-sdxl-1.0-small",
+ ... variant="fp16",
+ ... use_safetensors=True,
+ ... torch_dtype=torch.float16,
+ ... ).to("cuda")
+ >>> vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16).to("cuda")
+ >>> pipe = StableDiffusionXLControlNetImg2ImgPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-xl-base-1.0",
+ ... controlnet=controlnet,
+ ... vae=vae,
+ ... variant="fp16",
+ ... use_safetensors=True,
+ ... torch_dtype=torch.float16,
+ ... ).to("cuda")
+ >>> pipe.enable_model_cpu_offload()
+
+
+ >>> def get_depth_map(image):
+ ... image = feature_extractor(images=image, return_tensors="pt").pixel_values.to("cuda")
+ ... with torch.no_grad(), torch.autocast("cuda"):
+ ... depth_map = depth_estimator(image).predicted_depth
+
+ ... depth_map = torch.nn.functional.interpolate(
+ ... depth_map.unsqueeze(1),
+ ... size=(1024, 1024),
+ ... mode="bicubic",
+ ... align_corners=False,
+ ... )
+ ... depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
+ ... depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
+ ... depth_map = (depth_map - depth_min) / (depth_max - depth_min)
+ ... image = torch.cat([depth_map] * 3, dim=1)
+ ... image = image.permute(0, 2, 3, 1).cpu().numpy()[0]
+ ... image = Image.fromarray((image * 255.0).clip(0, 255).astype(np.uint8))
+ ... return image
+
+
+ >>> prompt = "A robot, 4k photo"
+ >>> image = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... ).resize((1024, 1024))
+ >>> controlnet_conditioning_scale = 0.5 # recommended for good generalization
+ >>> depth_image = get_depth_map(image)
+
+ >>> images = pipe(
+ ... prompt,
+ ... image=image,
+ ... control_image=depth_image,
+ ... strength=0.99,
+ ... num_inference_steps=50,
+ ... controlnet_conditioning_scale=controlnet_conditioning_scale,
+ ... ).images
+ >>> images[0].save(f"robot_cat.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class StableDiffusionXLControlNetImg2ImgPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ IPAdapterMixin,
+):
+ r"""
+ Pipeline for image-to-image generation using Stable Diffusion XL with ControlNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+ specifically the
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+ variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ tokenizer_2 (`CLIPTokenizer`):
+ Second Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+ Provides additional conditioning to the unet during the denoising process. If you set multiple ControlNets
+ as a list, the outputs from each ControlNet are added together to create one combined additional
+ conditioning.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`):
+ Whether the `unet` requires an `aesthetic_score` condition to be passed during inference. Also see the
+ config of `stabilityai/stable-diffusion-xl-refiner-1-0`.
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+ Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+ `stabilityai/stable-diffusion-xl-base-1-0`.
+ add_watermarker (`bool`, *optional*):
+ Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
+ watermark output images. If not defined, it will default to True if the package is installed, otherwise no
+ watermarker will be used.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+ _optional_components = [
+ "tokenizer",
+ "tokenizer_2",
+ "text_encoder",
+ "text_encoder_2",
+ "feature_extractor",
+ "image_encoder",
+ ]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
+ scheduler: KarrasDiffusionSchedulers,
+ requires_aesthetics_score: bool = False,
+ force_zeros_for_empty_prompt: bool = True,
+ add_watermarker: Optional[bool] = None,
+ feature_extractor: CLIPImageProcessor = None,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ ):
+ super().__init__()
+
+ if isinstance(controlnet, (list, tuple)):
+ controlnet = MultiControlNetModel(controlnet)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.control_image_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+ )
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+ if add_watermarker:
+ self.watermark = StableDiffusionXLWatermarker()
+ else:
+ self.watermark = None
+
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ image,
+ strength,
+ num_inference_steps,
+ callback_steps,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ pooled_prompt_embeds=None,
+ negative_pooled_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ controlnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+ if num_inference_steps is None:
+ raise ValueError("`num_inference_steps` cannot be None.")
+ elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0:
+ raise ValueError(
+ f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type"
+ f" {type(num_inference_steps)}."
+ )
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+ )
+
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+ )
+
+ # `prompt` needs more sophisticated handling when there are multiple
+ # conditionings.
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if isinstance(prompt, list):
+ logger.warning(
+ f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+ " prompts. The conditionings will be fixed across the prompts."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ self.check_image(image, prompt, prompt_embeds)
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if not isinstance(image, list):
+ raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+ # When `image` is a nested list:
+ # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+ elif any(isinstance(i, list) for i in image):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif len(image) != len(self.controlnet.nets):
+ raise ValueError(
+ f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+ )
+
+ for image_ in image:
+ self.check_image(image_, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `controlnet_conditioning_scale`
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ if not isinstance(controlnet_conditioning_scale, float):
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if isinstance(controlnet_conditioning_scale, list):
+ if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+ self.controlnet.nets
+ ):
+ raise ValueError(
+ "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+ " the same length as the number of controlnets"
+ )
+ else:
+ assert False
+
+ if not isinstance(control_guidance_start, (tuple, list)):
+ control_guidance_start = [control_guidance_start]
+
+ if not isinstance(control_guidance_end, (tuple, list)):
+ control_guidance_end = [control_guidance_end]
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if len(control_guidance_start) != len(self.controlnet.nets):
+ raise ValueError(
+ f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.check_image
+ def check_image(self, image, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet_sd_xl.StableDiffusionXLControlNetPipeline.prepare_image
+ def prepare_control_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.prepare_latents
+ def prepare_latents(
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True
+ ):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ # Offload text encoder if `enable_model_cpu_offload` was enabled
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.text_encoder_2.to("cpu")
+ torch.cuda.empty_cache()
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ if self.vae.config.force_upcast:
+ image = image.float()
+ self.vae.to(dtype=torch.float32)
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ if self.vae.config.force_upcast:
+ self.vae.to(dtype)
+
+ init_latents = init_latents.to(dtype)
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ if add_noise:
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+
+ latents = init_latents
+
+ return latents
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids
+ def _get_add_time_ids(
+ self,
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ aesthetic_score,
+ negative_aesthetic_score,
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype,
+ text_encoder_projection_dim=None,
+ ):
+ if self.config.requires_aesthetics_score:
+ add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,))
+ add_neg_time_ids = list(
+ negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,)
+ )
+ else:
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+ add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if (
+ expected_add_embed_dim > passed_add_embed_dim
+ and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim
+ ):
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model."
+ )
+ elif (
+ expected_add_embed_dim < passed_add_embed_dim
+ and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim
+ ):
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model."
+ )
+ elif expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype)
+
+ return add_time_ids, add_neg_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ image: PipelineImageInput = None,
+ control_image: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ strength: float = 0.8,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ controlnet_conditioning_scale: Union[float, List[float]] = 0.8,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ original_size: Tuple[int, int] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Tuple[int, int] = None,
+ negative_original_size: Optional[Tuple[int, int]] = None,
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+ negative_target_size: Optional[Tuple[int, int]] = None,
+ aesthetic_score: float = 6.0,
+ negative_aesthetic_score: float = 2.5,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The initial image will be used as the starting point for the image generation process. Can also accept
+ image latents as `image`, if passing latents directly, it will not be encoded again.
+ control_image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If
+ the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can
+ also be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If
+ height and/or width are passed, `image` is resized according to them. If multiple ControlNets are
+ specified in init, images must be passed as a list such that each element of the list can be correctly
+ batched for input to a single controlnet.
+ height (`int`, *optional*, defaults to the size of control_image):
+ The height in pixels of the generated image. Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ width (`int`, *optional*, defaults to the size of control_image):
+ The width in pixels of the generated image. Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
+ to the residual in the original unet. If multiple ControlNets are specified in init, you can set the
+ corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ In this mode, the ControlNet encoder will try best to recognize the content of the input image even if
+ you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the controlnet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the controlnet stops applying.
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a target image resolution. It should be as same
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ aesthetic_score (`float`, *optional*, defaults to 6.0):
+ Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
+ simulate an aesthetic score of the generated image by influencing the negative text condition.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple`
+ containing the output images.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+ control_guidance_start, control_guidance_end = (
+ mult * [control_guidance_start],
+ mult * [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ control_image,
+ strength,
+ num_inference_steps,
+ callback_steps,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ controlnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+ controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+ global_pool_conditions = (
+ controlnet.config.global_pool_conditions
+ if isinstance(controlnet, ControlNetModel)
+ else controlnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3.1. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt,
+ prompt_2,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 3.2 Encode ip_adapter_image
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image and controlnet_conditioning_image
+ image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+
+ if isinstance(controlnet, ControlNetModel):
+ control_image = self.prepare_control_image(
+ image=control_image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ height, width = control_image.shape[-2:]
+ elif isinstance(controlnet, MultiControlNetModel):
+ control_images = []
+
+ for control_image_ in control_image:
+ control_image_ = self.prepare_control_image(
+ image=control_image_,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+
+ control_images.append(control_image_)
+
+ control_image = control_images
+ height, width = control_image[0].shape[-2:]
+ else:
+ assert False
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ latents = self.prepare_latents(
+ image,
+ latent_timestep,
+ batch_size,
+ num_images_per_prompt,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ True,
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Create tensor stating which controlnets to keep
+ controlnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
+
+ # 7.2 Prepare added time ids & embeddings
+ if isinstance(control_image, list):
+ original_size = original_size or control_image[0].shape[-2:]
+ else:
+ original_size = original_size or control_image.shape[-2:]
+ target_size = target_size or (height, width)
+
+ if negative_original_size is None:
+ negative_original_size = original_size
+ if negative_target_size is None:
+ negative_target_size = target_size
+ add_text_embeds = pooled_prompt_embeds
+
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids, add_neg_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ aesthetic_score,
+ negative_aesthetic_score,
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+ add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+
+ # controlnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer ControlNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ controlnet_added_cond_kwargs = {
+ "text_embeds": add_text_embeds.chunk(2)[1],
+ "time_ids": add_time_ids.chunk(2)[1],
+ }
+ else:
+ control_model_input = latent_model_input
+ controlnet_prompt_embeds = prompt_embeds
+ controlnet_added_cond_kwargs = added_cond_kwargs
+
+ if isinstance(controlnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+ else:
+ controlnet_cond_scale = controlnet_conditioning_scale
+ if isinstance(controlnet_cond_scale, list):
+ controlnet_cond_scale = controlnet_cond_scale[0]
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=controlnet_prompt_embeds,
+ controlnet_cond=control_image,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ added_cond_kwargs=controlnet_added_cond_kwargs,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered ControlNet only for the conditional batch.
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ added_cond_kwargs["image_embeds"] = image_embeds
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and controlnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.controlnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ # unscale/denormalize the latents
+ # denormalize with the mean and std if available and not None
+ has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+ has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+ if has_latents_mean and has_latents_std:
+ latents_mean = (
+ torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents_std = (
+ torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+ else:
+ latents = latents / self.vae.config.scaling_factor
+
+ image = self.vae.decode(latents, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ image = latents
+ return StableDiffusionXLPipelineOutput(images=image)
+
+ # apply watermark if available
+ if self.watermark is not None:
+ image = self.watermark.apply_watermark(image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/controlnet/pipeline_flax_controlnet.py b/model/diffusers_c/pipelines/controlnet/pipeline_flax_controlnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..b7e9428fe4fc7b0948f34454d60742df63f4a640
--- /dev/null
+++ b/model/diffusers_c/pipelines/controlnet/pipeline_flax_controlnet.py
@@ -0,0 +1,532 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from functools import partial
+from typing import Dict, List, Optional, Union
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.core.frozen_dict import FrozenDict
+from flax.jax_utils import unreplicate
+from flax.training.common_utils import shard
+from PIL import Image
+from transformers import CLIPFeatureExtractor, CLIPTokenizer, FlaxCLIPTextModel
+
+from ...models import FlaxAutoencoderKL, FlaxControlNetModel, FlaxUNet2DConditionModel
+from ...schedulers import (
+ FlaxDDIMScheduler,
+ FlaxDPMSolverMultistepScheduler,
+ FlaxLMSDiscreteScheduler,
+ FlaxPNDMScheduler,
+)
+from ...utils import PIL_INTERPOLATION, logging, replace_example_docstring
+from ..pipeline_flax_utils import FlaxDiffusionPipeline
+from ..stable_diffusion import FlaxStableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker_flax import FlaxStableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+# Set to True to use python for loop instead of jax.fori_loop for easier debugging
+DEBUG = False
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import jax
+ >>> import numpy as np
+ >>> import jax.numpy as jnp
+ >>> from flax.jax_utils import replicate
+ >>> from flax.training.common_utils import shard
+ >>> from diffusers.utils import load_image, make_image_grid
+ >>> from PIL import Image
+ >>> from diffusers import FlaxStableDiffusionControlNetPipeline, FlaxControlNetModel
+
+
+ >>> def create_key(seed=0):
+ ... return jax.random.PRNGKey(seed)
+
+
+ >>> rng = create_key(0)
+
+ >>> # get canny image
+ >>> canny_image = load_image(
+ ... "https://huggingface.co/datasets/YiYiXu/test-doc-assets/resolve/main/blog_post_cell_10_output_0.jpeg"
+ ... )
+
+ >>> prompts = "best quality, extremely detailed"
+ >>> negative_prompts = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+ >>> # load control net and stable diffusion v1-5
+ >>> controlnet, controlnet_params = FlaxControlNetModel.from_pretrained(
+ ... "lllyasviel/sd-controlnet-canny", from_pt=True, dtype=jnp.float32
+ ... )
+ >>> pipe, params = FlaxStableDiffusionControlNetPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", controlnet=controlnet, revision="flax", dtype=jnp.float32
+ ... )
+ >>> params["controlnet"] = controlnet_params
+
+ >>> num_samples = jax.device_count()
+ >>> rng = jax.random.split(rng, jax.device_count())
+
+ >>> prompt_ids = pipe.prepare_text_inputs([prompts] * num_samples)
+ >>> negative_prompt_ids = pipe.prepare_text_inputs([negative_prompts] * num_samples)
+ >>> processed_image = pipe.prepare_image_inputs([canny_image] * num_samples)
+
+ >>> p_params = replicate(params)
+ >>> prompt_ids = shard(prompt_ids)
+ >>> negative_prompt_ids = shard(negative_prompt_ids)
+ >>> processed_image = shard(processed_image)
+
+ >>> output = pipe(
+ ... prompt_ids=prompt_ids,
+ ... image=processed_image,
+ ... params=p_params,
+ ... prng_seed=rng,
+ ... num_inference_steps=50,
+ ... neg_prompt_ids=negative_prompt_ids,
+ ... jit=True,
+ ... ).images
+
+ >>> output_images = pipe.numpy_to_pil(np.asarray(output.reshape((num_samples,) + output.shape[-3:])))
+ >>> output_images = make_image_grid(output_images, num_samples // 4, 4)
+ >>> output_images.save("generated_image.png")
+ ```
+"""
+
+
+class FlaxStableDiffusionControlNetPipeline(FlaxDiffusionPipeline):
+ r"""
+ Flax-based pipeline for text-to-image generation using Stable Diffusion with ControlNet Guidance.
+
+ This model inherits from [`FlaxDiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`FlaxAutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.FlaxCLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`FlaxUNet2DConditionModel`]):
+ A `FlaxUNet2DConditionModel` to denoise the encoded image latents.
+ controlnet ([`FlaxControlNetModel`]:
+ Provides additional conditioning to the `unet` during the denoising process.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`FlaxDDIMScheduler`], [`FlaxLMSDiscreteScheduler`], [`FlaxPNDMScheduler`], or
+ [`FlaxDPMSolverMultistepScheduler`].
+ safety_checker ([`FlaxStableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ def __init__(
+ self,
+ vae: FlaxAutoencoderKL,
+ text_encoder: FlaxCLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: FlaxUNet2DConditionModel,
+ controlnet: FlaxControlNetModel,
+ scheduler: Union[
+ FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler
+ ],
+ safety_checker: FlaxStableDiffusionSafetyChecker,
+ feature_extractor: CLIPFeatureExtractor,
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ super().__init__()
+ self.dtype = dtype
+
+ if safety_checker is None:
+ logger.warn(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ def prepare_text_inputs(self, prompt: Union[str, List[str]]):
+ if not isinstance(prompt, (str, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ text_input = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+
+ return text_input.input_ids
+
+ def prepare_image_inputs(self, image: Union[Image.Image, List[Image.Image]]):
+ if not isinstance(image, (Image.Image, list)):
+ raise ValueError(f"image has to be of type `PIL.Image.Image` or list but is {type(image)}")
+
+ if isinstance(image, Image.Image):
+ image = [image]
+
+ processed_images = jnp.concatenate([preprocess(img, jnp.float32) for img in image])
+
+ return processed_images
+
+ def _get_has_nsfw_concepts(self, features, params):
+ has_nsfw_concepts = self.safety_checker(features, params)
+ return has_nsfw_concepts
+
+ def _run_safety_checker(self, images, safety_model_params, jit=False):
+ # safety_model_params should already be replicated when jit is True
+ pil_images = [Image.fromarray(image) for image in images]
+ features = self.feature_extractor(pil_images, return_tensors="np").pixel_values
+
+ if jit:
+ features = shard(features)
+ has_nsfw_concepts = _p_get_has_nsfw_concepts(self, features, safety_model_params)
+ has_nsfw_concepts = unshard(has_nsfw_concepts)
+ safety_model_params = unreplicate(safety_model_params)
+ else:
+ has_nsfw_concepts = self._get_has_nsfw_concepts(features, safety_model_params)
+
+ images_was_copied = False
+ for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
+ if has_nsfw_concept:
+ if not images_was_copied:
+ images_was_copied = True
+ images = images.copy()
+
+ images[idx] = np.zeros(images[idx].shape, dtype=np.uint8) # black image
+
+ if any(has_nsfw_concepts):
+ warnings.warn(
+ "Potential NSFW content was detected in one or more images. A black image will be returned"
+ " instead. Try again with a different prompt and/or seed."
+ )
+
+ return images, has_nsfw_concepts
+
+ def _generate(
+ self,
+ prompt_ids: jnp.ndarray,
+ image: jnp.ndarray,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ num_inference_steps: int,
+ guidance_scale: float,
+ latents: Optional[jnp.ndarray] = None,
+ neg_prompt_ids: Optional[jnp.ndarray] = None,
+ controlnet_conditioning_scale: float = 1.0,
+ ):
+ height, width = image.shape[-2:]
+ if height % 64 != 0 or width % 64 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 64 but are {height} and {width}.")
+
+ # get prompt text embeddings
+ prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0]
+
+ # TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0`
+ # implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0`
+ batch_size = prompt_ids.shape[0]
+
+ max_length = prompt_ids.shape[-1]
+
+ if neg_prompt_ids is None:
+ uncond_input = self.tokenizer(
+ [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np"
+ ).input_ids
+ else:
+ uncond_input = neg_prompt_ids
+ negative_prompt_embeds = self.text_encoder(uncond_input, params=params["text_encoder"])[0]
+ context = jnp.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ image = jnp.concatenate([image] * 2)
+
+ latents_shape = (
+ batch_size,
+ self.unet.config.in_channels,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if latents is None:
+ latents = jax.random.normal(prng_seed, shape=latents_shape, dtype=jnp.float32)
+ else:
+ if latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+
+ def loop_body(step, args):
+ latents, scheduler_state = args
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ latents_input = jnp.concatenate([latents] * 2)
+
+ t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step]
+ timestep = jnp.broadcast_to(t, latents_input.shape[0])
+
+ latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t)
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet.apply(
+ {"params": params["controlnet"]},
+ jnp.array(latents_input),
+ jnp.array(timestep, dtype=jnp.int32),
+ encoder_hidden_states=context,
+ controlnet_cond=image,
+ conditioning_scale=controlnet_conditioning_scale,
+ return_dict=False,
+ )
+
+ # predict the noise residual
+ noise_pred = self.unet.apply(
+ {"params": params["unet"]},
+ jnp.array(latents_input),
+ jnp.array(timestep, dtype=jnp.int32),
+ encoder_hidden_states=context,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ ).sample
+
+ # perform guidance
+ noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple()
+ return latents, scheduler_state
+
+ scheduler_state = self.scheduler.set_timesteps(
+ params["scheduler"], num_inference_steps=num_inference_steps, shape=latents_shape
+ )
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * params["scheduler"].init_noise_sigma
+
+ if DEBUG:
+ # run with python for loop
+ for i in range(num_inference_steps):
+ latents, scheduler_state = loop_body(i, (latents, scheduler_state))
+ else:
+ latents, _ = jax.lax.fori_loop(0, num_inference_steps, loop_body, (latents, scheduler_state))
+
+ # scale and decode the image latents with vae
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample
+
+ image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1)
+ return image
+
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt_ids: jnp.ndarray,
+ image: jnp.ndarray,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ num_inference_steps: int = 50,
+ guidance_scale: Union[float, jnp.ndarray] = 7.5,
+ latents: jnp.ndarray = None,
+ neg_prompt_ids: jnp.ndarray = None,
+ controlnet_conditioning_scale: Union[float, jnp.ndarray] = 1.0,
+ return_dict: bool = True,
+ jit: bool = False,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt_ids (`jnp.ndarray`):
+ The prompt or prompts to guide the image generation.
+ image (`jnp.ndarray`):
+ Array representing the ControlNet input condition to provide guidance to the `unet` for generation.
+ params (`Dict` or `FrozenDict`):
+ Dictionary containing the model parameters/weights.
+ prng_seed (`jax.Array`):
+ Array containing random number generator key.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ latents (`jnp.ndarray`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ array is generated by sampling using the supplied random `generator`.
+ controlnet_conditioning_scale (`float` or `jnp.ndarray`, *optional*, defaults to 1.0):
+ The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+ to the residual in the original `unet`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] instead of
+ a plain tuple.
+ jit (`bool`, defaults to `False`):
+ Whether to run `pmap` versions of the generation and safety scoring functions.
+
+
+
+ This argument exists because `__call__` is not yet end-to-end pmap-able. It will be removed in a
+ future release.
+
+
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated images
+ and the second element is a list of `bool`s indicating whether the corresponding generated image
+ contains "not-safe-for-work" (nsfw) content.
+ """
+
+ height, width = image.shape[-2:]
+
+ if isinstance(guidance_scale, float):
+ # Convert to a tensor so each device gets a copy. Follow the prompt_ids for
+ # shape information, as they may be sharded (when `jit` is `True`), or not.
+ guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0])
+ if len(prompt_ids.shape) > 2:
+ # Assume sharded
+ guidance_scale = guidance_scale[:, None]
+
+ if isinstance(controlnet_conditioning_scale, float):
+ # Convert to a tensor so each device gets a copy. Follow the prompt_ids for
+ # shape information, as they may be sharded (when `jit` is `True`), or not.
+ controlnet_conditioning_scale = jnp.array([controlnet_conditioning_scale] * prompt_ids.shape[0])
+ if len(prompt_ids.shape) > 2:
+ # Assume sharded
+ controlnet_conditioning_scale = controlnet_conditioning_scale[:, None]
+
+ if jit:
+ images = _p_generate(
+ self,
+ prompt_ids,
+ image,
+ params,
+ prng_seed,
+ num_inference_steps,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ controlnet_conditioning_scale,
+ )
+ else:
+ images = self._generate(
+ prompt_ids,
+ image,
+ params,
+ prng_seed,
+ num_inference_steps,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ controlnet_conditioning_scale,
+ )
+
+ if self.safety_checker is not None:
+ safety_params = params["safety_checker"]
+ images_uint8_casted = (images * 255).round().astype("uint8")
+ num_devices, batch_size = images.shape[:2]
+
+ images_uint8_casted = np.asarray(images_uint8_casted).reshape(num_devices * batch_size, height, width, 3)
+ images_uint8_casted, has_nsfw_concept = self._run_safety_checker(images_uint8_casted, safety_params, jit)
+ images = np.array(images)
+
+ # block images
+ if any(has_nsfw_concept):
+ for i, is_nsfw in enumerate(has_nsfw_concept):
+ if is_nsfw:
+ images[i] = np.asarray(images_uint8_casted[i])
+
+ images = images.reshape(num_devices, batch_size, height, width, 3)
+ else:
+ images = np.asarray(images)
+ has_nsfw_concept = False
+
+ if not return_dict:
+ return (images, has_nsfw_concept)
+
+ return FlaxStableDiffusionPipelineOutput(images=images, nsfw_content_detected=has_nsfw_concept)
+
+
+# Static argnums are pipe, num_inference_steps. A change would trigger recompilation.
+# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`).
+@partial(
+ jax.pmap,
+ in_axes=(None, 0, 0, 0, 0, None, 0, 0, 0, 0),
+ static_broadcasted_argnums=(0, 5),
+)
+def _p_generate(
+ pipe,
+ prompt_ids,
+ image,
+ params,
+ prng_seed,
+ num_inference_steps,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ controlnet_conditioning_scale,
+):
+ return pipe._generate(
+ prompt_ids,
+ image,
+ params,
+ prng_seed,
+ num_inference_steps,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ controlnet_conditioning_scale,
+ )
+
+
+@partial(jax.pmap, static_broadcasted_argnums=(0,))
+def _p_get_has_nsfw_concepts(pipe, features, params):
+ return pipe._get_has_nsfw_concepts(features, params)
+
+
+def unshard(x: jnp.ndarray):
+ # einops.rearrange(x, 'd b ... -> (d b) ...')
+ num_devices, batch_size = x.shape[:2]
+ rest = x.shape[2:]
+ return x.reshape(num_devices * batch_size, *rest)
+
+
+def preprocess(image, dtype):
+ image = image.convert("RGB")
+ w, h = image.size
+ w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64
+ image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
+ image = jnp.array(image).astype(dtype) / 255.0
+ image = image[None].transpose(0, 3, 1, 2)
+ return image
diff --git a/model/diffusers_c/pipelines/controlnet_xs/__init__.py b/model/diffusers_c/pipelines/controlnet_xs/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/model/diffusers_c/pipelines/controlnet_xs/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/controlnet_xs/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..be758572656efc5343177700f60fa8c8cdb72b83
Binary files /dev/null and b/model/diffusers_c/pipelines/controlnet_xs/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/dance_diffusion/__init__.py b/model/diffusers_c/pipelines/dance_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..0d3e466dfa65b2e9890451607959ed45d092cae7
--- /dev/null
+++ b/model/diffusers_c/pipelines/dance_diffusion/__init__.py
@@ -0,0 +1,18 @@
+from typing import TYPE_CHECKING
+
+from ...utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {"pipeline_dance_diffusion": ["DanceDiffusionPipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_dance_diffusion import DanceDiffusionPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/dance_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/dance_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..98dd516919e9ae83dce6284caf3c3a0d608f3932
Binary files /dev/null and b/model/diffusers_c/pipelines/dance_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/dance_diffusion/__pycache__/pipeline_dance_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/dance_diffusion/__pycache__/pipeline_dance_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6f03b9270acf53a1cdcbf3453d2afbe5e2281e3e
Binary files /dev/null and b/model/diffusers_c/pipelines/dance_diffusion/__pycache__/pipeline_dance_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/dance_diffusion/pipeline_dance_diffusion.py b/model/diffusers_c/pipelines/dance_diffusion/pipeline_dance_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..bcd36c412b54ceb47df088a032161aa88d83950f
--- /dev/null
+++ b/model/diffusers_c/pipelines/dance_diffusion/pipeline_dance_diffusion.py
@@ -0,0 +1,156 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ...utils import logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class DanceDiffusionPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for audio generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ unet ([`UNet1DModel`]):
+ A `UNet1DModel` to denoise the encoded audio.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded audio latents. Can be one of
+ [`IPNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "unet"
+
+ def __init__(self, unet, scheduler):
+ super().__init__()
+ self.register_modules(unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ batch_size: int = 1,
+ num_inference_steps: int = 100,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ audio_length_in_s: Optional[float] = None,
+ return_dict: bool = True,
+ ) -> Union[AudioPipelineOutput, Tuple]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ batch_size (`int`, *optional*, defaults to 1):
+ The number of audio samples to generate.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher-quality audio sample at
+ the expense of slower inference.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ audio_length_in_s (`float`, *optional*, defaults to `self.unet.config.sample_size/self.unet.config.sample_rate`):
+ The length of the generated audio sample in seconds.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.AudioPipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ ```py
+ from diffusers import DiffusionPipeline
+ from scipy.io.wavfile import write
+
+ model_id = "harmonai/maestro-150k"
+ pipe = DiffusionPipeline.from_pretrained(model_id)
+ pipe = pipe.to("cuda")
+
+ audios = pipe(audio_length_in_s=4.0).audios
+
+ # To save locally
+ for i, audio in enumerate(audios):
+ write(f"maestro_test_{i}.wav", pipe.unet.sample_rate, audio.transpose())
+
+ # To dislay in google colab
+ import IPython.display as ipd
+
+ for audio in audios:
+ display(ipd.Audio(audio, rate=pipe.unet.sample_rate))
+ ```
+
+ Returns:
+ [`~pipelines.AudioPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.AudioPipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated audio.
+ """
+
+ if audio_length_in_s is None:
+ audio_length_in_s = self.unet.config.sample_size / self.unet.config.sample_rate
+
+ sample_size = audio_length_in_s * self.unet.config.sample_rate
+
+ down_scale_factor = 2 ** len(self.unet.up_blocks)
+ if sample_size < 3 * down_scale_factor:
+ raise ValueError(
+ f"{audio_length_in_s} is too small. Make sure it's bigger or equal to"
+ f" {3 * down_scale_factor / self.unet.config.sample_rate}."
+ )
+
+ original_sample_size = int(sample_size)
+ if sample_size % down_scale_factor != 0:
+ sample_size = (
+ (audio_length_in_s * self.unet.config.sample_rate) // down_scale_factor + 1
+ ) * down_scale_factor
+ logger.info(
+ f"{audio_length_in_s} is increased to {sample_size / self.unet.config.sample_rate} so that it can be handled"
+ f" by the model. It will be cut to {original_sample_size / self.unet.config.sample_rate} after the denoising"
+ " process."
+ )
+ sample_size = int(sample_size)
+
+ dtype = next(self.unet.parameters()).dtype
+ shape = (batch_size, self.unet.config.in_channels, sample_size)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ audio = randn_tensor(shape, generator=generator, device=self._execution_device, dtype=dtype)
+
+ # set step values
+ self.scheduler.set_timesteps(num_inference_steps, device=audio.device)
+ self.scheduler.timesteps = self.scheduler.timesteps.to(dtype)
+
+ for t in self.progress_bar(self.scheduler.timesteps):
+ # 1. predict noise model_output
+ model_output = self.unet(audio, t).sample
+
+ # 2. compute previous audio sample: x_t -> t_t-1
+ audio = self.scheduler.step(model_output, t, audio).prev_sample
+
+ audio = audio.clamp(-1, 1).float().cpu().numpy()
+
+ audio = audio[:, :, :original_sample_size]
+
+ if not return_dict:
+ return (audio,)
+
+ return AudioPipelineOutput(audios=audio)
diff --git a/model/diffusers_c/pipelines/ddim/__init__.py b/model/diffusers_c/pipelines/ddim/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..d9eede47c897370a23c47c05291690881c987025
--- /dev/null
+++ b/model/diffusers_c/pipelines/ddim/__init__.py
@@ -0,0 +1,18 @@
+from typing import TYPE_CHECKING
+
+from ...utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {"pipeline_ddim": ["DDIMPipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_ddim import DDIMPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/ddim/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/ddim/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..355c66f10389d3a0fb2ce1b29e9204331128a44b
Binary files /dev/null and b/model/diffusers_c/pipelines/ddim/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/ddim/__pycache__/pipeline_ddim.cpython-39.pyc b/model/diffusers_c/pipelines/ddim/__pycache__/pipeline_ddim.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1a37cf63e666f4ffe673a52b49eeb042f2bf29de
Binary files /dev/null and b/model/diffusers_c/pipelines/ddim/__pycache__/pipeline_ddim.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/ddim/pipeline_ddim.py b/model/diffusers_c/pipelines/ddim/pipeline_ddim.py
new file mode 100644
index 0000000000000000000000000000000000000000..a3b967ed369bd3e7287aa6dc628ecad6484a6359
--- /dev/null
+++ b/model/diffusers_c/pipelines/ddim/pipeline_ddim.py
@@ -0,0 +1,154 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ...schedulers import DDIMScheduler
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+class DDIMPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for image generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image. Can be one of
+ [`DDPMScheduler`], or [`DDIMScheduler`].
+ """
+
+ model_cpu_offload_seq = "unet"
+
+ def __init__(self, unet, scheduler):
+ super().__init__()
+
+ # make sure scheduler can always be converted to DDIM
+ scheduler = DDIMScheduler.from_config(scheduler.config)
+
+ self.register_modules(unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ batch_size: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ eta: float = 0.0,
+ num_inference_steps: int = 50,
+ use_clipped_model_output: Optional[bool] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ) -> Union[ImagePipelineOutput, Tuple]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ batch_size (`int`, *optional*, defaults to 1):
+ The number of images to generate.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. A value of `0` corresponds to
+ DDIM and `1` corresponds to DDPM.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ use_clipped_model_output (`bool`, *optional*, defaults to `None`):
+ If `True` or `False`, see documentation for [`DDIMScheduler.step`]. If `None`, nothing is passed
+ downstream to the scheduler (use `None` for schedulers which don't support this argument).
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ ```py
+ >>> from diffusers import DDIMPipeline
+ >>> import PIL.Image
+ >>> import numpy as np
+
+ >>> # load model and scheduler
+ >>> pipe = DDIMPipeline.from_pretrained("fusing/ddim-lsun-bedroom")
+
+ >>> # run pipeline in inference (sample random noise and denoise)
+ >>> image = pipe(eta=0.0, num_inference_steps=50)
+
+ >>> # process image to PIL
+ >>> image_processed = image.cpu().permute(0, 2, 3, 1)
+ >>> image_processed = (image_processed + 1.0) * 127.5
+ >>> image_processed = image_processed.numpy().astype(np.uint8)
+ >>> image_pil = PIL.Image.fromarray(image_processed[0])
+
+ >>> # save image
+ >>> image_pil.save("test.png")
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images
+ """
+
+ # Sample gaussian noise to begin loop
+ if isinstance(self.unet.config.sample_size, int):
+ image_shape = (
+ batch_size,
+ self.unet.config.in_channels,
+ self.unet.config.sample_size,
+ self.unet.config.sample_size,
+ )
+ else:
+ image_shape = (batch_size, self.unet.config.in_channels, *self.unet.config.sample_size)
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ image = randn_tensor(image_shape, generator=generator, device=self._execution_device, dtype=self.unet.dtype)
+
+ # set step values
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ for t in self.progress_bar(self.scheduler.timesteps):
+ # 1. predict noise model_output
+ model_output = self.unet(image, t).sample
+
+ # 2. predict previous mean of image x_t-1 and add variance depending on eta
+ # eta corresponds to η in paper and should be between [0, 1]
+ # do x_t -> x_t-1
+ image = self.scheduler.step(
+ model_output, t, image, eta=eta, use_clipped_model_output=use_clipped_model_output, generator=generator
+ ).prev_sample
+
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/ddpm/__init__.py b/model/diffusers_c/pipelines/ddpm/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..eb41dd1dcf642c791f3d7b0d985efcaf3e4a2c22
--- /dev/null
+++ b/model/diffusers_c/pipelines/ddpm/__init__.py
@@ -0,0 +1,22 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ _LazyModule,
+)
+
+
+_import_structure = {"pipeline_ddpm": ["DDPMPipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_ddpm import DDPMPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/ddpm/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/ddpm/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2c1cfe9deea742335c2fd695da9faf2c148e8634
Binary files /dev/null and b/model/diffusers_c/pipelines/ddpm/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/ddpm/__pycache__/pipeline_ddpm.cpython-39.pyc b/model/diffusers_c/pipelines/ddpm/__pycache__/pipeline_ddpm.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..222bd4a2d2b912a14e2e7dff1ee15fca3dc9bf4c
Binary files /dev/null and b/model/diffusers_c/pipelines/ddpm/__pycache__/pipeline_ddpm.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/ddpm/pipeline_ddpm.py b/model/diffusers_c/pipelines/ddpm/pipeline_ddpm.py
new file mode 100644
index 0000000000000000000000000000000000000000..093a3cdfe512a7fdda0529bc2c6d47721508620e
--- /dev/null
+++ b/model/diffusers_c/pipelines/ddpm/pipeline_ddpm.py
@@ -0,0 +1,127 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+class DDPMPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for image generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image. Can be one of
+ [`DDPMScheduler`], or [`DDIMScheduler`].
+ """
+
+ model_cpu_offload_seq = "unet"
+
+ def __init__(self, unet, scheduler):
+ super().__init__()
+ self.register_modules(unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ batch_size: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ num_inference_steps: int = 1000,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ) -> Union[ImagePipelineOutput, Tuple]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ batch_size (`int`, *optional*, defaults to 1):
+ The number of images to generate.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ num_inference_steps (`int`, *optional*, defaults to 1000):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ ```py
+ >>> from diffusers import DDPMPipeline
+
+ >>> # load model and scheduler
+ >>> pipe = DDPMPipeline.from_pretrained("google/ddpm-cat-256")
+
+ >>> # run pipeline in inference (sample random noise and denoise)
+ >>> image = pipe().images[0]
+
+ >>> # save image
+ >>> image.save("ddpm_generated_image.png")
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images
+ """
+ # Sample gaussian noise to begin loop
+ if isinstance(self.unet.config.sample_size, int):
+ image_shape = (
+ batch_size,
+ self.unet.config.in_channels,
+ self.unet.config.sample_size,
+ self.unet.config.sample_size,
+ )
+ else:
+ image_shape = (batch_size, self.unet.config.in_channels, *self.unet.config.sample_size)
+
+ if self.device.type == "mps":
+ # randn does not work reproducibly on mps
+ image = randn_tensor(image_shape, generator=generator)
+ image = image.to(self.device)
+ else:
+ image = randn_tensor(image_shape, generator=generator, device=self.device)
+
+ # set step values
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ for t in self.progress_bar(self.scheduler.timesteps):
+ # 1. predict noise model_output
+ model_output = self.unet(image, t).sample
+
+ # 2. compute previous image: x_t -> x_t-1
+ image = self.scheduler.step(model_output, t, image, generator=generator).prev_sample
+
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__init__.py b/model/diffusers_c/pipelines/deepfloyd_if/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..79aab1fb186a857dd0a3353c4b5905b4595b5b7b
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/__init__.py
@@ -0,0 +1,85 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {
+ "timesteps": [
+ "fast27_timesteps",
+ "smart100_timesteps",
+ "smart185_timesteps",
+ "smart27_timesteps",
+ "smart50_timesteps",
+ "super100_timesteps",
+ "super27_timesteps",
+ "super40_timesteps",
+ ]
+}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_if"] = ["IFPipeline"]
+ _import_structure["pipeline_if_img2img"] = ["IFImg2ImgPipeline"]
+ _import_structure["pipeline_if_img2img_superresolution"] = ["IFImg2ImgSuperResolutionPipeline"]
+ _import_structure["pipeline_if_inpainting"] = ["IFInpaintingPipeline"]
+ _import_structure["pipeline_if_inpainting_superresolution"] = ["IFInpaintingSuperResolutionPipeline"]
+ _import_structure["pipeline_if_superresolution"] = ["IFSuperResolutionPipeline"]
+ _import_structure["pipeline_output"] = ["IFPipelineOutput"]
+ _import_structure["safety_checker"] = ["IFSafetyChecker"]
+ _import_structure["watermark"] = ["IFWatermarker"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_if import IFPipeline
+ from .pipeline_if_img2img import IFImg2ImgPipeline
+ from .pipeline_if_img2img_superresolution import IFImg2ImgSuperResolutionPipeline
+ from .pipeline_if_inpainting import IFInpaintingPipeline
+ from .pipeline_if_inpainting_superresolution import IFInpaintingSuperResolutionPipeline
+ from .pipeline_if_superresolution import IFSuperResolutionPipeline
+ from .pipeline_output import IFPipelineOutput
+ from .safety_checker import IFSafetyChecker
+ from .timesteps import (
+ fast27_timesteps,
+ smart27_timesteps,
+ smart50_timesteps,
+ smart100_timesteps,
+ smart185_timesteps,
+ super27_timesteps,
+ super40_timesteps,
+ super100_timesteps,
+ )
+ from .watermark import IFWatermarker
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..42aaa41b79ad4aa30d7422d4645b3c9aea77ec80
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..00bea2703593cdab5c74ef266512cd8d61adb89b
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0c4c88c81589269efbaf2485139cd8f0a1b4e3ee
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_img2img_superresolution.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_img2img_superresolution.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..935b9aeb9338fcaba5193d882b5e08038fc36fd4
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_img2img_superresolution.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_inpainting.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_inpainting.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..80f0daf19fa1ebf4f6f7ad41615dc9ca8849707f
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_inpainting.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_inpainting_superresolution.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_inpainting_superresolution.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..66d2a151fa08bcba89874f90e712b1d218d43720
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_inpainting_superresolution.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_superresolution.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_superresolution.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1b2a10ae825152e0ef45394f751f92ab5cb3a332
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_if_superresolution.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_output.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_output.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2ec331aff9324f832b68d439d061ca5f70816ae4
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/pipeline_output.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/safety_checker.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/safety_checker.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..548bfef8d9c263dc70610ad118a023daabfe99c7
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/safety_checker.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/watermark.cpython-39.pyc b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/watermark.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f745b7270a238d7d359e2a2638f598fec0400584
Binary files /dev/null and b/model/diffusers_c/pipelines/deepfloyd_if/__pycache__/watermark.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if.py b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if.py
new file mode 100644
index 0000000000000000000000000000000000000000..64806d783d5173a3f12fd5f593e5dc87531d493d
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if.py
@@ -0,0 +1,788 @@
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import CLIPImageProcessor, T5EncoderModel, T5Tokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...models import UNet2DConditionModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ BACKENDS_MAPPING,
+ is_accelerate_available,
+ is_bs4_available,
+ is_ftfy_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import IFPipelineOutput
+from .safety_checker import IFSafetyChecker
+from .watermark import IFWatermarker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+if is_bs4_available():
+ from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+ import ftfy
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import IFPipeline, IFSuperResolutionPipeline, DiffusionPipeline
+ >>> from diffusers.utils import pt_to_pil
+ >>> import torch
+
+ >>> pipe = IFPipeline.from_pretrained("DeepFloyd/IF-I-XL-v1.0", variant="fp16", torch_dtype=torch.float16)
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = 'a photo of a kangaroo wearing an orange hoodie and blue sunglasses standing in front of the eiffel tower holding a sign that says "very deep learning"'
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
+
+ >>> image = pipe(prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds, output_type="pt").images
+
+ >>> # save intermediate image
+ >>> pil_image = pt_to_pil(image)
+ >>> pil_image[0].save("./if_stage_I.png")
+
+ >>> super_res_1_pipe = IFSuperResolutionPipeline.from_pretrained(
+ ... "DeepFloyd/IF-II-L-v1.0", text_encoder=None, variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> super_res_1_pipe.enable_model_cpu_offload()
+
+ >>> image = super_res_1_pipe(
+ ... image=image, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds, output_type="pt"
+ ... ).images
+
+ >>> # save intermediate image
+ >>> pil_image = pt_to_pil(image)
+ >>> pil_image[0].save("./if_stage_I.png")
+
+ >>> safety_modules = {
+ ... "feature_extractor": pipe.feature_extractor,
+ ... "safety_checker": pipe.safety_checker,
+ ... "watermarker": pipe.watermarker,
+ ... }
+ >>> super_res_2_pipe = DiffusionPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-x4-upscaler", **safety_modules, torch_dtype=torch.float16
+ ... )
+ >>> super_res_2_pipe.enable_model_cpu_offload()
+
+ >>> image = super_res_2_pipe(
+ ... prompt=prompt,
+ ... image=image,
+ ... ).images
+ >>> image[0].save("./if_stage_II.png")
+ ```
+"""
+
+
+class IFPipeline(DiffusionPipeline, LoraLoaderMixin):
+ tokenizer: T5Tokenizer
+ text_encoder: T5EncoderModel
+
+ unet: UNet2DConditionModel
+ scheduler: DDPMScheduler
+
+ feature_extractor: Optional[CLIPImageProcessor]
+ safety_checker: Optional[IFSafetyChecker]
+
+ watermarker: Optional[IFWatermarker]
+
+ bad_punct_regex = re.compile(
+ r"["
+ + "#®•©™&@·º½¾¿¡§~"
+ + r"\)"
+ + r"\("
+ + r"\]"
+ + r"\["
+ + r"\}"
+ + r"\{"
+ + r"\|"
+ + "\\"
+ + r"\/"
+ + r"\*"
+ + r"]{1,}"
+ ) # noqa
+
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor", "watermarker"]
+ model_cpu_offload_seq = "text_encoder->unet"
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ safety_checker: Optional[IFSafetyChecker],
+ feature_extractor: Optional[CLIPImageProcessor],
+ watermarker: Optional[IFWatermarker],
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ watermarker=watermarker,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ def remove_all_hooks(self):
+ if is_accelerate_available():
+ from accelerate.hooks import remove_hook_from_module
+ else:
+ raise ImportError("Please install accelerate via `pip install accelerate`")
+
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
+ if model is not None:
+ remove_hook_from_module(model, recurse=True)
+
+ self.unet_offload_hook = None
+ self.text_encoder_offload_hook = None
+ self.final_offload_hook = None
+
+ @torch.no_grad()
+ def encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ do_classifier_free_guidance: bool = True,
+ num_images_per_prompt: int = 1,
+ device: Optional[torch.device] = None,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ clean_caption: bool = False,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ clean_caption (bool, defaults to `False`):
+ If `True`, the function will preprocess and clean the provided caption before encoding.
+ """
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
+ max_length = 77
+
+ if prompt_embeds is None:
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {max_length} tokens: {removed_text}"
+ )
+
+ attention_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ dtype = self.unet.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ else:
+ negative_prompt_embeds = None
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+ image, nsfw_detected, watermark_detected = self.safety_checker(
+ images=image,
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
+ )
+ else:
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+
+ return image, nsfw_detected, watermark_detected
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def prepare_intermediate_images(self, batch_size, num_channels, height, width, dtype, device, generator):
+ shape = (batch_size, num_channels, height, width)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ intermediate_images = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ intermediate_images = intermediate_images * self.scheduler.init_noise_sigma
+ return intermediate_images
+
+ def _text_preprocessing(self, text, clean_caption=False):
+ if clean_caption and not is_bs4_available():
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if clean_caption and not is_ftfy_available():
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if not isinstance(text, (tuple, list)):
+ text = [text]
+
+ def process(text: str):
+ if clean_caption:
+ text = self._clean_caption(text)
+ text = self._clean_caption(text)
+ else:
+ text = text.lower().strip()
+ return text
+
+ return [process(t) for t in text]
+
+ def _clean_caption(self, caption):
+ caption = str(caption)
+ caption = ul.unquote_plus(caption)
+ caption = caption.strip().lower()
+ caption = re.sub("", "person", caption)
+ # urls:
+ caption = re.sub(
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ caption = re.sub(
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ # html:
+ caption = BeautifulSoup(caption, features="html.parser").text
+
+ # @
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+ # 31C0—31EF CJK Strokes
+ # 31F0—31FF Katakana Phonetic Extensions
+ # 3200—32FF Enclosed CJK Letters and Months
+ # 3300—33FF CJK Compatibility
+ # 3400—4DBF CJK Unified Ideographs Extension A
+ # 4DC0—4DFF Yijing Hexagram Symbols
+ # 4E00—9FFF CJK Unified Ideographs
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+ #######################################################
+
+ # все виды тире / all types of dash --> "-"
+ caption = re.sub(
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
+ "-",
+ caption,
+ )
+
+ # кавычки к одному стандарту
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
+ caption = re.sub(r"[‘’]", "'", caption)
+
+ # "
+ caption = re.sub(r""?", "", caption)
+ # &
+ caption = re.sub(r"&", "", caption)
+
+ # ip adresses:
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+ # article ids:
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+ # \n
+ caption = re.sub(r"\\n", " ", caption)
+
+ # "#123"
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
+ # "#12345.."
+ caption = re.sub(r"#\d{5,}\b", "", caption)
+ # "123456.."
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
+ # filenames:
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+ #
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
+
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
+
+ # this-is-my-cute-cat / this_is_my_cute_cat
+ regex2 = re.compile(r"(?:\-|\_)")
+ if len(re.findall(regex2, caption)) > 3:
+ caption = re.sub(regex2, " ", caption)
+
+ caption = ftfy.fix_text(caption)
+ caption = html.unescape(html.unescape(caption))
+
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
+
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
+
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+ caption = re.sub(r"\s+", " ", caption)
+
+ caption.strip()
+
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+ caption = re.sub(r"^\.\S+$", "", caption)
+
+ return caption.strip()
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ num_inference_steps: int = 100,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ clean_caption: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size):
+ The width in pixels of the generated image.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ clean_caption (`bool`, *optional*, defaults to `True`):
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
+ prompt.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
+ or watermarked content, according to the `safety_checker`.
+ """
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+ # 2. Define call parameters
+ height = height or self.unet.config.sample_size
+ width = width or self.unet.config.sample_size
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ do_classifier_free_guidance,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clean_caption=clean_caption,
+ )
+
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare intermediate images
+ intermediate_images = self.prepare_intermediate_images(
+ batch_size * num_images_per_prompt,
+ self.unet.config.in_channels,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # HACK: see comment in `enable_model_cpu_offload`
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
+ self.text_encoder_offload_hook.offload()
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ model_input = (
+ torch.cat([intermediate_images] * 2) if do_classifier_free_guidance else intermediate_images
+ )
+ model_input = self.scheduler.scale_model_input(model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1], dim=1)
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1], dim=1)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
+
+ if self.scheduler.config.variance_type not in ["learned", "learned_range"]:
+ noise_pred, _ = noise_pred.split(model_input.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ intermediate_images = self.scheduler.step(
+ noise_pred, t, intermediate_images, **extra_step_kwargs, return_dict=False
+ )[0]
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, intermediate_images)
+
+ image = intermediate_images
+
+ if output_type == "pil":
+ # 8. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 9. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # 10. Convert to PIL
+ image = self.numpy_to_pil(image)
+
+ # 11. Apply watermark
+ if self.watermarker is not None:
+ image = self.watermarker.apply_watermark(image, self.unet.config.sample_size)
+ elif output_type == "pt":
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+ else:
+ # 8. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 9. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, nsfw_detected, watermark_detected)
+
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_img2img.py b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..6ec4ce6f11f97f1f8ec1b364cb00e60ad50ec6a9
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_img2img.py
@@ -0,0 +1,910 @@
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, T5EncoderModel, T5Tokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...models import UNet2DConditionModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ BACKENDS_MAPPING,
+ PIL_INTERPOLATION,
+ is_accelerate_available,
+ is_bs4_available,
+ is_ftfy_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import IFPipelineOutput
+from .safety_checker import IFSafetyChecker
+from .watermark import IFWatermarker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+if is_bs4_available():
+ from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+ import ftfy
+
+
+def resize(images: PIL.Image.Image, img_size: int) -> PIL.Image.Image:
+ w, h = images.size
+
+ coef = w / h
+
+ w, h = img_size, img_size
+
+ if coef >= 1:
+ w = int(round(img_size / 8 * coef) * 8)
+ else:
+ h = int(round(img_size / 8 / coef) * 8)
+
+ images = images.resize((w, h), resample=PIL_INTERPOLATION["bicubic"], reducing_gap=None)
+
+ return images
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import IFImg2ImgPipeline, IFImg2ImgSuperResolutionPipeline, DiffusionPipeline
+ >>> from diffusers.utils import pt_to_pil
+ >>> import torch
+ >>> from PIL import Image
+ >>> import requests
+ >>> from io import BytesIO
+
+ >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+ >>> response = requests.get(url)
+ >>> original_image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> original_image = original_image.resize((768, 512))
+
+ >>> pipe = IFImg2ImgPipeline.from_pretrained(
+ ... "DeepFloyd/IF-I-XL-v1.0",
+ ... variant="fp16",
+ ... torch_dtype=torch.float16,
+ ... )
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = "A fantasy landscape in style minecraft"
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
+
+ >>> image = pipe(
+ ... image=original_image,
+ ... prompt_embeds=prompt_embeds,
+ ... negative_prompt_embeds=negative_embeds,
+ ... output_type="pt",
+ ... ).images
+
+ >>> # save intermediate image
+ >>> pil_image = pt_to_pil(image)
+ >>> pil_image[0].save("./if_stage_I.png")
+
+ >>> super_res_1_pipe = IFImg2ImgSuperResolutionPipeline.from_pretrained(
+ ... "DeepFloyd/IF-II-L-v1.0",
+ ... text_encoder=None,
+ ... variant="fp16",
+ ... torch_dtype=torch.float16,
+ ... )
+ >>> super_res_1_pipe.enable_model_cpu_offload()
+
+ >>> image = super_res_1_pipe(
+ ... image=image,
+ ... original_image=original_image,
+ ... prompt_embeds=prompt_embeds,
+ ... negative_prompt_embeds=negative_embeds,
+ ... ).images
+ >>> image[0].save("./if_stage_II.png")
+ ```
+"""
+
+
+class IFImg2ImgPipeline(DiffusionPipeline, LoraLoaderMixin):
+ tokenizer: T5Tokenizer
+ text_encoder: T5EncoderModel
+
+ unet: UNet2DConditionModel
+ scheduler: DDPMScheduler
+
+ feature_extractor: Optional[CLIPImageProcessor]
+ safety_checker: Optional[IFSafetyChecker]
+
+ watermarker: Optional[IFWatermarker]
+
+ bad_punct_regex = re.compile(
+ r"["
+ + "#®•©™&@·º½¾¿¡§~"
+ + r"\)"
+ + r"\("
+ + r"\]"
+ + r"\["
+ + r"\}"
+ + r"\{"
+ + r"\|"
+ + "\\"
+ + r"\/"
+ + r"\*"
+ + r"]{1,}"
+ ) # noqa
+
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor", "watermarker"]
+ model_cpu_offload_seq = "text_encoder->unet"
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ safety_checker: Optional[IFSafetyChecker],
+ feature_extractor: Optional[CLIPImageProcessor],
+ watermarker: Optional[IFWatermarker],
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ watermarker=watermarker,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.remove_all_hooks
+ def remove_all_hooks(self):
+ if is_accelerate_available():
+ from accelerate.hooks import remove_hook_from_module
+ else:
+ raise ImportError("Please install accelerate via `pip install accelerate`")
+
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
+ if model is not None:
+ remove_hook_from_module(model, recurse=True)
+
+ self.unet_offload_hook = None
+ self.text_encoder_offload_hook = None
+ self.final_offload_hook = None
+
+ @torch.no_grad()
+ def encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ do_classifier_free_guidance: bool = True,
+ num_images_per_prompt: int = 1,
+ device: Optional[torch.device] = None,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ clean_caption: bool = False,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ clean_caption (bool, defaults to `False`):
+ If `True`, the function will preprocess and clean the provided caption before encoding.
+ """
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
+ max_length = 77
+
+ if prompt_embeds is None:
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {max_length} tokens: {removed_text}"
+ )
+
+ attention_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ dtype = self.unet.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ else:
+ negative_prompt_embeds = None
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+ image, nsfw_detected, watermark_detected = self.safety_checker(
+ images=image,
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
+ )
+ else:
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+
+ return image, nsfw_detected, watermark_detected
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ batch_size,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if isinstance(image, list):
+ check_image_type = image[0]
+ else:
+ check_image_type = image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(image, list):
+ image_batch_size = len(image)
+ elif isinstance(image, torch.Tensor):
+ image_batch_size = image.shape[0]
+ elif isinstance(image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(image, np.ndarray):
+ image_batch_size = image.shape[0]
+ else:
+ assert False
+
+ if batch_size != image_batch_size:
+ raise ValueError(f"image batch size: {image_batch_size} must be same as prompt batch size {batch_size}")
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+ def _text_preprocessing(self, text, clean_caption=False):
+ if clean_caption and not is_bs4_available():
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if clean_caption and not is_ftfy_available():
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if not isinstance(text, (tuple, list)):
+ text = [text]
+
+ def process(text: str):
+ if clean_caption:
+ text = self._clean_caption(text)
+ text = self._clean_caption(text)
+ else:
+ text = text.lower().strip()
+ return text
+
+ return [process(t) for t in text]
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+ def _clean_caption(self, caption):
+ caption = str(caption)
+ caption = ul.unquote_plus(caption)
+ caption = caption.strip().lower()
+ caption = re.sub("", "person", caption)
+ # urls:
+ caption = re.sub(
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ caption = re.sub(
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ # html:
+ caption = BeautifulSoup(caption, features="html.parser").text
+
+ # @
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+ # 31C0—31EF CJK Strokes
+ # 31F0—31FF Katakana Phonetic Extensions
+ # 3200—32FF Enclosed CJK Letters and Months
+ # 3300—33FF CJK Compatibility
+ # 3400—4DBF CJK Unified Ideographs Extension A
+ # 4DC0—4DFF Yijing Hexagram Symbols
+ # 4E00—9FFF CJK Unified Ideographs
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+ #######################################################
+
+ # все виды тире / all types of dash --> "-"
+ caption = re.sub(
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
+ "-",
+ caption,
+ )
+
+ # кавычки к одному стандарту
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
+ caption = re.sub(r"[‘’]", "'", caption)
+
+ # "
+ caption = re.sub(r""?", "", caption)
+ # &
+ caption = re.sub(r"&", "", caption)
+
+ # ip adresses:
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+ # article ids:
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+ # \n
+ caption = re.sub(r"\\n", " ", caption)
+
+ # "#123"
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
+ # "#12345.."
+ caption = re.sub(r"#\d{5,}\b", "", caption)
+ # "123456.."
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
+ # filenames:
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+ #
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
+
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
+
+ # this-is-my-cute-cat / this_is_my_cute_cat
+ regex2 = re.compile(r"(?:\-|\_)")
+ if len(re.findall(regex2, caption)) > 3:
+ caption = re.sub(regex2, " ", caption)
+
+ caption = ftfy.fix_text(caption)
+ caption = html.unescape(html.unescape(caption))
+
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
+
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
+
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+ caption = re.sub(r"\s+", " ", caption)
+
+ caption.strip()
+
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+ caption = re.sub(r"^\.\S+$", "", caption)
+
+ return caption.strip()
+
+ def preprocess_image(self, image: PIL.Image.Image) -> torch.Tensor:
+ if not isinstance(image, list):
+ image = [image]
+
+ def numpy_to_pt(images):
+ if images.ndim == 3:
+ images = images[..., None]
+
+ images = torch.from_numpy(images.transpose(0, 3, 1, 2))
+ return images
+
+ if isinstance(image[0], PIL.Image.Image):
+ new_image = []
+
+ for image_ in image:
+ image_ = image_.convert("RGB")
+ image_ = resize(image_, self.unet.sample_size)
+ image_ = np.array(image_)
+ image_ = image_.astype(np.float32)
+ image_ = image_ / 127.5 - 1
+ new_image.append(image_)
+
+ image = new_image
+
+ image = np.stack(image, axis=0) # to np
+ image = numpy_to_pt(image) # to pt
+
+ elif isinstance(image[0], np.ndarray):
+ image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
+ image = numpy_to_pt(image)
+
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
+
+ return image
+
+ def get_timesteps(self, num_inference_steps, strength):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_intermediate_images(
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None
+ ):
+ _, channels, height, width = image.shape
+
+ batch_size = batch_size * num_images_per_prompt
+
+ shape = (batch_size, channels, height, width)
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
+ image = self.scheduler.add_noise(image, noise, timestep)
+
+ return image
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: Union[
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
+ ] = None,
+ strength: float = 0.7,
+ num_inference_steps: int = 80,
+ timesteps: List[int] = None,
+ guidance_scale: float = 10.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ clean_caption: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ strength (`float`, *optional*, defaults to 0.7):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 80):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 10.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ clean_caption (`bool`, *optional*, defaults to `True`):
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
+ prompt.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
+ or watermarked content, according to the `safety_checker`.
+ """
+ # 1. Check inputs. Raise error if not correct
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ self.check_inputs(
+ prompt, image, batch_size, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ # 2. Define call parameters
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ do_classifier_free_guidance,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clean_caption=clean_caption,
+ )
+
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ dtype = prompt_embeds.dtype
+
+ # 4. Prepare timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)
+
+ # 5. Prepare intermediate images
+ image = self.preprocess_image(image)
+ image = image.to(device=device, dtype=dtype)
+
+ noise_timestep = timesteps[0:1]
+ noise_timestep = noise_timestep.repeat(batch_size * num_images_per_prompt)
+
+ intermediate_images = self.prepare_intermediate_images(
+ image, noise_timestep, batch_size, num_images_per_prompt, dtype, device, generator
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # HACK: see comment in `enable_model_cpu_offload`
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
+ self.text_encoder_offload_hook.offload()
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ model_input = (
+ torch.cat([intermediate_images] * 2) if do_classifier_free_guidance else intermediate_images
+ )
+ model_input = self.scheduler.scale_model_input(model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1], dim=1)
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1], dim=1)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
+
+ if self.scheduler.config.variance_type not in ["learned", "learned_range"]:
+ noise_pred, _ = noise_pred.split(model_input.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ intermediate_images = self.scheduler.step(
+ noise_pred, t, intermediate_images, **extra_step_kwargs, return_dict=False
+ )[0]
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, intermediate_images)
+
+ image = intermediate_images
+
+ if output_type == "pil":
+ # 8. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 9. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # 10. Convert to PIL
+ image = self.numpy_to_pil(image)
+
+ # 11. Apply watermark
+ if self.watermarker is not None:
+ self.watermarker.apply_watermark(image, self.unet.config.sample_size)
+ elif output_type == "pt":
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+ else:
+ # 8. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 9. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, nsfw_detected, watermark_detected)
+
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_img2img_superresolution.py b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_img2img_superresolution.py
new file mode 100644
index 0000000000000000000000000000000000000000..d59c2b533dc110399147a28de5428a0a78105572
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_img2img_superresolution.py
@@ -0,0 +1,1029 @@
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, T5EncoderModel, T5Tokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...models import UNet2DConditionModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ BACKENDS_MAPPING,
+ PIL_INTERPOLATION,
+ is_accelerate_available,
+ is_bs4_available,
+ is_ftfy_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import IFPipelineOutput
+from .safety_checker import IFSafetyChecker
+from .watermark import IFWatermarker
+
+
+if is_bs4_available():
+ from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+ import ftfy
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.resize
+def resize(images: PIL.Image.Image, img_size: int) -> PIL.Image.Image:
+ w, h = images.size
+
+ coef = w / h
+
+ w, h = img_size, img_size
+
+ if coef >= 1:
+ w = int(round(img_size / 8 * coef) * 8)
+ else:
+ h = int(round(img_size / 8 / coef) * 8)
+
+ images = images.resize((w, h), resample=PIL_INTERPOLATION["bicubic"], reducing_gap=None)
+
+ return images
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import IFImg2ImgPipeline, IFImg2ImgSuperResolutionPipeline, DiffusionPipeline
+ >>> from diffusers.utils import pt_to_pil
+ >>> import torch
+ >>> from PIL import Image
+ >>> import requests
+ >>> from io import BytesIO
+
+ >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+ >>> response = requests.get(url)
+ >>> original_image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> original_image = original_image.resize((768, 512))
+
+ >>> pipe = IFImg2ImgPipeline.from_pretrained(
+ ... "DeepFloyd/IF-I-XL-v1.0",
+ ... variant="fp16",
+ ... torch_dtype=torch.float16,
+ ... )
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = "A fantasy landscape in style minecraft"
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
+
+ >>> image = pipe(
+ ... image=original_image,
+ ... prompt_embeds=prompt_embeds,
+ ... negative_prompt_embeds=negative_embeds,
+ ... output_type="pt",
+ ... ).images
+
+ >>> # save intermediate image
+ >>> pil_image = pt_to_pil(image)
+ >>> pil_image[0].save("./if_stage_I.png")
+
+ >>> super_res_1_pipe = IFImg2ImgSuperResolutionPipeline.from_pretrained(
+ ... "DeepFloyd/IF-II-L-v1.0",
+ ... text_encoder=None,
+ ... variant="fp16",
+ ... torch_dtype=torch.float16,
+ ... )
+ >>> super_res_1_pipe.enable_model_cpu_offload()
+
+ >>> image = super_res_1_pipe(
+ ... image=image,
+ ... original_image=original_image,
+ ... prompt_embeds=prompt_embeds,
+ ... negative_prompt_embeds=negative_embeds,
+ ... ).images
+ >>> image[0].save("./if_stage_II.png")
+ ```
+"""
+
+
+class IFImg2ImgSuperResolutionPipeline(DiffusionPipeline, LoraLoaderMixin):
+ tokenizer: T5Tokenizer
+ text_encoder: T5EncoderModel
+
+ unet: UNet2DConditionModel
+ scheduler: DDPMScheduler
+ image_noising_scheduler: DDPMScheduler
+
+ feature_extractor: Optional[CLIPImageProcessor]
+ safety_checker: Optional[IFSafetyChecker]
+
+ watermarker: Optional[IFWatermarker]
+
+ bad_punct_regex = re.compile(
+ r"["
+ + "#®•©™&@·º½¾¿¡§~"
+ + r"\)"
+ + r"\("
+ + r"\]"
+ + r"\["
+ + r"\}"
+ + r"\{"
+ + r"\|"
+ + "\\"
+ + r"\/"
+ + r"\*"
+ + r"]{1,}"
+ ) # noqa
+
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor"]
+ model_cpu_offload_seq = "text_encoder->unet"
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ image_noising_scheduler: DDPMScheduler,
+ safety_checker: Optional[IFSafetyChecker],
+ feature_extractor: Optional[CLIPImageProcessor],
+ watermarker: Optional[IFWatermarker],
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ if unet.config.in_channels != 6:
+ logger.warn(
+ "It seems like you have loaded a checkpoint that shall not be used for super resolution from {unet.config._name_or_path} as it accepts {unet.config.in_channels} input channels instead of 6. Please make sure to pass a super resolution checkpoint as the `'unet'`: IFSuperResolutionPipeline.from_pretrained(unet=super_resolution_unet, ...)`."
+ )
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ image_noising_scheduler=image_noising_scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ watermarker=watermarker,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.remove_all_hooks
+ def remove_all_hooks(self):
+ if is_accelerate_available():
+ from accelerate.hooks import remove_hook_from_module
+ else:
+ raise ImportError("Please install accelerate via `pip install accelerate`")
+
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
+ if model is not None:
+ remove_hook_from_module(model, recurse=True)
+
+ self.unet_offload_hook = None
+ self.text_encoder_offload_hook = None
+ self.final_offload_hook = None
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+ def _text_preprocessing(self, text, clean_caption=False):
+ if clean_caption and not is_bs4_available():
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if clean_caption and not is_ftfy_available():
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if not isinstance(text, (tuple, list)):
+ text = [text]
+
+ def process(text: str):
+ if clean_caption:
+ text = self._clean_caption(text)
+ text = self._clean_caption(text)
+ else:
+ text = text.lower().strip()
+ return text
+
+ return [process(t) for t in text]
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+ def _clean_caption(self, caption):
+ caption = str(caption)
+ caption = ul.unquote_plus(caption)
+ caption = caption.strip().lower()
+ caption = re.sub("", "person", caption)
+ # urls:
+ caption = re.sub(
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ caption = re.sub(
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ # html:
+ caption = BeautifulSoup(caption, features="html.parser").text
+
+ # @
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+ # 31C0—31EF CJK Strokes
+ # 31F0—31FF Katakana Phonetic Extensions
+ # 3200—32FF Enclosed CJK Letters and Months
+ # 3300—33FF CJK Compatibility
+ # 3400—4DBF CJK Unified Ideographs Extension A
+ # 4DC0—4DFF Yijing Hexagram Symbols
+ # 4E00—9FFF CJK Unified Ideographs
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+ #######################################################
+
+ # все виды тире / all types of dash --> "-"
+ caption = re.sub(
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
+ "-",
+ caption,
+ )
+
+ # кавычки к одному стандарту
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
+ caption = re.sub(r"[‘’]", "'", caption)
+
+ # "
+ caption = re.sub(r""?", "", caption)
+ # &
+ caption = re.sub(r"&", "", caption)
+
+ # ip adresses:
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+ # article ids:
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+ # \n
+ caption = re.sub(r"\\n", " ", caption)
+
+ # "#123"
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
+ # "#12345.."
+ caption = re.sub(r"#\d{5,}\b", "", caption)
+ # "123456.."
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
+ # filenames:
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+ #
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
+
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
+
+ # this-is-my-cute-cat / this_is_my_cute_cat
+ regex2 = re.compile(r"(?:\-|\_)")
+ if len(re.findall(regex2, caption)) > 3:
+ caption = re.sub(regex2, " ", caption)
+
+ caption = ftfy.fix_text(caption)
+ caption = html.unescape(html.unescape(caption))
+
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
+
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
+
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+ caption = re.sub(r"\s+", " ", caption)
+
+ caption.strip()
+
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+ caption = re.sub(r"^\.\S+$", "", caption)
+
+ return caption.strip()
+
+ @torch.no_grad()
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ do_classifier_free_guidance: bool = True,
+ num_images_per_prompt: int = 1,
+ device: Optional[torch.device] = None,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ clean_caption: bool = False,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ clean_caption (bool, defaults to `False`):
+ If `True`, the function will preprocess and clean the provided caption before encoding.
+ """
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
+ max_length = 77
+
+ if prompt_embeds is None:
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {max_length} tokens: {removed_text}"
+ )
+
+ attention_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ dtype = self.unet.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ else:
+ negative_prompt_embeds = None
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+ image, nsfw_detected, watermark_detected = self.safety_checker(
+ images=image,
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
+ )
+ else:
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+
+ return image, nsfw_detected, watermark_detected
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ original_image,
+ batch_size,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # image
+
+ if isinstance(image, list):
+ check_image_type = image[0]
+ else:
+ check_image_type = image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(image, list):
+ image_batch_size = len(image)
+ elif isinstance(image, torch.Tensor):
+ image_batch_size = image.shape[0]
+ elif isinstance(image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(image, np.ndarray):
+ image_batch_size = image.shape[0]
+ else:
+ assert False
+
+ if batch_size != image_batch_size:
+ raise ValueError(f"image batch size: {image_batch_size} must be same as prompt batch size {batch_size}")
+
+ # original_image
+
+ if isinstance(original_image, list):
+ check_image_type = original_image[0]
+ else:
+ check_image_type = original_image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`original_image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(original_image, list):
+ image_batch_size = len(original_image)
+ elif isinstance(original_image, torch.Tensor):
+ image_batch_size = original_image.shape[0]
+ elif isinstance(original_image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(original_image, np.ndarray):
+ image_batch_size = original_image.shape[0]
+ else:
+ assert False
+
+ if batch_size != image_batch_size:
+ raise ValueError(
+ f"original_image batch size: {image_batch_size} must be same as prompt batch size {batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.preprocess_image with preprocess_image -> preprocess_original_image
+ def preprocess_original_image(self, image: PIL.Image.Image) -> torch.Tensor:
+ if not isinstance(image, list):
+ image = [image]
+
+ def numpy_to_pt(images):
+ if images.ndim == 3:
+ images = images[..., None]
+
+ images = torch.from_numpy(images.transpose(0, 3, 1, 2))
+ return images
+
+ if isinstance(image[0], PIL.Image.Image):
+ new_image = []
+
+ for image_ in image:
+ image_ = image_.convert("RGB")
+ image_ = resize(image_, self.unet.sample_size)
+ image_ = np.array(image_)
+ image_ = image_.astype(np.float32)
+ image_ = image_ / 127.5 - 1
+ new_image.append(image_)
+
+ image = new_image
+
+ image = np.stack(image, axis=0) # to np
+ image = numpy_to_pt(image) # to pt
+
+ elif isinstance(image[0], np.ndarray):
+ image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
+ image = numpy_to_pt(image)
+
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
+
+ return image
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_superresolution.IFSuperResolutionPipeline.preprocess_image
+ def preprocess_image(self, image: PIL.Image.Image, num_images_per_prompt, device) -> torch.Tensor:
+ if not isinstance(image, torch.Tensor) and not isinstance(image, list):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ image = [np.array(i).astype(np.float32) / 127.5 - 1.0 for i in image]
+
+ image = np.stack(image, axis=0) # to np
+ image = torch.from_numpy(image.transpose(0, 3, 1, 2))
+ elif isinstance(image[0], np.ndarray):
+ image = np.stack(image, axis=0) # to np
+ if image.ndim == 5:
+ image = image[0]
+
+ image = torch.from_numpy(image.transpose(0, 3, 1, 2))
+ elif isinstance(image, list) and isinstance(image[0], torch.Tensor):
+ dims = image[0].ndim
+
+ if dims == 3:
+ image = torch.stack(image, dim=0)
+ elif dims == 4:
+ image = torch.concat(image, dim=0)
+ else:
+ raise ValueError(f"Image must have 3 or 4 dimensions, instead got {dims}")
+
+ image = image.to(device=device, dtype=self.unet.dtype)
+
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
+
+ return image
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.prepare_intermediate_images
+ def prepare_intermediate_images(
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None
+ ):
+ _, channels, height, width = image.shape
+
+ batch_size = batch_size * num_images_per_prompt
+
+ shape = (batch_size, channels, height, width)
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
+ image = self.scheduler.add_noise(image, noise, timestep)
+
+ return image
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image: Union[PIL.Image.Image, np.ndarray, torch.FloatTensor],
+ original_image: Union[
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
+ ] = None,
+ strength: float = 0.8,
+ prompt: Union[str, List[str]] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 4.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ noise_level: int = 250,
+ clean_caption: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ original_image (`torch.FloatTensor` or `PIL.Image.Image`):
+ The original image that `image` was varied from.
+ strength (`float`, *optional*, defaults to 0.8):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ noise_level (`int`, *optional*, defaults to 250):
+ The amount of noise to add to the upscaled image. Must be in the range `[0, 1000)`
+ clean_caption (`bool`, *optional*, defaults to `True`):
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
+ prompt.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
+ or watermarked content, according to the `safety_checker`.
+ """
+ # 1. Check inputs. Raise error if not correct
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ self.check_inputs(
+ prompt,
+ image,
+ original_image,
+ batch_size,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ do_classifier_free_guidance,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clean_caption=clean_caption,
+ )
+
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ dtype = prompt_embeds.dtype
+
+ # 4. Prepare timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)
+
+ # 5. prepare original image
+ original_image = self.preprocess_original_image(original_image)
+ original_image = original_image.to(device=device, dtype=dtype)
+
+ # 6. Prepare intermediate images
+ noise_timestep = timesteps[0:1]
+ noise_timestep = noise_timestep.repeat(batch_size * num_images_per_prompt)
+
+ intermediate_images = self.prepare_intermediate_images(
+ original_image,
+ noise_timestep,
+ batch_size,
+ num_images_per_prompt,
+ dtype,
+ device,
+ generator,
+ )
+
+ # 7. Prepare upscaled image and noise level
+ _, _, height, width = original_image.shape
+
+ image = self.preprocess_image(image, num_images_per_prompt, device)
+
+ upscaled = F.interpolate(image, (height, width), mode="bilinear", align_corners=True)
+
+ noise_level = torch.tensor([noise_level] * upscaled.shape[0], device=upscaled.device)
+ noise = randn_tensor(upscaled.shape, generator=generator, device=upscaled.device, dtype=upscaled.dtype)
+ upscaled = self.image_noising_scheduler.add_noise(upscaled, noise, timesteps=noise_level)
+
+ if do_classifier_free_guidance:
+ noise_level = torch.cat([noise_level] * 2)
+
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # HACK: see comment in `enable_model_cpu_offload`
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
+ self.text_encoder_offload_hook.offload()
+
+ # 9. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ model_input = torch.cat([intermediate_images, upscaled], dim=1)
+
+ model_input = torch.cat([model_input] * 2) if do_classifier_free_guidance else model_input
+ model_input = self.scheduler.scale_model_input(model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ class_labels=noise_level,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1] // 2, dim=1)
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1] // 2, dim=1)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
+
+ if self.scheduler.config.variance_type not in ["learned", "learned_range"]:
+ noise_pred, _ = noise_pred.split(intermediate_images.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ intermediate_images = self.scheduler.step(
+ noise_pred, t, intermediate_images, **extra_step_kwargs, return_dict=False
+ )[0]
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, intermediate_images)
+
+ image = intermediate_images
+
+ if output_type == "pil":
+ # 10. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 11. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # 12. Convert to PIL
+ image = self.numpy_to_pil(image)
+
+ # 13. Apply watermark
+ if self.watermarker is not None:
+ self.watermarker.apply_watermark(image, self.unet.config.sample_size)
+ elif output_type == "pt":
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+ else:
+ # 10. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 11. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, nsfw_detected, watermark_detected)
+
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_inpainting.py b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_inpainting.py
new file mode 100644
index 0000000000000000000000000000000000000000..1dbb5e92ec4c6f4e6f90ca55ee4b1f6d2cc53352
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_inpainting.py
@@ -0,0 +1,1030 @@
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, T5EncoderModel, T5Tokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...models import UNet2DConditionModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ BACKENDS_MAPPING,
+ PIL_INTERPOLATION,
+ is_accelerate_available,
+ is_bs4_available,
+ is_ftfy_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import IFPipelineOutput
+from .safety_checker import IFSafetyChecker
+from .watermark import IFWatermarker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+if is_bs4_available():
+ from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+ import ftfy
+
+
+# Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.resize
+def resize(images: PIL.Image.Image, img_size: int) -> PIL.Image.Image:
+ w, h = images.size
+
+ coef = w / h
+
+ w, h = img_size, img_size
+
+ if coef >= 1:
+ w = int(round(img_size / 8 * coef) * 8)
+ else:
+ h = int(round(img_size / 8 / coef) * 8)
+
+ images = images.resize((w, h), resample=PIL_INTERPOLATION["bicubic"], reducing_gap=None)
+
+ return images
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import IFInpaintingPipeline, IFInpaintingSuperResolutionPipeline, DiffusionPipeline
+ >>> from diffusers.utils import pt_to_pil
+ >>> import torch
+ >>> from PIL import Image
+ >>> import requests
+ >>> from io import BytesIO
+
+ >>> url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/if/person.png"
+ >>> response = requests.get(url)
+ >>> original_image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> original_image = original_image
+
+ >>> url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/if/glasses_mask.png"
+ >>> response = requests.get(url)
+ >>> mask_image = Image.open(BytesIO(response.content))
+ >>> mask_image = mask_image
+
+ >>> pipe = IFInpaintingPipeline.from_pretrained(
+ ... "DeepFloyd/IF-I-XL-v1.0", variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = "blue sunglasses"
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
+
+ >>> image = pipe(
+ ... image=original_image,
+ ... mask_image=mask_image,
+ ... prompt_embeds=prompt_embeds,
+ ... negative_prompt_embeds=negative_embeds,
+ ... output_type="pt",
+ ... ).images
+
+ >>> # save intermediate image
+ >>> pil_image = pt_to_pil(image)
+ >>> pil_image[0].save("./if_stage_I.png")
+
+ >>> super_res_1_pipe = IFInpaintingSuperResolutionPipeline.from_pretrained(
+ ... "DeepFloyd/IF-II-L-v1.0", text_encoder=None, variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> super_res_1_pipe.enable_model_cpu_offload()
+
+ >>> image = super_res_1_pipe(
+ ... image=image,
+ ... mask_image=mask_image,
+ ... original_image=original_image,
+ ... prompt_embeds=prompt_embeds,
+ ... negative_prompt_embeds=negative_embeds,
+ ... ).images
+ >>> image[0].save("./if_stage_II.png")
+ ```
+"""
+
+
+class IFInpaintingPipeline(DiffusionPipeline, LoraLoaderMixin):
+ tokenizer: T5Tokenizer
+ text_encoder: T5EncoderModel
+
+ unet: UNet2DConditionModel
+ scheduler: DDPMScheduler
+
+ feature_extractor: Optional[CLIPImageProcessor]
+ safety_checker: Optional[IFSafetyChecker]
+
+ watermarker: Optional[IFWatermarker]
+
+ bad_punct_regex = re.compile(
+ r"["
+ + "#®•©™&@·º½¾¿¡§~"
+ + r"\)"
+ + r"\("
+ + r"\]"
+ + r"\["
+ + r"\}"
+ + r"\{"
+ + r"\|"
+ + "\\"
+ + r"\/"
+ + r"\*"
+ + r"]{1,}"
+ ) # noqa
+
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor", "watermarker"]
+ model_cpu_offload_seq = "text_encoder->unet"
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ safety_checker: Optional[IFSafetyChecker],
+ feature_extractor: Optional[CLIPImageProcessor],
+ watermarker: Optional[IFWatermarker],
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ watermarker=watermarker,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.remove_all_hooks
+ def remove_all_hooks(self):
+ if is_accelerate_available():
+ from accelerate.hooks import remove_hook_from_module
+ else:
+ raise ImportError("Please install accelerate via `pip install accelerate`")
+
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
+ if model is not None:
+ remove_hook_from_module(model, recurse=True)
+
+ self.unet_offload_hook = None
+ self.text_encoder_offload_hook = None
+ self.final_offload_hook = None
+
+ @torch.no_grad()
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ do_classifier_free_guidance: bool = True,
+ num_images_per_prompt: int = 1,
+ device: Optional[torch.device] = None,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ clean_caption: bool = False,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ clean_caption (bool, defaults to `False`):
+ If `True`, the function will preprocess and clean the provided caption before encoding.
+ """
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
+ max_length = 77
+
+ if prompt_embeds is None:
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {max_length} tokens: {removed_text}"
+ )
+
+ attention_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ dtype = self.unet.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ else:
+ negative_prompt_embeds = None
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+ image, nsfw_detected, watermark_detected = self.safety_checker(
+ images=image,
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
+ )
+ else:
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+
+ return image, nsfw_detected, watermark_detected
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ mask_image,
+ batch_size,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # image
+
+ if isinstance(image, list):
+ check_image_type = image[0]
+ else:
+ check_image_type = image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(image, list):
+ image_batch_size = len(image)
+ elif isinstance(image, torch.Tensor):
+ image_batch_size = image.shape[0]
+ elif isinstance(image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(image, np.ndarray):
+ image_batch_size = image.shape[0]
+ else:
+ assert False
+
+ if batch_size != image_batch_size:
+ raise ValueError(f"image batch size: {image_batch_size} must be same as prompt batch size {batch_size}")
+
+ # mask_image
+
+ if isinstance(mask_image, list):
+ check_image_type = mask_image[0]
+ else:
+ check_image_type = mask_image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`mask_image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(mask_image, list):
+ image_batch_size = len(mask_image)
+ elif isinstance(mask_image, torch.Tensor):
+ image_batch_size = mask_image.shape[0]
+ elif isinstance(mask_image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(mask_image, np.ndarray):
+ image_batch_size = mask_image.shape[0]
+ else:
+ assert False
+
+ if image_batch_size != 1 and batch_size != image_batch_size:
+ raise ValueError(
+ f"mask_image batch size: {image_batch_size} must be `1` or the same as prompt batch size {batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+ def _text_preprocessing(self, text, clean_caption=False):
+ if clean_caption and not is_bs4_available():
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if clean_caption and not is_ftfy_available():
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if not isinstance(text, (tuple, list)):
+ text = [text]
+
+ def process(text: str):
+ if clean_caption:
+ text = self._clean_caption(text)
+ text = self._clean_caption(text)
+ else:
+ text = text.lower().strip()
+ return text
+
+ return [process(t) for t in text]
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+ def _clean_caption(self, caption):
+ caption = str(caption)
+ caption = ul.unquote_plus(caption)
+ caption = caption.strip().lower()
+ caption = re.sub("", "person", caption)
+ # urls:
+ caption = re.sub(
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ caption = re.sub(
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ # html:
+ caption = BeautifulSoup(caption, features="html.parser").text
+
+ # @
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+ # 31C0—31EF CJK Strokes
+ # 31F0—31FF Katakana Phonetic Extensions
+ # 3200—32FF Enclosed CJK Letters and Months
+ # 3300—33FF CJK Compatibility
+ # 3400—4DBF CJK Unified Ideographs Extension A
+ # 4DC0—4DFF Yijing Hexagram Symbols
+ # 4E00—9FFF CJK Unified Ideographs
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+ #######################################################
+
+ # все виды тире / all types of dash --> "-"
+ caption = re.sub(
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
+ "-",
+ caption,
+ )
+
+ # кавычки к одному стандарту
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
+ caption = re.sub(r"[‘’]", "'", caption)
+
+ # "
+ caption = re.sub(r""?", "", caption)
+ # &
+ caption = re.sub(r"&", "", caption)
+
+ # ip adresses:
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+ # article ids:
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+ # \n
+ caption = re.sub(r"\\n", " ", caption)
+
+ # "#123"
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
+ # "#12345.."
+ caption = re.sub(r"#\d{5,}\b", "", caption)
+ # "123456.."
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
+ # filenames:
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+ #
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
+
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
+
+ # this-is-my-cute-cat / this_is_my_cute_cat
+ regex2 = re.compile(r"(?:\-|\_)")
+ if len(re.findall(regex2, caption)) > 3:
+ caption = re.sub(regex2, " ", caption)
+
+ caption = ftfy.fix_text(caption)
+ caption = html.unescape(html.unescape(caption))
+
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
+
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
+
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+ caption = re.sub(r"\s+", " ", caption)
+
+ caption.strip()
+
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+ caption = re.sub(r"^\.\S+$", "", caption)
+
+ return caption.strip()
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.preprocess_image
+ def preprocess_image(self, image: PIL.Image.Image) -> torch.Tensor:
+ if not isinstance(image, list):
+ image = [image]
+
+ def numpy_to_pt(images):
+ if images.ndim == 3:
+ images = images[..., None]
+
+ images = torch.from_numpy(images.transpose(0, 3, 1, 2))
+ return images
+
+ if isinstance(image[0], PIL.Image.Image):
+ new_image = []
+
+ for image_ in image:
+ image_ = image_.convert("RGB")
+ image_ = resize(image_, self.unet.sample_size)
+ image_ = np.array(image_)
+ image_ = image_.astype(np.float32)
+ image_ = image_ / 127.5 - 1
+ new_image.append(image_)
+
+ image = new_image
+
+ image = np.stack(image, axis=0) # to np
+ image = numpy_to_pt(image) # to pt
+
+ elif isinstance(image[0], np.ndarray):
+ image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
+ image = numpy_to_pt(image)
+
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
+
+ return image
+
+ def preprocess_mask_image(self, mask_image) -> torch.Tensor:
+ if not isinstance(mask_image, list):
+ mask_image = [mask_image]
+
+ if isinstance(mask_image[0], torch.Tensor):
+ mask_image = torch.cat(mask_image, axis=0) if mask_image[0].ndim == 4 else torch.stack(mask_image, axis=0)
+
+ if mask_image.ndim == 2:
+ # Batch and add channel dim for single mask
+ mask_image = mask_image.unsqueeze(0).unsqueeze(0)
+ elif mask_image.ndim == 3 and mask_image.shape[0] == 1:
+ # Single mask, the 0'th dimension is considered to be
+ # the existing batch size of 1
+ mask_image = mask_image.unsqueeze(0)
+ elif mask_image.ndim == 3 and mask_image.shape[0] != 1:
+ # Batch of mask, the 0'th dimension is considered to be
+ # the batching dimension
+ mask_image = mask_image.unsqueeze(1)
+
+ mask_image[mask_image < 0.5] = 0
+ mask_image[mask_image >= 0.5] = 1
+
+ elif isinstance(mask_image[0], PIL.Image.Image):
+ new_mask_image = []
+
+ for mask_image_ in mask_image:
+ mask_image_ = mask_image_.convert("L")
+ mask_image_ = resize(mask_image_, self.unet.sample_size)
+ mask_image_ = np.array(mask_image_)
+ mask_image_ = mask_image_[None, None, :]
+ new_mask_image.append(mask_image_)
+
+ mask_image = new_mask_image
+
+ mask_image = np.concatenate(mask_image, axis=0)
+ mask_image = mask_image.astype(np.float32) / 255.0
+ mask_image[mask_image < 0.5] = 0
+ mask_image[mask_image >= 0.5] = 1
+ mask_image = torch.from_numpy(mask_image)
+
+ elif isinstance(mask_image[0], np.ndarray):
+ mask_image = np.concatenate([m[None, None, :] for m in mask_image], axis=0)
+
+ mask_image[mask_image < 0.5] = 0
+ mask_image[mask_image >= 0.5] = 1
+ mask_image = torch.from_numpy(mask_image)
+
+ return mask_image
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_intermediate_images(
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, mask_image, generator=None
+ ):
+ image_batch_size, channels, height, width = image.shape
+
+ batch_size = batch_size * num_images_per_prompt
+
+ shape = (batch_size, channels, height, width)
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
+ noised_image = self.scheduler.add_noise(image, noise, timestep)
+
+ image = (1 - mask_image) * image + mask_image * noised_image
+
+ return image
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: Union[
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
+ ] = None,
+ mask_image: Union[
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
+ ] = None,
+ strength: float = 1.0,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ clean_caption: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ mask_image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+ repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
+ to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
+ instead of 3, so the expected shape would be `(B, H, W, 1)`.
+ strength (`float`, *optional*, defaults to 1.0):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ clean_caption (`bool`, *optional*, defaults to `True`):
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
+ prompt.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
+ or watermarked content, according to the `safety_checker`.
+ """
+ # 1. Check inputs. Raise error if not correct
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ self.check_inputs(
+ prompt,
+ image,
+ mask_image,
+ batch_size,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ do_classifier_free_guidance,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clean_caption=clean_caption,
+ )
+
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ dtype = prompt_embeds.dtype
+
+ # 4. Prepare timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)
+
+ # 5. Prepare intermediate images
+ image = self.preprocess_image(image)
+ image = image.to(device=device, dtype=dtype)
+
+ mask_image = self.preprocess_mask_image(mask_image)
+ mask_image = mask_image.to(device=device, dtype=dtype)
+
+ if mask_image.shape[0] == 1:
+ mask_image = mask_image.repeat_interleave(batch_size * num_images_per_prompt, dim=0)
+ else:
+ mask_image = mask_image.repeat_interleave(num_images_per_prompt, dim=0)
+
+ noise_timestep = timesteps[0:1]
+ noise_timestep = noise_timestep.repeat(batch_size * num_images_per_prompt)
+
+ intermediate_images = self.prepare_intermediate_images(
+ image, noise_timestep, batch_size, num_images_per_prompt, dtype, device, mask_image, generator
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # HACK: see comment in `enable_model_cpu_offload`
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
+ self.text_encoder_offload_hook.offload()
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ model_input = (
+ torch.cat([intermediate_images] * 2) if do_classifier_free_guidance else intermediate_images
+ )
+ model_input = self.scheduler.scale_model_input(model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1], dim=1)
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1], dim=1)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
+
+ if self.scheduler.config.variance_type not in ["learned", "learned_range"]:
+ noise_pred, _ = noise_pred.split(model_input.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ prev_intermediate_images = intermediate_images
+
+ intermediate_images = self.scheduler.step(
+ noise_pred, t, intermediate_images, **extra_step_kwargs, return_dict=False
+ )[0]
+
+ intermediate_images = (1 - mask_image) * prev_intermediate_images + mask_image * intermediate_images
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, intermediate_images)
+
+ image = intermediate_images
+
+ if output_type == "pil":
+ # 8. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 9. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # 10. Convert to PIL
+ image = self.numpy_to_pil(image)
+
+ # 11. Apply watermark
+ if self.watermarker is not None:
+ self.watermarker.apply_watermark(image, self.unet.config.sample_size)
+ elif output_type == "pt":
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+ else:
+ # 8. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 9. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, nsfw_detected, watermark_detected)
+
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_inpainting_superresolution.py b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_inpainting_superresolution.py
new file mode 100644
index 0000000000000000000000000000000000000000..cb9200cffce536de25adad3b762c0f22ae667b60
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_inpainting_superresolution.py
@@ -0,0 +1,1137 @@
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, T5EncoderModel, T5Tokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...models import UNet2DConditionModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ BACKENDS_MAPPING,
+ PIL_INTERPOLATION,
+ is_accelerate_available,
+ is_bs4_available,
+ is_ftfy_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import IFPipelineOutput
+from .safety_checker import IFSafetyChecker
+from .watermark import IFWatermarker
+
+
+if is_bs4_available():
+ from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+ import ftfy
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.resize
+def resize(images: PIL.Image.Image, img_size: int) -> PIL.Image.Image:
+ w, h = images.size
+
+ coef = w / h
+
+ w, h = img_size, img_size
+
+ if coef >= 1:
+ w = int(round(img_size / 8 * coef) * 8)
+ else:
+ h = int(round(img_size / 8 / coef) * 8)
+
+ images = images.resize((w, h), resample=PIL_INTERPOLATION["bicubic"], reducing_gap=None)
+
+ return images
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import IFInpaintingPipeline, IFInpaintingSuperResolutionPipeline, DiffusionPipeline
+ >>> from diffusers.utils import pt_to_pil
+ >>> import torch
+ >>> from PIL import Image
+ >>> import requests
+ >>> from io import BytesIO
+
+ >>> url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/if/person.png"
+ >>> response = requests.get(url)
+ >>> original_image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> original_image = original_image
+
+ >>> url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/if/glasses_mask.png"
+ >>> response = requests.get(url)
+ >>> mask_image = Image.open(BytesIO(response.content))
+ >>> mask_image = mask_image
+
+ >>> pipe = IFInpaintingPipeline.from_pretrained(
+ ... "DeepFloyd/IF-I-XL-v1.0", variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = "blue sunglasses"
+
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
+ >>> image = pipe(
+ ... image=original_image,
+ ... mask_image=mask_image,
+ ... prompt_embeds=prompt_embeds,
+ ... negative_prompt_embeds=negative_embeds,
+ ... output_type="pt",
+ ... ).images
+
+ >>> # save intermediate image
+ >>> pil_image = pt_to_pil(image)
+ >>> pil_image[0].save("./if_stage_I.png")
+
+ >>> super_res_1_pipe = IFInpaintingSuperResolutionPipeline.from_pretrained(
+ ... "DeepFloyd/IF-II-L-v1.0", text_encoder=None, variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> super_res_1_pipe.enable_model_cpu_offload()
+
+ >>> image = super_res_1_pipe(
+ ... image=image,
+ ... mask_image=mask_image,
+ ... original_image=original_image,
+ ... prompt_embeds=prompt_embeds,
+ ... negative_prompt_embeds=negative_embeds,
+ ... ).images
+ >>> image[0].save("./if_stage_II.png")
+ ```
+ """
+
+
+class IFInpaintingSuperResolutionPipeline(DiffusionPipeline, LoraLoaderMixin):
+ tokenizer: T5Tokenizer
+ text_encoder: T5EncoderModel
+
+ unet: UNet2DConditionModel
+ scheduler: DDPMScheduler
+ image_noising_scheduler: DDPMScheduler
+
+ feature_extractor: Optional[CLIPImageProcessor]
+ safety_checker: Optional[IFSafetyChecker]
+
+ watermarker: Optional[IFWatermarker]
+
+ bad_punct_regex = re.compile(
+ r"["
+ + "#®•©™&@·º½¾¿¡§~"
+ + r"\)"
+ + r"\("
+ + r"\]"
+ + r"\["
+ + r"\}"
+ + r"\{"
+ + r"\|"
+ + "\\"
+ + r"\/"
+ + r"\*"
+ + r"]{1,}"
+ ) # noqa
+
+ model_cpu_offload_seq = "text_encoder->unet"
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor", "watermarker"]
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ image_noising_scheduler: DDPMScheduler,
+ safety_checker: Optional[IFSafetyChecker],
+ feature_extractor: Optional[CLIPImageProcessor],
+ watermarker: Optional[IFWatermarker],
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ if unet.config.in_channels != 6:
+ logger.warn(
+ "It seems like you have loaded a checkpoint that shall not be used for super resolution from {unet.config._name_or_path} as it accepts {unet.config.in_channels} input channels instead of 6. Please make sure to pass a super resolution checkpoint as the `'unet'`: IFSuperResolutionPipeline.from_pretrained(unet=super_resolution_unet, ...)`."
+ )
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ image_noising_scheduler=image_noising_scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ watermarker=watermarker,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.remove_all_hooks
+ def remove_all_hooks(self):
+ if is_accelerate_available():
+ from accelerate.hooks import remove_hook_from_module
+ else:
+ raise ImportError("Please install accelerate via `pip install accelerate`")
+
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
+ if model is not None:
+ remove_hook_from_module(model, recurse=True)
+
+ self.unet_offload_hook = None
+ self.text_encoder_offload_hook = None
+ self.final_offload_hook = None
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+ def _text_preprocessing(self, text, clean_caption=False):
+ if clean_caption and not is_bs4_available():
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if clean_caption and not is_ftfy_available():
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if not isinstance(text, (tuple, list)):
+ text = [text]
+
+ def process(text: str):
+ if clean_caption:
+ text = self._clean_caption(text)
+ text = self._clean_caption(text)
+ else:
+ text = text.lower().strip()
+ return text
+
+ return [process(t) for t in text]
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+ def _clean_caption(self, caption):
+ caption = str(caption)
+ caption = ul.unquote_plus(caption)
+ caption = caption.strip().lower()
+ caption = re.sub("", "person", caption)
+ # urls:
+ caption = re.sub(
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ caption = re.sub(
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ # html:
+ caption = BeautifulSoup(caption, features="html.parser").text
+
+ # @
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+ # 31C0—31EF CJK Strokes
+ # 31F0—31FF Katakana Phonetic Extensions
+ # 3200—32FF Enclosed CJK Letters and Months
+ # 3300—33FF CJK Compatibility
+ # 3400—4DBF CJK Unified Ideographs Extension A
+ # 4DC0—4DFF Yijing Hexagram Symbols
+ # 4E00—9FFF CJK Unified Ideographs
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+ #######################################################
+
+ # все виды тире / all types of dash --> "-"
+ caption = re.sub(
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
+ "-",
+ caption,
+ )
+
+ # кавычки к одному стандарту
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
+ caption = re.sub(r"[‘’]", "'", caption)
+
+ # "
+ caption = re.sub(r""?", "", caption)
+ # &
+ caption = re.sub(r"&", "", caption)
+
+ # ip adresses:
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+ # article ids:
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+ # \n
+ caption = re.sub(r"\\n", " ", caption)
+
+ # "#123"
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
+ # "#12345.."
+ caption = re.sub(r"#\d{5,}\b", "", caption)
+ # "123456.."
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
+ # filenames:
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+ #
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
+
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
+
+ # this-is-my-cute-cat / this_is_my_cute_cat
+ regex2 = re.compile(r"(?:\-|\_)")
+ if len(re.findall(regex2, caption)) > 3:
+ caption = re.sub(regex2, " ", caption)
+
+ caption = ftfy.fix_text(caption)
+ caption = html.unescape(html.unescape(caption))
+
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
+
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
+
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+ caption = re.sub(r"\s+", " ", caption)
+
+ caption.strip()
+
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+ caption = re.sub(r"^\.\S+$", "", caption)
+
+ return caption.strip()
+
+ @torch.no_grad()
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ do_classifier_free_guidance: bool = True,
+ num_images_per_prompt: int = 1,
+ device: Optional[torch.device] = None,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ clean_caption: bool = False,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ clean_caption (bool, defaults to `False`):
+ If `True`, the function will preprocess and clean the provided caption before encoding.
+ """
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
+ max_length = 77
+
+ if prompt_embeds is None:
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {max_length} tokens: {removed_text}"
+ )
+
+ attention_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ dtype = self.unet.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ else:
+ negative_prompt_embeds = None
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+ image, nsfw_detected, watermark_detected = self.safety_checker(
+ images=image,
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
+ )
+ else:
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+
+ return image, nsfw_detected, watermark_detected
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ original_image,
+ mask_image,
+ batch_size,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # image
+
+ if isinstance(image, list):
+ check_image_type = image[0]
+ else:
+ check_image_type = image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(image, list):
+ image_batch_size = len(image)
+ elif isinstance(image, torch.Tensor):
+ image_batch_size = image.shape[0]
+ elif isinstance(image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(image, np.ndarray):
+ image_batch_size = image.shape[0]
+ else:
+ assert False
+
+ if batch_size != image_batch_size:
+ raise ValueError(f"image batch size: {image_batch_size} must be same as prompt batch size {batch_size}")
+
+ # original_image
+
+ if isinstance(original_image, list):
+ check_image_type = original_image[0]
+ else:
+ check_image_type = original_image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`original_image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(original_image, list):
+ image_batch_size = len(original_image)
+ elif isinstance(original_image, torch.Tensor):
+ image_batch_size = original_image.shape[0]
+ elif isinstance(original_image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(original_image, np.ndarray):
+ image_batch_size = original_image.shape[0]
+ else:
+ assert False
+
+ if batch_size != image_batch_size:
+ raise ValueError(
+ f"original_image batch size: {image_batch_size} must be same as prompt batch size {batch_size}"
+ )
+
+ # mask_image
+
+ if isinstance(mask_image, list):
+ check_image_type = mask_image[0]
+ else:
+ check_image_type = mask_image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`mask_image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(mask_image, list):
+ image_batch_size = len(mask_image)
+ elif isinstance(mask_image, torch.Tensor):
+ image_batch_size = mask_image.shape[0]
+ elif isinstance(mask_image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(mask_image, np.ndarray):
+ image_batch_size = mask_image.shape[0]
+ else:
+ assert False
+
+ if image_batch_size != 1 and batch_size != image_batch_size:
+ raise ValueError(
+ f"mask_image batch size: {image_batch_size} must be `1` or the same as prompt batch size {batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.preprocess_image with preprocess_image -> preprocess_original_image
+ def preprocess_original_image(self, image: PIL.Image.Image) -> torch.Tensor:
+ if not isinstance(image, list):
+ image = [image]
+
+ def numpy_to_pt(images):
+ if images.ndim == 3:
+ images = images[..., None]
+
+ images = torch.from_numpy(images.transpose(0, 3, 1, 2))
+ return images
+
+ if isinstance(image[0], PIL.Image.Image):
+ new_image = []
+
+ for image_ in image:
+ image_ = image_.convert("RGB")
+ image_ = resize(image_, self.unet.sample_size)
+ image_ = np.array(image_)
+ image_ = image_.astype(np.float32)
+ image_ = image_ / 127.5 - 1
+ new_image.append(image_)
+
+ image = new_image
+
+ image = np.stack(image, axis=0) # to np
+ image = numpy_to_pt(image) # to pt
+
+ elif isinstance(image[0], np.ndarray):
+ image = np.concatenate(image, axis=0) if image[0].ndim == 4 else np.stack(image, axis=0)
+ image = numpy_to_pt(image)
+
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
+
+ return image
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_superresolution.IFSuperResolutionPipeline.preprocess_image
+ def preprocess_image(self, image: PIL.Image.Image, num_images_per_prompt, device) -> torch.Tensor:
+ if not isinstance(image, torch.Tensor) and not isinstance(image, list):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ image = [np.array(i).astype(np.float32) / 127.5 - 1.0 for i in image]
+
+ image = np.stack(image, axis=0) # to np
+ image = torch.from_numpy(image.transpose(0, 3, 1, 2))
+ elif isinstance(image[0], np.ndarray):
+ image = np.stack(image, axis=0) # to np
+ if image.ndim == 5:
+ image = image[0]
+
+ image = torch.from_numpy(image.transpose(0, 3, 1, 2))
+ elif isinstance(image, list) and isinstance(image[0], torch.Tensor):
+ dims = image[0].ndim
+
+ if dims == 3:
+ image = torch.stack(image, dim=0)
+ elif dims == 4:
+ image = torch.concat(image, dim=0)
+ else:
+ raise ValueError(f"Image must have 3 or 4 dimensions, instead got {dims}")
+
+ image = image.to(device=device, dtype=self.unet.dtype)
+
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
+
+ return image
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_inpainting.IFInpaintingPipeline.preprocess_mask_image
+ def preprocess_mask_image(self, mask_image) -> torch.Tensor:
+ if not isinstance(mask_image, list):
+ mask_image = [mask_image]
+
+ if isinstance(mask_image[0], torch.Tensor):
+ mask_image = torch.cat(mask_image, axis=0) if mask_image[0].ndim == 4 else torch.stack(mask_image, axis=0)
+
+ if mask_image.ndim == 2:
+ # Batch and add channel dim for single mask
+ mask_image = mask_image.unsqueeze(0).unsqueeze(0)
+ elif mask_image.ndim == 3 and mask_image.shape[0] == 1:
+ # Single mask, the 0'th dimension is considered to be
+ # the existing batch size of 1
+ mask_image = mask_image.unsqueeze(0)
+ elif mask_image.ndim == 3 and mask_image.shape[0] != 1:
+ # Batch of mask, the 0'th dimension is considered to be
+ # the batching dimension
+ mask_image = mask_image.unsqueeze(1)
+
+ mask_image[mask_image < 0.5] = 0
+ mask_image[mask_image >= 0.5] = 1
+
+ elif isinstance(mask_image[0], PIL.Image.Image):
+ new_mask_image = []
+
+ for mask_image_ in mask_image:
+ mask_image_ = mask_image_.convert("L")
+ mask_image_ = resize(mask_image_, self.unet.sample_size)
+ mask_image_ = np.array(mask_image_)
+ mask_image_ = mask_image_[None, None, :]
+ new_mask_image.append(mask_image_)
+
+ mask_image = new_mask_image
+
+ mask_image = np.concatenate(mask_image, axis=0)
+ mask_image = mask_image.astype(np.float32) / 255.0
+ mask_image[mask_image < 0.5] = 0
+ mask_image[mask_image >= 0.5] = 1
+ mask_image = torch.from_numpy(mask_image)
+
+ elif isinstance(mask_image[0], np.ndarray):
+ mask_image = np.concatenate([m[None, None, :] for m in mask_image], axis=0)
+
+ mask_image[mask_image < 0.5] = 0
+ mask_image[mask_image >= 0.5] = 1
+ mask_image = torch.from_numpy(mask_image)
+
+ return mask_image
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_img2img.IFImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if_inpainting.IFInpaintingPipeline.prepare_intermediate_images
+ def prepare_intermediate_images(
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, mask_image, generator=None
+ ):
+ image_batch_size, channels, height, width = image.shape
+
+ batch_size = batch_size * num_images_per_prompt
+
+ shape = (batch_size, channels, height, width)
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
+ noised_image = self.scheduler.add_noise(image, noise, timestep)
+
+ image = (1 - mask_image) * image + mask_image * noised_image
+
+ return image
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image: Union[PIL.Image.Image, np.ndarray, torch.FloatTensor],
+ original_image: Union[
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
+ ] = None,
+ mask_image: Union[
+ PIL.Image.Image, torch.Tensor, np.ndarray, List[PIL.Image.Image], List[torch.Tensor], List[np.ndarray]
+ ] = None,
+ strength: float = 0.8,
+ prompt: Union[str, List[str]] = None,
+ num_inference_steps: int = 100,
+ timesteps: List[int] = None,
+ guidance_scale: float = 4.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ noise_level: int = 0,
+ clean_caption: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ original_image (`torch.FloatTensor` or `PIL.Image.Image`):
+ The original image that `image` was varied from.
+ mask_image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+ repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
+ to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
+ instead of 3, so the expected shape would be `(B, H, W, 1)`.
+ strength (`float`, *optional*, defaults to 0.8):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ noise_level (`int`, *optional*, defaults to 0):
+ The amount of noise to add to the upscaled image. Must be in the range `[0, 1000)`
+ clean_caption (`bool`, *optional*, defaults to `True`):
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
+ prompt.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
+ or watermarked content, according to the `safety_checker`.
+ """
+ # 1. Check inputs. Raise error if not correct
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ self.check_inputs(
+ prompt,
+ image,
+ original_image,
+ mask_image,
+ batch_size,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ do_classifier_free_guidance,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clean_caption=clean_caption,
+ )
+
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ dtype = prompt_embeds.dtype
+
+ # 4. Prepare timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)
+
+ # 5. prepare original image
+ original_image = self.preprocess_original_image(original_image)
+ original_image = original_image.to(device=device, dtype=dtype)
+
+ # 6. prepare mask image
+ mask_image = self.preprocess_mask_image(mask_image)
+ mask_image = mask_image.to(device=device, dtype=dtype)
+
+ if mask_image.shape[0] == 1:
+ mask_image = mask_image.repeat_interleave(batch_size * num_images_per_prompt, dim=0)
+ else:
+ mask_image = mask_image.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # 6. Prepare intermediate images
+ noise_timestep = timesteps[0:1]
+ noise_timestep = noise_timestep.repeat(batch_size * num_images_per_prompt)
+
+ intermediate_images = self.prepare_intermediate_images(
+ original_image,
+ noise_timestep,
+ batch_size,
+ num_images_per_prompt,
+ dtype,
+ device,
+ mask_image,
+ generator,
+ )
+
+ # 7. Prepare upscaled image and noise level
+ _, _, height, width = original_image.shape
+
+ image = self.preprocess_image(image, num_images_per_prompt, device)
+
+ upscaled = F.interpolate(image, (height, width), mode="bilinear", align_corners=True)
+
+ noise_level = torch.tensor([noise_level] * upscaled.shape[0], device=upscaled.device)
+ noise = randn_tensor(upscaled.shape, generator=generator, device=upscaled.device, dtype=upscaled.dtype)
+ upscaled = self.image_noising_scheduler.add_noise(upscaled, noise, timesteps=noise_level)
+
+ if do_classifier_free_guidance:
+ noise_level = torch.cat([noise_level] * 2)
+
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # HACK: see comment in `enable_model_cpu_offload`
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
+ self.text_encoder_offload_hook.offload()
+
+ # 9. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ model_input = torch.cat([intermediate_images, upscaled], dim=1)
+
+ model_input = torch.cat([model_input] * 2) if do_classifier_free_guidance else model_input
+ model_input = self.scheduler.scale_model_input(model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ class_labels=noise_level,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1] // 2, dim=1)
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1] // 2, dim=1)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
+
+ if self.scheduler.config.variance_type not in ["learned", "learned_range"]:
+ noise_pred, _ = noise_pred.split(intermediate_images.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ prev_intermediate_images = intermediate_images
+
+ intermediate_images = self.scheduler.step(
+ noise_pred, t, intermediate_images, **extra_step_kwargs, return_dict=False
+ )[0]
+
+ intermediate_images = (1 - mask_image) * prev_intermediate_images + mask_image * intermediate_images
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, intermediate_images)
+
+ image = intermediate_images
+
+ if output_type == "pil":
+ # 10. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 11. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # 12. Convert to PIL
+ image = self.numpy_to_pil(image)
+
+ # 13. Apply watermark
+ if self.watermarker is not None:
+ self.watermarker.apply_watermark(image, self.unet.config.sample_size)
+ elif output_type == "pt":
+ nsfw_detected = None
+ watermark_detected = None
+
+ else:
+ # 10. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 11. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, nsfw_detected, watermark_detected)
+
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_superresolution.py b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_superresolution.py
new file mode 100644
index 0000000000000000000000000000000000000000..2b48f5887c29cf575fd892f2dc7ecbdadaa70a07
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_if_superresolution.py
@@ -0,0 +1,885 @@
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, T5EncoderModel, T5Tokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...models import UNet2DConditionModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ BACKENDS_MAPPING,
+ is_accelerate_available,
+ is_bs4_available,
+ is_ftfy_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_output import IFPipelineOutput
+from .safety_checker import IFSafetyChecker
+from .watermark import IFWatermarker
+
+
+if is_bs4_available():
+ from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+ import ftfy
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import IFPipeline, IFSuperResolutionPipeline, DiffusionPipeline
+ >>> from diffusers.utils import pt_to_pil
+ >>> import torch
+
+ >>> pipe = IFPipeline.from_pretrained("DeepFloyd/IF-I-XL-v1.0", variant="fp16", torch_dtype=torch.float16)
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = 'a photo of a kangaroo wearing an orange hoodie and blue sunglasses standing in front of the eiffel tower holding a sign that says "very deep learning"'
+ >>> prompt_embeds, negative_embeds = pipe.encode_prompt(prompt)
+
+ >>> image = pipe(prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds, output_type="pt").images
+
+ >>> # save intermediate image
+ >>> pil_image = pt_to_pil(image)
+ >>> pil_image[0].save("./if_stage_I.png")
+
+ >>> super_res_1_pipe = IFSuperResolutionPipeline.from_pretrained(
+ ... "DeepFloyd/IF-II-L-v1.0", text_encoder=None, variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> super_res_1_pipe.enable_model_cpu_offload()
+
+ >>> image = super_res_1_pipe(
+ ... image=image, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_embeds
+ ... ).images
+ >>> image[0].save("./if_stage_II.png")
+ ```
+"""
+
+
+class IFSuperResolutionPipeline(DiffusionPipeline, LoraLoaderMixin):
+ tokenizer: T5Tokenizer
+ text_encoder: T5EncoderModel
+
+ unet: UNet2DConditionModel
+ scheduler: DDPMScheduler
+ image_noising_scheduler: DDPMScheduler
+
+ feature_extractor: Optional[CLIPImageProcessor]
+ safety_checker: Optional[IFSafetyChecker]
+
+ watermarker: Optional[IFWatermarker]
+
+ bad_punct_regex = re.compile(
+ r"["
+ + "#®•©™&@·º½¾¿¡§~"
+ + r"\)"
+ + r"\("
+ + r"\]"
+ + r"\["
+ + r"\}"
+ + r"\{"
+ + r"\|"
+ + "\\"
+ + r"\/"
+ + r"\*"
+ + r"]{1,}"
+ ) # noqa
+
+ _optional_components = ["tokenizer", "text_encoder", "safety_checker", "feature_extractor", "watermarker"]
+ model_cpu_offload_seq = "text_encoder->unet"
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ image_noising_scheduler: DDPMScheduler,
+ safety_checker: Optional[IFSafetyChecker],
+ feature_extractor: Optional[CLIPImageProcessor],
+ watermarker: Optional[IFWatermarker],
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the IF license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ if unet.config.in_channels != 6:
+ logger.warn(
+ "It seems like you have loaded a checkpoint that shall not be used for super resolution from {unet.config._name_or_path} as it accepts {unet.config.in_channels} input channels instead of 6. Please make sure to pass a super resolution checkpoint as the `'unet'`: IFSuperResolutionPipeline.from_pretrained(unet=super_resolution_unet, ...)`."
+ )
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ image_noising_scheduler=image_noising_scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ watermarker=watermarker,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.remove_all_hooks
+ def remove_all_hooks(self):
+ if is_accelerate_available():
+ from accelerate.hooks import remove_hook_from_module
+ else:
+ raise ImportError("Please install accelerate via `pip install accelerate`")
+
+ for model in [self.text_encoder, self.unet, self.safety_checker]:
+ if model is not None:
+ remove_hook_from_module(model, recurse=True)
+
+ self.unet_offload_hook = None
+ self.text_encoder_offload_hook = None
+ self.final_offload_hook = None
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+ def _text_preprocessing(self, text, clean_caption=False):
+ if clean_caption and not is_bs4_available():
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if clean_caption and not is_ftfy_available():
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if not isinstance(text, (tuple, list)):
+ text = [text]
+
+ def process(text: str):
+ if clean_caption:
+ text = self._clean_caption(text)
+ text = self._clean_caption(text)
+ else:
+ text = text.lower().strip()
+ return text
+
+ return [process(t) for t in text]
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+ def _clean_caption(self, caption):
+ caption = str(caption)
+ caption = ul.unquote_plus(caption)
+ caption = caption.strip().lower()
+ caption = re.sub("", "person", caption)
+ # urls:
+ caption = re.sub(
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ caption = re.sub(
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ # html:
+ caption = BeautifulSoup(caption, features="html.parser").text
+
+ # @
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+ # 31C0—31EF CJK Strokes
+ # 31F0—31FF Katakana Phonetic Extensions
+ # 3200—32FF Enclosed CJK Letters and Months
+ # 3300—33FF CJK Compatibility
+ # 3400—4DBF CJK Unified Ideographs Extension A
+ # 4DC0—4DFF Yijing Hexagram Symbols
+ # 4E00—9FFF CJK Unified Ideographs
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+ #######################################################
+
+ # все виды тире / all types of dash --> "-"
+ caption = re.sub(
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
+ "-",
+ caption,
+ )
+
+ # кавычки к одному стандарту
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
+ caption = re.sub(r"[‘’]", "'", caption)
+
+ # "
+ caption = re.sub(r""?", "", caption)
+ # &
+ caption = re.sub(r"&", "", caption)
+
+ # ip adresses:
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+ # article ids:
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+ # \n
+ caption = re.sub(r"\\n", " ", caption)
+
+ # "#123"
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
+ # "#12345.."
+ caption = re.sub(r"#\d{5,}\b", "", caption)
+ # "123456.."
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
+ # filenames:
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+ #
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
+
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
+
+ # this-is-my-cute-cat / this_is_my_cute_cat
+ regex2 = re.compile(r"(?:\-|\_)")
+ if len(re.findall(regex2, caption)) > 3:
+ caption = re.sub(regex2, " ", caption)
+
+ caption = ftfy.fix_text(caption)
+ caption = html.unescape(html.unescape(caption))
+
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
+
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
+
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+ caption = re.sub(r"\s+", " ", caption)
+
+ caption.strip()
+
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+ caption = re.sub(r"^\.\S+$", "", caption)
+
+ return caption.strip()
+
+ @torch.no_grad()
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ do_classifier_free_guidance: bool = True,
+ num_images_per_prompt: int = 1,
+ device: Optional[torch.device] = None,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ clean_caption: bool = False,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ clean_caption (bool, defaults to `False`):
+ If `True`, the function will preprocess and clean the provided caption before encoding.
+ """
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # while T5 can handle much longer input sequences than 77, the text encoder was trained with a max length of 77 for IF
+ max_length = 77
+
+ if prompt_embeds is None:
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {max_length} tokens: {removed_text}"
+ )
+
+ attention_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ dtype = self.unet.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ else:
+ negative_prompt_embeds = None
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+ image, nsfw_detected, watermark_detected = self.safety_checker(
+ images=image,
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
+ )
+ else:
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+
+ return image, nsfw_detected, watermark_detected
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ batch_size,
+ noise_level,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps:
+ raise ValueError(
+ f"`noise_level`: {noise_level} must be a valid timestep in `self.noising_scheduler`, [0, {self.image_noising_scheduler.config.num_train_timesteps})"
+ )
+
+ if isinstance(image, list):
+ check_image_type = image[0]
+ else:
+ check_image_type = image
+
+ if (
+ not isinstance(check_image_type, torch.Tensor)
+ and not isinstance(check_image_type, PIL.Image.Image)
+ and not isinstance(check_image_type, np.ndarray)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, or List[...] but is"
+ f" {type(check_image_type)}"
+ )
+
+ if isinstance(image, list):
+ image_batch_size = len(image)
+ elif isinstance(image, torch.Tensor):
+ image_batch_size = image.shape[0]
+ elif isinstance(image, PIL.Image.Image):
+ image_batch_size = 1
+ elif isinstance(image, np.ndarray):
+ image_batch_size = image.shape[0]
+ else:
+ assert False
+
+ if batch_size != image_batch_size:
+ raise ValueError(f"image batch size: {image_batch_size} must be same as prompt batch size {batch_size}")
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline.prepare_intermediate_images
+ def prepare_intermediate_images(self, batch_size, num_channels, height, width, dtype, device, generator):
+ shape = (batch_size, num_channels, height, width)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ intermediate_images = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ intermediate_images = intermediate_images * self.scheduler.init_noise_sigma
+ return intermediate_images
+
+ def preprocess_image(self, image, num_images_per_prompt, device):
+ if not isinstance(image, torch.Tensor) and not isinstance(image, list):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ image = [np.array(i).astype(np.float32) / 127.5 - 1.0 for i in image]
+
+ image = np.stack(image, axis=0) # to np
+ image = torch.from_numpy(image.transpose(0, 3, 1, 2))
+ elif isinstance(image[0], np.ndarray):
+ image = np.stack(image, axis=0) # to np
+ if image.ndim == 5:
+ image = image[0]
+
+ image = torch.from_numpy(image.transpose(0, 3, 1, 2))
+ elif isinstance(image, list) and isinstance(image[0], torch.Tensor):
+ dims = image[0].ndim
+
+ if dims == 3:
+ image = torch.stack(image, dim=0)
+ elif dims == 4:
+ image = torch.concat(image, dim=0)
+ else:
+ raise ValueError(f"Image must have 3 or 4 dimensions, instead got {dims}")
+
+ image = image.to(device=device, dtype=self.unet.dtype)
+
+ image = image.repeat_interleave(num_images_per_prompt, dim=0)
+
+ return image
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: int = None,
+ width: int = None,
+ image: Union[PIL.Image.Image, np.ndarray, torch.FloatTensor] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 4.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ noise_level: int = 250,
+ clean_caption: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ height (`int`, *optional*, defaults to None):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to None):
+ The width in pixels of the generated image.
+ image (`PIL.Image.Image`, `np.ndarray`, `torch.FloatTensor`):
+ The image to be upscaled.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*, defaults to None):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ noise_level (`int`, *optional*, defaults to 250):
+ The amount of noise to add to the upscaled image. Must be in the range `[0, 1000)`
+ clean_caption (`bool`, *optional*, defaults to `True`):
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
+ prompt.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.IFPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
+ returning a tuple, the first element is a list with the generated images, and the second element is a list
+ of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
+ or watermarked content, according to the `safety_checker`.
+ """
+ # 1. Check inputs. Raise error if not correct
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ self.check_inputs(
+ prompt,
+ image,
+ batch_size,
+ noise_level,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+
+ height = height or self.unet.config.sample_size
+ width = width or self.unet.config.sample_size
+
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ do_classifier_free_guidance,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clean_caption=clean_caption,
+ )
+
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare intermediate images
+ num_channels = self.unet.config.in_channels // 2
+ intermediate_images = self.prepare_intermediate_images(
+ batch_size * num_images_per_prompt,
+ num_channels,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Prepare upscaled image and noise level
+ image = self.preprocess_image(image, num_images_per_prompt, device)
+ upscaled = F.interpolate(image, (height, width), mode="bilinear", align_corners=True)
+
+ noise_level = torch.tensor([noise_level] * upscaled.shape[0], device=upscaled.device)
+ noise = randn_tensor(upscaled.shape, generator=generator, device=upscaled.device, dtype=upscaled.dtype)
+ upscaled = self.image_noising_scheduler.add_noise(upscaled, noise, timesteps=noise_level)
+
+ if do_classifier_free_guidance:
+ noise_level = torch.cat([noise_level] * 2)
+
+ # HACK: see comment in `enable_model_cpu_offload`
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
+ self.text_encoder_offload_hook.offload()
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ model_input = torch.cat([intermediate_images, upscaled], dim=1)
+
+ model_input = torch.cat([model_input] * 2) if do_classifier_free_guidance else model_input
+ model_input = self.scheduler.scale_model_input(model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ class_labels=noise_level,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred_uncond, _ = noise_pred_uncond.split(model_input.shape[1] // 2, dim=1)
+ noise_pred_text, predicted_variance = noise_pred_text.split(model_input.shape[1] // 2, dim=1)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
+
+ if self.scheduler.config.variance_type not in ["learned", "learned_range"]:
+ noise_pred, _ = noise_pred.split(intermediate_images.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ intermediate_images = self.scheduler.step(
+ noise_pred, t, intermediate_images, **extra_step_kwargs, return_dict=False
+ )[0]
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, intermediate_images)
+
+ image = intermediate_images
+
+ if output_type == "pil":
+ # 9. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 10. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # 11. Convert to PIL
+ image = self.numpy_to_pil(image)
+
+ # 12. Apply watermark
+ if self.watermarker is not None:
+ self.watermarker.apply_watermark(image, self.unet.config.sample_size)
+ elif output_type == "pt":
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+ else:
+ # 9. Post-processing
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ # 10. Run safety checker
+ image, nsfw_detected, watermark_detected = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, nsfw_detected, watermark_detected)
+
+ return IFPipelineOutput(images=image, nsfw_detected=nsfw_detected, watermark_detected=watermark_detected)
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/pipeline_output.py b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_output.py
new file mode 100644
index 0000000000000000000000000000000000000000..7f39ab5ba70ccbcaa1ca10438fe829d243277e06
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/pipeline_output.py
@@ -0,0 +1,28 @@
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL.Image
+
+from ...utils import BaseOutput
+
+
+@dataclass
+class IFPipelineOutput(BaseOutput):
+ """
+ Args:
+ Output class for Stable Diffusion pipelines.
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+ num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+ nsfw_detected (`List[bool]`)
+ List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content or a watermark. `None` if safety checking could not be performed.
+ watermark_detected (`List[bool]`)
+ List of flags denoting whether the corresponding generated image likely has a watermark. `None` if safety
+ checking could not be performed.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+ nsfw_detected: Optional[List[bool]]
+ watermark_detected: Optional[List[bool]]
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/safety_checker.py b/model/diffusers_c/pipelines/deepfloyd_if/safety_checker.py
new file mode 100644
index 0000000000000000000000000000000000000000..8ffeed580bbea1514b11bf7a168a952328d8f424
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/safety_checker.py
@@ -0,0 +1,59 @@
+import numpy as np
+import torch
+import torch.nn as nn
+from transformers import CLIPConfig, CLIPVisionModelWithProjection, PreTrainedModel
+
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class IFSafetyChecker(PreTrainedModel):
+ config_class = CLIPConfig
+
+ _no_split_modules = ["CLIPEncoderLayer"]
+
+ def __init__(self, config: CLIPConfig):
+ super().__init__(config)
+
+ self.vision_model = CLIPVisionModelWithProjection(config.vision_config)
+
+ self.p_head = nn.Linear(config.vision_config.projection_dim, 1)
+ self.w_head = nn.Linear(config.vision_config.projection_dim, 1)
+
+ @torch.no_grad()
+ def forward(self, clip_input, images, p_threshold=0.5, w_threshold=0.5):
+ image_embeds = self.vision_model(clip_input)[0]
+
+ nsfw_detected = self.p_head(image_embeds)
+ nsfw_detected = nsfw_detected.flatten()
+ nsfw_detected = nsfw_detected > p_threshold
+ nsfw_detected = nsfw_detected.tolist()
+
+ if any(nsfw_detected):
+ logger.warning(
+ "Potential NSFW content was detected in one or more images. A black image will be returned instead."
+ " Try again with a different prompt and/or seed."
+ )
+
+ for idx, nsfw_detected_ in enumerate(nsfw_detected):
+ if nsfw_detected_:
+ images[idx] = np.zeros(images[idx].shape)
+
+ watermark_detected = self.w_head(image_embeds)
+ watermark_detected = watermark_detected.flatten()
+ watermark_detected = watermark_detected > w_threshold
+ watermark_detected = watermark_detected.tolist()
+
+ if any(watermark_detected):
+ logger.warning(
+ "Potential watermarked content was detected in one or more images. A black image will be returned instead."
+ " Try again with a different prompt and/or seed."
+ )
+
+ for idx, watermark_detected_ in enumerate(watermark_detected):
+ if watermark_detected_:
+ images[idx] = np.zeros(images[idx].shape)
+
+ return images, nsfw_detected, watermark_detected
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/timesteps.py b/model/diffusers_c/pipelines/deepfloyd_if/timesteps.py
new file mode 100644
index 0000000000000000000000000000000000000000..d44285c017bbb2ccffa4ae86dd77792a048625d9
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/timesteps.py
@@ -0,0 +1,579 @@
+fast27_timesteps = [
+ 999,
+ 800,
+ 799,
+ 600,
+ 599,
+ 500,
+ 400,
+ 399,
+ 377,
+ 355,
+ 333,
+ 311,
+ 288,
+ 266,
+ 244,
+ 222,
+ 200,
+ 199,
+ 177,
+ 155,
+ 133,
+ 111,
+ 88,
+ 66,
+ 44,
+ 22,
+ 0,
+]
+
+smart27_timesteps = [
+ 999,
+ 976,
+ 952,
+ 928,
+ 905,
+ 882,
+ 858,
+ 857,
+ 810,
+ 762,
+ 715,
+ 714,
+ 572,
+ 429,
+ 428,
+ 286,
+ 285,
+ 238,
+ 190,
+ 143,
+ 142,
+ 118,
+ 95,
+ 71,
+ 47,
+ 24,
+ 0,
+]
+
+smart50_timesteps = [
+ 999,
+ 988,
+ 977,
+ 966,
+ 955,
+ 944,
+ 933,
+ 922,
+ 911,
+ 900,
+ 899,
+ 879,
+ 859,
+ 840,
+ 820,
+ 800,
+ 799,
+ 766,
+ 733,
+ 700,
+ 699,
+ 650,
+ 600,
+ 599,
+ 500,
+ 499,
+ 400,
+ 399,
+ 350,
+ 300,
+ 299,
+ 266,
+ 233,
+ 200,
+ 199,
+ 179,
+ 159,
+ 140,
+ 120,
+ 100,
+ 99,
+ 88,
+ 77,
+ 66,
+ 55,
+ 44,
+ 33,
+ 22,
+ 11,
+ 0,
+]
+
+smart100_timesteps = [
+ 999,
+ 995,
+ 992,
+ 989,
+ 985,
+ 981,
+ 978,
+ 975,
+ 971,
+ 967,
+ 964,
+ 961,
+ 957,
+ 956,
+ 951,
+ 947,
+ 942,
+ 937,
+ 933,
+ 928,
+ 923,
+ 919,
+ 914,
+ 913,
+ 908,
+ 903,
+ 897,
+ 892,
+ 887,
+ 881,
+ 876,
+ 871,
+ 870,
+ 864,
+ 858,
+ 852,
+ 846,
+ 840,
+ 834,
+ 828,
+ 827,
+ 820,
+ 813,
+ 806,
+ 799,
+ 792,
+ 785,
+ 784,
+ 777,
+ 770,
+ 763,
+ 756,
+ 749,
+ 742,
+ 741,
+ 733,
+ 724,
+ 716,
+ 707,
+ 699,
+ 698,
+ 688,
+ 677,
+ 666,
+ 656,
+ 655,
+ 645,
+ 634,
+ 623,
+ 613,
+ 612,
+ 598,
+ 584,
+ 570,
+ 569,
+ 555,
+ 541,
+ 527,
+ 526,
+ 505,
+ 484,
+ 483,
+ 462,
+ 440,
+ 439,
+ 396,
+ 395,
+ 352,
+ 351,
+ 308,
+ 307,
+ 264,
+ 263,
+ 220,
+ 219,
+ 176,
+ 132,
+ 88,
+ 44,
+ 0,
+]
+
+smart185_timesteps = [
+ 999,
+ 997,
+ 995,
+ 992,
+ 990,
+ 988,
+ 986,
+ 984,
+ 981,
+ 979,
+ 977,
+ 975,
+ 972,
+ 970,
+ 968,
+ 966,
+ 964,
+ 961,
+ 959,
+ 957,
+ 956,
+ 954,
+ 951,
+ 949,
+ 946,
+ 944,
+ 941,
+ 939,
+ 936,
+ 934,
+ 931,
+ 929,
+ 926,
+ 924,
+ 921,
+ 919,
+ 916,
+ 914,
+ 913,
+ 910,
+ 907,
+ 905,
+ 902,
+ 899,
+ 896,
+ 893,
+ 891,
+ 888,
+ 885,
+ 882,
+ 879,
+ 877,
+ 874,
+ 871,
+ 870,
+ 867,
+ 864,
+ 861,
+ 858,
+ 855,
+ 852,
+ 849,
+ 846,
+ 843,
+ 840,
+ 837,
+ 834,
+ 831,
+ 828,
+ 827,
+ 824,
+ 821,
+ 817,
+ 814,
+ 811,
+ 808,
+ 804,
+ 801,
+ 798,
+ 795,
+ 791,
+ 788,
+ 785,
+ 784,
+ 780,
+ 777,
+ 774,
+ 770,
+ 766,
+ 763,
+ 760,
+ 756,
+ 752,
+ 749,
+ 746,
+ 742,
+ 741,
+ 737,
+ 733,
+ 730,
+ 726,
+ 722,
+ 718,
+ 714,
+ 710,
+ 707,
+ 703,
+ 699,
+ 698,
+ 694,
+ 690,
+ 685,
+ 681,
+ 677,
+ 673,
+ 669,
+ 664,
+ 660,
+ 656,
+ 655,
+ 650,
+ 646,
+ 641,
+ 636,
+ 632,
+ 627,
+ 622,
+ 618,
+ 613,
+ 612,
+ 607,
+ 602,
+ 596,
+ 591,
+ 586,
+ 580,
+ 575,
+ 570,
+ 569,
+ 563,
+ 557,
+ 551,
+ 545,
+ 539,
+ 533,
+ 527,
+ 526,
+ 519,
+ 512,
+ 505,
+ 498,
+ 491,
+ 484,
+ 483,
+ 474,
+ 466,
+ 457,
+ 449,
+ 440,
+ 439,
+ 428,
+ 418,
+ 407,
+ 396,
+ 395,
+ 381,
+ 366,
+ 352,
+ 351,
+ 330,
+ 308,
+ 307,
+ 286,
+ 264,
+ 263,
+ 242,
+ 220,
+ 219,
+ 176,
+ 175,
+ 132,
+ 131,
+ 88,
+ 44,
+ 0,
+]
+
+super27_timesteps = [
+ 999,
+ 991,
+ 982,
+ 974,
+ 966,
+ 958,
+ 950,
+ 941,
+ 933,
+ 925,
+ 916,
+ 908,
+ 900,
+ 899,
+ 874,
+ 850,
+ 825,
+ 800,
+ 799,
+ 700,
+ 600,
+ 500,
+ 400,
+ 300,
+ 200,
+ 100,
+ 0,
+]
+
+super40_timesteps = [
+ 999,
+ 992,
+ 985,
+ 978,
+ 971,
+ 964,
+ 957,
+ 949,
+ 942,
+ 935,
+ 928,
+ 921,
+ 914,
+ 907,
+ 900,
+ 899,
+ 879,
+ 859,
+ 840,
+ 820,
+ 800,
+ 799,
+ 766,
+ 733,
+ 700,
+ 699,
+ 650,
+ 600,
+ 599,
+ 500,
+ 499,
+ 400,
+ 399,
+ 300,
+ 299,
+ 200,
+ 199,
+ 100,
+ 99,
+ 0,
+]
+
+super100_timesteps = [
+ 999,
+ 996,
+ 992,
+ 989,
+ 985,
+ 982,
+ 979,
+ 975,
+ 972,
+ 968,
+ 965,
+ 961,
+ 958,
+ 955,
+ 951,
+ 948,
+ 944,
+ 941,
+ 938,
+ 934,
+ 931,
+ 927,
+ 924,
+ 920,
+ 917,
+ 914,
+ 910,
+ 907,
+ 903,
+ 900,
+ 899,
+ 891,
+ 884,
+ 876,
+ 869,
+ 861,
+ 853,
+ 846,
+ 838,
+ 830,
+ 823,
+ 815,
+ 808,
+ 800,
+ 799,
+ 788,
+ 777,
+ 766,
+ 755,
+ 744,
+ 733,
+ 722,
+ 711,
+ 700,
+ 699,
+ 688,
+ 677,
+ 666,
+ 655,
+ 644,
+ 633,
+ 622,
+ 611,
+ 600,
+ 599,
+ 585,
+ 571,
+ 557,
+ 542,
+ 528,
+ 514,
+ 500,
+ 499,
+ 485,
+ 471,
+ 457,
+ 442,
+ 428,
+ 414,
+ 400,
+ 399,
+ 379,
+ 359,
+ 340,
+ 320,
+ 300,
+ 299,
+ 279,
+ 259,
+ 240,
+ 220,
+ 200,
+ 199,
+ 166,
+ 133,
+ 100,
+ 99,
+ 66,
+ 33,
+ 0,
+]
diff --git a/model/diffusers_c/pipelines/deepfloyd_if/watermark.py b/model/diffusers_c/pipelines/deepfloyd_if/watermark.py
new file mode 100644
index 0000000000000000000000000000000000000000..ca10413de1370775842edd42668863d52192a718
--- /dev/null
+++ b/model/diffusers_c/pipelines/deepfloyd_if/watermark.py
@@ -0,0 +1,46 @@
+from typing import List
+
+import PIL.Image
+import torch
+from PIL import Image
+
+from ...configuration_utils import ConfigMixin
+from ...models.modeling_utils import ModelMixin
+from ...utils import PIL_INTERPOLATION
+
+
+class IFWatermarker(ModelMixin, ConfigMixin):
+ def __init__(self):
+ super().__init__()
+
+ self.register_buffer("watermark_image", torch.zeros((62, 62, 4)))
+ self.watermark_image_as_pil = None
+
+ def apply_watermark(self, images: List[PIL.Image.Image], sample_size=None):
+ # copied from https://github.com/deep-floyd/IF/blob/b77482e36ca2031cb94dbca1001fc1e6400bf4ab/deepfloyd_if/modules/base.py#L287
+
+ h = images[0].height
+ w = images[0].width
+
+ sample_size = sample_size or h
+
+ coef = min(h / sample_size, w / sample_size)
+ img_h, img_w = (int(h / coef), int(w / coef)) if coef < 1 else (h, w)
+
+ S1, S2 = 1024**2, img_w * img_h
+ K = (S2 / S1) ** 0.5
+ wm_size, wm_x, wm_y = int(K * 62), img_w - int(14 * K), img_h - int(14 * K)
+
+ if self.watermark_image_as_pil is None:
+ watermark_image = self.watermark_image.to(torch.uint8).cpu().numpy()
+ watermark_image = Image.fromarray(watermark_image, mode="RGBA")
+ self.watermark_image_as_pil = watermark_image
+
+ wm_img = self.watermark_image_as_pil.resize(
+ (wm_size, wm_size), PIL_INTERPOLATION["bicubic"], reducing_gap=None
+ )
+
+ for pil_img in images:
+ pil_img.paste(wm_img, box=(wm_x - wm_size, wm_y - wm_size, wm_x, wm_y), mask=wm_img.split()[-1])
+
+ return images
diff --git a/model/diffusers_c/pipelines/deprecated/README.md b/model/diffusers_c/pipelines/deprecated/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..f389266ed5e673831e7580b517ca361f000eb8ca
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/README.md
@@ -0,0 +1,3 @@
+# Deprecated Pipelines
+
+This folder contains pipelines that have very low usage as measured by model downloads, issues and PRs. While you can still use the pipelines just as before, we will stop testing the pipelines and will not accept any changes to existing files.
diff --git a/model/diffusers_c/pipelines/deprecated/__init__.py b/model/diffusers_c/pipelines/deprecated/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..9936323170adbceac2c5c25e3881ea731d8602e1
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/__init__.py
@@ -0,0 +1,153 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_librosa_available,
+ is_note_seq_available,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_pt_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_pt_objects))
+else:
+ _import_structure["latent_diffusion_uncond"] = ["LDMPipeline"]
+ _import_structure["pndm"] = ["PNDMPipeline"]
+ _import_structure["repaint"] = ["RePaintPipeline"]
+ _import_structure["score_sde_ve"] = ["ScoreSdeVePipeline"]
+ _import_structure["stochastic_karras_ve"] = ["KarrasVePipeline"]
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["alt_diffusion"] = [
+ "AltDiffusionImg2ImgPipeline",
+ "AltDiffusionPipeline",
+ "AltDiffusionPipelineOutput",
+ ]
+ _import_structure["versatile_diffusion"] = [
+ "VersatileDiffusionDualGuidedPipeline",
+ "VersatileDiffusionImageVariationPipeline",
+ "VersatileDiffusionPipeline",
+ "VersatileDiffusionTextToImagePipeline",
+ ]
+ _import_structure["vq_diffusion"] = ["VQDiffusionPipeline"]
+ _import_structure["stable_diffusion_variants"] = [
+ "CycleDiffusionPipeline",
+ "StableDiffusionInpaintPipelineLegacy",
+ "StableDiffusionPix2PixZeroPipeline",
+ "StableDiffusionParadigmsPipeline",
+ "StableDiffusionModelEditingPipeline",
+ ]
+
+try:
+ if not (is_torch_available() and is_librosa_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_librosa_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_librosa_objects))
+
+else:
+ _import_structure["audio_diffusion"] = ["AudioDiffusionPipeline", "Mel"]
+
+try:
+ if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_transformers_and_torch_and_note_seq_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_transformers_and_torch_and_note_seq_objects))
+
+else:
+ _import_structure["spectrogram_diffusion"] = ["MidiProcessor", "SpectrogramDiffusionPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_pt_objects import *
+
+ else:
+ from .latent_diffusion_uncond import LDMPipeline
+ from .pndm import PNDMPipeline
+ from .repaint import RePaintPipeline
+ from .score_sde_ve import ScoreSdeVePipeline
+ from .stochastic_karras_ve import KarrasVePipeline
+
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+
+ else:
+ from .alt_diffusion import AltDiffusionImg2ImgPipeline, AltDiffusionPipeline, AltDiffusionPipelineOutput
+ from .audio_diffusion import AudioDiffusionPipeline, Mel
+ from .spectrogram_diffusion import SpectrogramDiffusionPipeline
+ from .stable_diffusion_variants import (
+ CycleDiffusionPipeline,
+ StableDiffusionInpaintPipelineLegacy,
+ StableDiffusionModelEditingPipeline,
+ StableDiffusionParadigmsPipeline,
+ StableDiffusionPix2PixZeroPipeline,
+ )
+ from .stochastic_karras_ve import KarrasVePipeline
+ from .versatile_diffusion import (
+ VersatileDiffusionDualGuidedPipeline,
+ VersatileDiffusionImageVariationPipeline,
+ VersatileDiffusionPipeline,
+ VersatileDiffusionTextToImagePipeline,
+ )
+ from .vq_diffusion import VQDiffusionPipeline
+
+ try:
+ if not (is_torch_available() and is_librosa_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_librosa_objects import *
+ else:
+ from .audio_diffusion import AudioDiffusionPipeline, Mel
+
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_transformers_and_torch_and_note_seq_objects import * # noqa F403
+ else:
+ from .spectrogram_diffusion import (
+ MidiProcessor,
+ SpectrogramDiffusionPipeline,
+ )
+
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/deprecated/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..078d1dea128f26602f179e22784ef9f81a5d1524
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/__init__.py b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..71fa15b3feff08dc4008d1fa02ba61ad1300efed
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__init__.py
@@ -0,0 +1,53 @@
+from typing import TYPE_CHECKING
+
+from ....utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ....utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["modeling_roberta_series"] = ["RobertaSeriesModelWithTransformation"]
+ _import_structure["pipeline_alt_diffusion"] = ["AltDiffusionPipeline"]
+ _import_structure["pipeline_alt_diffusion_img2img"] = ["AltDiffusionImg2ImgPipeline"]
+
+ _import_structure["pipeline_output"] = ["AltDiffusionPipelineOutput"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ....utils.dummy_torch_and_transformers_objects import *
+
+ else:
+ from .modeling_roberta_series import RobertaSeriesModelWithTransformation
+ from .pipeline_alt_diffusion import AltDiffusionPipeline
+ from .pipeline_alt_diffusion_img2img import AltDiffusionImg2ImgPipeline
+ from .pipeline_output import AltDiffusionPipelineOutput
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d95d617025e36e25b24ef249c274408ae3cd2611
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/modeling_roberta_series.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/modeling_roberta_series.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f9bca59a0e4d71d58c7a852e2af14df85b4c34e1
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/modeling_roberta_series.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_alt_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_alt_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..39f55743b13fdce6ddacfea59328e0f785a96c79
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_alt_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_alt_diffusion_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_alt_diffusion_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1648eaff1ca97f7202ceb1269f4937150d277874
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_alt_diffusion_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_output.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_output.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f22b8ede4b88d4531db6eaac7af0cdfc4160052f
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/alt_diffusion/__pycache__/pipeline_output.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/modeling_roberta_series.py b/model/diffusers_c/pipelines/deprecated/alt_diffusion/modeling_roberta_series.py
new file mode 100644
index 0000000000000000000000000000000000000000..f73ef15d7de7948a9cbad246027ca71f4a6db198
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/alt_diffusion/modeling_roberta_series.py
@@ -0,0 +1,124 @@
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import torch
+from torch import nn
+from transformers import RobertaPreTrainedModel, XLMRobertaConfig, XLMRobertaModel
+from transformers.utils import ModelOutput
+
+
+@dataclass
+class TransformationModelOutput(ModelOutput):
+ """
+ Base class for text model's outputs that also contains a pooling of the last hidden states.
+
+ Args:
+ text_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
+ The text embeddings obtained by applying the projection layer to the pooler_output.
+ last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+ Sequence of hidden-states at the output of the last layer of the model.
+ hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+ Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+ one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+ Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+ attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+ Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+ sequence_length)`.
+
+ Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+ heads.
+ """
+
+ projection_state: Optional[torch.FloatTensor] = None
+ last_hidden_state: torch.FloatTensor = None
+ hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+ attentions: Optional[Tuple[torch.FloatTensor]] = None
+
+
+class RobertaSeriesConfig(XLMRobertaConfig):
+ def __init__(
+ self,
+ pad_token_id=1,
+ bos_token_id=0,
+ eos_token_id=2,
+ project_dim=512,
+ pooler_fn="cls",
+ learn_encoder=False,
+ use_attention_mask=True,
+ **kwargs,
+ ):
+ super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+ self.project_dim = project_dim
+ self.pooler_fn = pooler_fn
+ self.learn_encoder = learn_encoder
+ self.use_attention_mask = use_attention_mask
+
+
+class RobertaSeriesModelWithTransformation(RobertaPreTrainedModel):
+ _keys_to_ignore_on_load_unexpected = [r"pooler", r"logit_scale"]
+ _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+ base_model_prefix = "roberta"
+ config_class = RobertaSeriesConfig
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.roberta = XLMRobertaModel(config)
+ self.transformation = nn.Linear(config.hidden_size, config.project_dim)
+ self.has_pre_transformation = getattr(config, "has_pre_transformation", False)
+ if self.has_pre_transformation:
+ self.transformation_pre = nn.Linear(config.hidden_size, config.project_dim)
+ self.pre_LN = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+ self.post_init()
+
+ def forward(
+ self,
+ input_ids: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ token_type_ids: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.Tensor] = None,
+ head_mask: Optional[torch.Tensor] = None,
+ inputs_embeds: Optional[torch.Tensor] = None,
+ encoder_hidden_states: Optional[torch.Tensor] = None,
+ encoder_attention_mask: Optional[torch.Tensor] = None,
+ output_attentions: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ ):
+ r""" """
+
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+ outputs = self.base_model(
+ input_ids=input_ids,
+ attention_mask=attention_mask,
+ token_type_ids=token_type_ids,
+ position_ids=position_ids,
+ head_mask=head_mask,
+ inputs_embeds=inputs_embeds,
+ encoder_hidden_states=encoder_hidden_states,
+ encoder_attention_mask=encoder_attention_mask,
+ output_attentions=output_attentions,
+ output_hidden_states=True if self.has_pre_transformation else output_hidden_states,
+ return_dict=return_dict,
+ )
+
+ if self.has_pre_transformation:
+ sequence_output2 = outputs["hidden_states"][-2]
+ sequence_output2 = self.pre_LN(sequence_output2)
+ projection_state2 = self.transformation_pre(sequence_output2)
+
+ return TransformationModelOutput(
+ projection_state=projection_state2,
+ last_hidden_state=outputs.last_hidden_state,
+ hidden_states=outputs.hidden_states,
+ attentions=outputs.attentions,
+ )
+ else:
+ projection_state = self.transformation(outputs.last_hidden_state)
+ return TransformationModelOutput(
+ projection_state=projection_state,
+ last_hidden_state=outputs.last_hidden_state,
+ hidden_states=outputs.hidden_states,
+ attentions=outputs.attentions,
+ )
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_alt_diffusion.py b/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_alt_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..e4583699e79e54aaccab60d496293af2b45ee846
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_alt_diffusion.py
@@ -0,0 +1,946 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection, XLMRobertaTokenizer
+
+from ....configuration_utils import FrozenDict
+from ....image_processor import PipelineImageInput, VaeImageProcessor
+from ....loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ....models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ....models.lora import adjust_lora_scale_text_encoder
+from ....schedulers import KarrasDiffusionSchedulers
+from ....utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from .modeling_roberta_series import RobertaSeriesModelWithTransformation
+from .pipeline_output import AltDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import AltDiffusionPipeline
+
+ >>> pipe = AltDiffusionPipeline.from_pretrained("BAAI/AltDiffusion-m9", torch_dtype=torch.float16)
+ >>> pipe = pipe.to("cuda")
+
+ >>> # "dark elf princess, highly detailed, d & d, fantasy, highly detailed, digital painting, trending on artstation, concept art, sharp focus, illustration, art by artgerm and greg rutkowski and fuji choko and viktoria gavrilenko and hoang lap"
+ >>> prompt = "黑暗精灵公主，非常详细，幻想，非常详细，数字绘画，概念艺术，敏锐的焦点，插图"
+ >>> image = pipe(prompt).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class AltDiffusionPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Alt Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.RobertaSeriesModelWithTransformation`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.XLMRobertaTokenizer`]):
+ A `XLMRobertaTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: RobertaSeriesModelWithTransformation,
+ tokenizer: XLMRobertaTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Alt Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def guidance_rescale(self):
+ return self._guidance_rescale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.AltDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ guidance_rescale (`float`, *optional*, defaults to 0.0):
+ Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
+ using zero terminal SNR.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.AltDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.AltDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+ # to deal with lora scaling and other possible forward hooks
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._guidance_rescale = guidance_rescale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None:
+ output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+ image_embeds, negative_image_embeds = self.encode_image(
+ ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+ )
+ if self.do_classifier_free_guidance:
+ image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 6.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
+ # 6.2 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return AltDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_alt_diffusion_img2img.py b/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_alt_diffusion_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..156e52c249d9f73d8cbc3a10c7624257071b16eb
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_alt_diffusion_img2img.py
@@ -0,0 +1,1018 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection, XLMRobertaTokenizer
+
+from ....configuration_utils import FrozenDict
+from ....image_processor import PipelineImageInput, VaeImageProcessor
+from ....loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ....models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ....models.lora import adjust_lora_scale_text_encoder
+from ....schedulers import KarrasDiffusionSchedulers
+from ....utils import (
+ PIL_INTERPOLATION,
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from .modeling_roberta_series import RobertaSeriesModelWithTransformation
+from .pipeline_output import AltDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import requests
+ >>> import torch
+ >>> from PIL import Image
+ >>> from io import BytesIO
+
+ >>> from diffusers import AltDiffusionImg2ImgPipeline
+
+ >>> device = "cuda"
+ >>> model_id_or_path = "BAAI/AltDiffusion-m9"
+ >>> pipe = AltDiffusionImg2ImgPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
+ >>> pipe = pipe.to(device)
+
+ >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+
+ >>> response = requests.get(url)
+ >>> init_image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> init_image = init_image.resize((768, 512))
+
+ >>> # "A fantasy landscape, trending on artstation"
+ >>> prompt = "幻想风景, artstation"
+
+ >>> images = pipe(prompt=prompt, image=init_image, strength=0.75, guidance_scale=7.5).images
+ >>> images[0].save("幻想风景.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
+def preprocess(image):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class AltDiffusionImg2ImgPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-guided image-to-image generation using Alt Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.RobertaSeriesModelWithTransformation`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.XLMRobertaTokenizer`]):
+ A `XLMRobertaTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: RobertaSeriesModelWithTransformation,
+ tokenizer: XLMRobertaTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Alt Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+
+ return latents
+
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ strength: float = 0.8,
+ num_inference_steps: Optional[int] = 50,
+ timesteps: List[int] = None,
+ guidance_scale: Optional[float] = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: Optional[float] = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: int = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
+ numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
+ or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
+ list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
+ latents as `image`, but if passing latents directly it is not encoded again.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.AltDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.AltDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.AltDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None:
+ output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+ image_embeds, negative_image_embeds = self.encode_image(
+ ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+ )
+ if self.do_classifier_free_guidance:
+ image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
+ # 4. Preprocess image
+ image = self.image_processor.preprocess(image)
+
+ # 5. set timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+ # 6. Prepare latent variables
+ latents = self.prepare_latents(
+ image,
+ latent_timestep,
+ batch_size,
+ num_images_per_prompt,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
+ # 7.2 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return AltDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_output.py b/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_output.py
new file mode 100644
index 0000000000000000000000000000000000000000..dd174ae3c21fe4110babd503f0418366472059ff
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/alt_diffusion/pipeline_output.py
@@ -0,0 +1,28 @@
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL.Image
+
+from ....utils import (
+ BaseOutput,
+)
+
+
+@dataclass
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_output.StableDiffusionPipelineOutput with Stable->Alt
+class AltDiffusionPipelineOutput(BaseOutput):
+ """
+ Output class for Alt Diffusion pipelines.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ nsfw_content_detected (`List[bool]`)
+ List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or
+ `None` if safety checking could not be performed.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+ nsfw_content_detected: Optional[List[bool]]
diff --git a/model/diffusers_c/pipelines/deprecated/audio_diffusion/__init__.py b/model/diffusers_c/pipelines/deprecated/audio_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..3127951863a7db3f9dd8e42ac5ab64fa9ac3ec0c
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/audio_diffusion/__init__.py
@@ -0,0 +1,23 @@
+from typing import TYPE_CHECKING
+
+from ....utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {
+ "mel": ["Mel"],
+ "pipeline_audio_diffusion": ["AudioDiffusionPipeline"],
+}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .mel import Mel
+ from .pipeline_audio_diffusion import AudioDiffusionPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/deprecated/audio_diffusion/mel.py b/model/diffusers_c/pipelines/deprecated/audio_diffusion/mel.py
new file mode 100644
index 0000000000000000000000000000000000000000..3426c3ad0428efc69f2f8dee7a620df3c78b1405
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/audio_diffusion/mel.py
@@ -0,0 +1,179 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import numpy as np # noqa: E402
+
+from ....configuration_utils import ConfigMixin, register_to_config
+from ....schedulers.scheduling_utils import SchedulerMixin
+
+
+try:
+ import librosa # noqa: E402
+
+ _librosa_can_be_imported = True
+ _import_error = ""
+except Exception as e:
+ _librosa_can_be_imported = False
+ _import_error = (
+ f"Cannot import librosa because {e}. Make sure to correctly install librosa to be able to install it."
+ )
+
+
+from PIL import Image # noqa: E402
+
+
+class Mel(ConfigMixin, SchedulerMixin):
+ """
+ Parameters:
+ x_res (`int`):
+ x resolution of spectrogram (time).
+ y_res (`int`):
+ y resolution of spectrogram (frequency bins).
+ sample_rate (`int`):
+ Sample rate of audio.
+ n_fft (`int`):
+ Number of Fast Fourier Transforms.
+ hop_length (`int`):
+ Hop length (a higher number is recommended if `y_res` < 256).
+ top_db (`int`):
+ Loudest decibel value.
+ n_iter (`int`):
+ Number of iterations for Griffin-Lim Mel inversion.
+ """
+
+ config_name = "mel_config.json"
+
+ @register_to_config
+ def __init__(
+ self,
+ x_res: int = 256,
+ y_res: int = 256,
+ sample_rate: int = 22050,
+ n_fft: int = 2048,
+ hop_length: int = 512,
+ top_db: int = 80,
+ n_iter: int = 32,
+ ):
+ self.hop_length = hop_length
+ self.sr = sample_rate
+ self.n_fft = n_fft
+ self.top_db = top_db
+ self.n_iter = n_iter
+ self.set_resolution(x_res, y_res)
+ self.audio = None
+
+ if not _librosa_can_be_imported:
+ raise ValueError(_import_error)
+
+ def set_resolution(self, x_res: int, y_res: int):
+ """Set resolution.
+
+ Args:
+ x_res (`int`):
+ x resolution of spectrogram (time).
+ y_res (`int`):
+ y resolution of spectrogram (frequency bins).
+ """
+ self.x_res = x_res
+ self.y_res = y_res
+ self.n_mels = self.y_res
+ self.slice_size = self.x_res * self.hop_length - 1
+
+ def load_audio(self, audio_file: str = None, raw_audio: np.ndarray = None):
+ """Load audio.
+
+ Args:
+ audio_file (`str`):
+ An audio file that must be on disk due to [Librosa](https://librosa.org/) limitation.
+ raw_audio (`np.ndarray`):
+ The raw audio file as a NumPy array.
+ """
+ if audio_file is not None:
+ self.audio, _ = librosa.load(audio_file, mono=True, sr=self.sr)
+ else:
+ self.audio = raw_audio
+
+ # Pad with silence if necessary.
+ if len(self.audio) < self.x_res * self.hop_length:
+ self.audio = np.concatenate([self.audio, np.zeros((self.x_res * self.hop_length - len(self.audio),))])
+
+ def get_number_of_slices(self) -> int:
+ """Get number of slices in audio.
+
+ Returns:
+ `int`:
+ Number of spectograms audio can be sliced into.
+ """
+ return len(self.audio) // self.slice_size
+
+ def get_audio_slice(self, slice: int = 0) -> np.ndarray:
+ """Get slice of audio.
+
+ Args:
+ slice (`int`):
+ Slice number of audio (out of `get_number_of_slices()`).
+
+ Returns:
+ `np.ndarray`:
+ The audio slice as a NumPy array.
+ """
+ return self.audio[self.slice_size * slice : self.slice_size * (slice + 1)]
+
+ def get_sample_rate(self) -> int:
+ """Get sample rate.
+
+ Returns:
+ `int`:
+ Sample rate of audio.
+ """
+ return self.sr
+
+ def audio_slice_to_image(self, slice: int) -> Image.Image:
+ """Convert slice of audio to spectrogram.
+
+ Args:
+ slice (`int`):
+ Slice number of audio to convert (out of `get_number_of_slices()`).
+
+ Returns:
+ `PIL Image`:
+ A grayscale image of `x_res x y_res`.
+ """
+ S = librosa.feature.melspectrogram(
+ y=self.get_audio_slice(slice), sr=self.sr, n_fft=self.n_fft, hop_length=self.hop_length, n_mels=self.n_mels
+ )
+ log_S = librosa.power_to_db(S, ref=np.max, top_db=self.top_db)
+ bytedata = (((log_S + self.top_db) * 255 / self.top_db).clip(0, 255) + 0.5).astype(np.uint8)
+ image = Image.fromarray(bytedata)
+ return image
+
+ def image_to_audio(self, image: Image.Image) -> np.ndarray:
+ """Converts spectrogram to audio.
+
+ Args:
+ image (`PIL Image`):
+ An grayscale image of `x_res x y_res`.
+
+ Returns:
+ audio (`np.ndarray`):
+ The audio as a NumPy array.
+ """
+ bytedata = np.frombuffer(image.tobytes(), dtype="uint8").reshape((image.height, image.width))
+ log_S = bytedata.astype("float") * self.top_db / 255 - self.top_db
+ S = librosa.db_to_power(log_S)
+ audio = librosa.feature.inverse.mel_to_audio(
+ S, sr=self.sr, n_fft=self.n_fft, hop_length=self.hop_length, n_iter=self.n_iter
+ )
+ return audio
diff --git a/model/diffusers_c/pipelines/deprecated/audio_diffusion/pipeline_audio_diffusion.py b/model/diffusers_c/pipelines/deprecated/audio_diffusion/pipeline_audio_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..47044e050acf7221e3da74d78defa042d5d583f2
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/audio_diffusion/pipeline_audio_diffusion.py
@@ -0,0 +1,329 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from math import acos, sin
+from typing import List, Tuple, Union
+
+import numpy as np
+import torch
+from PIL import Image
+
+from ....models import AutoencoderKL, UNet2DConditionModel
+from ....schedulers import DDIMScheduler, DDPMScheduler
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import AudioPipelineOutput, BaseOutput, DiffusionPipeline, ImagePipelineOutput
+from .mel import Mel
+
+
+class AudioDiffusionPipeline(DiffusionPipeline):
+ """
+ Pipeline for audio diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ vqae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ mel ([`Mel`]):
+ Transform audio into a spectrogram.
+ scheduler ([`DDIMScheduler`] or [`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`] or [`DDPMScheduler`].
+ """
+
+ _optional_components = ["vqvae"]
+
+ def __init__(
+ self,
+ vqvae: AutoencoderKL,
+ unet: UNet2DConditionModel,
+ mel: Mel,
+ scheduler: Union[DDIMScheduler, DDPMScheduler],
+ ):
+ super().__init__()
+ self.register_modules(unet=unet, scheduler=scheduler, mel=mel, vqvae=vqvae)
+
+ def get_default_steps(self) -> int:
+ """Returns default number of steps recommended for inference.
+
+ Returns:
+ `int`:
+ The number of steps.
+ """
+ return 50 if isinstance(self.scheduler, DDIMScheduler) else 1000
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ batch_size: int = 1,
+ audio_file: str = None,
+ raw_audio: np.ndarray = None,
+ slice: int = 0,
+ start_step: int = 0,
+ steps: int = None,
+ generator: torch.Generator = None,
+ mask_start_secs: float = 0,
+ mask_end_secs: float = 0,
+ step_generator: torch.Generator = None,
+ eta: float = 0,
+ noise: torch.Tensor = None,
+ encoding: torch.Tensor = None,
+ return_dict=True,
+ ) -> Union[
+ Union[AudioPipelineOutput, ImagePipelineOutput],
+ Tuple[List[Image.Image], Tuple[int, List[np.ndarray]]],
+ ]:
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ batch_size (`int`):
+ Number of samples to generate.
+ audio_file (`str`):
+ An audio file that must be on disk due to [Librosa](https://librosa.org/) limitation.
+ raw_audio (`np.ndarray`):
+ The raw audio file as a NumPy array.
+ slice (`int`):
+ Slice number of audio to convert.
+ start_step (int):
+ Step to start diffusion from.
+ steps (`int`):
+ Number of denoising steps (defaults to `50` for DDIM and `1000` for DDPM).
+ generator (`torch.Generator`):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ mask_start_secs (`float`):
+ Number of seconds of audio to mask (not generate) at start.
+ mask_end_secs (`float`):
+ Number of seconds of audio to mask (not generate) at end.
+ step_generator (`torch.Generator`):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) used to denoise.
+ None
+ eta (`float`):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ noise (`torch.Tensor`):
+ A noise tensor of shape `(batch_size, 1, height, width)` or `None`.
+ encoding (`torch.Tensor`):
+ A tensor for [`UNet2DConditionModel`] of shape `(batch_size, seq_length, cross_attention_dim)`.
+ return_dict (`bool`):
+ Whether or not to return a [`AudioPipelineOutput`], [`ImagePipelineOutput`] or a plain tuple.
+
+ Examples:
+
+ For audio diffusion:
+
+ ```py
+ import torch
+ from IPython.display import Audio
+ from diffusers import DiffusionPipeline
+
+ device = "cuda" if torch.cuda.is_available() else "cpu"
+ pipe = DiffusionPipeline.from_pretrained("teticio/audio-diffusion-256").to(device)
+
+ output = pipe()
+ display(output.images[0])
+ display(Audio(output.audios[0], rate=mel.get_sample_rate()))
+ ```
+
+ For latent audio diffusion:
+
+ ```py
+ import torch
+ from IPython.display import Audio
+ from diffusers import DiffusionPipeline
+
+ device = "cuda" if torch.cuda.is_available() else "cpu"
+ pipe = DiffusionPipeline.from_pretrained("teticio/latent-audio-diffusion-256").to(device)
+
+ output = pipe()
+ display(output.images[0])
+ display(Audio(output.audios[0], rate=pipe.mel.get_sample_rate()))
+ ```
+
+ For other tasks like variation, inpainting, outpainting, etc:
+
+ ```py
+ output = pipe(
+ raw_audio=output.audios[0, 0],
+ start_step=int(pipe.get_default_steps() / 2),
+ mask_start_secs=1,
+ mask_end_secs=1,
+ )
+ display(output.images[0])
+ display(Audio(output.audios[0], rate=pipe.mel.get_sample_rate()))
+ ```
+
+ Returns:
+ `List[PIL Image]`:
+ A list of Mel spectrograms (`float`, `List[np.ndarray]`) with the sample rate and raw audio.
+ """
+
+ steps = steps or self.get_default_steps()
+ self.scheduler.set_timesteps(steps)
+ step_generator = step_generator or generator
+ # For backwards compatibility
+ if isinstance(self.unet.config.sample_size, int):
+ self.unet.config.sample_size = (self.unet.config.sample_size, self.unet.config.sample_size)
+ if noise is None:
+ noise = randn_tensor(
+ (
+ batch_size,
+ self.unet.config.in_channels,
+ self.unet.config.sample_size[0],
+ self.unet.config.sample_size[1],
+ ),
+ generator=generator,
+ device=self.device,
+ )
+ images = noise
+ mask = None
+
+ if audio_file is not None or raw_audio is not None:
+ self.mel.load_audio(audio_file, raw_audio)
+ input_image = self.mel.audio_slice_to_image(slice)
+ input_image = np.frombuffer(input_image.tobytes(), dtype="uint8").reshape(
+ (input_image.height, input_image.width)
+ )
+ input_image = (input_image / 255) * 2 - 1
+ input_images = torch.tensor(input_image[np.newaxis, :, :], dtype=torch.float).to(self.device)
+
+ if self.vqvae is not None:
+ input_images = self.vqvae.encode(torch.unsqueeze(input_images, 0)).latent_dist.sample(
+ generator=generator
+ )[0]
+ input_images = self.vqvae.config.scaling_factor * input_images
+
+ if start_step > 0:
+ images[0, 0] = self.scheduler.add_noise(input_images, noise, self.scheduler.timesteps[start_step - 1])
+
+ pixels_per_second = (
+ self.unet.config.sample_size[1] * self.mel.get_sample_rate() / self.mel.x_res / self.mel.hop_length
+ )
+ mask_start = int(mask_start_secs * pixels_per_second)
+ mask_end = int(mask_end_secs * pixels_per_second)
+ mask = self.scheduler.add_noise(input_images, noise, torch.tensor(self.scheduler.timesteps[start_step:]))
+
+ for step, t in enumerate(self.progress_bar(self.scheduler.timesteps[start_step:])):
+ if isinstance(self.unet, UNet2DConditionModel):
+ model_output = self.unet(images, t, encoding)["sample"]
+ else:
+ model_output = self.unet(images, t)["sample"]
+
+ if isinstance(self.scheduler, DDIMScheduler):
+ images = self.scheduler.step(
+ model_output=model_output,
+ timestep=t,
+ sample=images,
+ eta=eta,
+ generator=step_generator,
+ )["prev_sample"]
+ else:
+ images = self.scheduler.step(
+ model_output=model_output,
+ timestep=t,
+ sample=images,
+ generator=step_generator,
+ )["prev_sample"]
+
+ if mask is not None:
+ if mask_start > 0:
+ images[:, :, :, :mask_start] = mask[:, step, :, :mask_start]
+ if mask_end > 0:
+ images[:, :, :, -mask_end:] = mask[:, step, :, -mask_end:]
+
+ if self.vqvae is not None:
+ # 0.18215 was scaling factor used in training to ensure unit variance
+ images = 1 / self.vqvae.config.scaling_factor * images
+ images = self.vqvae.decode(images)["sample"]
+
+ images = (images / 2 + 0.5).clamp(0, 1)
+ images = images.cpu().permute(0, 2, 3, 1).numpy()
+ images = (images * 255).round().astype("uint8")
+ images = list(
+ (Image.fromarray(_[:, :, 0]) for _ in images)
+ if images.shape[3] == 1
+ else (Image.fromarray(_, mode="RGB").convert("L") for _ in images)
+ )
+
+ audios = [self.mel.image_to_audio(_) for _ in images]
+ if not return_dict:
+ return images, (self.mel.get_sample_rate(), audios)
+
+ return BaseOutput(**AudioPipelineOutput(np.array(audios)[:, np.newaxis, :]), **ImagePipelineOutput(images))
+
+ @torch.no_grad()
+ def encode(self, images: List[Image.Image], steps: int = 50) -> np.ndarray:
+ """
+ Reverse the denoising step process to recover a noisy image from the generated image.
+
+ Args:
+ images (`List[PIL Image]`):
+ List of images to encode.
+ steps (`int`):
+ Number of encoding steps to perform (defaults to `50`).
+
+ Returns:
+ `np.ndarray`:
+ A noise tensor of shape `(batch_size, 1, height, width)`.
+ """
+
+ # Only works with DDIM as this method is deterministic
+ assert isinstance(self.scheduler, DDIMScheduler)
+ self.scheduler.set_timesteps(steps)
+ sample = np.array(
+ [np.frombuffer(image.tobytes(), dtype="uint8").reshape((1, image.height, image.width)) for image in images]
+ )
+ sample = (sample / 255) * 2 - 1
+ sample = torch.Tensor(sample).to(self.device)
+
+ for t in self.progress_bar(torch.flip(self.scheduler.timesteps, (0,))):
+ prev_timestep = t - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps
+ alpha_prod_t = self.scheduler.alphas_cumprod[t]
+ alpha_prod_t_prev = (
+ self.scheduler.alphas_cumprod[prev_timestep]
+ if prev_timestep >= 0
+ else self.scheduler.final_alpha_cumprod
+ )
+ beta_prod_t = 1 - alpha_prod_t
+ model_output = self.unet(sample, t)["sample"]
+ pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * model_output
+ sample = (sample - pred_sample_direction) * alpha_prod_t_prev ** (-0.5)
+ sample = sample * alpha_prod_t ** (0.5) + beta_prod_t ** (0.5) * model_output
+
+ return sample
+
+ @staticmethod
+ def slerp(x0: torch.Tensor, x1: torch.Tensor, alpha: float) -> torch.Tensor:
+ """Spherical Linear intERPolation.
+
+ Args:
+ x0 (`torch.Tensor`):
+ The first tensor to interpolate between.
+ x1 (`torch.Tensor`):
+ Second tensor to interpolate between.
+ alpha (`float`):
+ Interpolation between 0 and 1
+
+ Returns:
+ `torch.Tensor`:
+ The interpolated tensor.
+ """
+
+ theta = acos(torch.dot(torch.flatten(x0), torch.flatten(x1)) / torch.norm(x0) / torch.norm(x1))
+ return sin((1 - alpha) * theta) * x0 / sin(theta) + sin(alpha * theta) * x1 / sin(theta)
diff --git a/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__init__.py b/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..214f5bbca969f9ae0629578c72aaf339f86ded88
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__init__.py
@@ -0,0 +1,18 @@
+from typing import TYPE_CHECKING
+
+from ....utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {"pipeline_latent_diffusion_uncond": ["LDMPipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_latent_diffusion_uncond import LDMPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d06ea9901896b43d52b0a2fb986feee71681691c
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__pycache__/pipeline_latent_diffusion_uncond.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__pycache__/pipeline_latent_diffusion_uncond.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5768bc301882de2645ebc04d4edc58b1ec7cb5c3
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/__pycache__/pipeline_latent_diffusion_uncond.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/pipeline_latent_diffusion_uncond.py b/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/pipeline_latent_diffusion_uncond.py
new file mode 100644
index 0000000000000000000000000000000000000000..7fe5d59f771da48ea3c7fb549dbb9762cac4620e
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/latent_diffusion_uncond/pipeline_latent_diffusion_uncond.py
@@ -0,0 +1,130 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ....models import UNet2DModel, VQModel
+from ....schedulers import DDIMScheduler
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+class LDMPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for unconditional image generation using latent diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ vqvae ([`VQModel`]):
+ Vector-quantized (VQ) model to encode and decode images to and from latent representations.
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ [`DDIMScheduler`] is used in combination with `unet` to denoise the encoded image latents.
+ """
+
+ def __init__(self, vqvae: VQModel, unet: UNet2DModel, scheduler: DDIMScheduler):
+ super().__init__()
+ self.register_modules(vqvae=vqvae, unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ batch_size: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ eta: float = 0.0,
+ num_inference_steps: int = 50,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ **kwargs,
+ ) -> Union[Tuple, ImagePipelineOutput]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ batch_size (`int`, *optional*, defaults to 1):
+ Number of images to generate.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ ```py
+ >>> from diffusers import LDMPipeline
+
+ >>> # load model and scheduler
+ >>> pipe = LDMPipeline.from_pretrained("CompVis/ldm-celebahq-256")
+
+ >>> # run pipeline in inference (sample random noise and denoise)
+ >>> image = pipe().images[0]
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images
+ """
+
+ latents = randn_tensor(
+ (batch_size, self.unet.config.in_channels, self.unet.config.sample_size, self.unet.config.sample_size),
+ generator=generator,
+ )
+ latents = latents.to(self.device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+
+ extra_kwargs = {}
+ if accepts_eta:
+ extra_kwargs["eta"] = eta
+
+ for t in self.progress_bar(self.scheduler.timesteps):
+ latent_model_input = self.scheduler.scale_model_input(latents, t)
+ # predict the noise residual
+ noise_prediction = self.unet(latent_model_input, t).sample
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_prediction, t, latents, **extra_kwargs).prev_sample
+
+ # adjust latents with inverse of vae scale
+ latents = latents / self.vqvae.config.scaling_factor
+ # decode the image latents with the VAE
+ image = self.vqvae.decode(latents).sample
+
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/deprecated/pndm/__init__.py b/model/diffusers_c/pipelines/deprecated/pndm/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5e3bdba74079d77576655e22b43014a0438a9c2e
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/pndm/__init__.py
@@ -0,0 +1,18 @@
+from typing import TYPE_CHECKING
+
+from ....utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {"pipeline_pndm": ["PNDMPipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_pndm import PNDMPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/deprecated/pndm/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/pndm/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1184719cc1a969cd9d6b9cee91d5ff8fe173dda8
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/pndm/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/pndm/__pycache__/pipeline_pndm.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/pndm/__pycache__/pipeline_pndm.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ffa4079f2c7f8598ab9b55f74e08931bf96e6ae8
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/pndm/__pycache__/pipeline_pndm.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/pndm/pipeline_pndm.py b/model/diffusers_c/pipelines/deprecated/pndm/pipeline_pndm.py
new file mode 100644
index 0000000000000000000000000000000000000000..ef78af1940cedd07718a9e05d99da50e38a220d3
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/pndm/pipeline_pndm.py
@@ -0,0 +1,121 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ....models import UNet2DModel
+from ....schedulers import PNDMScheduler
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+class PNDMPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for unconditional image generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image latents.
+ scheduler ([`PNDMScheduler`]):
+ A `PNDMScheduler` to be used in combination with `unet` to denoise the encoded image.
+ """
+
+ unet: UNet2DModel
+ scheduler: PNDMScheduler
+
+ def __init__(self, unet: UNet2DModel, scheduler: PNDMScheduler):
+ super().__init__()
+
+ scheduler = PNDMScheduler.from_config(scheduler.config)
+
+ self.register_modules(unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ batch_size: int = 1,
+ num_inference_steps: int = 50,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ **kwargs,
+ ) -> Union[ImagePipelineOutput, Tuple]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ batch_size (`int`, `optional`, defaults to 1):
+ The number of images to generate.
+ num_inference_steps (`int`, `optional`, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator`, `optional`):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ output_type (`str`, `optional`, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ ```py
+ >>> from diffusers import PNDMPipeline
+
+ >>> # load model and scheduler
+ >>> pndm = PNDMPipeline.from_pretrained("google/ddpm-cifar10-32")
+
+ >>> # run pipeline in inference (sample random noise and denoise)
+ >>> image = pndm().images[0]
+
+ >>> # save image
+ >>> image.save("pndm_generated_image.png")
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+ # For more information on the sampling method you can take a look at Algorithm 2 of
+ # the official paper: https://arxiv.org/pdf/2202.09778.pdf
+
+ # Sample gaussian noise to begin loop
+ image = randn_tensor(
+ (batch_size, self.unet.config.in_channels, self.unet.config.sample_size, self.unet.config.sample_size),
+ generator=generator,
+ device=self.device,
+ )
+
+ self.scheduler.set_timesteps(num_inference_steps)
+ for t in self.progress_bar(self.scheduler.timesteps):
+ model_output = self.unet(image, t).sample
+
+ image = self.scheduler.step(model_output, t, image).prev_sample
+
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/deprecated/repaint/__init__.py b/model/diffusers_c/pipelines/deprecated/repaint/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..2c6b04af52d40e8a2bfa2aa5812b9fb8b1da06f5
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/repaint/__init__.py
@@ -0,0 +1,19 @@
+from typing import TYPE_CHECKING
+
+from ....utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {"pipeline_repaint": ["RePaintPipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_repaint import RePaintPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/deprecated/repaint/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/repaint/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9a06b5089cff767a54a4a589e469e3811fc7a5f1
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/repaint/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/repaint/__pycache__/pipeline_repaint.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/repaint/__pycache__/pipeline_repaint.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5c01c6ee3098c760192d5347ad433cb9a4c18438
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/repaint/__pycache__/pipeline_repaint.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/repaint/pipeline_repaint.py b/model/diffusers_c/pipelines/deprecated/repaint/pipeline_repaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..c03a3d8fc3b206b1db3b7b839082aa66aad028dc
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/repaint/pipeline_repaint.py
@@ -0,0 +1,230 @@
+# Copyright 2024 ETH Zurich Computer Vision Lab and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+
+from ....models import UNet2DModel
+from ....schedulers import RePaintScheduler
+from ....utils import PIL_INTERPOLATION, deprecate, logging
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
+def _preprocess_image(image: Union[List, PIL.Image.Image, torch.Tensor]):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+def _preprocess_mask(mask: Union[List, PIL.Image.Image, torch.Tensor]):
+ if isinstance(mask, torch.Tensor):
+ return mask
+ elif isinstance(mask, PIL.Image.Image):
+ mask = [mask]
+
+ if isinstance(mask[0], PIL.Image.Image):
+ w, h = mask[0].size
+ w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32
+ mask = [np.array(m.convert("L").resize((w, h), resample=PIL_INTERPOLATION["nearest"]))[None, :] for m in mask]
+ mask = np.concatenate(mask, axis=0)
+ mask = mask.astype(np.float32) / 255.0
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+ mask = torch.from_numpy(mask)
+ elif isinstance(mask[0], torch.Tensor):
+ mask = torch.cat(mask, dim=0)
+ return mask
+
+
+class RePaintPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for image inpainting using RePaint.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image latents.
+ scheduler ([`RePaintScheduler`]):
+ A `RePaintScheduler` to be used in combination with `unet` to denoise the encoded image.
+ """
+
+ unet: UNet2DModel
+ scheduler: RePaintScheduler
+ model_cpu_offload_seq = "unet"
+
+ def __init__(self, unet, scheduler):
+ super().__init__()
+ self.register_modules(unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image: Union[torch.Tensor, PIL.Image.Image],
+ mask_image: Union[torch.Tensor, PIL.Image.Image],
+ num_inference_steps: int = 250,
+ eta: float = 0.0,
+ jump_length: int = 10,
+ jump_n_sample: int = 10,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ) -> Union[ImagePipelineOutput, Tuple]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
+ The original image to inpaint on.
+ mask_image (`torch.FloatTensor` or `PIL.Image.Image`):
+ The mask_image where 0.0 define which part of the original image to inpaint.
+ num_inference_steps (`int`, *optional*, defaults to 1000):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ eta (`float`):
+ The weight of the added noise in a diffusion step. Its value is between 0.0 and 1.0; 0.0 corresponds to
+ DDIM and 1.0 is the DDPM scheduler.
+ jump_length (`int`, *optional*, defaults to 10):
+ The number of steps taken forward in time before going backward in time for a single jump ("j" in
+ RePaint paper). Take a look at Figure 9 and 10 in the [paper](https://arxiv.org/pdf/2201.09865.pdf).
+ jump_n_sample (`int`, *optional*, defaults to 10):
+ The number of times to make a forward time jump for a given chosen time sample. Take a look at Figure 9
+ and 10 in the [paper](https://arxiv.org/pdf/2201.09865.pdf).
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ output_type (`str`, `optional`, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ ```py
+ >>> from io import BytesIO
+ >>> import torch
+ >>> import PIL
+ >>> import requests
+ >>> from diffusers import RePaintPipeline, RePaintScheduler
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/repaint/celeba_hq_256.png"
+ >>> mask_url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/repaint/mask_256.png"
+
+ >>> # Load the original image and the mask as PIL images
+ >>> original_image = download_image(img_url).resize((256, 256))
+ >>> mask_image = download_image(mask_url).resize((256, 256))
+
+ >>> # Load the RePaint scheduler and pipeline based on a pretrained DDPM model
+ >>> scheduler = RePaintScheduler.from_pretrained("google/ddpm-ema-celebahq-256")
+ >>> pipe = RePaintPipeline.from_pretrained("google/ddpm-ema-celebahq-256", scheduler=scheduler)
+ >>> pipe = pipe.to("cuda")
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(0)
+ >>> output = pipe(
+ ... image=original_image,
+ ... mask_image=mask_image,
+ ... num_inference_steps=250,
+ ... eta=0.0,
+ ... jump_length=10,
+ ... jump_n_sample=10,
+ ... generator=generator,
+ ... )
+ >>> inpainted_image = output.images[0]
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+
+ original_image = image
+
+ original_image = _preprocess_image(original_image)
+ original_image = original_image.to(device=self._execution_device, dtype=self.unet.dtype)
+ mask_image = _preprocess_mask(mask_image)
+ mask_image = mask_image.to(device=self._execution_device, dtype=self.unet.dtype)
+
+ batch_size = original_image.shape[0]
+
+ # sample gaussian noise to begin the loop
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ image_shape = original_image.shape
+ image = randn_tensor(image_shape, generator=generator, device=self._execution_device, dtype=self.unet.dtype)
+
+ # set step values
+ self.scheduler.set_timesteps(num_inference_steps, jump_length, jump_n_sample, self._execution_device)
+ self.scheduler.eta = eta
+
+ t_last = self.scheduler.timesteps[0] + 1
+ generator = generator[0] if isinstance(generator, list) else generator
+ for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+ if t < t_last:
+ # predict the noise residual
+ model_output = self.unet(image, t).sample
+ # compute previous image: x_t -> x_t-1
+ image = self.scheduler.step(model_output, t, image, original_image, mask_image, generator).prev_sample
+
+ else:
+ # compute the reverse: x_t-1 -> x_t
+ image = self.scheduler.undo_step(image, t_last, generator)
+ t_last = t
+
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/deprecated/score_sde_ve/__init__.py b/model/diffusers_c/pipelines/deprecated/score_sde_ve/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..87c167c3dbd26e0408a41ef197a42dc5eb7038d7
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/score_sde_ve/__init__.py
@@ -0,0 +1,19 @@
+from typing import TYPE_CHECKING
+
+from ....utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {"pipeline_score_sde_ve": ["ScoreSdeVePipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_score_sde_ve import ScoreSdeVePipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/deprecated/score_sde_ve/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/score_sde_ve/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1251ef262be21e74b8b28b4d428ef53c785b5f19
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/score_sde_ve/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/score_sde_ve/__pycache__/pipeline_score_sde_ve.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/score_sde_ve/__pycache__/pipeline_score_sde_ve.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b7ffbba08ad831cf3de68ab6bcd1e04eccb77d3d
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/score_sde_ve/__pycache__/pipeline_score_sde_ve.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/score_sde_ve/pipeline_score_sde_ve.py b/model/diffusers_c/pipelines/deprecated/score_sde_ve/pipeline_score_sde_ve.py
new file mode 100644
index 0000000000000000000000000000000000000000..b0bb114a81b75a979b49c0884e921a33beedff4f
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/score_sde_ve/pipeline_score_sde_ve.py
@@ -0,0 +1,109 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ....models import UNet2DModel
+from ....schedulers import ScoreSdeVeScheduler
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+class ScoreSdeVePipeline(DiffusionPipeline):
+ r"""
+ Pipeline for unconditional image generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image.
+ scheduler ([`ScoreSdeVeScheduler`]):
+ A `ScoreSdeVeScheduler` to be used in combination with `unet` to denoise the encoded image.
+ """
+
+ unet: UNet2DModel
+ scheduler: ScoreSdeVeScheduler
+
+ def __init__(self, unet: UNet2DModel, scheduler: ScoreSdeVeScheduler):
+ super().__init__()
+ self.register_modules(unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ batch_size: int = 1,
+ num_inference_steps: int = 2000,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ **kwargs,
+ ) -> Union[ImagePipelineOutput, Tuple]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ batch_size (`int`, *optional*, defaults to 1):
+ The number of images to generate.
+ generator (`torch.Generator`, `optional`):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ output_type (`str`, `optional`, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+
+ img_size = self.unet.config.sample_size
+ shape = (batch_size, 3, img_size, img_size)
+
+ model = self.unet
+
+ sample = randn_tensor(shape, generator=generator) * self.scheduler.init_noise_sigma
+ sample = sample.to(self.device)
+
+ self.scheduler.set_timesteps(num_inference_steps)
+ self.scheduler.set_sigmas(num_inference_steps)
+
+ for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+ sigma_t = self.scheduler.sigmas[i] * torch.ones(shape[0], device=self.device)
+
+ # correction step
+ for _ in range(self.scheduler.config.correct_steps):
+ model_output = self.unet(sample, sigma_t).sample
+ sample = self.scheduler.step_correct(model_output, sample, generator=generator).prev_sample
+
+ # prediction step
+ model_output = model(sample, sigma_t).sample
+ output = self.scheduler.step_pred(model_output, t, sample, generator=generator)
+
+ sample, sample_mean = output.prev_sample, output.prev_sample_mean
+
+ sample = sample_mean.clamp(0, 1)
+ sample = sample.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "pil":
+ sample = self.numpy_to_pil(sample)
+
+ if not return_dict:
+ return (sample,)
+
+ return ImagePipelineOutput(images=sample)
diff --git a/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/__init__.py b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..150954baa0eb8f8a7216b4891effc14a71e21b1b
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/__init__.py
@@ -0,0 +1,75 @@
+# flake8: noqa
+from typing import TYPE_CHECKING
+from ....utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ _LazyModule,
+ is_note_seq_available,
+ OptionalDependencyNotAvailable,
+ is_torch_available,
+ is_transformers_available,
+ get_objects_from_module,
+)
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ....utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["continous_encoder"] = ["SpectrogramContEncoder"]
+ _import_structure["notes_encoder"] = ["SpectrogramNotesEncoder"]
+ _import_structure["pipeline_spectrogram_diffusion"] = [
+ "SpectrogramContEncoder",
+ "SpectrogramDiffusionPipeline",
+ "T5FilmDecoder",
+ ]
+try:
+ if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ....utils import dummy_transformers_and_torch_and_note_seq_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_transformers_and_torch_and_note_seq_objects))
+else:
+ _import_structure["midi_utils"] = ["MidiProcessor"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ....utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_spectrogram_diffusion import SpectrogramDiffusionPipeline
+ from .pipeline_spectrogram_diffusion import SpectrogramContEncoder
+ from .pipeline_spectrogram_diffusion import SpectrogramNotesEncoder
+ from .pipeline_spectrogram_diffusion import T5FilmDecoder
+
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_note_seq_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ....utils.dummy_transformers_and_torch_and_note_seq_objects import *
+
+ else:
+ from .midi_utils import MidiProcessor
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/continuous_encoder.py b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/continuous_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..8664c2fb67113972976a24aa980d3aa7778b807e
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/continuous_encoder.py
@@ -0,0 +1,92 @@
+# Copyright 2022 The Music Spectrogram Diffusion Authors.
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from transformers.modeling_utils import ModuleUtilsMixin
+from transformers.models.t5.modeling_t5 import (
+ T5Block,
+ T5Config,
+ T5LayerNorm,
+)
+
+from ....configuration_utils import ConfigMixin, register_to_config
+from ....models import ModelMixin
+
+
+class SpectrogramContEncoder(ModelMixin, ConfigMixin, ModuleUtilsMixin):
+ @register_to_config
+ def __init__(
+ self,
+ input_dims: int,
+ targets_context_length: int,
+ d_model: int,
+ dropout_rate: float,
+ num_layers: int,
+ num_heads: int,
+ d_kv: int,
+ d_ff: int,
+ feed_forward_proj: str,
+ is_decoder: bool = False,
+ ):
+ super().__init__()
+
+ self.input_proj = nn.Linear(input_dims, d_model, bias=False)
+
+ self.position_encoding = nn.Embedding(targets_context_length, d_model)
+ self.position_encoding.weight.requires_grad = False
+
+ self.dropout_pre = nn.Dropout(p=dropout_rate)
+
+ t5config = T5Config(
+ d_model=d_model,
+ num_heads=num_heads,
+ d_kv=d_kv,
+ d_ff=d_ff,
+ feed_forward_proj=feed_forward_proj,
+ dropout_rate=dropout_rate,
+ is_decoder=is_decoder,
+ is_encoder_decoder=False,
+ )
+ self.encoders = nn.ModuleList()
+ for lyr_num in range(num_layers):
+ lyr = T5Block(t5config)
+ self.encoders.append(lyr)
+
+ self.layer_norm = T5LayerNorm(d_model)
+ self.dropout_post = nn.Dropout(p=dropout_rate)
+
+ def forward(self, encoder_inputs, encoder_inputs_mask):
+ x = self.input_proj(encoder_inputs)
+
+ # terminal relative positional encodings
+ max_positions = encoder_inputs.shape[1]
+ input_positions = torch.arange(max_positions, device=encoder_inputs.device)
+
+ seq_lens = encoder_inputs_mask.sum(-1)
+ input_positions = torch.roll(input_positions.unsqueeze(0), tuple(seq_lens.tolist()), dims=0)
+ x += self.position_encoding(input_positions)
+
+ x = self.dropout_pre(x)
+
+ # inverted the attention mask
+ input_shape = encoder_inputs.size()
+ extended_attention_mask = self.get_extended_attention_mask(encoder_inputs_mask, input_shape)
+
+ for lyr in self.encoders:
+ x = lyr(x, extended_attention_mask)[0]
+ x = self.layer_norm(x)
+
+ return self.dropout_post(x), encoder_inputs_mask
diff --git a/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/midi_utils.py b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/midi_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..e777e844935e76427cd92ead59b7d23eb8b1cbe5
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/midi_utils.py
@@ -0,0 +1,667 @@
+# Copyright 2022 The Music Spectrogram Diffusion Authors.
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import dataclasses
+import math
+import os
+from typing import Any, Callable, List, Mapping, MutableMapping, Optional, Sequence, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from ....utils import is_note_seq_available
+from .pipeline_spectrogram_diffusion import TARGET_FEATURE_LENGTH
+
+
+if is_note_seq_available():
+ import note_seq
+else:
+ raise ImportError("Please install note-seq via `pip install note-seq`")
+
+
+INPUT_FEATURE_LENGTH = 2048
+
+SAMPLE_RATE = 16000
+HOP_SIZE = 320
+FRAME_RATE = int(SAMPLE_RATE // HOP_SIZE)
+
+DEFAULT_STEPS_PER_SECOND = 100
+DEFAULT_MAX_SHIFT_SECONDS = 10
+DEFAULT_NUM_VELOCITY_BINS = 1
+
+SLAKH_CLASS_PROGRAMS = {
+ "Acoustic Piano": 0,
+ "Electric Piano": 4,
+ "Chromatic Percussion": 8,
+ "Organ": 16,
+ "Acoustic Guitar": 24,
+ "Clean Electric Guitar": 26,
+ "Distorted Electric Guitar": 29,
+ "Acoustic Bass": 32,
+ "Electric Bass": 33,
+ "Violin": 40,
+ "Viola": 41,
+ "Cello": 42,
+ "Contrabass": 43,
+ "Orchestral Harp": 46,
+ "Timpani": 47,
+ "String Ensemble": 48,
+ "Synth Strings": 50,
+ "Choir and Voice": 52,
+ "Orchestral Hit": 55,
+ "Trumpet": 56,
+ "Trombone": 57,
+ "Tuba": 58,
+ "French Horn": 60,
+ "Brass Section": 61,
+ "Soprano/Alto Sax": 64,
+ "Tenor Sax": 66,
+ "Baritone Sax": 67,
+ "Oboe": 68,
+ "English Horn": 69,
+ "Bassoon": 70,
+ "Clarinet": 71,
+ "Pipe": 73,
+ "Synth Lead": 80,
+ "Synth Pad": 88,
+}
+
+
+@dataclasses.dataclass
+class NoteRepresentationConfig:
+ """Configuration note representations."""
+
+ onsets_only: bool
+ include_ties: bool
+
+
+@dataclasses.dataclass
+class NoteEventData:
+ pitch: int
+ velocity: Optional[int] = None
+ program: Optional[int] = None
+ is_drum: Optional[bool] = None
+ instrument: Optional[int] = None
+
+
+@dataclasses.dataclass
+class NoteEncodingState:
+ """Encoding state for note transcription, keeping track of active pitches."""
+
+ # velocity bin for active pitches and programs
+ active_pitches: MutableMapping[Tuple[int, int], int] = dataclasses.field(default_factory=dict)
+
+
+@dataclasses.dataclass
+class EventRange:
+ type: str
+ min_value: int
+ max_value: int
+
+
+@dataclasses.dataclass
+class Event:
+ type: str
+ value: int
+
+
+class Tokenizer:
+ def __init__(self, regular_ids: int):
+ # The special tokens: 0=PAD, 1=EOS, and 2=UNK
+ self._num_special_tokens = 3
+ self._num_regular_tokens = regular_ids
+
+ def encode(self, token_ids):
+ encoded = []
+ for token_id in token_ids:
+ if not 0 <= token_id < self._num_regular_tokens:
+ raise ValueError(
+ f"token_id {token_id} does not fall within valid range of [0, {self._num_regular_tokens})"
+ )
+ encoded.append(token_id + self._num_special_tokens)
+
+ # Add EOS token
+ encoded.append(1)
+
+ # Pad to till INPUT_FEATURE_LENGTH
+ encoded = encoded + [0] * (INPUT_FEATURE_LENGTH - len(encoded))
+
+ return encoded
+
+
+class Codec:
+ """Encode and decode events.
+
+ Useful for declaring what certain ranges of a vocabulary should be used for. This is intended to be used from
+ Python before encoding or after decoding with GenericTokenVocabulary. This class is more lightweight and does not
+ include things like EOS or UNK token handling.
+
+ To ensure that 'shift' events are always the first block of the vocab and start at 0, that event type is required
+ and specified separately.
+ """
+
+ def __init__(self, max_shift_steps: int, steps_per_second: float, event_ranges: List[EventRange]):
+ """Define Codec.
+
+ Args:
+ max_shift_steps: Maximum number of shift steps that can be encoded.
+ steps_per_second: Shift steps will be interpreted as having a duration of
+ 1 / steps_per_second.
+ event_ranges: Other supported event types and their ranges.
+ """
+ self.steps_per_second = steps_per_second
+ self._shift_range = EventRange(type="shift", min_value=0, max_value=max_shift_steps)
+ self._event_ranges = [self._shift_range] + event_ranges
+ # Ensure all event types have unique names.
+ assert len(self._event_ranges) == len({er.type for er in self._event_ranges})
+
+ @property
+ def num_classes(self) -> int:
+ return sum(er.max_value - er.min_value + 1 for er in self._event_ranges)
+
+ # The next couple methods are simplified special case methods just for shift
+ # events that are intended to be used from within autograph functions.
+
+ def is_shift_event_index(self, index: int) -> bool:
+ return (self._shift_range.min_value <= index) and (index <= self._shift_range.max_value)
+
+ @property
+ def max_shift_steps(self) -> int:
+ return self._shift_range.max_value
+
+ def encode_event(self, event: Event) -> int:
+ """Encode an event to an index."""
+ offset = 0
+ for er in self._event_ranges:
+ if event.type == er.type:
+ if not er.min_value <= event.value <= er.max_value:
+ raise ValueError(
+ f"Event value {event.value} is not within valid range "
+ f"[{er.min_value}, {er.max_value}] for type {event.type}"
+ )
+ return offset + event.value - er.min_value
+ offset += er.max_value - er.min_value + 1
+
+ raise ValueError(f"Unknown event type: {event.type}")
+
+ def event_type_range(self, event_type: str) -> Tuple[int, int]:
+ """Return [min_id, max_id] for an event type."""
+ offset = 0
+ for er in self._event_ranges:
+ if event_type == er.type:
+ return offset, offset + (er.max_value - er.min_value)
+ offset += er.max_value - er.min_value + 1
+
+ raise ValueError(f"Unknown event type: {event_type}")
+
+ def decode_event_index(self, index: int) -> Event:
+ """Decode an event index to an Event."""
+ offset = 0
+ for er in self._event_ranges:
+ if offset <= index <= offset + er.max_value - er.min_value:
+ return Event(type=er.type, value=er.min_value + index - offset)
+ offset += er.max_value - er.min_value + 1
+
+ raise ValueError(f"Unknown event index: {index}")
+
+
+@dataclasses.dataclass
+class ProgramGranularity:
+ # both tokens_map_fn and program_map_fn should be idempotent
+ tokens_map_fn: Callable[[Sequence[int], Codec], Sequence[int]]
+ program_map_fn: Callable[[int], int]
+
+
+def drop_programs(tokens, codec: Codec):
+ """Drops program change events from a token sequence."""
+ min_program_id, max_program_id = codec.event_type_range("program")
+ return tokens[(tokens < min_program_id) | (tokens > max_program_id)]
+
+
+def programs_to_midi_classes(tokens, codec):
+ """Modifies program events to be the first program in the MIDI class."""
+ min_program_id, max_program_id = codec.event_type_range("program")
+ is_program = (tokens >= min_program_id) & (tokens <= max_program_id)
+ return np.where(is_program, min_program_id + 8 * ((tokens - min_program_id) // 8), tokens)
+
+
+PROGRAM_GRANULARITIES = {
+ # "flat" granularity; drop program change tokens and set NoteSequence
+ # programs to zero
+ "flat": ProgramGranularity(tokens_map_fn=drop_programs, program_map_fn=lambda program: 0),
+ # map each program to the first program in its MIDI class
+ "midi_class": ProgramGranularity(
+ tokens_map_fn=programs_to_midi_classes, program_map_fn=lambda program: 8 * (program // 8)
+ ),
+ # leave programs as is
+ "full": ProgramGranularity(tokens_map_fn=lambda tokens, codec: tokens, program_map_fn=lambda program: program),
+}
+
+
+def frame(signal, frame_length, frame_step, pad_end=False, pad_value=0, axis=-1):
+ """
+ equivalent of tf.signal.frame
+ """
+ signal_length = signal.shape[axis]
+ if pad_end:
+ frames_overlap = frame_length - frame_step
+ rest_samples = np.abs(signal_length - frames_overlap) % np.abs(frame_length - frames_overlap)
+ pad_size = int(frame_length - rest_samples)
+
+ if pad_size != 0:
+ pad_axis = [0] * signal.ndim
+ pad_axis[axis] = pad_size
+ signal = F.pad(signal, pad_axis, "constant", pad_value)
+ frames = signal.unfold(axis, frame_length, frame_step)
+ return frames
+
+
+def program_to_slakh_program(program):
+ # this is done very hackily, probably should use a custom mapping
+ for slakh_program in sorted(SLAKH_CLASS_PROGRAMS.values(), reverse=True):
+ if program >= slakh_program:
+ return slakh_program
+
+
+def audio_to_frames(
+ samples,
+ hop_size: int,
+ frame_rate: int,
+) -> Tuple[Sequence[Sequence[int]], torch.Tensor]:
+ """Convert audio samples to non-overlapping frames and frame times."""
+ frame_size = hop_size
+ samples = np.pad(samples, [0, frame_size - len(samples) % frame_size], mode="constant")
+
+ # Split audio into frames.
+ frames = frame(
+ torch.Tensor(samples).unsqueeze(0),
+ frame_length=frame_size,
+ frame_step=frame_size,
+ pad_end=False, # TODO check why its off by 1 here when True
+ )
+
+ num_frames = len(samples) // frame_size
+
+ times = np.arange(num_frames) / frame_rate
+ return frames, times
+
+
+def note_sequence_to_onsets_and_offsets_and_programs(
+ ns: note_seq.NoteSequence,
+) -> Tuple[Sequence[float], Sequence[NoteEventData]]:
+ """Extract onset & offset times and pitches & programs from a NoteSequence.
+
+ The onset & offset times will not necessarily be in sorted order.
+
+ Args:
+ ns: NoteSequence from which to extract onsets and offsets.
+
+ Returns:
+ times: A list of note onset and offset times. values: A list of NoteEventData objects where velocity is zero for
+ note
+ offsets.
+ """
+ # Sort by program and pitch and put offsets before onsets as a tiebreaker for
+ # subsequent stable sort.
+ notes = sorted(ns.notes, key=lambda note: (note.is_drum, note.program, note.pitch))
+ times = [note.end_time for note in notes if not note.is_drum] + [note.start_time for note in notes]
+ values = [
+ NoteEventData(pitch=note.pitch, velocity=0, program=note.program, is_drum=False)
+ for note in notes
+ if not note.is_drum
+ ] + [
+ NoteEventData(pitch=note.pitch, velocity=note.velocity, program=note.program, is_drum=note.is_drum)
+ for note in notes
+ ]
+ return times, values
+
+
+def num_velocity_bins_from_codec(codec: Codec):
+ """Get number of velocity bins from event codec."""
+ lo, hi = codec.event_type_range("velocity")
+ return hi - lo
+
+
+# segment an array into segments of length n
+def segment(a, n):
+ return [a[i : i + n] for i in range(0, len(a), n)]
+
+
+def velocity_to_bin(velocity, num_velocity_bins):
+ if velocity == 0:
+ return 0
+ else:
+ return math.ceil(num_velocity_bins * velocity / note_seq.MAX_MIDI_VELOCITY)
+
+
+def note_event_data_to_events(
+ state: Optional[NoteEncodingState],
+ value: NoteEventData,
+ codec: Codec,
+) -> Sequence[Event]:
+ """Convert note event data to a sequence of events."""
+ if value.velocity is None:
+ # onsets only, no program or velocity
+ return [Event("pitch", value.pitch)]
+ else:
+ num_velocity_bins = num_velocity_bins_from_codec(codec)
+ velocity_bin = velocity_to_bin(value.velocity, num_velocity_bins)
+ if value.program is None:
+ # onsets + offsets + velocities only, no programs
+ if state is not None:
+ state.active_pitches[(value.pitch, 0)] = velocity_bin
+ return [Event("velocity", velocity_bin), Event("pitch", value.pitch)]
+ else:
+ if value.is_drum:
+ # drum events use a separate vocabulary
+ return [Event("velocity", velocity_bin), Event("drum", value.pitch)]
+ else:
+ # program + velocity + pitch
+ if state is not None:
+ state.active_pitches[(value.pitch, value.program)] = velocity_bin
+ return [
+ Event("program", value.program),
+ Event("velocity", velocity_bin),
+ Event("pitch", value.pitch),
+ ]
+
+
+def note_encoding_state_to_events(state: NoteEncodingState) -> Sequence[Event]:
+ """Output program and pitch events for active notes plus a final tie event."""
+ events = []
+ for pitch, program in sorted(state.active_pitches.keys(), key=lambda k: k[::-1]):
+ if state.active_pitches[(pitch, program)]:
+ events += [Event("program", program), Event("pitch", pitch)]
+ events.append(Event("tie", 0))
+ return events
+
+
+def encode_and_index_events(
+ state, event_times, event_values, codec, frame_times, encode_event_fn, encoding_state_to_events_fn=None
+):
+ """Encode a sequence of timed events and index to audio frame times.
+
+ Encodes time shifts as repeated single step shifts for later run length encoding.
+
+ Optionally, also encodes a sequence of "state events", keeping track of the current encoding state at each audio
+ frame. This can be used e.g. to prepend events representing the current state to a targets segment.
+
+ Args:
+ state: Initial event encoding state.
+ event_times: Sequence of event times.
+ event_values: Sequence of event values.
+ encode_event_fn: Function that transforms event value into a sequence of one
+ or more Event objects.
+ codec: An Codec object that maps Event objects to indices.
+ frame_times: Time for every audio frame.
+ encoding_state_to_events_fn: Function that transforms encoding state into a
+ sequence of one or more Event objects.
+
+ Returns:
+ events: Encoded events and shifts. event_start_indices: Corresponding start event index for every audio frame.
+ Note: one event can correspond to multiple audio indices due to sampling rate differences. This makes
+ splitting sequences tricky because the same event can appear at the end of one sequence and the beginning of
+ another.
+ event_end_indices: Corresponding end event index for every audio frame. Used
+ to ensure when slicing that one chunk ends where the next begins. Should always be true that
+ event_end_indices[i] = event_start_indices[i + 1].
+ state_events: Encoded "state" events representing the encoding state before
+ each event.
+ state_event_indices: Corresponding state event index for every audio frame.
+ """
+ indices = np.argsort(event_times, kind="stable")
+ event_steps = [round(event_times[i] * codec.steps_per_second) for i in indices]
+ event_values = [event_values[i] for i in indices]
+
+ events = []
+ state_events = []
+ event_start_indices = []
+ state_event_indices = []
+
+ cur_step = 0
+ cur_event_idx = 0
+ cur_state_event_idx = 0
+
+ def fill_event_start_indices_to_cur_step():
+ while (
+ len(event_start_indices) < len(frame_times)
+ and frame_times[len(event_start_indices)] < cur_step / codec.steps_per_second
+ ):
+ event_start_indices.append(cur_event_idx)
+ state_event_indices.append(cur_state_event_idx)
+
+ for event_step, event_value in zip(event_steps, event_values):
+ while event_step > cur_step:
+ events.append(codec.encode_event(Event(type="shift", value=1)))
+ cur_step += 1
+ fill_event_start_indices_to_cur_step()
+ cur_event_idx = len(events)
+ cur_state_event_idx = len(state_events)
+ if encoding_state_to_events_fn:
+ # Dump state to state events *before* processing the next event, because
+ # we want to capture the state prior to the occurrence of the event.
+ for e in encoding_state_to_events_fn(state):
+ state_events.append(codec.encode_event(e))
+
+ for e in encode_event_fn(state, event_value, codec):
+ events.append(codec.encode_event(e))
+
+ # After the last event, continue filling out the event_start_indices array.
+ # The inequality is not strict because if our current step lines up exactly
+ # with (the start of) an audio frame, we need to add an additional shift event
+ # to "cover" that frame.
+ while cur_step / codec.steps_per_second <= frame_times[-1]:
+ events.append(codec.encode_event(Event(type="shift", value=1)))
+ cur_step += 1
+ fill_event_start_indices_to_cur_step()
+ cur_event_idx = len(events)
+
+ # Now fill in event_end_indices. We need this extra array to make sure that
+ # when we slice events, each slice ends exactly where the subsequent slice
+ # begins.
+ event_end_indices = event_start_indices[1:] + [len(events)]
+
+ events = np.array(events).astype(np.int32)
+ state_events = np.array(state_events).astype(np.int32)
+ event_start_indices = segment(np.array(event_start_indices).astype(np.int32), TARGET_FEATURE_LENGTH)
+ event_end_indices = segment(np.array(event_end_indices).astype(np.int32), TARGET_FEATURE_LENGTH)
+ state_event_indices = segment(np.array(state_event_indices).astype(np.int32), TARGET_FEATURE_LENGTH)
+
+ outputs = []
+ for start_indices, end_indices, event_indices in zip(event_start_indices, event_end_indices, state_event_indices):
+ outputs.append(
+ {
+ "inputs": events,
+ "event_start_indices": start_indices,
+ "event_end_indices": end_indices,
+ "state_events": state_events,
+ "state_event_indices": event_indices,
+ }
+ )
+
+ return outputs
+
+
+def extract_sequence_with_indices(features, state_events_end_token=None, feature_key="inputs"):
+ """Extract target sequence corresponding to audio token segment."""
+ features = features.copy()
+ start_idx = features["event_start_indices"][0]
+ end_idx = features["event_end_indices"][-1]
+
+ features[feature_key] = features[feature_key][start_idx:end_idx]
+
+ if state_events_end_token is not None:
+ # Extract the state events corresponding to the audio start token, and
+ # prepend them to the targets array.
+ state_event_start_idx = features["state_event_indices"][0]
+ state_event_end_idx = state_event_start_idx + 1
+ while features["state_events"][state_event_end_idx - 1] != state_events_end_token:
+ state_event_end_idx += 1
+ features[feature_key] = np.concatenate(
+ [
+ features["state_events"][state_event_start_idx:state_event_end_idx],
+ features[feature_key],
+ ],
+ axis=0,
+ )
+
+ return features
+
+
+def map_midi_programs(
+ feature, codec: Codec, granularity_type: str = "full", feature_key: str = "inputs"
+) -> Mapping[str, Any]:
+ """Apply MIDI program map to token sequences."""
+ granularity = PROGRAM_GRANULARITIES[granularity_type]
+
+ feature[feature_key] = granularity.tokens_map_fn(feature[feature_key], codec)
+ return feature
+
+
+def run_length_encode_shifts_fn(
+ features,
+ codec: Codec,
+ feature_key: str = "inputs",
+ state_change_event_types: Sequence[str] = (),
+) -> Callable[[Mapping[str, Any]], Mapping[str, Any]]:
+ """Return a function that run-length encodes shifts for a given codec.
+
+ Args:
+ codec: The Codec to use for shift events.
+ feature_key: The feature key for which to run-length encode shifts.
+ state_change_event_types: A list of event types that represent state
+ changes; tokens corresponding to these event types will be interpreted as state changes and redundant ones
+ will be removed.
+
+ Returns:
+ A preprocessing function that run-length encodes single-step shifts.
+ """
+ state_change_event_ranges = [codec.event_type_range(event_type) for event_type in state_change_event_types]
+
+ def run_length_encode_shifts(features: MutableMapping[str, Any]) -> Mapping[str, Any]:
+ """Combine leading/interior shifts, trim trailing shifts.
+
+ Args:
+ features: Dict of features to process.
+
+ Returns:
+ A dict of features.
+ """
+ events = features[feature_key]
+
+ shift_steps = 0
+ total_shift_steps = 0
+ output = np.array([], dtype=np.int32)
+
+ current_state = np.zeros(len(state_change_event_ranges), dtype=np.int32)
+
+ for event in events:
+ if codec.is_shift_event_index(event):
+ shift_steps += 1
+ total_shift_steps += 1
+
+ else:
+ # If this event is a state change and has the same value as the current
+ # state, we can skip it entirely.
+ is_redundant = False
+ for i, (min_index, max_index) in enumerate(state_change_event_ranges):
+ if (min_index <= event) and (event <= max_index):
+ if current_state[i] == event:
+ is_redundant = True
+ current_state[i] = event
+ if is_redundant:
+ continue
+
+ # Once we've reached a non-shift event, RLE all previous shift events
+ # before outputting the non-shift event.
+ if shift_steps > 0:
+ shift_steps = total_shift_steps
+ while shift_steps > 0:
+ output_steps = np.minimum(codec.max_shift_steps, shift_steps)
+ output = np.concatenate([output, [output_steps]], axis=0)
+ shift_steps -= output_steps
+ output = np.concatenate([output, [event]], axis=0)
+
+ features[feature_key] = output
+ return features
+
+ return run_length_encode_shifts(features)
+
+
+def note_representation_processor_chain(features, codec: Codec, note_representation_config: NoteRepresentationConfig):
+ tie_token = codec.encode_event(Event("tie", 0))
+ state_events_end_token = tie_token if note_representation_config.include_ties else None
+
+ features = extract_sequence_with_indices(
+ features, state_events_end_token=state_events_end_token, feature_key="inputs"
+ )
+
+ features = map_midi_programs(features, codec)
+
+ features = run_length_encode_shifts_fn(features, codec, state_change_event_types=["velocity", "program"])
+
+ return features
+
+
+class MidiProcessor:
+ def __init__(self):
+ self.codec = Codec(
+ max_shift_steps=DEFAULT_MAX_SHIFT_SECONDS * DEFAULT_STEPS_PER_SECOND,
+ steps_per_second=DEFAULT_STEPS_PER_SECOND,
+ event_ranges=[
+ EventRange("pitch", note_seq.MIN_MIDI_PITCH, note_seq.MAX_MIDI_PITCH),
+ EventRange("velocity", 0, DEFAULT_NUM_VELOCITY_BINS),
+ EventRange("tie", 0, 0),
+ EventRange("program", note_seq.MIN_MIDI_PROGRAM, note_seq.MAX_MIDI_PROGRAM),
+ EventRange("drum", note_seq.MIN_MIDI_PITCH, note_seq.MAX_MIDI_PITCH),
+ ],
+ )
+ self.tokenizer = Tokenizer(self.codec.num_classes)
+ self.note_representation_config = NoteRepresentationConfig(onsets_only=False, include_ties=True)
+
+ def __call__(self, midi: Union[bytes, os.PathLike, str]):
+ if not isinstance(midi, bytes):
+ with open(midi, "rb") as f:
+ midi = f.read()
+
+ ns = note_seq.midi_to_note_sequence(midi)
+ ns_sus = note_seq.apply_sustain_control_changes(ns)
+
+ for note in ns_sus.notes:
+ if not note.is_drum:
+ note.program = program_to_slakh_program(note.program)
+
+ samples = np.zeros(int(ns_sus.total_time * SAMPLE_RATE))
+
+ _, frame_times = audio_to_frames(samples, HOP_SIZE, FRAME_RATE)
+ times, values = note_sequence_to_onsets_and_offsets_and_programs(ns_sus)
+
+ events = encode_and_index_events(
+ state=NoteEncodingState(),
+ event_times=times,
+ event_values=values,
+ frame_times=frame_times,
+ codec=self.codec,
+ encode_event_fn=note_event_data_to_events,
+ encoding_state_to_events_fn=note_encoding_state_to_events,
+ )
+
+ events = [
+ note_representation_processor_chain(event, self.codec, self.note_representation_config) for event in events
+ ]
+ input_tokens = [self.tokenizer.encode(event["inputs"]) for event in events]
+
+ return input_tokens
diff --git a/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/notes_encoder.py b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/notes_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..1259f0bf056aa40ad111b4f282210687430ad68f
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/notes_encoder.py
@@ -0,0 +1,86 @@
+# Copyright 2022 The Music Spectrogram Diffusion Authors.
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from transformers.modeling_utils import ModuleUtilsMixin
+from transformers.models.t5.modeling_t5 import T5Block, T5Config, T5LayerNorm
+
+from ....configuration_utils import ConfigMixin, register_to_config
+from ....models import ModelMixin
+
+
+class SpectrogramNotesEncoder(ModelMixin, ConfigMixin, ModuleUtilsMixin):
+ @register_to_config
+ def __init__(
+ self,
+ max_length: int,
+ vocab_size: int,
+ d_model: int,
+ dropout_rate: float,
+ num_layers: int,
+ num_heads: int,
+ d_kv: int,
+ d_ff: int,
+ feed_forward_proj: str,
+ is_decoder: bool = False,
+ ):
+ super().__init__()
+
+ self.token_embedder = nn.Embedding(vocab_size, d_model)
+
+ self.position_encoding = nn.Embedding(max_length, d_model)
+ self.position_encoding.weight.requires_grad = False
+
+ self.dropout_pre = nn.Dropout(p=dropout_rate)
+
+ t5config = T5Config(
+ vocab_size=vocab_size,
+ d_model=d_model,
+ num_heads=num_heads,
+ d_kv=d_kv,
+ d_ff=d_ff,
+ dropout_rate=dropout_rate,
+ feed_forward_proj=feed_forward_proj,
+ is_decoder=is_decoder,
+ is_encoder_decoder=False,
+ )
+
+ self.encoders = nn.ModuleList()
+ for lyr_num in range(num_layers):
+ lyr = T5Block(t5config)
+ self.encoders.append(lyr)
+
+ self.layer_norm = T5LayerNorm(d_model)
+ self.dropout_post = nn.Dropout(p=dropout_rate)
+
+ def forward(self, encoder_input_tokens, encoder_inputs_mask):
+ x = self.token_embedder(encoder_input_tokens)
+
+ seq_length = encoder_input_tokens.shape[1]
+ inputs_positions = torch.arange(seq_length, device=encoder_input_tokens.device)
+ x += self.position_encoding(inputs_positions)
+
+ x = self.dropout_pre(x)
+
+ # inverted the attention mask
+ input_shape = encoder_input_tokens.size()
+ extended_attention_mask = self.get_extended_attention_mask(encoder_inputs_mask, input_shape)
+
+ for lyr in self.encoders:
+ x = lyr(x, extended_attention_mask)[0]
+ x = self.layer_norm(x)
+
+ return self.dropout_post(x), encoder_inputs_mask
diff --git a/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/pipeline_spectrogram_diffusion.py b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/pipeline_spectrogram_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..496a1f7658f16b2d9a5b7e2fe99b8c51168d9ffd
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/spectrogram_diffusion/pipeline_spectrogram_diffusion.py
@@ -0,0 +1,269 @@
+# Copyright 2022 The Music Spectrogram Diffusion Authors.
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Any, Callable, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ....models import T5FilmDecoder
+from ....schedulers import DDPMScheduler
+from ....utils import is_onnx_available, logging
+from ....utils.torch_utils import randn_tensor
+
+
+if is_onnx_available():
+ from ...onnx_utils import OnnxRuntimeModel
+
+from ...pipeline_utils import AudioPipelineOutput, DiffusionPipeline
+from .continuous_encoder import SpectrogramContEncoder
+from .notes_encoder import SpectrogramNotesEncoder
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+TARGET_FEATURE_LENGTH = 256
+
+
+class SpectrogramDiffusionPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for unconditional audio generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ notes_encoder ([`SpectrogramNotesEncoder`]):
+ continuous_encoder ([`SpectrogramContEncoder`]):
+ decoder ([`T5FilmDecoder`]):
+ A [`T5FilmDecoder`] to denoise the encoded audio latents.
+ scheduler ([`DDPMScheduler`]):
+ A scheduler to be used in combination with `decoder` to denoise the encoded audio latents.
+ melgan ([`OnnxRuntimeModel`]):
+ """
+
+ _optional_components = ["melgan"]
+
+ def __init__(
+ self,
+ notes_encoder: SpectrogramNotesEncoder,
+ continuous_encoder: SpectrogramContEncoder,
+ decoder: T5FilmDecoder,
+ scheduler: DDPMScheduler,
+ melgan: OnnxRuntimeModel if is_onnx_available() else Any,
+ ) -> None:
+ super().__init__()
+
+ # From MELGAN
+ self.min_value = math.log(1e-5) # Matches MelGAN training.
+ self.max_value = 4.0 # Largest value for most examples
+ self.n_dims = 128
+
+ self.register_modules(
+ notes_encoder=notes_encoder,
+ continuous_encoder=continuous_encoder,
+ decoder=decoder,
+ scheduler=scheduler,
+ melgan=melgan,
+ )
+
+ def scale_features(self, features, output_range=(-1.0, 1.0), clip=False):
+ """Linearly scale features to network outputs range."""
+ min_out, max_out = output_range
+ if clip:
+ features = torch.clip(features, self.min_value, self.max_value)
+ # Scale to [0, 1].
+ zero_one = (features - self.min_value) / (self.max_value - self.min_value)
+ # Scale to [min_out, max_out].
+ return zero_one * (max_out - min_out) + min_out
+
+ def scale_to_features(self, outputs, input_range=(-1.0, 1.0), clip=False):
+ """Invert by linearly scaling network outputs to features range."""
+ min_out, max_out = input_range
+ outputs = torch.clip(outputs, min_out, max_out) if clip else outputs
+ # Scale to [0, 1].
+ zero_one = (outputs - min_out) / (max_out - min_out)
+ # Scale to [self.min_value, self.max_value].
+ return zero_one * (self.max_value - self.min_value) + self.min_value
+
+ def encode(self, input_tokens, continuous_inputs, continuous_mask):
+ tokens_mask = input_tokens > 0
+ tokens_encoded, tokens_mask = self.notes_encoder(
+ encoder_input_tokens=input_tokens, encoder_inputs_mask=tokens_mask
+ )
+
+ continuous_encoded, continuous_mask = self.continuous_encoder(
+ encoder_inputs=continuous_inputs, encoder_inputs_mask=continuous_mask
+ )
+
+ return [(tokens_encoded, tokens_mask), (continuous_encoded, continuous_mask)]
+
+ def decode(self, encodings_and_masks, input_tokens, noise_time):
+ timesteps = noise_time
+ if not torch.is_tensor(timesteps):
+ timesteps = torch.tensor([timesteps], dtype=torch.long, device=input_tokens.device)
+ elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(input_tokens.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps * torch.ones(input_tokens.shape[0], dtype=timesteps.dtype, device=timesteps.device)
+
+ logits = self.decoder(
+ encodings_and_masks=encodings_and_masks, decoder_input_tokens=input_tokens, decoder_noise_time=timesteps
+ )
+ return logits
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ input_tokens: List[List[int]],
+ generator: Optional[torch.Generator] = None,
+ num_inference_steps: int = 100,
+ return_dict: bool = True,
+ output_type: str = "numpy",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ) -> Union[AudioPipelineOutput, Tuple]:
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ input_tokens (`List[List[int]]`):
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality audio at the
+ expense of slower inference.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.AudioPipelineOutput`] instead of a plain tuple.
+ output_type (`str`, *optional*, defaults to `"numpy"`):
+ The output format of the generated audio.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Example:
+
+ ```py
+ >>> from diffusers import SpectrogramDiffusionPipeline, MidiProcessor
+
+ >>> pipe = SpectrogramDiffusionPipeline.from_pretrained("google/music-spectrogram-diffusion")
+ >>> pipe = pipe.to("cuda")
+ >>> processor = MidiProcessor()
+
+ >>> # Download MIDI from: wget http://www.piano-midi.de/midis/beethoven/beethoven_hammerklavier_2.mid
+ >>> output = pipe(processor("beethoven_hammerklavier_2.mid"))
+
+ >>> audio = output.audios[0]
+ ```
+
+ Returns:
+ [`pipelines.AudioPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`pipelines.AudioPipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated audio.
+ """
+
+ pred_mel = np.zeros([1, TARGET_FEATURE_LENGTH, self.n_dims], dtype=np.float32)
+ full_pred_mel = np.zeros([1, 0, self.n_dims], np.float32)
+ ones = torch.ones((1, TARGET_FEATURE_LENGTH), dtype=bool, device=self.device)
+
+ for i, encoder_input_tokens in enumerate(input_tokens):
+ if i == 0:
+ encoder_continuous_inputs = torch.from_numpy(pred_mel[:1].copy()).to(
+ device=self.device, dtype=self.decoder.dtype
+ )
+ # The first chunk has no previous context.
+ encoder_continuous_mask = torch.zeros((1, TARGET_FEATURE_LENGTH), dtype=bool, device=self.device)
+ else:
+ # The full song pipeline does not feed in a context feature, so the mask
+ # will be all 0s after the feature converter. Because we know we're
+ # feeding in a full context chunk from the previous prediction, set it
+ # to all 1s.
+ encoder_continuous_mask = ones
+
+ encoder_continuous_inputs = self.scale_features(
+ encoder_continuous_inputs, output_range=[-1.0, 1.0], clip=True
+ )
+
+ encodings_and_masks = self.encode(
+ input_tokens=torch.IntTensor([encoder_input_tokens]).to(device=self.device),
+ continuous_inputs=encoder_continuous_inputs,
+ continuous_mask=encoder_continuous_mask,
+ )
+
+ # Sample encoder_continuous_inputs shaped gaussian noise to begin loop
+ x = randn_tensor(
+ shape=encoder_continuous_inputs.shape,
+ generator=generator,
+ device=self.device,
+ dtype=self.decoder.dtype,
+ )
+
+ # set step values
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ # Denoising diffusion loop
+ for j, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+ output = self.decode(
+ encodings_and_masks=encodings_and_masks,
+ input_tokens=x,
+ noise_time=t / self.scheduler.config.num_train_timesteps, # rescale to [0, 1)
+ )
+
+ # Compute previous output: x_t -> x_t-1
+ x = self.scheduler.step(output, t, x, generator=generator).prev_sample
+
+ mel = self.scale_to_features(x, input_range=[-1.0, 1.0])
+ encoder_continuous_inputs = mel[:1]
+ pred_mel = mel.cpu().float().numpy()
+
+ full_pred_mel = np.concatenate([full_pred_mel, pred_mel[:1]], axis=1)
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ callback(i, full_pred_mel)
+
+ logger.info("Generated segment", i)
+
+ if output_type == "numpy" and not is_onnx_available():
+ raise ValueError(
+ "Cannot return output in 'np' format if ONNX is not available. Make sure to have ONNX installed or set 'output_type' to 'mel'."
+ )
+ elif output_type == "numpy" and self.melgan is None:
+ raise ValueError(
+ "Cannot return output in 'np' format if melgan component is not defined. Make sure to define `self.melgan` or set 'output_type' to 'mel'."
+ )
+
+ if output_type == "numpy":
+ output = self.melgan(input_features=full_pred_mel.astype(np.float32))
+ else:
+ output = full_pred_mel
+
+ if not return_dict:
+ return (output,)
+
+ return AudioPipelineOutput(audios=output)
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__init__.py b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..36cf1a33ce6ada8e718aabadb9a706737aee30bd
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__init__.py
@@ -0,0 +1,55 @@
+from typing import TYPE_CHECKING
+
+from ....utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ....utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_cycle_diffusion"] = ["CycleDiffusionPipeline"]
+ _import_structure["pipeline_stable_diffusion_inpaint_legacy"] = ["StableDiffusionInpaintPipelineLegacy"]
+ _import_structure["pipeline_stable_diffusion_model_editing"] = ["StableDiffusionModelEditingPipeline"]
+
+ _import_structure["pipeline_stable_diffusion_paradigms"] = ["StableDiffusionParadigmsPipeline"]
+ _import_structure["pipeline_stable_diffusion_pix2pix_zero"] = ["StableDiffusionPix2PixZeroPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ....utils.dummy_torch_and_transformers_objects import *
+
+ else:
+ from .pipeline_cycle_diffusion import CycleDiffusionPipeline
+ from .pipeline_stable_diffusion_inpaint_legacy import StableDiffusionInpaintPipelineLegacy
+ from .pipeline_stable_diffusion_model_editing import StableDiffusionModelEditingPipeline
+ from .pipeline_stable_diffusion_paradigms import StableDiffusionParadigmsPipeline
+ from .pipeline_stable_diffusion_pix2pix_zero import StableDiffusionPix2PixZeroPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ce4e46740d3aa11dca46ba986c2bf715d72864ad
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_cycle_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_cycle_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..21038d87bf3f0981bfdca76b343d23345aab0e45
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_cycle_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_inpaint_legacy.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_inpaint_legacy.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d340bf76dc94e4dbf7ed39ebec751b2305060a73
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_inpaint_legacy.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_model_editing.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_model_editing.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5c078646db623f018c31307303795e64808fefea
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_model_editing.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_paradigms.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_paradigms.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..969deade4c9b3b08268c2d2016847bbba30220f9
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_paradigms.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_pix2pix_zero.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_pix2pix_zero.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..addc583d543dd29a2e3b137fceb44c86f6527431
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/__pycache__/pipeline_stable_diffusion_pix2pix_zero.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_cycle_diffusion.py b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_cycle_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..0581effef2fe33e4d76e1e5305f11c15e2d3f22d
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_cycle_diffusion.py
@@ -0,0 +1,948 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ....configuration_utils import FrozenDict
+from ....image_processor import PipelineImageInput, VaeImageProcessor
+from ....loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ....models import AutoencoderKL, UNet2DConditionModel
+from ....models.lora import adjust_lora_scale_text_encoder
+from ....schedulers import DDIMScheduler
+from ....utils import PIL_INTERPOLATION, USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline
+from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
+def preprocess(image):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+def posterior_sample(scheduler, latents, timestep, clean_latents, generator, eta):
+ # 1. get previous step value (=t-1)
+ prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps
+
+ if prev_timestep <= 0:
+ return clean_latents
+
+ # 2. compute alphas, betas
+ alpha_prod_t = scheduler.alphas_cumprod[timestep]
+ alpha_prod_t_prev = (
+ scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
+ )
+
+ variance = scheduler._get_variance(timestep, prev_timestep)
+ std_dev_t = eta * variance ** (0.5)
+
+ # direction pointing to x_t
+ e_t = (latents - alpha_prod_t ** (0.5) * clean_latents) / (1 - alpha_prod_t) ** (0.5)
+ dir_xt = (1.0 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * e_t
+ noise = std_dev_t * randn_tensor(
+ clean_latents.shape, dtype=clean_latents.dtype, device=clean_latents.device, generator=generator
+ )
+ prev_latents = alpha_prod_t_prev ** (0.5) * clean_latents + dir_xt + noise
+
+ return prev_latents
+
+
+def compute_noise(scheduler, prev_latents, latents, timestep, noise_pred, eta):
+ # 1. get previous step value (=t-1)
+ prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps
+
+ # 2. compute alphas, betas
+ alpha_prod_t = scheduler.alphas_cumprod[timestep]
+ alpha_prod_t_prev = (
+ scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
+ )
+
+ beta_prod_t = 1 - alpha_prod_t
+
+ # 3. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)
+
+ # 4. Clip "predicted x_0"
+ if scheduler.config.clip_sample:
+ pred_original_sample = torch.clamp(pred_original_sample, -1, 1)
+
+ # 5. compute variance: "sigma_t(η)" -> see formula (16)
+ # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
+ variance = scheduler._get_variance(timestep, prev_timestep)
+ std_dev_t = eta * variance ** (0.5)
+
+ # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * noise_pred
+
+ noise = (prev_latents - (alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction)) / (
+ variance ** (0.5) * eta
+ )
+ return noise
+
+
+class CycleDiffusionPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+ r"""
+ Pipeline for text-guided image to image generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can only be an
+ instance of [`DDIMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: DDIMScheduler,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def check_inputs(
+ self,
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = image.shape[0]
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(image.shape[0])
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt * num_images_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0)
+
+ # add noise to latents using the timestep
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ clean_latents = init_latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+
+ return latents, clean_latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ source_prompt: Union[str, List[str]],
+ image: PipelineImageInput = None,
+ strength: float = 0.8,
+ num_inference_steps: Optional[int] = 50,
+ guidance_scale: Optional[float] = 7.5,
+ source_guidance_scale: Optional[float] = 1,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: Optional[float] = 0.1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`torch.FloatTensor` `np.ndarray`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image` or tensor representing an image batch to be used as the starting point. Can also accept image
+ latents as `image`, but if passing latents directly it is not encoded again.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ source_guidance_scale (`float`, *optional*, defaults to 1):
+ Guidance scale for the source prompt. This is useful to control the amount of influence the source
+ prompt has for encoding.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Example:
+
+ ```py
+ import requests
+ import torch
+ from PIL import Image
+ from io import BytesIO
+
+ from diffusers import CycleDiffusionPipeline, DDIMScheduler
+
+ # load the pipeline
+ # make sure you're logged in with `huggingface-cli login`
+ model_id_or_path = "CompVis/stable-diffusion-v1-4"
+ scheduler = DDIMScheduler.from_pretrained(model_id_or_path, subfolder="scheduler")
+ pipe = CycleDiffusionPipeline.from_pretrained(model_id_or_path, scheduler=scheduler).to("cuda")
+
+ # let's download an initial image
+ url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/An%20astronaut%20riding%20a%20horse.png"
+ response = requests.get(url)
+ init_image = Image.open(BytesIO(response.content)).convert("RGB")
+ init_image = init_image.resize((512, 512))
+ init_image.save("horse.png")
+
+ # let's specify a prompt
+ source_prompt = "An astronaut riding a horse"
+ prompt = "An astronaut riding an elephant"
+
+ # call the pipeline
+ image = pipe(
+ prompt=prompt,
+ source_prompt=source_prompt,
+ image=init_image,
+ num_inference_steps=100,
+ eta=0.1,
+ strength=0.8,
+ guidance_scale=2,
+ source_guidance_scale=1,
+ ).images[0]
+
+ image.save("horse_to_elephant.png")
+
+ # let's try another example
+ # See more samples at the original repo: https://github.com/ChenWu98/cycle-diffusion
+ url = (
+ "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/A%20black%20colored%20car.png"
+ )
+ response = requests.get(url)
+ init_image = Image.open(BytesIO(response.content)).convert("RGB")
+ init_image = init_image.resize((512, 512))
+ init_image.save("black.png")
+
+ source_prompt = "A black colored car"
+ prompt = "A blue colored car"
+
+ # call the pipeline
+ torch.manual_seed(0)
+ image = pipe(
+ prompt=prompt,
+ source_prompt=source_prompt,
+ image=init_image,
+ num_inference_steps=100,
+ eta=0.1,
+ strength=0.85,
+ guidance_scale=3,
+ source_guidance_scale=1,
+ ).images[0]
+
+ image.save("black_to_blue.png")
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 1. Check inputs
+ self.check_inputs(prompt, strength, callback_steps)
+
+ # 2. Define call parameters
+ batch_size = 1 if isinstance(prompt, str) else len(prompt)
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ prompt_embeds=prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_skip,
+ )
+ source_prompt_embeds_tuple = self.encode_prompt(
+ source_prompt, device, num_images_per_prompt, do_classifier_free_guidance, None, clip_skip=clip_skip
+ )
+ if prompt_embeds_tuple[1] is not None:
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+ else:
+ prompt_embeds = prompt_embeds_tuple[0]
+ if source_prompt_embeds_tuple[1] is not None:
+ source_prompt_embeds = torch.cat([source_prompt_embeds_tuple[1], source_prompt_embeds_tuple[0]])
+ else:
+ source_prompt_embeds = source_prompt_embeds_tuple[0]
+
+ # 4. Preprocess image
+ image = self.image_processor.preprocess(image)
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+ # 6. Prepare latent variables
+ latents, clean_latents = self.prepare_latents(
+ image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
+ )
+ source_latents = latents
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+ generator = extra_step_kwargs.pop("generator", None)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ source_latent_model_input = (
+ torch.cat([source_latents] * 2) if do_classifier_free_guidance else source_latents
+ )
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+ source_latent_model_input = self.scheduler.scale_model_input(source_latent_model_input, t)
+
+ # predict the noise residual
+ if do_classifier_free_guidance:
+ concat_latent_model_input = torch.stack(
+ [
+ source_latent_model_input[0],
+ latent_model_input[0],
+ source_latent_model_input[1],
+ latent_model_input[1],
+ ],
+ dim=0,
+ )
+ concat_prompt_embeds = torch.stack(
+ [
+ source_prompt_embeds[0],
+ prompt_embeds[0],
+ source_prompt_embeds[1],
+ prompt_embeds[1],
+ ],
+ dim=0,
+ )
+ else:
+ concat_latent_model_input = torch.cat(
+ [
+ source_latent_model_input,
+ latent_model_input,
+ ],
+ dim=0,
+ )
+ concat_prompt_embeds = torch.cat(
+ [
+ source_prompt_embeds,
+ prompt_embeds,
+ ],
+ dim=0,
+ )
+
+ concat_noise_pred = self.unet(
+ concat_latent_model_input,
+ t,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_hidden_states=concat_prompt_embeds,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ (
+ source_noise_pred_uncond,
+ noise_pred_uncond,
+ source_noise_pred_text,
+ noise_pred_text,
+ ) = concat_noise_pred.chunk(4, dim=0)
+
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ source_noise_pred = source_noise_pred_uncond + source_guidance_scale * (
+ source_noise_pred_text - source_noise_pred_uncond
+ )
+
+ else:
+ (source_noise_pred, noise_pred) = concat_noise_pred.chunk(2, dim=0)
+
+ # Sample source_latents from the posterior distribution.
+ prev_source_latents = posterior_sample(
+ self.scheduler, source_latents, t, clean_latents, generator=generator, **extra_step_kwargs
+ )
+ # Compute noise.
+ noise = compute_noise(
+ self.scheduler, prev_source_latents, source_latents, t, source_noise_pred, **extra_step_kwargs
+ )
+ source_latents = prev_source_latents
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred, t, latents, variance_noise=noise, **extra_step_kwargs
+ ).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 9. Post-processing
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_onnx_stable_diffusion_inpaint_legacy.py b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_onnx_stable_diffusion_inpaint_legacy.py
new file mode 100644
index 0000000000000000000000000000000000000000..0aa5e68bfcb4b276cafea63fce6e2a5f3e2f79c2
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_onnx_stable_diffusion_inpaint_legacy.py
@@ -0,0 +1,542 @@
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTokenizer
+
+from ....configuration_utils import FrozenDict
+from ....schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ....utils import deprecate, logging
+from ...onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel
+from ...pipeline_utils import DiffusionPipeline
+from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def preprocess(image):
+ w, h = image.size
+ w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32
+ image = image.resize((w, h), resample=PIL.Image.LANCZOS)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image[None].transpose(0, 3, 1, 2)
+ return 2.0 * image - 1.0
+
+
+def preprocess_mask(mask, scale_factor=8):
+ mask = mask.convert("L")
+ w, h = mask.size
+ w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32
+ mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL.Image.NEAREST)
+ mask = np.array(mask).astype(np.float32) / 255.0
+ mask = np.tile(mask, (4, 1, 1))
+ mask = mask[None].transpose(0, 1, 2, 3) # what does this step do?
+ mask = 1 - mask # repaint white, keep black
+ return mask
+
+
+class OnnxStableDiffusionInpaintPipelineLegacy(DiffusionPipeline):
+ r"""
+ Pipeline for text-guided image inpainting using Stable Diffusion. This is a *legacy feature* for Onnx pipelines to
+ provide compatibility with StableDiffusionInpaintPipelineLegacy and may be removed in the future.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ """
+
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _is_onnx = True
+
+ vae_encoder: OnnxRuntimeModel
+ vae_decoder: OnnxRuntimeModel
+ text_encoder: OnnxRuntimeModel
+ tokenizer: CLIPTokenizer
+ unet: OnnxRuntimeModel
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+ safety_checker: OnnxRuntimeModel
+ feature_extractor: CLIPImageProcessor
+
+ def __init__(
+ self,
+ vae_encoder: OnnxRuntimeModel,
+ vae_decoder: OnnxRuntimeModel,
+ text_encoder: OnnxRuntimeModel,
+ tokenizer: CLIPTokenizer,
+ unet: OnnxRuntimeModel,
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+ safety_checker: OnnxRuntimeModel,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae_encoder=vae_encoder,
+ vae_decoder=vae_decoder,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ num_images_per_prompt: Optional[int],
+ do_classifier_free_guidance: bool,
+ negative_prompt: Optional[str],
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids
+
+ if not np.array_equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0]
+
+ prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt] * batch_size
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0]
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ def check_inputs(
+ self,
+ prompt,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[np.ndarray, PIL.Image.Image] = None,
+ mask_image: Union[np.ndarray, PIL.Image.Image] = None,
+ strength: float = 0.8,
+ num_inference_steps: Optional[int] = 50,
+ guidance_scale: Optional[float] = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: Optional[float] = 0.0,
+ generator: Optional[np.random.RandomState] = None,
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`nd.ndarray` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process. This is the image whose masked region will be inpainted.
+ mask_image (`nd.ndarray` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+ replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
+ PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should
+ contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`.uu
+ strength (`float`, *optional*, defaults to 0.8):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter will be modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (?) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`np.random.RandomState`, *optional*):
+ A np.random.RandomState to make generation deterministic.
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+
+ # check inputs. Raise error if not correct
+ self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+ # define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if generator is None:
+ generator = np.random
+
+ # set timesteps
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ if isinstance(image, PIL.Image.Image):
+ image = preprocess(image)
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ prompt_embeds = self._encode_prompt(
+ prompt,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ latents_dtype = prompt_embeds.dtype
+ image = image.astype(latents_dtype)
+
+ # encode the init image into latents and scale the latents
+ init_latents = self.vae_encoder(sample=image)[0]
+ init_latents = 0.18215 * init_latents
+
+ # Expand init_latents for batch_size and num_images_per_prompt
+ init_latents = np.concatenate([init_latents] * num_images_per_prompt, axis=0)
+ init_latents_orig = init_latents
+
+ # preprocess mask
+ if not isinstance(mask_image, np.ndarray):
+ mask_image = preprocess_mask(mask_image, 8)
+ mask_image = mask_image.astype(latents_dtype)
+ mask = np.concatenate([mask_image] * num_images_per_prompt, axis=0)
+
+ # check sizes
+ if not mask.shape == init_latents.shape:
+ raise ValueError("The mask and image should be the same size!")
+
+ # get the original timestep using init_timestep
+ offset = self.scheduler.config.get("steps_offset", 0)
+ init_timestep = int(num_inference_steps * strength) + offset
+ init_timestep = min(init_timestep, num_inference_steps)
+
+ timesteps = self.scheduler.timesteps.numpy()[-init_timestep]
+ timesteps = np.array([timesteps] * batch_size * num_images_per_prompt)
+
+ # add noise to latents using the timesteps
+ noise = generator.randn(*init_latents.shape).astype(latents_dtype)
+ init_latents = self.scheduler.add_noise(
+ torch.from_numpy(init_latents), torch.from_numpy(noise), torch.from_numpy(timesteps)
+ )
+ init_latents = init_latents.numpy()
+
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (?) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to ? in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ latents = init_latents
+
+ t_start = max(num_inference_steps - init_timestep + offset, 0)
+ timesteps = self.scheduler.timesteps[t_start:].numpy()
+ timestep_dtype = next(
+ (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)"
+ )
+ timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]
+
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ timestep = np.array([t], dtype=timestep_dtype)
+ noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)[
+ 0
+ ]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs
+ ).prev_sample
+
+ latents = latents.numpy()
+
+ init_latents_proper = self.scheduler.add_noise(
+ torch.from_numpy(init_latents_orig), torch.from_numpy(noise), torch.from_numpy(np.array([t]))
+ )
+
+ init_latents_proper = init_latents_proper.numpy()
+
+ latents = (init_latents_proper * mask) + (latents * (1 - mask))
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ latents = 1 / 0.18215 * latents
+ # image = self.vae_decoder(latent_sample=latents)[0]
+ # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
+ image = np.concatenate(
+ [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])]
+ )
+
+ image = np.clip(image / 2 + 0.5, 0, 1)
+ image = image.transpose((0, 2, 3, 1))
+
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(
+ self.numpy_to_pil(image), return_tensors="np"
+ ).pixel_values.astype(image.dtype)
+ # There will throw an error if use safety_checker batchsize>1
+ images, has_nsfw_concept = [], []
+ for i in range(image.shape[0]):
+ image_i, has_nsfw_concept_i = self.safety_checker(
+ clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
+ )
+ images.append(image_i)
+ has_nsfw_concept.append(has_nsfw_concept_i[0])
+ image = np.concatenate(images)
+ else:
+ has_nsfw_concept = None
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_inpaint_legacy.py b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_inpaint_legacy.py
new file mode 100644
index 0000000000000000000000000000000000000000..980adf2737630c6fec4fd35b799dd1d64022025f
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_inpaint_legacy.py
@@ -0,0 +1,786 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ....configuration_utils import FrozenDict
+from ....image_processor import VaeImageProcessor
+from ....loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ....models import AutoencoderKL, UNet2DConditionModel
+from ....models.lora import adjust_lora_scale_text_encoder
+from ....schedulers import KarrasDiffusionSchedulers
+from ....utils import PIL_INTERPOLATION, USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline
+from ...stable_diffusion import StableDiffusionPipelineOutput
+from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__)
+
+
+def preprocess_image(image, batch_size):
+ w, h = image.size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+ image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
+ image = np.array(image).astype(np.float32) / 255.0
+ image = np.vstack([image[None].transpose(0, 3, 1, 2)] * batch_size)
+ image = torch.from_numpy(image)
+ return 2.0 * image - 1.0
+
+
+def preprocess_mask(mask, batch_size, scale_factor=8):
+ if not isinstance(mask, torch.FloatTensor):
+ mask = mask.convert("L")
+ w, h = mask.size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+ mask = mask.resize((w // scale_factor, h // scale_factor), resample=PIL_INTERPOLATION["nearest"])
+ mask = np.array(mask).astype(np.float32) / 255.0
+ mask = np.tile(mask, (4, 1, 1))
+ mask = np.vstack([mask[None]] * batch_size)
+ mask = 1 - mask # repaint white, keep black
+ mask = torch.from_numpy(mask)
+ return mask
+
+ else:
+ valid_mask_channel_sizes = [1, 3]
+ # if mask channel is fourth tensor dimension, permute dimensions to pytorch standard (B, C, H, W)
+ if mask.shape[3] in valid_mask_channel_sizes:
+ mask = mask.permute(0, 3, 1, 2)
+ elif mask.shape[1] not in valid_mask_channel_sizes:
+ raise ValueError(
+ f"Mask channel dimension of size in {valid_mask_channel_sizes} should be second or fourth dimension,"
+ f" but received mask of shape {tuple(mask.shape)}"
+ )
+ # (potentially) reduce mask channel dimension from 3 to 1 for broadcasting to latent shape
+ mask = mask.mean(dim=1, keepdim=True)
+ h, w = mask.shape[-2:]
+ h, w = (x - x % 8 for x in (h, w)) # resize to integer multiple of 8
+ mask = torch.nn.functional.interpolate(mask, (h // scale_factor, w // scale_factor))
+ return mask
+
+
+class StableDiffusionInpaintPipelineLegacy(
+ DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
+ r"""
+ Pipeline for text-guided image inpainting using Stable Diffusion. *This is an experimental feature*.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ In addition the pipeline inherits the following loading methods:
+ - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
+ - *LoRA*: [`loaders.LoraLoaderMixin.load_lora_weights`]
+ - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]
+
+ as well as the following saving methods:
+ - *LoRA*: [`loaders.LoraLoaderMixin.save_lora_weights`]
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["feature_extractor"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ deprecation_message = (
+ f"The class {self.__class__} is deprecated and will be removed in v1.0.0. You can achieve exactly the same functionality"
+ "by loading your model into `StableDiffusionInpaintPipeline` instead. See https://github.com/huggingface/diffusers/pull/3533"
+ "for more information."
+ )
+ deprecate("legacy is outdated", "1.0.0", deprecation_message, standard_warn=False)
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(self, image, timestep, num_images_per_prompt, dtype, device, generator):
+ image = image.to(device=device, dtype=dtype)
+ init_latent_dist = self.vae.encode(image).latent_dist
+ init_latents = init_latent_dist.sample(generator=generator)
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ # Expand init_latents for batch_size and num_images_per_prompt
+ init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0)
+ init_latents_orig = init_latents
+
+ # add noise to latents using the timesteps
+ noise = randn_tensor(init_latents.shape, generator=generator, device=device, dtype=dtype)
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+ return latents, init_latents_orig, noise
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ strength: float = 0.8,
+ num_inference_steps: Optional[int] = 50,
+ guidance_scale: Optional[float] = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ add_predicted_noise: Optional[bool] = False,
+ eta: Optional[float] = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process. This is the image whose masked region will be inpainted.
+ mask_image (`torch.FloatTensor` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+ replaced by noise and therefore repainted, while black pixels will be preserved. If `mask_image` is a
+ PIL image, it will be converted to a single channel (luminance) before use. If mask is a tensor, the
+ expected shape should be either `(B, H, W, C)` or `(B, C, H, W)`, where C is 1 or 3.
+ strength (`float`, *optional*, defaults to 0.8):
+ Conceptually, indicates how much to inpaint the masked area. Must be between 0 and 1. When `strength`
+ is 1, the denoising process will be run on the masked area for the full number of iterations specified
+ in `num_inference_steps`. `image` will be used as a reference for the masked area, adding more noise to
+ that region the larger the `strength`. If `strength` is 0, no inpainting will occur.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The reference number of denoising steps. More denoising steps usually lead to a higher quality image at
+ the expense of slower inference. This parameter will be modulated by `strength`, as explained above.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. Ignored when not using guidance (i.e., ignored if `guidance_scale`
+ is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ add_predicted_noise (`bool`, *optional*, defaults to True):
+ Use predicted noise instead of random noise when constructing noisy versions of the original image in
+ the reverse diffusion process
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+ # 1. Check inputs
+ self.check_inputs(prompt, strength, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Preprocess image and mask
+ if not isinstance(image, torch.FloatTensor):
+ image = preprocess_image(image, batch_size)
+
+ mask_image = preprocess_mask(mask_image, batch_size, self.vae_scale_factor)
+
+ # 5. set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+ # 6. Prepare latent variables
+ # encode the init image into latents and scale the latents
+ latents, init_latents_orig, noise = self.prepare_latents(
+ image, latent_timestep, num_images_per_prompt, prompt_embeds.dtype, device, generator
+ )
+
+ # 7. Prepare mask latent
+ mask = mask_image.to(device=device, dtype=latents.dtype)
+ mask = torch.cat([mask] * num_images_per_prompt)
+
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 9. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+ # masking
+ if add_predicted_noise:
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_orig, noise_pred_uncond, torch.tensor([t])
+ )
+ else:
+ init_latents_proper = self.scheduler.add_noise(init_latents_orig, noise, torch.tensor([t]))
+
+ latents = (init_latents_proper * mask) + (latents * (1 - mask))
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # use original latents corresponding to unmasked portions of the image
+ latents = (init_latents_orig * mask) + (latents * (1 - mask))
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py
new file mode 100644
index 0000000000000000000000000000000000000000..dee93fc2eb5343a33612e57045035179457f1919
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_model_editing.py
@@ -0,0 +1,824 @@
+# Copyright 2024 TIME Authors and The HuggingFace Team. All rights reserved."
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+from ....image_processor import VaeImageProcessor
+from ....loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ....models import AutoencoderKL, UNet2DConditionModel
+from ....models.lora import adjust_lora_scale_text_encoder
+from ....schedulers import PNDMScheduler
+from ....schedulers.scheduling_utils import SchedulerMixin
+from ....utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+AUGS_CONST = ["A photo of ", "An image of ", "A picture of "]
+
+
+class StableDiffusionModelEditingPipeline(
+ DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin
+):
+ r"""
+ Pipeline for text-to-image model editing.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPFeatureExtractor`]):
+ A `CLIPFeatureExtractor` to extract features from generated images; used as inputs to the `safety_checker`.
+ with_to_k ([`bool`]):
+ Whether to edit the key projection matrices along with the value projection matrices.
+ with_augs ([`list`]):
+ Textual augmentations to apply while editing the text-to-image model. Set to `[]` for no augmentations.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: SchedulerMixin,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPFeatureExtractor,
+ requires_safety_checker: bool = True,
+ with_to_k: bool = True,
+ with_augs: list = AUGS_CONST,
+ ):
+ super().__init__()
+
+ if isinstance(scheduler, PNDMScheduler):
+ logger.error("PNDMScheduler for this pipeline is currently not supported.")
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ self.with_to_k = with_to_k
+ self.with_augs = with_augs
+
+ # get cross-attention layers
+ ca_layers = []
+
+ def append_ca(net_):
+ if net_.__class__.__name__ == "CrossAttention":
+ ca_layers.append(net_)
+ elif hasattr(net_, "children"):
+ for net__ in net_.children():
+ append_ca(net__)
+
+ # recursively find all cross-attention layers in unet
+ for net in self.unet.named_children():
+ if "down" in net[0]:
+ append_ca(net[1])
+ elif "up" in net[0]:
+ append_ca(net[1])
+ elif "mid" in net[0]:
+ append_ca(net[1])
+
+ # get projection matrices
+ self.ca_clip_layers = [l for l in ca_layers if l.to_v.in_features == 768]
+ self.projection_matrices = [l.to_v for l in self.ca_clip_layers]
+ self.og_matrices = [copy.deepcopy(l.to_v) for l in self.ca_clip_layers]
+ if self.with_to_k:
+ self.projection_matrices = self.projection_matrices + [l.to_k for l in self.ca_clip_layers]
+ self.og_matrices = self.og_matrices + [copy.deepcopy(l.to_k) for l in self.ca_clip_layers]
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def edit_model(
+ self,
+ source_prompt: str,
+ destination_prompt: str,
+ lamb: float = 0.1,
+ restart_params: bool = True,
+ ):
+ r"""
+ Apply model editing via closed-form solution (see Eq. 5 in the TIME [paper](https://arxiv.org/abs/2303.08084)).
+
+ Args:
+ source_prompt (`str`):
+ The source prompt containing the concept to be edited.
+ destination_prompt (`str`):
+ The destination prompt. Must contain all words from `source_prompt` with additional ones to specify the
+ target edit.
+ lamb (`float`, *optional*, defaults to 0.1):
+ The lambda parameter specifying the regularization intesity. Smaller values increase the editing power.
+ restart_params (`bool`, *optional*, defaults to True):
+ Restart the model parameters to their pre-trained version before editing. This is done to avoid edit
+ compounding. When it is `False`, edits accumulate.
+ """
+
+ # restart LDM parameters
+ if restart_params:
+ num_ca_clip_layers = len(self.ca_clip_layers)
+ for idx_, l in enumerate(self.ca_clip_layers):
+ l.to_v = copy.deepcopy(self.og_matrices[idx_])
+ self.projection_matrices[idx_] = l.to_v
+ if self.with_to_k:
+ l.to_k = copy.deepcopy(self.og_matrices[num_ca_clip_layers + idx_])
+ self.projection_matrices[num_ca_clip_layers + idx_] = l.to_k
+
+ # set up sentences
+ old_texts = [source_prompt]
+ new_texts = [destination_prompt]
+ # add augmentations
+ base = old_texts[0] if old_texts[0][0:1] != "A" else "a" + old_texts[0][1:]
+ for aug in self.with_augs:
+ old_texts.append(aug + base)
+ base = new_texts[0] if new_texts[0][0:1] != "A" else "a" + new_texts[0][1:]
+ for aug in self.with_augs:
+ new_texts.append(aug + base)
+
+ # prepare input k* and v*
+ old_embs, new_embs = [], []
+ for old_text, new_text in zip(old_texts, new_texts):
+ text_input = self.tokenizer(
+ [old_text, new_text],
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_embeddings = self.text_encoder(text_input.input_ids.to(self.device))[0]
+ old_emb, new_emb = text_embeddings
+ old_embs.append(old_emb)
+ new_embs.append(new_emb)
+
+ # identify corresponding destinations for each token in old_emb
+ idxs_replaces = []
+ for old_text, new_text in zip(old_texts, new_texts):
+ tokens_a = self.tokenizer(old_text).input_ids
+ tokens_b = self.tokenizer(new_text).input_ids
+ tokens_a = [self.tokenizer.encode("a ")[1] if self.tokenizer.decode(t) == "an" else t for t in tokens_a]
+ tokens_b = [self.tokenizer.encode("a ")[1] if self.tokenizer.decode(t) == "an" else t for t in tokens_b]
+ num_orig_tokens = len(tokens_a)
+ idxs_replace = []
+ j = 0
+ for i in range(num_orig_tokens):
+ curr_token = tokens_a[i]
+ while tokens_b[j] != curr_token:
+ j += 1
+ idxs_replace.append(j)
+ j += 1
+ while j < 77:
+ idxs_replace.append(j)
+ j += 1
+ while len(idxs_replace) < 77:
+ idxs_replace.append(76)
+ idxs_replaces.append(idxs_replace)
+
+ # prepare batch: for each pair of setences, old context and new values
+ contexts, valuess = [], []
+ for old_emb, new_emb, idxs_replace in zip(old_embs, new_embs, idxs_replaces):
+ context = old_emb.detach()
+ values = []
+ with torch.no_grad():
+ for layer in self.projection_matrices:
+ values.append(layer(new_emb[idxs_replace]).detach())
+ contexts.append(context)
+ valuess.append(values)
+
+ # edit the model
+ for layer_num in range(len(self.projection_matrices)):
+ # mat1 = \lambda W + \sum{v k^T}
+ mat1 = lamb * self.projection_matrices[layer_num].weight
+
+ # mat2 = \lambda I + \sum{k k^T}
+ mat2 = lamb * torch.eye(
+ self.projection_matrices[layer_num].weight.shape[1],
+ device=self.projection_matrices[layer_num].weight.device,
+ )
+
+ # aggregate sums for mat1, mat2
+ for context, values in zip(contexts, valuess):
+ context_vector = context.reshape(context.shape[0], context.shape[1], 1)
+ context_vector_T = context.reshape(context.shape[0], 1, context.shape[1])
+ value_vector = values[layer_num].reshape(values[layer_num].shape[0], values[layer_num].shape[1], 1)
+ for_mat1 = (value_vector @ context_vector_T).sum(dim=0)
+ for_mat2 = (context_vector @ context_vector_T).sum(dim=0)
+ mat1 += for_mat1
+ mat2 += for_mat2
+
+ # update projection matrix
+ self.projection_matrices[layer_num].weight = torch.nn.Parameter(mat1 @ torch.inverse(mat2))
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+
+ Examples:
+
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionModelEditingPipeline
+
+ >>> model_ckpt = "CompVis/stable-diffusion-v1-4"
+ >>> pipe = StableDiffusionModelEditingPipeline.from_pretrained(model_ckpt)
+
+ >>> pipe = pipe.to("cuda")
+
+ >>> source_prompt = "A pack of roses"
+ >>> destination_prompt = "A pack of blue roses"
+ >>> pipe.edit_model(source_prompt, destination_prompt)
+
+ >>> prompt = "A field of roses"
+ >>> image = pipe(prompt).images[0]
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_paradigms.py b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_paradigms.py
new file mode 100644
index 0000000000000000000000000000000000000000..ddc866ef9b863b2cacbf33110c9cd824fb75799b
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_paradigms.py
@@ -0,0 +1,786 @@
+# Copyright 2024 ParaDiGMS authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ....image_processor import VaeImageProcessor
+from ....loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ....models import AutoencoderKL, UNet2DConditionModel
+from ....models.lora import adjust_lora_scale_text_encoder
+from ....schedulers import KarrasDiffusionSchedulers
+from ....utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import DDPMParallelScheduler
+ >>> from diffusers import StableDiffusionParadigmsPipeline
+
+ >>> scheduler = DDPMParallelScheduler.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="scheduler")
+
+ >>> pipe = StableDiffusionParadigmsPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", scheduler=scheduler, torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> ngpu, batch_per_device = torch.cuda.device_count(), 5
+ >>> pipe.wrapped_unet = torch.nn.DataParallel(pipe.unet, device_ids=[d for d in range(ngpu)])
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> image = pipe(prompt, parallel=ngpu * batch_per_device, num_inference_steps=1000).images[0]
+ ```
+"""
+
+
+class StableDiffusionParadigmsPipeline(
+ DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
+ r"""
+ Pipeline for text-to-image generation using a parallelized version of Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # attribute to wrap the unet with torch.nn.DataParallel when running multiple denoising steps on multiple GPUs
+ self.wrapped_unet = self.unet
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def _cumsum(self, input, dim, debug=False):
+ if debug:
+ # cumsum_cuda_kernel does not have a deterministic implementation
+ # so perform cumsum on cpu for debugging purposes
+ return torch.cumsum(input.cpu().float(), dim=dim).to(input.device)
+ else:
+ return torch.cumsum(input, dim=dim)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ parallel: int = 10,
+ tolerance: float = 0.1,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ debug: bool = False,
+ clip_skip: int = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ parallel (`int`, *optional*, defaults to 10):
+ The batch size to use when doing parallel sampling. More parallelism may lead to faster inference but
+ requires higher memory usage and can also require more total FLOPs.
+ tolerance (`float`, *optional*, defaults to 0.1):
+ The error tolerance for determining when to slide the batch window forward for parallel sampling. Lower
+ tolerance usually leads to less or no degradation. Higher tolerance is faster but can risk degradation
+ of sample quality. The tolerance is specified as a ratio of the scheduler's noise magnitude.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ debug (`bool`, *optional*, defaults to `False`):
+ Whether or not to run in debug mode. In debug mode, `torch.cumsum` is evaluated using the CPU.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+ extra_step_kwargs.pop("generator", None)
+
+ # # 7. Denoising loop
+ scheduler = self.scheduler
+ parallel = min(parallel, len(scheduler.timesteps))
+
+ begin_idx = 0
+ end_idx = parallel
+ latents_time_evolution_buffer = torch.stack([latents] * (len(scheduler.timesteps) + 1))
+
+ # We must make sure the noise of stochastic schedulers such as DDPM is sampled only once per timestep.
+ # Sampling inside the parallel denoising loop will mess this up, so we pre-sample the noise vectors outside the denoising loop.
+ noise_array = torch.zeros_like(latents_time_evolution_buffer)
+ for j in range(len(scheduler.timesteps)):
+ base_noise = randn_tensor(
+ shape=latents.shape, generator=generator, device=latents.device, dtype=prompt_embeds.dtype
+ )
+ noise = (self.scheduler._get_variance(scheduler.timesteps[j]) ** 0.5) * base_noise
+ noise_array[j] = noise.clone()
+
+ # We specify the error tolerance as a ratio of the scheduler's noise magnitude. We similarly compute the error tolerance
+ # outside of the denoising loop to avoid recomputing it at every step.
+ # We will be dividing the norm of the noise, so we store its inverse here to avoid a division at every step.
+ inverse_variance_norm = 1.0 / torch.tensor(
+ [scheduler._get_variance(scheduler.timesteps[j]) for j in range(len(scheduler.timesteps))] + [0]
+ ).to(noise_array.device)
+ latent_dim = noise_array[0, 0].numel()
+ inverse_variance_norm = inverse_variance_norm[:, None] / latent_dim
+
+ scaled_tolerance = tolerance**2
+
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ steps = 0
+ while begin_idx < len(scheduler.timesteps):
+ # these have shape (parallel_dim, 2*batch_size, ...)
+ # parallel_len is at most parallel, but could be less if we are at the end of the timesteps
+ # we are processing batch window of timesteps spanning [begin_idx, end_idx)
+ parallel_len = end_idx - begin_idx
+
+ block_prompt_embeds = torch.stack([prompt_embeds] * parallel_len)
+ block_latents = latents_time_evolution_buffer[begin_idx:end_idx]
+ block_t = scheduler.timesteps[begin_idx:end_idx, None].repeat(1, batch_size * num_images_per_prompt)
+ t_vec = block_t
+ if do_classifier_free_guidance:
+ t_vec = t_vec.repeat(1, 2)
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = (
+ torch.cat([block_latents] * 2, dim=1) if do_classifier_free_guidance else block_latents
+ )
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t_vec)
+
+ # if parallel_len is small, no need to use multiple GPUs
+ net = self.wrapped_unet if parallel_len > 3 else self.unet
+ # predict the noise residual, shape is now [parallel_len * 2 * batch_size * num_images_per_prompt, ...]
+ model_output = net(
+ latent_model_input.flatten(0, 1),
+ t_vec.flatten(0, 1),
+ encoder_hidden_states=block_prompt_embeds.flatten(0, 1),
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ per_latent_shape = model_output.shape[1:]
+ if do_classifier_free_guidance:
+ model_output = model_output.reshape(
+ parallel_len, 2, batch_size * num_images_per_prompt, *per_latent_shape
+ )
+ noise_pred_uncond, noise_pred_text = model_output[:, 0], model_output[:, 1]
+ model_output = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ model_output = model_output.reshape(
+ parallel_len * batch_size * num_images_per_prompt, *per_latent_shape
+ )
+
+ block_latents_denoise = scheduler.batch_step_no_noise(
+ model_output=model_output,
+ timesteps=block_t.flatten(0, 1),
+ sample=block_latents.flatten(0, 1),
+ **extra_step_kwargs,
+ ).reshape(block_latents.shape)
+
+ # back to shape (parallel_dim, batch_size, ...)
+ # now we want to add the pre-sampled noise
+ # parallel sampling algorithm requires computing the cumulative drift from the beginning
+ # of the window, so we need to compute cumulative sum of the deltas and the pre-sampled noises.
+ delta = block_latents_denoise - block_latents
+ cumulative_delta = self._cumsum(delta, dim=0, debug=debug)
+ cumulative_noise = self._cumsum(noise_array[begin_idx:end_idx], dim=0, debug=debug)
+
+ # if we are using an ODE-like scheduler (like DDIM), we don't want to add noise
+ if scheduler._is_ode_scheduler:
+ cumulative_noise = 0
+
+ block_latents_new = (
+ latents_time_evolution_buffer[begin_idx][None,] + cumulative_delta + cumulative_noise
+ )
+ cur_error = torch.linalg.norm(
+ (block_latents_new - latents_time_evolution_buffer[begin_idx + 1 : end_idx + 1]).reshape(
+ parallel_len, batch_size * num_images_per_prompt, -1
+ ),
+ dim=-1,
+ ).pow(2)
+ error_ratio = cur_error * inverse_variance_norm[begin_idx + 1 : end_idx + 1]
+
+ # find the first index of the vector error_ratio that is greater than error tolerance
+ # we can shift the window for the next iteration up to this index
+ error_ratio = torch.nn.functional.pad(
+ error_ratio, (0, 0, 0, 1), value=1e9
+ ) # handle the case when everything is below ratio, by padding the end of parallel_len dimension
+ any_error_at_time = torch.max(error_ratio > scaled_tolerance, dim=1).values.int()
+ ind = torch.argmax(any_error_at_time).item()
+
+ # compute the new begin and end idxs for the window
+ new_begin_idx = begin_idx + min(1 + ind, parallel)
+ new_end_idx = min(new_begin_idx + parallel, len(scheduler.timesteps))
+
+ # store the computed latents for the current window in the global buffer
+ latents_time_evolution_buffer[begin_idx + 1 : end_idx + 1] = block_latents_new
+ # initialize the new sliding window latents with the end of the current window,
+ # should be better than random initialization
+ latents_time_evolution_buffer[end_idx : new_end_idx + 1] = latents_time_evolution_buffer[end_idx][
+ None,
+ ]
+
+ steps += 1
+
+ progress_bar.update(new_begin_idx - begin_idx)
+ if callback is not None and steps % callback_steps == 0:
+ callback(begin_idx, block_t[begin_idx], latents_time_evolution_buffer[begin_idx])
+
+ begin_idx = new_begin_idx
+ end_idx = new_end_idx
+
+ latents = latents_time_evolution_buffer[-1]
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_pix2pix_zero.py b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_pix2pix_zero.py
new file mode 100644
index 0000000000000000000000000000000000000000..c819e5728181aa69816ab7c2c73afd91032a47b9
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stable_diffusion_variants/pipeline_stable_diffusion_pix2pix_zero.py
@@ -0,0 +1,1304 @@
+# Copyright 2024 Pix2Pix Zero Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import (
+ BlipForConditionalGeneration,
+ BlipProcessor,
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTokenizer,
+)
+
+from ....image_processor import PipelineImageInput, VaeImageProcessor
+from ....loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ....models import AutoencoderKL, UNet2DConditionModel
+from ....models.attention_processor import Attention
+from ....models.lora import adjust_lora_scale_text_encoder
+from ....schedulers import DDIMScheduler, DDPMScheduler, EulerAncestralDiscreteScheduler, LMSDiscreteScheduler
+from ....schedulers.scheduling_ddim_inverse import DDIMInverseScheduler
+from ....utils import (
+ PIL_INTERPOLATION,
+ USE_PEFT_BACKEND,
+ BaseOutput,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ...stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from ...stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class Pix2PixInversionPipelineOutput(BaseOutput, TextualInversionLoaderMixin):
+ """
+ Output class for Stable Diffusion pipelines.
+
+ Args:
+ latents (`torch.FloatTensor`)
+ inverted latents tensor
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+ num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+ """
+
+ latents: torch.FloatTensor
+ images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import requests
+ >>> import torch
+
+ >>> from diffusers import DDIMScheduler, StableDiffusionPix2PixZeroPipeline
+
+
+ >>> def download(embedding_url, local_filepath):
+ ... r = requests.get(embedding_url)
+ ... with open(local_filepath, "wb") as f:
+ ... f.write(r.content)
+
+
+ >>> model_ckpt = "CompVis/stable-diffusion-v1-4"
+ >>> pipeline = StableDiffusionPix2PixZeroPipeline.from_pretrained(model_ckpt, torch_dtype=torch.float16)
+ >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+ >>> pipeline.to("cuda")
+
+ >>> prompt = "a high resolution painting of a cat in the style of van gough"
+ >>> source_emb_url = "https://hf.co/datasets/sayakpaul/sample-datasets/resolve/main/cat.pt"
+ >>> target_emb_url = "https://hf.co/datasets/sayakpaul/sample-datasets/resolve/main/dog.pt"
+
+ >>> for url in [source_emb_url, target_emb_url]:
+ ... download(url, url.split("/")[-1])
+
+ >>> src_embeds = torch.load(source_emb_url.split("/")[-1])
+ >>> target_embeds = torch.load(target_emb_url.split("/")[-1])
+ >>> images = pipeline(
+ ... prompt,
+ ... source_embeds=src_embeds,
+ ... target_embeds=target_embeds,
+ ... num_inference_steps=50,
+ ... cross_attention_guidance_amount=0.15,
+ ... ).images
+
+ >>> images[0].save("edited_image_dog.png")
+ ```
+"""
+
+EXAMPLE_INVERT_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from transformers import BlipForConditionalGeneration, BlipProcessor
+ >>> from diffusers import DDIMScheduler, DDIMInverseScheduler, StableDiffusionPix2PixZeroPipeline
+
+ >>> import requests
+ >>> from PIL import Image
+
+ >>> captioner_id = "Salesforce/blip-image-captioning-base"
+ >>> processor = BlipProcessor.from_pretrained(captioner_id)
+ >>> model = BlipForConditionalGeneration.from_pretrained(
+ ... captioner_id, torch_dtype=torch.float16, low_cpu_mem_usage=True
+ ... )
+
+ >>> sd_model_ckpt = "CompVis/stable-diffusion-v1-4"
+ >>> pipeline = StableDiffusionPix2PixZeroPipeline.from_pretrained(
+ ... sd_model_ckpt,
+ ... caption_generator=model,
+ ... caption_processor=processor,
+ ... torch_dtype=torch.float16,
+ ... safety_checker=None,
+ ... )
+
+ >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+ >>> pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
+ >>> pipeline.enable_model_cpu_offload()
+
+ >>> img_url = "https://github.com/pix2pixzero/pix2pix-zero/raw/main/assets/test_images/cats/cat_6.png"
+
+ >>> raw_image = Image.open(requests.get(img_url, stream=True).raw).convert("RGB").resize((512, 512))
+ >>> # generate caption
+ >>> caption = pipeline.generate_caption(raw_image)
+
+ >>> # "a photography of a cat with flowers and dai dai daie - daie - daie kasaii"
+ >>> inv_latents = pipeline.invert(caption, image=raw_image).latents
+ >>> # we need to generate source and target embeds
+
+ >>> source_prompts = ["a cat sitting on the street", "a cat playing in the field", "a face of a cat"]
+
+ >>> target_prompts = ["a dog sitting on the street", "a dog playing in the field", "a face of a dog"]
+
+ >>> source_embeds = pipeline.get_embeds(source_prompts)
+ >>> target_embeds = pipeline.get_embeds(target_prompts)
+ >>> # the latents can then be used to edit a real image
+ >>> # when using Stable Diffusion 2 or other models that use v-prediction
+ >>> # set `cross_attention_guidance_amount` to 0.01 or less to avoid input latent gradient explosion
+
+ >>> image = pipeline(
+ ... caption,
+ ... source_embeds=source_embeds,
+ ... target_embeds=target_embeds,
+ ... num_inference_steps=50,
+ ... cross_attention_guidance_amount=0.15,
+ ... generator=generator,
+ ... latents=inv_latents,
+ ... negative_prompt=caption,
+ ... ).images[0]
+ >>> image.save("edited_image.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
+def preprocess(image):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+def prepare_unet(unet: UNet2DConditionModel):
+ """Modifies the UNet (`unet`) to perform Pix2Pix Zero optimizations."""
+ pix2pix_zero_attn_procs = {}
+ for name in unet.attn_processors.keys():
+ module_name = name.replace(".processor", "")
+ module = unet.get_submodule(module_name)
+ if "attn2" in name:
+ pix2pix_zero_attn_procs[name] = Pix2PixZeroAttnProcessor(is_pix2pix_zero=True)
+ module.requires_grad_(True)
+ else:
+ pix2pix_zero_attn_procs[name] = Pix2PixZeroAttnProcessor(is_pix2pix_zero=False)
+ module.requires_grad_(False)
+
+ unet.set_attn_processor(pix2pix_zero_attn_procs)
+ return unet
+
+
+class Pix2PixZeroL2Loss:
+ def __init__(self):
+ self.loss = 0.0
+
+ def compute_loss(self, predictions, targets):
+ self.loss += ((predictions - targets) ** 2).sum((1, 2)).mean(0)
+
+
+class Pix2PixZeroAttnProcessor:
+ """An attention processor class to store the attention weights.
+ In Pix2Pix Zero, it happens during computations in the cross-attention blocks."""
+
+ def __init__(self, is_pix2pix_zero=False):
+ self.is_pix2pix_zero = is_pix2pix_zero
+ if self.is_pix2pix_zero:
+ self.reference_cross_attn_map = {}
+
+ def __call__(
+ self,
+ attn: Attention,
+ hidden_states,
+ encoder_hidden_states=None,
+ attention_mask=None,
+ timestep=None,
+ loss=None,
+ ):
+ batch_size, sequence_length, _ = hidden_states.shape
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ query = attn.to_q(hidden_states)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ query = attn.head_to_batch_dim(query)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ attention_probs = attn.get_attention_scores(query, key, attention_mask)
+ if self.is_pix2pix_zero and timestep is not None:
+ # new bookkeeping to save the attention weights.
+ if loss is None:
+ self.reference_cross_attn_map[timestep.item()] = attention_probs.detach().cpu()
+ # compute loss
+ elif loss is not None:
+ prev_attn_probs = self.reference_cross_attn_map.pop(timestep.item())
+ loss.compute_loss(attention_probs, prev_attn_probs.to(attention_probs.device))
+
+ hidden_states = torch.bmm(attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ return hidden_states
+
+
+class StableDiffusionPix2PixZeroPipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+ Pipeline for pixel-level image editing using Pix2Pix Zero. Based on Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], [`EulerAncestralDiscreteScheduler`], or [`DDPMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ requires_safety_checker (bool):
+ Whether the pipeline requires a safety checker. We recommend setting it to True if you're using the
+ pipeline publicly.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = [
+ "safety_checker",
+ "feature_extractor",
+ "caption_generator",
+ "caption_processor",
+ "inverse_scheduler",
+ ]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: Union[DDPMScheduler, DDIMScheduler, EulerAncestralDiscreteScheduler, LMSDiscreteScheduler],
+ feature_extractor: CLIPImageProcessor,
+ safety_checker: StableDiffusionSafetyChecker,
+ inverse_scheduler: DDIMInverseScheduler,
+ caption_generator: BlipForConditionalGeneration,
+ caption_processor: BlipProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ caption_processor=caption_processor,
+ caption_generator=caption_generator,
+ inverse_scheduler=inverse_scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ source_embeds,
+ target_embeds,
+ callback_steps,
+ prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if source_embeds is None and target_embeds is None:
+ raise ValueError("`source_embeds` and `target_embeds` cannot be undefined.")
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def generate_caption(self, images):
+ """Generates caption for a given image."""
+ text = "a photography of"
+
+ prev_device = self.caption_generator.device
+
+ device = self._execution_device
+ inputs = self.caption_processor(images, text, return_tensors="pt").to(
+ device=device, dtype=self.caption_generator.dtype
+ )
+ self.caption_generator.to(device)
+ outputs = self.caption_generator.generate(**inputs, max_new_tokens=128)
+
+ # offload caption generator
+ self.caption_generator.to(prev_device)
+
+ caption = self.caption_processor.batch_decode(outputs, skip_special_tokens=True)[0]
+ return caption
+
+ def construct_direction(self, embs_source: torch.Tensor, embs_target: torch.Tensor):
+ """Constructs the edit direction to steer the image generation process semantically."""
+ return (embs_target.mean(0) - embs_source.mean(0)).unsqueeze(0)
+
+ @torch.no_grad()
+ def get_embeds(self, prompt: List[str], batch_size: int = 16) -> torch.FloatTensor:
+ num_prompts = len(prompt)
+ embeds = []
+ for i in range(0, num_prompts, batch_size):
+ prompt_slice = prompt[i : i + batch_size]
+
+ input_ids = self.tokenizer(
+ prompt_slice,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ ).input_ids
+
+ input_ids = input_ids.to(self.text_encoder.device)
+ embeds.append(self.text_encoder(input_ids)[0])
+
+ return torch.cat(embeds, dim=0).mean(0)[None]
+
+ def prepare_image_latents(self, image, batch_size, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ if image.shape[1] == 4:
+ latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if isinstance(generator, list):
+ latents = [
+ self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
+ ]
+ latents = torch.cat(latents, dim=0)
+ else:
+ latents = self.vae.encode(image).latent_dist.sample(generator)
+
+ latents = self.vae.config.scaling_factor * latents
+
+ if batch_size != latents.shape[0]:
+ if batch_size % latents.shape[0] == 0:
+ # expand image_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_latents_per_image = batch_size // latents.shape[0]
+ latents = torch.cat([latents] * additional_latents_per_image, dim=0)
+ else:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ latents = torch.cat([latents], dim=0)
+
+ return latents
+
+ def get_epsilon(self, model_output: torch.Tensor, sample: torch.Tensor, timestep: int):
+ pred_type = self.inverse_scheduler.config.prediction_type
+ alpha_prod_t = self.inverse_scheduler.alphas_cumprod[timestep]
+
+ beta_prod_t = 1 - alpha_prod_t
+
+ if pred_type == "epsilon":
+ return model_output
+ elif pred_type == "sample":
+ return (sample - alpha_prod_t ** (0.5) * model_output) / beta_prod_t ** (0.5)
+ elif pred_type == "v_prediction":
+ return (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {pred_type} must be one of `epsilon`, `sample`, or `v_prediction`"
+ )
+
+ def auto_corr_loss(self, hidden_states, generator=None):
+ reg_loss = 0.0
+ for i in range(hidden_states.shape[0]):
+ for j in range(hidden_states.shape[1]):
+ noise = hidden_states[i : i + 1, j : j + 1, :, :]
+ while True:
+ roll_amount = torch.randint(noise.shape[2] // 2, (1,), generator=generator).item()
+ reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=2)).mean() ** 2
+ reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=3)).mean() ** 2
+
+ if noise.shape[2] <= 8:
+ break
+ noise = F.avg_pool2d(noise, kernel_size=2)
+ return reg_loss
+
+ def kl_divergence(self, hidden_states):
+ mean = hidden_states.mean()
+ var = hidden_states.var()
+ return var + mean**2 - 1 - torch.log(var + 1e-7)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ source_embeds: torch.Tensor = None,
+ target_embeds: torch.Tensor = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ cross_attention_guidance_amount: float = 0.1,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ source_embeds (`torch.Tensor`):
+ Source concept embeddings. Generation of the embeddings as per the [original
+ paper](https://arxiv.org/abs/2302.03027). Used in discovering the edit direction.
+ target_embeds (`torch.Tensor`):
+ Target concept embeddings. Generation of the embeddings as per the [original
+ paper](https://arxiv.org/abs/2302.03027). Used in discovering the edit direction.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ cross_attention_guidance_amount (`float`, defaults to 0.1):
+ Amount of guidance needed from the reference cross-attention maps.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+ # 0. Define the spatial resolutions.
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ source_embeds,
+ target_embeds,
+ callback_steps,
+ prompt_embeds,
+ )
+
+ # 3. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+ if cross_attention_kwargs is None:
+ cross_attention_kwargs = {}
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Generate the inverted noise from the input image or any other image
+ # generated from the input prompt.
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+ latents_init = latents.clone()
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 8. Rejig the UNet so that we can obtain the cross-attenion maps and
+ # use them for guiding the subsequent image generation.
+ self.unet = prepare_unet(self.unet)
+
+ # 7. Denoising loop where we obtain the cross-attention maps.
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs={"timestep": t},
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 8. Compute the edit directions.
+ edit_direction = self.construct_direction(source_embeds, target_embeds).to(prompt_embeds.device)
+
+ # 9. Edit the prompt embeddings as per the edit directions discovered.
+ prompt_embeds_edit = prompt_embeds.clone()
+ prompt_embeds_edit[1:2] += edit_direction
+
+ # 10. Second denoising loop to generate the edited image.
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ latents = latents_init
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # we want to learn the latent such that it steers the generation
+ # process towards the edited direction, so make the make initial
+ # noise learnable
+ x_in = latent_model_input.detach().clone()
+ x_in.requires_grad = True
+
+ # optimizer
+ opt = torch.optim.SGD([x_in], lr=cross_attention_guidance_amount)
+
+ with torch.enable_grad():
+ # initialize loss
+ loss = Pix2PixZeroL2Loss()
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ x_in,
+ t,
+ encoder_hidden_states=prompt_embeds_edit.detach(),
+ cross_attention_kwargs={"timestep": t, "loss": loss},
+ ).sample
+
+ loss.loss.backward(retain_graph=False)
+ opt.step()
+
+ # recompute the noise
+ noise_pred = self.unet(
+ x_in.detach(),
+ t,
+ encoder_hidden_states=prompt_embeds_edit,
+ cross_attention_kwargs={"timestep": None},
+ ).sample
+
+ latents = x_in.detach().chunk(2)[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_INVERT_DOC_STRING)
+ def invert(
+ self,
+ prompt: Optional[str] = None,
+ image: PipelineImageInput = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ cross_attention_guidance_amount: float = 0.1,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ lambda_auto_corr: float = 20.0,
+ lambda_kl: float = 20.0,
+ num_reg_steps: int = 5,
+ num_auto_corr_rolls: int = 5,
+ ):
+ r"""
+ Function used to generate inverted latents given a prompt and image.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ image (`torch.FloatTensor` `np.ndarray`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, or tensor representing an image batch which will be used for conditioning. Can also accept
+ image latents as `image`, if passing latents directly, it will not be encoded again.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 1):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ cross_attention_guidance_amount (`float`, defaults to 0.1):
+ Amount of guidance needed from the reference cross-attention maps.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ lambda_auto_corr (`float`, *optional*, defaults to 20.0):
+ Lambda parameter to control auto correction
+ lambda_kl (`float`, *optional*, defaults to 20.0):
+ Lambda parameter to control Kullback–Leibler divergence output
+ num_reg_steps (`int`, *optional*, defaults to 5):
+ Number of regularization loss steps
+ num_auto_corr_rolls (`int`, *optional*, defaults to 5):
+ Number of auto correction roll steps
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.pipeline_stable_diffusion_pix2pix_zero.Pix2PixInversionPipelineOutput`] or
+ `tuple`:
+ [`~pipelines.stable_diffusion.pipeline_stable_diffusion_pix2pix_zero.Pix2PixInversionPipelineOutput`] if
+ `return_dict` is True, otherwise a `tuple. When returning a tuple, the first element is the inverted
+ latents tensor and then second is the corresponding decoded image.
+ """
+ # 1. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+ if cross_attention_kwargs is None:
+ cross_attention_kwargs = {}
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Preprocess image
+ image = self.image_processor.preprocess(image)
+
+ # 4. Prepare latent variables
+ latents = self.prepare_image_latents(image, batch_size, self.vae.dtype, device, generator)
+
+ # 5. Encode input prompt
+ num_images_per_prompt = 1
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ prompt_embeds=prompt_embeds,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.inverse_scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.inverse_scheduler.timesteps
+
+ # 6. Rejig the UNet so that we can obtain the cross-attenion maps and
+ # use them for guiding the subsequent image generation.
+ self.unet = prepare_unet(self.unet)
+
+ # 7. Denoising loop where we obtain the cross-attention maps.
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.inverse_scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.inverse_scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs={"timestep": t},
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # regularization of the noise prediction
+ with torch.enable_grad():
+ for _ in range(num_reg_steps):
+ if lambda_auto_corr > 0:
+ for _ in range(num_auto_corr_rolls):
+ var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True)
+
+ # Derive epsilon from model output before regularizing to IID standard normal
+ var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t)
+
+ l_ac = self.auto_corr_loss(var_epsilon, generator=generator)
+ l_ac.backward()
+
+ grad = var.grad.detach() / num_auto_corr_rolls
+ noise_pred = noise_pred - lambda_auto_corr * grad
+
+ if lambda_kl > 0:
+ var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True)
+
+ # Derive epsilon from model output before regularizing to IID standard normal
+ var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t)
+
+ l_kld = self.kl_divergence(var_epsilon)
+ l_kld.backward()
+
+ grad = var.grad.detach()
+ noise_pred = noise_pred - lambda_kl * grad
+
+ noise_pred = noise_pred.detach()
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.inverse_scheduler.step(noise_pred, t, latents).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or (
+ (i + 1) > num_warmup_steps and (i + 1) % self.inverse_scheduler.order == 0
+ ):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ inverted_latents = latents.detach().clone()
+
+ # 8. Post-processing
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (inverted_latents, image)
+
+ return Pix2PixInversionPipelineOutput(latents=inverted_latents, images=image)
diff --git a/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__init__.py b/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..15c9a8c27f98dd7e1913bd57dfd5e8dae71172b4
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__init__.py
@@ -0,0 +1,19 @@
+from typing import TYPE_CHECKING
+
+from ....utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {"pipeline_stochastic_karras_ve": ["KarrasVePipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_stochastic_karras_ve import KarrasVePipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7f832655aa3fc0c726f09dfe9c8443d54829b994
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__pycache__/pipeline_stochastic_karras_ve.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__pycache__/pipeline_stochastic_karras_ve.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..418f911ab2de7ae481a5cdbabf148e9a185fa2c5
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/__pycache__/pipeline_stochastic_karras_ve.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/pipeline_stochastic_karras_ve.py b/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/pipeline_stochastic_karras_ve.py
new file mode 100644
index 0000000000000000000000000000000000000000..023edb4ce4bd766b364aaeafcdfb2a2c1eaf9980
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/stochastic_karras_ve/pipeline_stochastic_karras_ve.py
@@ -0,0 +1,128 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ....models import UNet2DModel
+from ....schedulers import KarrasVeScheduler
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+class KarrasVePipeline(DiffusionPipeline):
+ r"""
+ Pipeline for unconditional image generation.
+
+ Parameters:
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image.
+ scheduler ([`KarrasVeScheduler`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image.
+ """
+
+ # add type hints for linting
+ unet: UNet2DModel
+ scheduler: KarrasVeScheduler
+
+ def __init__(self, unet: UNet2DModel, scheduler: KarrasVeScheduler):
+ super().__init__()
+ self.register_modules(unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ batch_size: int = 1,
+ num_inference_steps: int = 50,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ **kwargs,
+ ) -> Union[Tuple, ImagePipelineOutput]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ batch_size (`int`, *optional*, defaults to 1):
+ The number of images to generate.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+
+ img_size = self.unet.config.sample_size
+ shape = (batch_size, 3, img_size, img_size)
+
+ model = self.unet
+
+ # sample x_0 ~ N(0, sigma_0^2 * I)
+ sample = randn_tensor(shape, generator=generator, device=self.device) * self.scheduler.init_noise_sigma
+
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ for t in self.progress_bar(self.scheduler.timesteps):
+ # here sigma_t == t_i from the paper
+ sigma = self.scheduler.schedule[t]
+ sigma_prev = self.scheduler.schedule[t - 1] if t > 0 else 0
+
+ # 1. Select temporarily increased noise level sigma_hat
+ # 2. Add new noise to move from sample_i to sample_hat
+ sample_hat, sigma_hat = self.scheduler.add_noise_to_input(sample, sigma, generator=generator)
+
+ # 3. Predict the noise residual given the noise magnitude `sigma_hat`
+ # The model inputs and output are adjusted by following eq. (213) in [1].
+ model_output = (sigma_hat / 2) * model((sample_hat + 1) / 2, sigma_hat / 2).sample
+
+ # 4. Evaluate dx/dt at sigma_hat
+ # 5. Take Euler step from sigma to sigma_prev
+ step_output = self.scheduler.step(model_output, sigma_hat, sigma_prev, sample_hat)
+
+ if sigma_prev != 0:
+ # 6. Apply 2nd order correction
+ # The model inputs and output are adjusted by following eq. (213) in [1].
+ model_output = (sigma_prev / 2) * model((step_output.prev_sample + 1) / 2, sigma_prev / 2).sample
+ step_output = self.scheduler.step_correct(
+ model_output,
+ sigma_hat,
+ sigma_prev,
+ sample_hat,
+ step_output.prev_sample,
+ step_output["derivative"],
+ )
+ sample = step_output.prev_sample
+
+ sample = (sample / 2 + 0.5).clamp(0, 1)
+ image = sample.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__init__.py b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8ea6ef6e2f65b96aebebdf72cb80135003e4f08d
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__init__.py
@@ -0,0 +1,71 @@
+from typing import TYPE_CHECKING
+
+from ....utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ is_torch_available,
+ is_transformers_available,
+ is_transformers_version,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ....utils.dummy_torch_and_transformers_objects import (
+ VersatileDiffusionDualGuidedPipeline,
+ VersatileDiffusionImageVariationPipeline,
+ VersatileDiffusionPipeline,
+ VersatileDiffusionTextToImagePipeline,
+ )
+
+ _dummy_objects.update(
+ {
+ "VersatileDiffusionDualGuidedPipeline": VersatileDiffusionDualGuidedPipeline,
+ "VersatileDiffusionImageVariationPipeline": VersatileDiffusionImageVariationPipeline,
+ "VersatileDiffusionPipeline": VersatileDiffusionPipeline,
+ "VersatileDiffusionTextToImagePipeline": VersatileDiffusionTextToImagePipeline,
+ }
+ )
+else:
+ _import_structure["modeling_text_unet"] = ["UNetFlatConditionModel"]
+ _import_structure["pipeline_versatile_diffusion"] = ["VersatileDiffusionPipeline"]
+ _import_structure["pipeline_versatile_diffusion_dual_guided"] = ["VersatileDiffusionDualGuidedPipeline"]
+ _import_structure["pipeline_versatile_diffusion_image_variation"] = ["VersatileDiffusionImageVariationPipeline"]
+ _import_structure["pipeline_versatile_diffusion_text_to_image"] = ["VersatileDiffusionTextToImagePipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ....utils.dummy_torch_and_transformers_objects import (
+ VersatileDiffusionDualGuidedPipeline,
+ VersatileDiffusionImageVariationPipeline,
+ VersatileDiffusionPipeline,
+ VersatileDiffusionTextToImagePipeline,
+ )
+ else:
+ from .pipeline_versatile_diffusion import VersatileDiffusionPipeline
+ from .pipeline_versatile_diffusion_dual_guided import VersatileDiffusionDualGuidedPipeline
+ from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline
+ from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..534fc3a280b8d78e99fca4d75397d343a6b920d1
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/modeling_text_unet.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/modeling_text_unet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..306e6e6e5d9931573cca0b04e620e04428a2b094
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/modeling_text_unet.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..09a317d2b5b4d20137c25f3871a5e6beeb3c2b29
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_dual_guided.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_dual_guided.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5c87e867c2254ab7f55984c9f130ceeb08fce647
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_dual_guided.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_image_variation.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_image_variation.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..508292cf61ca27efdc2141d5d6559b0edebb1b26
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_image_variation.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_text_to_image.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_text_to_image.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d90fa908e86324da4e0796973e39fac32c617cde
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/__pycache__/pipeline_versatile_diffusion_text_to_image.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py
new file mode 100644
index 0000000000000000000000000000000000000000..97c533be58645977e92f2177866b9df2efbe1a82
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py
@@ -0,0 +1,2498 @@
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from diffusers.utils import deprecate
+
+from ....configuration_utils import ConfigMixin, register_to_config
+from ....models import ModelMixin
+from ....models.activations import get_activation
+from ....models.attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ Attention,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnAddedKVProcessor2_0,
+ AttnProcessor,
+)
+from ....models.embeddings import (
+ GaussianFourierProjection,
+ ImageHintTimeEmbedding,
+ ImageProjection,
+ ImageTimeEmbedding,
+ TextImageProjection,
+ TextImageTimeEmbedding,
+ TextTimeEmbedding,
+ TimestepEmbedding,
+ Timesteps,
+)
+from ....models.resnet import ResnetBlockCondNorm2D
+from ....models.transformers.dual_transformer_2d import DualTransformer2DModel
+from ....models.transformers.transformer_2d import Transformer2DModel
+from ....models.unets.unet_2d_condition import UNet2DConditionOutput
+from ....utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
+from ....utils.torch_utils import apply_freeu
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def get_down_block(
+ down_block_type,
+ num_layers,
+ in_channels,
+ out_channels,
+ temb_channels,
+ add_downsample,
+ resnet_eps,
+ resnet_act_fn,
+ num_attention_heads,
+ transformer_layers_per_block,
+ attention_type,
+ attention_head_dim,
+ resnet_groups=None,
+ cross_attention_dim=None,
+ downsample_padding=None,
+ dual_cross_attention=False,
+ use_linear_projection=False,
+ only_cross_attention=False,
+ upcast_attention=False,
+ resnet_time_scale_shift="default",
+ resnet_skip_time_act=False,
+ resnet_out_scale_factor=1.0,
+ cross_attention_norm=None,
+ dropout=0.0,
+):
+ down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+ if down_block_type == "DownBlockFlat":
+ return DownBlockFlat(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif down_block_type == "CrossAttnDownBlockFlat":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockFlat")
+ return CrossAttnDownBlockFlat(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_downsample=add_downsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ downsample_padding=downsample_padding,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ raise ValueError(f"{down_block_type} is not supported.")
+
+
+def get_up_block(
+ up_block_type,
+ num_layers,
+ in_channels,
+ out_channels,
+ prev_output_channel,
+ temb_channels,
+ add_upsample,
+ resnet_eps,
+ resnet_act_fn,
+ num_attention_heads,
+ transformer_layers_per_block,
+ resolution_idx,
+ attention_type,
+ attention_head_dim,
+ resnet_groups=None,
+ cross_attention_dim=None,
+ dual_cross_attention=False,
+ use_linear_projection=False,
+ only_cross_attention=False,
+ upcast_attention=False,
+ resnet_time_scale_shift="default",
+ resnet_skip_time_act=False,
+ resnet_out_scale_factor=1.0,
+ cross_attention_norm=None,
+ dropout=0.0,
+):
+ up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+ if up_block_type == "UpBlockFlat":
+ return UpBlockFlat(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ elif up_block_type == "CrossAttnUpBlockFlat":
+ if cross_attention_dim is None:
+ raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockFlat")
+ return CrossAttnUpBlockFlat(
+ num_layers=num_layers,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ prev_output_channel=prev_output_channel,
+ temb_channels=temb_channels,
+ dropout=dropout,
+ add_upsample=add_upsample,
+ resnet_eps=resnet_eps,
+ resnet_act_fn=resnet_act_fn,
+ resnet_groups=resnet_groups,
+ cross_attention_dim=cross_attention_dim,
+ num_attention_heads=num_attention_heads,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ )
+ raise ValueError(f"{up_block_type} is not supported.")
+
+
+class FourierEmbedder(nn.Module):
+ def __init__(self, num_freqs=64, temperature=100):
+ super().__init__()
+
+ self.num_freqs = num_freqs
+ self.temperature = temperature
+
+ freq_bands = temperature ** (torch.arange(num_freqs) / num_freqs)
+ freq_bands = freq_bands[None, None, None]
+ self.register_buffer("freq_bands", freq_bands, persistent=False)
+
+ def __call__(self, x):
+ x = self.freq_bands * x.unsqueeze(-1)
+ return torch.stack((x.sin(), x.cos()), dim=-1).permute(0, 1, 3, 4, 2).reshape(*x.shape[:2], -1)
+
+
+class GLIGENTextBoundingboxProjection(nn.Module):
+ def __init__(self, positive_len, out_dim, feature_type, fourier_freqs=8):
+ super().__init__()
+ self.positive_len = positive_len
+ self.out_dim = out_dim
+
+ self.fourier_embedder = FourierEmbedder(num_freqs=fourier_freqs)
+ self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy
+
+ if isinstance(out_dim, tuple):
+ out_dim = out_dim[0]
+
+ if feature_type == "text-only":
+ self.linears = nn.Sequential(
+ nn.Linear(self.positive_len + self.position_dim, 512),
+ nn.SiLU(),
+ nn.Linear(512, 512),
+ nn.SiLU(),
+ nn.Linear(512, out_dim),
+ )
+ self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+
+ elif feature_type == "text-image":
+ self.linears_text = nn.Sequential(
+ nn.Linear(self.positive_len + self.position_dim, 512),
+ nn.SiLU(),
+ nn.Linear(512, 512),
+ nn.SiLU(),
+ nn.Linear(512, out_dim),
+ )
+ self.linears_image = nn.Sequential(
+ nn.Linear(self.positive_len + self.position_dim, 512),
+ nn.SiLU(),
+ nn.Linear(512, 512),
+ nn.SiLU(),
+ nn.Linear(512, out_dim),
+ )
+ self.null_text_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+ self.null_image_feature = torch.nn.Parameter(torch.zeros([self.positive_len]))
+
+ self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim]))
+
+ def forward(
+ self,
+ boxes,
+ masks,
+ positive_embeddings=None,
+ phrases_masks=None,
+ image_masks=None,
+ phrases_embeddings=None,
+ image_embeddings=None,
+ ):
+ masks = masks.unsqueeze(-1)
+
+ xyxy_embedding = self.fourier_embedder(boxes)
+ xyxy_null = self.null_position_feature.view(1, 1, -1)
+ xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null
+
+ if positive_embeddings:
+ positive_null = self.null_positive_feature.view(1, 1, -1)
+ positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null
+
+ objs = self.linears(torch.cat([positive_embeddings, xyxy_embedding], dim=-1))
+ else:
+ phrases_masks = phrases_masks.unsqueeze(-1)
+ image_masks = image_masks.unsqueeze(-1)
+
+ text_null = self.null_text_feature.view(1, 1, -1)
+ image_null = self.null_image_feature.view(1, 1, -1)
+
+ phrases_embeddings = phrases_embeddings * phrases_masks + (1 - phrases_masks) * text_null
+ image_embeddings = image_embeddings * image_masks + (1 - image_masks) * image_null
+
+ objs_text = self.linears_text(torch.cat([phrases_embeddings, xyxy_embedding], dim=-1))
+ objs_image = self.linears_image(torch.cat([image_embeddings, xyxy_embedding], dim=-1))
+ objs = torch.cat([objs_text, objs_image], dim=1)
+
+ return objs
+
+
+class UNetFlatConditionModel(ModelMixin, ConfigMixin):
+ r"""
+ A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+ shaped output.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+ for all models (such as downloading or saving).
+
+ Parameters:
+ sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+ Height and width of input/output sample.
+ in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+ out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+ center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+ flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
+ Whether to flip the sin to cos in the time embedding.
+ freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+ down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "DownBlockFlat")`):
+ The tuple of downsample blocks to use.
+ mid_block_type (`str`, *optional*, defaults to `"UNetMidBlockFlatCrossAttn"`):
+ Block type for middle of UNet, it can be one of `UNetMidBlockFlatCrossAttn`, `UNetMidBlockFlat`, or
+ `UNetMidBlockFlatSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+ up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat")`):
+ The tuple of upsample blocks to use.
+ only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+ Whether to include self-attention in the basic transformer blocks, see
+ [`~models.attention.BasicTransformerBlock`].
+ block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+ The tuple of output channels for each block.
+ layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+ downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+ mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+ norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+ If `None`, normalization and activation layers is skipped in post-processing.
+ norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+ cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+ The dimension of the cross attention features.
+ transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+ [`~models.unet_2d_blocks.CrossAttnDownBlockFlat`], [`~models.unet_2d_blocks.CrossAttnUpBlockFlat`],
+ [`~models.unet_2d_blocks.UNetMidBlockFlatCrossAttn`].
+ reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+ The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+ blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+ [`~models.unet_2d_blocks.CrossAttnDownBlockFlat`], [`~models.unet_2d_blocks.CrossAttnUpBlockFlat`],
+ [`~models.unet_2d_blocks.UNetMidBlockFlatCrossAttn`].
+ encoder_hid_dim (`int`, *optional*, defaults to None):
+ If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+ dimension to `cross_attention_dim`.
+ encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+ If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+ embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+ attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+ num_attention_heads (`int`, *optional*):
+ The number of attention heads. If not defined, defaults to `attention_head_dim`
+ resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+ for ResNet blocks (see [`~models.resnet.ResnetBlockFlat`]). Choose from `default` or `scale_shift`.
+ class_embed_type (`str`, *optional*, defaults to `None`):
+ The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+ `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+ addition_embed_type (`str`, *optional*, defaults to `None`):
+ Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+ "text". "text" will use the `TextTimeEmbedding` layer.
+ addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+ Dimension for the timestep embeddings.
+ num_class_embeds (`int`, *optional*, defaults to `None`):
+ Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+ class conditioning with `class_embed_type` equal to `None`.
+ time_embedding_type (`str`, *optional*, defaults to `positional`):
+ The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+ time_embedding_dim (`int`, *optional*, defaults to `None`):
+ An optional override for the dimension of the projected time embedding.
+ time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+ Optional activation function to use only once on the time embeddings before they are passed to the rest of
+ the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+ timestep_post_act (`str`, *optional*, defaults to `None`):
+ The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+ time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+ The dimension of `cond_proj` layer in the timestep embedding.
+ conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`,
+ *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`,
+ *optional*): The dimension of the `class_labels` input when
+ `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+ class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+ embeddings with the class embeddings.
+ mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+ Whether to use cross attention with the mid block when using the `UNetMidBlockFlatSimpleCrossAttn`. If
+ `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+ `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+ otherwise.
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ sample_size: Optional[int] = None,
+ in_channels: int = 4,
+ out_channels: int = 4,
+ center_input_sample: bool = False,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str] = (
+ "CrossAttnDownBlockFlat",
+ "CrossAttnDownBlockFlat",
+ "CrossAttnDownBlockFlat",
+ "DownBlockFlat",
+ ),
+ mid_block_type: Optional[str] = "UNetMidBlockFlatCrossAttn",
+ up_block_types: Tuple[str] = (
+ "UpBlockFlat",
+ "CrossAttnUpBlockFlat",
+ "CrossAttnUpBlockFlat",
+ "CrossAttnUpBlockFlat",
+ ),
+ only_cross_attention: Union[bool, Tuple[bool]] = False,
+ block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+ layers_per_block: Union[int, Tuple[int]] = 2,
+ downsample_padding: int = 1,
+ mid_block_scale_factor: float = 1,
+ dropout: float = 0.0,
+ act_fn: str = "silu",
+ norm_num_groups: Optional[int] = 32,
+ norm_eps: float = 1e-5,
+ cross_attention_dim: Union[int, Tuple[int]] = 1280,
+ transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+ reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+ encoder_hid_dim: Optional[int] = None,
+ encoder_hid_dim_type: Optional[str] = None,
+ attention_head_dim: Union[int, Tuple[int]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ class_embed_type: Optional[str] = None,
+ addition_embed_type: Optional[str] = None,
+ addition_time_embed_dim: Optional[int] = None,
+ num_class_embeds: Optional[int] = None,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ resnet_skip_time_act: bool = False,
+ resnet_out_scale_factor: int = 1.0,
+ time_embedding_type: str = "positional",
+ time_embedding_dim: Optional[int] = None,
+ time_embedding_act_fn: Optional[str] = None,
+ timestep_post_act: Optional[str] = None,
+ time_cond_proj_dim: Optional[int] = None,
+ conv_in_kernel: int = 3,
+ conv_out_kernel: int = 3,
+ projection_class_embeddings_input_dim: Optional[int] = None,
+ attention_type: str = "default",
+ class_embeddings_concat: bool = False,
+ mid_block_only_cross_attention: Optional[bool] = None,
+ cross_attention_norm: Optional[str] = None,
+ addition_embed_type_num_heads=64,
+ ):
+ super().__init__()
+
+ self.sample_size = sample_size
+
+ if num_attention_heads is not None:
+ raise ValueError(
+ "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+ )
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ if len(down_block_types) != len(up_block_types):
+ raise ValueError(
+ f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+ )
+
+ if len(block_out_channels) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+ )
+
+ if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+ raise ValueError(
+ f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+ )
+ if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+ for layer_number_per_block in transformer_layers_per_block:
+ if isinstance(layer_number_per_block, list):
+ raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+ # input
+ conv_in_padding = (conv_in_kernel - 1) // 2
+ self.conv_in = LinearMultiDim(
+ in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+ )
+
+ # time
+ if time_embedding_type == "fourier":
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+ if time_embed_dim % 2 != 0:
+ raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+ self.time_proj = GaussianFourierProjection(
+ time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+ )
+ timestep_input_dim = time_embed_dim
+ elif time_embedding_type == "positional":
+ time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+ self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+ timestep_input_dim = block_out_channels[0]
+ else:
+ raise ValueError(
+ f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+ )
+
+ self.time_embedding = TimestepEmbedding(
+ timestep_input_dim,
+ time_embed_dim,
+ act_fn=act_fn,
+ post_act_fn=timestep_post_act,
+ cond_proj_dim=time_cond_proj_dim,
+ )
+
+ if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+ encoder_hid_dim_type = "text_proj"
+ self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+ logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+ if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+ )
+
+ if encoder_hid_dim_type == "text_proj":
+ self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+ elif encoder_hid_dim_type == "text_image_proj":
+ # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)`
+ self.encoder_hid_proj = TextImageProjection(
+ text_embed_dim=encoder_hid_dim,
+ image_embed_dim=cross_attention_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+ elif encoder_hid_dim_type == "image_proj":
+ # Kandinsky 2.2
+ self.encoder_hid_proj = ImageProjection(
+ image_embed_dim=encoder_hid_dim,
+ cross_attention_dim=cross_attention_dim,
+ )
+ elif encoder_hid_dim_type is not None:
+ raise ValueError(
+ f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
+ )
+ else:
+ self.encoder_hid_proj = None
+
+ # class embedding
+ if class_embed_type is None and num_class_embeds is not None:
+ self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+ elif class_embed_type == "timestep":
+ self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+ elif class_embed_type == "identity":
+ self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+ elif class_embed_type == "projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+ # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+ # 2. it projects from an arbitrary input dimension.
+ #
+ # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+ # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+ # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+ self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ elif class_embed_type == "simple_projection":
+ if projection_class_embeddings_input_dim is None:
+ raise ValueError(
+ "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+ )
+ self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+ else:
+ self.class_embedding = None
+
+ if addition_embed_type == "text":
+ if encoder_hid_dim is not None:
+ text_time_embedding_from_dim = encoder_hid_dim
+ else:
+ text_time_embedding_from_dim = cross_attention_dim
+
+ self.add_embedding = TextTimeEmbedding(
+ text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+ )
+ elif addition_embed_type == "text_image":
+ # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+ # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+ # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)`
+ self.add_embedding = TextImageTimeEmbedding(
+ text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+ )
+ elif addition_embed_type == "text_time":
+ self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+ self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+ elif addition_embed_type == "image":
+ # Kandinsky 2.2
+ self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+ elif addition_embed_type == "image_hint":
+ # Kandinsky 2.2 ControlNet
+ self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+ elif addition_embed_type is not None:
+ raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+ if time_embedding_act_fn is None:
+ self.time_embed_act = None
+ else:
+ self.time_embed_act = get_activation(time_embedding_act_fn)
+
+ self.down_blocks = nn.ModuleList([])
+ self.up_blocks = nn.ModuleList([])
+
+ if isinstance(only_cross_attention, bool):
+ if mid_block_only_cross_attention is None:
+ mid_block_only_cross_attention = only_cross_attention
+
+ only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+ if mid_block_only_cross_attention is None:
+ mid_block_only_cross_attention = False
+
+ if isinstance(num_attention_heads, int):
+ num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+ if isinstance(attention_head_dim, int):
+ attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+ if isinstance(layers_per_block, int):
+ layers_per_block = [layers_per_block] * len(down_block_types)
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+ if class_embeddings_concat:
+ # The time embeddings are concatenated with the class embeddings. The dimension of the
+ # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+ # regular time embeddings
+ blocks_time_embed_dim = time_embed_dim * 2
+ else:
+ blocks_time_embed_dim = time_embed_dim
+
+ # down
+ output_channel = block_out_channels[0]
+ for i, down_block_type in enumerate(down_block_types):
+ input_channel = output_channel
+ output_channel = block_out_channels[i]
+ is_final_block = i == len(block_out_channels) - 1
+
+ down_block = get_down_block(
+ down_block_type,
+ num_layers=layers_per_block[i],
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_downsample=not is_final_block,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim[i],
+ num_attention_heads=num_attention_heads[i],
+ downsample_padding=downsample_padding,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ resnet_out_scale_factor=resnet_out_scale_factor,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ dropout=dropout,
+ )
+ self.down_blocks.append(down_block)
+
+ # mid
+ if mid_block_type == "UNetMidBlockFlatCrossAttn":
+ self.mid_block = UNetMidBlockFlatCrossAttn(
+ transformer_layers_per_block=transformer_layers_per_block[-1],
+ in_channels=block_out_channels[-1],
+ temb_channels=blocks_time_embed_dim,
+ dropout=dropout,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ cross_attention_dim=cross_attention_dim[-1],
+ num_attention_heads=num_attention_heads[-1],
+ resnet_groups=norm_num_groups,
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ elif mid_block_type == "UNetMidBlockFlatSimpleCrossAttn":
+ self.mid_block = UNetMidBlockFlatSimpleCrossAttn(
+ in_channels=block_out_channels[-1],
+ temb_channels=blocks_time_embed_dim,
+ dropout=dropout,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ cross_attention_dim=cross_attention_dim[-1],
+ attention_head_dim=attention_head_dim[-1],
+ resnet_groups=norm_num_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ skip_time_act=resnet_skip_time_act,
+ only_cross_attention=mid_block_only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ )
+ elif mid_block_type == "UNetMidBlockFlat":
+ self.mid_block = UNetMidBlockFlat(
+ in_channels=block_out_channels[-1],
+ temb_channels=blocks_time_embed_dim,
+ dropout=dropout,
+ num_layers=0,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ output_scale_factor=mid_block_scale_factor,
+ resnet_groups=norm_num_groups,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ add_attention=False,
+ )
+ elif mid_block_type is None:
+ self.mid_block = None
+ else:
+ raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+ # count how many layers upsample the images
+ self.num_upsamplers = 0
+
+ # up
+ reversed_block_out_channels = list(reversed(block_out_channels))
+ reversed_num_attention_heads = list(reversed(num_attention_heads))
+ reversed_layers_per_block = list(reversed(layers_per_block))
+ reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+ reversed_transformer_layers_per_block = (
+ list(reversed(transformer_layers_per_block))
+ if reverse_transformer_layers_per_block is None
+ else reverse_transformer_layers_per_block
+ )
+ only_cross_attention = list(reversed(only_cross_attention))
+
+ output_channel = reversed_block_out_channels[0]
+ for i, up_block_type in enumerate(up_block_types):
+ is_final_block = i == len(block_out_channels) - 1
+
+ prev_output_channel = output_channel
+ output_channel = reversed_block_out_channels[i]
+ input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+ # add upsample block for all BUT final layer
+ if not is_final_block:
+ add_upsample = True
+ self.num_upsamplers += 1
+ else:
+ add_upsample = False
+
+ up_block = get_up_block(
+ up_block_type,
+ num_layers=reversed_layers_per_block[i] + 1,
+ transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+ in_channels=input_channel,
+ out_channels=output_channel,
+ prev_output_channel=prev_output_channel,
+ temb_channels=blocks_time_embed_dim,
+ add_upsample=add_upsample,
+ resnet_eps=norm_eps,
+ resnet_act_fn=act_fn,
+ resolution_idx=i,
+ resnet_groups=norm_num_groups,
+ cross_attention_dim=reversed_cross_attention_dim[i],
+ num_attention_heads=reversed_num_attention_heads[i],
+ dual_cross_attention=dual_cross_attention,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ attention_type=attention_type,
+ resnet_skip_time_act=resnet_skip_time_act,
+ resnet_out_scale_factor=resnet_out_scale_factor,
+ cross_attention_norm=cross_attention_norm,
+ attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+ dropout=dropout,
+ )
+ self.up_blocks.append(up_block)
+ prev_output_channel = output_channel
+
+ # out
+ if norm_num_groups is not None:
+ self.conv_norm_out = nn.GroupNorm(
+ num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+ )
+
+ self.conv_act = get_activation(act_fn)
+
+ else:
+ self.conv_norm_out = None
+ self.conv_act = None
+
+ conv_out_padding = (conv_out_kernel - 1) // 2
+ self.conv_out = LinearMultiDim(
+ block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+ )
+
+ if attention_type in ["gated", "gated-text-image"]:
+ positive_len = 768
+ if isinstance(cross_attention_dim, int):
+ positive_len = cross_attention_dim
+ elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
+ positive_len = cross_attention_dim[0]
+
+ feature_type = "text-only" if attention_type == "gated" else "text-image"
+ self.position_net = GLIGENTextBoundingboxProjection(
+ positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+ )
+
+ @property
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ def set_attention_slice(self, slice_size):
+ r"""
+ Enable sliced attention computation.
+
+ When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+ several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+ Args:
+ slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+ When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+ `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+ """
+ sliceable_head_dims = []
+
+ def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+ if hasattr(module, "set_attention_slice"):
+ sliceable_head_dims.append(module.sliceable_head_dim)
+
+ for child in module.children():
+ fn_recursive_retrieve_sliceable_dims(child)
+
+ # retrieve number of attention layers
+ for module in self.children():
+ fn_recursive_retrieve_sliceable_dims(module)
+
+ num_sliceable_layers = len(sliceable_head_dims)
+
+ if slice_size == "auto":
+ # half the attention head size is usually a good trade-off between
+ # speed and memory
+ slice_size = [dim // 2 for dim in sliceable_head_dims]
+ elif slice_size == "max":
+ # make smallest slice possible
+ slice_size = num_sliceable_layers * [1]
+
+ slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+ if len(slice_size) != len(sliceable_head_dims):
+ raise ValueError(
+ f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+ f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+ )
+
+ for i in range(len(slice_size)):
+ size = slice_size[i]
+ dim = sliceable_head_dims[i]
+ if size is not None and size > dim:
+ raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+ # Recursively walk through all the children.
+ # Any children which exposes the set_attention_slice method
+ # gets the message
+ def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+ if hasattr(module, "set_attention_slice"):
+ module.set_attention_slice(slice_size.pop())
+
+ for child in module.children():
+ fn_recursive_set_attention_slice(child, slice_size)
+
+ reversed_slice_size = list(reversed(slice_size))
+ for module in self.children():
+ fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if hasattr(module, "gradient_checkpointing"):
+ module.gradient_checkpointing = value
+
+ def enable_freeu(self, s1, s2, b1, b2):
+ r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+ The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+ Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+ are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+ Args:
+ s1 (`float`):
+ Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ s2 (`float`):
+ Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+ mitigate the "oversmoothing effect" in the enhanced denoising process.
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+ """
+ for i, upsample_block in enumerate(self.up_blocks):
+ setattr(upsample_block, "s1", s1)
+ setattr(upsample_block, "s2", s2)
+ setattr(upsample_block, "b1", b1)
+ setattr(upsample_block, "b2", b2)
+
+ def disable_freeu(self):
+ """Disables the FreeU mechanism."""
+ freeu_keys = {"s1", "s2", "b1", "b2"}
+ for i, upsample_block in enumerate(self.up_blocks):
+ for k in freeu_keys:
+ if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+ setattr(upsample_block, k, None)
+
+ def fuse_qkv_projections(self):
+ """
+ Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is 🧪 experimental.
+
+
+ """
+ self.original_attn_processors = None
+
+ for _, attn_processor in self.attn_processors.items():
+ if "Added" in str(attn_processor.__class__.__name__):
+ raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+ self.original_attn_processors = self.attn_processors
+
+ for module in self.modules():
+ if isinstance(module, Attention):
+ module.fuse_projections(fuse=True)
+
+ def unfuse_qkv_projections(self):
+ """Disables the fused QKV projection if enabled.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ """
+ if self.original_attn_processors is not None:
+ self.set_attn_processor(self.original_attn_processors)
+
+ def unload_lora(self):
+ """Unloads LoRA weights."""
+ deprecate(
+ "unload_lora",
+ "0.28.0",
+ "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().",
+ )
+ for module in self.modules():
+ if hasattr(module, "set_lora_layer"):
+ module.set_lora_layer(None)
+
+ def forward(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ class_labels: Optional[torch.Tensor] = None,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+ down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+ mid_block_additional_residual: Optional[torch.Tensor] = None,
+ down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+ encoder_attention_mask: Optional[torch.Tensor] = None,
+ return_dict: bool = True,
+ ) -> Union[UNet2DConditionOutput, Tuple]:
+ r"""
+ The [`UNetFlatConditionModel`] forward method.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The noisy input tensor with the following shape `(batch, channel, height, width)`.
+ timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+ encoder_hidden_states (`torch.FloatTensor`):
+ The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+ class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+ Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+ timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+ Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+ through the `self.time_embedding` layer to obtain the timestep embeddings.
+ attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+ An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+ is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+ negative values to the attention scores corresponding to "discard" tokens.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ added_cond_kwargs: (`dict`, *optional*):
+ A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+ are passed along to the UNet blocks.
+ down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+ A tuple of tensors that if specified are added to the residuals of down unet blocks.
+ mid_block_additional_residual: (`torch.Tensor`, *optional*):
+ A tensor that if specified is added to the residual of the middle unet block.
+ encoder_attention_mask (`torch.Tensor`):
+ A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+ `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+ which adds large negative values to the attention scores corresponding to "discard" tokens.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+ tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
+ added_cond_kwargs: (`dict`, *optional*):
+ A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+ are passed along to the UNet blocks.
+ down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+ additional residuals to be added to UNet long skip connections from down blocks to up blocks for
+ example from ControlNet side model(s)
+ mid_block_additional_residual (`torch.Tensor`, *optional*):
+ additional residual to be added to UNet mid block output, for example from ControlNet side model
+ down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+ additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+
+ Returns:
+ [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+ If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+ a `tuple` is returned where the first element is the sample tensor.
+ """
+ # By default samples have to be AT least a multiple of the overall upsampling factor.
+ # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+ # However, the upsampling interpolation output size can be forced to fit any upsampling size
+ # on the fly if necessary.
+ default_overall_up_factor = 2**self.num_upsamplers
+
+ # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+ forward_upsample_size = False
+ upsample_size = None
+
+ for dim in sample.shape[-2:]:
+ if dim % default_overall_up_factor != 0:
+ # Forward upsample size to force interpolation output size.
+ forward_upsample_size = True
+ break
+
+ # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+ # expects mask of shape:
+ # [batch, key_tokens]
+ # adds singleton query_tokens dimension:
+ # [batch, 1, key_tokens]
+ # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+ # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+ # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+ if attention_mask is not None:
+ # assume that mask is expressed as:
+ # (1 = keep, 0 = discard)
+ # convert mask into a bias that can be added to attention scores:
+ # (keep = +0, discard = -10000.0)
+ attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+ attention_mask = attention_mask.unsqueeze(1)
+
+ # convert encoder_attention_mask to a bias the same way we do for attention_mask
+ if encoder_attention_mask is not None:
+ encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+ encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+ # 0. center input if necessary
+ if self.config.center_input_sample:
+ sample = 2 * sample - 1.0
+
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # but time_embedding might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ t_emb = t_emb.to(dtype=sample.dtype)
+
+ emb = self.time_embedding(t_emb, timestep_cond)
+ aug_emb = None
+
+ if self.class_embedding is not None:
+ if class_labels is None:
+ raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+ if self.config.class_embed_type == "timestep":
+ class_labels = self.time_proj(class_labels)
+
+ # `Timesteps` does not contain any weights and will always return f32 tensors
+ # there might be better ways to encapsulate this.
+ class_labels = class_labels.to(dtype=sample.dtype)
+
+ class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+
+ if self.config.class_embeddings_concat:
+ emb = torch.cat([emb, class_emb], dim=-1)
+ else:
+ emb = emb + class_emb
+
+ if self.config.addition_embed_type == "text":
+ aug_emb = self.add_embedding(encoder_hidden_states)
+ elif self.config.addition_embed_type == "text_image":
+ # Kandinsky 2.1 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+ )
+
+ image_embs = added_cond_kwargs.get("image_embeds")
+ text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+ aug_emb = self.add_embedding(text_embs, image_embs)
+ elif self.config.addition_embed_type == "text_time":
+ # SDXL - style
+ if "text_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+ )
+ text_embeds = added_cond_kwargs.get("text_embeds")
+ if "time_ids" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+ )
+ time_ids = added_cond_kwargs.get("time_ids")
+ time_embeds = self.add_time_proj(time_ids.flatten())
+ time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+ add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+ add_embeds = add_embeds.to(emb.dtype)
+ aug_emb = self.add_embedding(add_embeds)
+ elif self.config.addition_embed_type == "image":
+ # Kandinsky 2.2 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+ )
+ image_embs = added_cond_kwargs.get("image_embeds")
+ aug_emb = self.add_embedding(image_embs)
+ elif self.config.addition_embed_type == "image_hint":
+ # Kandinsky 2.2 - style
+ if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+ )
+ image_embs = added_cond_kwargs.get("image_embeds")
+ hint = added_cond_kwargs.get("hint")
+ aug_emb, hint = self.add_embedding(image_embs, hint)
+ sample = torch.cat([sample, hint], dim=1)
+
+ emb = emb + aug_emb if aug_emb is not None else emb
+
+ if self.time_embed_act is not None:
+ emb = self.time_embed_act(emb)
+
+ if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+ encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+ elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+ # Kadinsky 2.1 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+ elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+ # Kandinsky 2.2 - style
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+ elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+ if "image_embeds" not in added_cond_kwargs:
+ raise ValueError(
+ f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`"
+ )
+ image_embeds = added_cond_kwargs.get("image_embeds")
+ image_embeds = self.encoder_hid_proj(image_embeds)
+ encoder_hidden_states = (encoder_hidden_states, image_embeds)
+
+ # 2. pre-process
+ sample = self.conv_in(sample)
+
+ # 2.5 GLIGEN position net
+ if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+ cross_attention_kwargs = cross_attention_kwargs.copy()
+ gligen_args = cross_attention_kwargs.pop("gligen")
+ cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+ # 3. down
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ if USE_PEFT_BACKEND:
+ # weight the lora layers by setting `lora_scale` for each PEFT layer
+ scale_lora_layers(self, lora_scale)
+
+ is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+ # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+ is_adapter = down_intrablock_additional_residuals is not None
+ # maintain backward compatibility for legacy usage, where
+ # T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+ # but can only use one or the other
+ if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+ deprecate(
+ "T2I should not use down_block_additional_residuals",
+ "1.3.0",
+ "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+ and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \
+ for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+ standard_warn=False,
+ )
+ down_intrablock_additional_residuals = down_block_additional_residuals
+ is_adapter = True
+
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ # For t2i-adapter CrossAttnDownBlockFlat
+ additional_residuals = {}
+ if is_adapter and len(down_intrablock_additional_residuals) > 0:
+ additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ **additional_residuals,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale)
+ if is_adapter and len(down_intrablock_additional_residuals) > 0:
+ sample += down_intrablock_additional_residuals.pop(0)
+
+ down_block_res_samples += res_samples
+
+ if is_controlnet:
+ new_down_block_res_samples = ()
+
+ for down_block_res_sample, down_block_additional_residual in zip(
+ down_block_res_samples, down_block_additional_residuals
+ ):
+ down_block_res_sample = down_block_res_sample + down_block_additional_residual
+ new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+ down_block_res_samples = new_down_block_res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+ else:
+ sample = self.mid_block(sample, emb)
+
+ # To support T2I-Adapter-XL
+ if (
+ is_adapter
+ and len(down_intrablock_additional_residuals) > 0
+ and sample.shape == down_intrablock_additional_residuals[0].shape
+ ):
+ sample += down_intrablock_additional_residuals.pop(0)
+
+ if is_controlnet:
+ sample = sample + mid_block_additional_residual
+
+ # 5. up
+ for i, upsample_block in enumerate(self.up_blocks):
+ is_final_block = i == len(self.up_blocks) - 1
+
+ res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+ down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+ # if we have not reached the final block and need to forward the
+ # upsample size, we do it here
+ if not is_final_block and forward_upsample_size:
+ upsample_size = down_block_res_samples[-1].shape[2:]
+
+ if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ upsample_size=upsample_size,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ )
+ else:
+ sample = upsample_block(
+ hidden_states=sample,
+ temb=emb,
+ res_hidden_states_tuple=res_samples,
+ upsample_size=upsample_size,
+ scale=lora_scale,
+ )
+
+ # 6. post-process
+ if self.conv_norm_out:
+ sample = self.conv_norm_out(sample)
+ sample = self.conv_act(sample)
+ sample = self.conv_out(sample)
+
+ if USE_PEFT_BACKEND:
+ # remove `lora_scale` from each PEFT layer
+ unscale_lora_layers(self, lora_scale)
+
+ if not return_dict:
+ return (sample,)
+
+ return UNet2DConditionOutput(sample=sample)
+
+
+class LinearMultiDim(nn.Linear):
+ def __init__(self, in_features, out_features=None, second_dim=4, *args, **kwargs):
+ in_features = [in_features, second_dim, 1] if isinstance(in_features, int) else list(in_features)
+ if out_features is None:
+ out_features = in_features
+ out_features = [out_features, second_dim, 1] if isinstance(out_features, int) else list(out_features)
+ self.in_features_multidim = in_features
+ self.out_features_multidim = out_features
+ super().__init__(np.array(in_features).prod(), np.array(out_features).prod())
+
+ def forward(self, input_tensor, *args, **kwargs):
+ shape = input_tensor.shape
+ n_dim = len(self.in_features_multidim)
+ input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_features)
+ output_tensor = super().forward(input_tensor)
+ output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_features_multidim)
+ return output_tensor
+
+
+class ResnetBlockFlat(nn.Module):
+ def __init__(
+ self,
+ *,
+ in_channels,
+ out_channels=None,
+ dropout=0.0,
+ temb_channels=512,
+ groups=32,
+ groups_out=None,
+ pre_norm=True,
+ eps=1e-6,
+ time_embedding_norm="default",
+ use_in_shortcut=None,
+ second_dim=4,
+ **kwargs,
+ ):
+ super().__init__()
+ self.pre_norm = pre_norm
+ self.pre_norm = True
+
+ in_channels = [in_channels, second_dim, 1] if isinstance(in_channels, int) else list(in_channels)
+ self.in_channels_prod = np.array(in_channels).prod()
+ self.channels_multidim = in_channels
+
+ if out_channels is not None:
+ out_channels = [out_channels, second_dim, 1] if isinstance(out_channels, int) else list(out_channels)
+ out_channels_prod = np.array(out_channels).prod()
+ self.out_channels_multidim = out_channels
+ else:
+ out_channels_prod = self.in_channels_prod
+ self.out_channels_multidim = self.channels_multidim
+ self.time_embedding_norm = time_embedding_norm
+
+ if groups_out is None:
+ groups_out = groups
+
+ self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=self.in_channels_prod, eps=eps, affine=True)
+ self.conv1 = torch.nn.Conv2d(self.in_channels_prod, out_channels_prod, kernel_size=1, padding=0)
+
+ if temb_channels is not None:
+ self.time_emb_proj = torch.nn.Linear(temb_channels, out_channels_prod)
+ else:
+ self.time_emb_proj = None
+
+ self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels_prod, eps=eps, affine=True)
+ self.dropout = torch.nn.Dropout(dropout)
+ self.conv2 = torch.nn.Conv2d(out_channels_prod, out_channels_prod, kernel_size=1, padding=0)
+
+ self.nonlinearity = nn.SiLU()
+
+ self.use_in_shortcut = (
+ self.in_channels_prod != out_channels_prod if use_in_shortcut is None else use_in_shortcut
+ )
+
+ self.conv_shortcut = None
+ if self.use_in_shortcut:
+ self.conv_shortcut = torch.nn.Conv2d(
+ self.in_channels_prod, out_channels_prod, kernel_size=1, stride=1, padding=0
+ )
+
+ def forward(self, input_tensor, temb):
+ shape = input_tensor.shape
+ n_dim = len(self.channels_multidim)
+ input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_channels_prod, 1, 1)
+ input_tensor = input_tensor.view(-1, self.in_channels_prod, 1, 1)
+
+ hidden_states = input_tensor
+
+ hidden_states = self.norm1(hidden_states)
+ hidden_states = self.nonlinearity(hidden_states)
+ hidden_states = self.conv1(hidden_states)
+
+ if temb is not None:
+ temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None]
+ hidden_states = hidden_states + temb
+
+ hidden_states = self.norm2(hidden_states)
+ hidden_states = self.nonlinearity(hidden_states)
+
+ hidden_states = self.dropout(hidden_states)
+ hidden_states = self.conv2(hidden_states)
+
+ if self.conv_shortcut is not None:
+ input_tensor = self.conv_shortcut(input_tensor)
+
+ output_tensor = input_tensor + hidden_states
+
+ output_tensor = output_tensor.view(*shape[0:-n_dim], -1)
+ output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_channels_multidim)
+
+ return output_tensor
+
+
+class DownBlockFlat(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_downsample: bool = True,
+ downsample_padding: int = 1,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlockFlat(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ LinearMultiDim(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ for resnet in self.resnets:
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale=scale)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+class CrossAttnDownBlockFlat(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ downsample_padding: int = 1,
+ add_downsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ for i in range(num_layers):
+ in_channels = in_channels if i == 0 else out_channels
+ resnets.append(
+ ResnetBlockFlat(
+ in_channels=in_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_downsample:
+ self.downsamplers = nn.ModuleList(
+ [
+ LinearMultiDim(
+ out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+ )
+ ]
+ )
+ else:
+ self.downsamplers = None
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ additional_residuals: Optional[torch.FloatTensor] = None,
+ ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
+ output_states = ()
+
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+ blocks = list(zip(self.resnets, self.attentions))
+
+ for i, (resnet, attn) in enumerate(blocks):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+
+ # apply additional residuals to the output of the last pair of resnet and attention blocks
+ if i == len(blocks) - 1 and additional_residuals is not None:
+ hidden_states = hidden_states + additional_residuals
+
+ output_states = output_states + (hidden_states,)
+
+ if self.downsamplers is not None:
+ for downsampler in self.downsamplers:
+ hidden_states = downsampler(hidden_states, scale=lora_scale)
+
+ output_states = output_states + (hidden_states,)
+
+ return hidden_states, output_states
+
+
+# Copied from diffusers.models.unets.unet_2d_blocks.UpBlock2D with UpBlock2D->UpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim
+class UpBlockFlat(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ prev_output_channel: int,
+ out_channels: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ ):
+ super().__init__()
+ resnets = []
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlockFlat(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ upsample_size: Optional[int] = None,
+ scale: float = 1.0,
+ ) -> torch.FloatTensor:
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+
+ for resnet in self.resnets:
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
+ )
+ else:
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet), hidden_states, temb
+ )
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=scale)
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size, scale=scale)
+
+ return hidden_states
+
+
+# Copied from diffusers.models.unets.unet_2d_blocks.CrossAttnUpBlock2D with CrossAttnUpBlock2D->CrossAttnUpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim
+class CrossAttnUpBlockFlat(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ out_channels: int,
+ prev_output_channel: int,
+ temb_channels: int,
+ resolution_idx: Optional[int] = None,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ cross_attention_dim: int = 1280,
+ output_scale_factor: float = 1.0,
+ add_upsample: bool = True,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+ resnets = []
+ attentions = []
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ for i in range(num_layers):
+ res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+ resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+ resnets.append(
+ ResnetBlockFlat(
+ in_channels=resnet_in_channels + res_skip_channels,
+ out_channels=out_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ out_channels // num_attention_heads,
+ in_channels=out_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ if add_upsample:
+ self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)])
+ else:
+ self.upsamplers = None
+
+ self.gradient_checkpointing = False
+ self.resolution_idx = resolution_idx
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ upsample_size: Optional[int] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ is_freeu_enabled = (
+ getattr(self, "s1", None)
+ and getattr(self, "s2", None)
+ and getattr(self, "b1", None)
+ and getattr(self, "b2", None)
+ )
+
+ for resnet, attn in zip(self.resnets, self.attentions):
+ # pop res hidden states
+ res_hidden_states = res_hidden_states_tuple[-1]
+ res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+
+ # FreeU: Only operate on the first two stages
+ if is_freeu_enabled:
+ hidden_states, res_hidden_states = apply_freeu(
+ self.resolution_idx,
+ hidden_states,
+ res_hidden_states,
+ s1=self.s1,
+ s2=self.s2,
+ b1=self.b1,
+ b2=self.b2,
+ )
+
+ hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ else:
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+
+ if self.upsamplers is not None:
+ for upsampler in self.upsamplers:
+ hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale)
+
+ return hidden_states
+
+
+# Copied from diffusers.models.unets.unet_2d_blocks.UNetMidBlock2D with UNetMidBlock2D->UNetMidBlockFlat, ResnetBlock2D->ResnetBlockFlat
+class UNetMidBlockFlat(nn.Module):
+ """
+ A 2D UNet mid-block [`UNetMidBlockFlat`] with multiple residual blocks and optional attention blocks.
+
+ Args:
+ in_channels (`int`): The number of input channels.
+ temb_channels (`int`): The number of temporal embedding channels.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout rate.
+ num_layers (`int`, *optional*, defaults to 1): The number of residual blocks.
+ resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
+ resnet_time_scale_shift (`str`, *optional*, defaults to `default`):
+ The type of normalization to apply to the time embeddings. This can help to improve the performance of the
+ model on tasks with long-range temporal dependencies.
+ resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks.
+ resnet_groups (`int`, *optional*, defaults to 32):
+ The number of groups to use in the group normalization layers of the resnet blocks.
+ attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks.
+ resnet_pre_norm (`bool`, *optional*, defaults to `True`):
+ Whether to use pre-normalization for the resnet blocks.
+ add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks.
+ attention_head_dim (`int`, *optional*, defaults to 1):
+ Dimension of a single attention head. The number of attention heads is determined based on this value and
+ the number of input channels.
+ output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor.
+
+ Returns:
+ `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
+ in_channels, height, width)`.
+
+ """
+
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default", # default, spatial
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ attn_groups: Optional[int] = None,
+ resnet_pre_norm: bool = True,
+ add_attention: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ ):
+ super().__init__()
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+ self.add_attention = add_attention
+
+ if attn_groups is None:
+ attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None
+
+ # there is always at least one resnet
+ if resnet_time_scale_shift == "spatial":
+ resnets = [
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ ]
+ else:
+ resnets = [
+ ResnetBlockFlat(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ attentions = []
+
+ if attention_head_dim is None:
+ logger.warn(
+ f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}."
+ )
+ attention_head_dim = in_channels
+
+ for _ in range(num_layers):
+ if self.add_attention:
+ attentions.append(
+ Attention(
+ in_channels,
+ heads=in_channels // attention_head_dim,
+ dim_head=attention_head_dim,
+ rescale_output_factor=output_scale_factor,
+ eps=resnet_eps,
+ norm_num_groups=attn_groups,
+ spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None,
+ residual_connection=True,
+ bias=True,
+ upcast_softmax=True,
+ _from_deprecated_attn_block=True,
+ )
+ )
+ else:
+ attentions.append(None)
+
+ if resnet_time_scale_shift == "spatial":
+ resnets.append(
+ ResnetBlockCondNorm2D(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm="spatial",
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ )
+ )
+ else:
+ resnets.append(
+ ResnetBlockFlat(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor:
+ hidden_states = self.resnets[0](hidden_states, temb)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ if attn is not None:
+ hidden_states = attn(hidden_states, temb=temb)
+ hidden_states = resnet(hidden_states, temb)
+
+ return hidden_states
+
+
+# Copied from diffusers.models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn with UNetMidBlock2DCrossAttn->UNetMidBlockFlatCrossAttn, ResnetBlock2D->ResnetBlockFlat
+class UNetMidBlockFlatCrossAttn(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ num_attention_heads: int = 1,
+ output_scale_factor: float = 1.0,
+ cross_attention_dim: int = 1280,
+ dual_cross_attention: bool = False,
+ use_linear_projection: bool = False,
+ upcast_attention: bool = False,
+ attention_type: str = "default",
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+ self.num_attention_heads = num_attention_heads
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ # support for variable transformer layers per block
+ if isinstance(transformer_layers_per_block, int):
+ transformer_layers_per_block = [transformer_layers_per_block] * num_layers
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlockFlat(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ ]
+ attentions = []
+
+ for i in range(num_layers):
+ if not dual_cross_attention:
+ attentions.append(
+ Transformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=transformer_layers_per_block[i],
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type=attention_type,
+ )
+ )
+ else:
+ attentions.append(
+ DualTransformer2DModel(
+ num_attention_heads,
+ in_channels // num_attention_heads,
+ in_channels=in_channels,
+ num_layers=1,
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ )
+ )
+ resnets.append(
+ ResnetBlockFlat(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ self.gradient_checkpointing = False
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module, return_dict=None):
+ def custom_forward(*inputs):
+ if return_dict is not None:
+ return module(*inputs, return_dict=return_dict)
+ else:
+ return module(*inputs)
+
+ return custom_forward
+
+ ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(resnet),
+ hidden_states,
+ temb,
+ **ckpt_kwargs,
+ )
+ else:
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ cross_attention_kwargs=cross_attention_kwargs,
+ attention_mask=attention_mask,
+ encoder_attention_mask=encoder_attention_mask,
+ return_dict=False,
+ )[0]
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
+
+
+# Copied from diffusers.models.unets.unet_2d_blocks.UNetMidBlock2DSimpleCrossAttn with UNetMidBlock2DSimpleCrossAttn->UNetMidBlockFlatSimpleCrossAttn, ResnetBlock2D->ResnetBlockFlat
+class UNetMidBlockFlatSimpleCrossAttn(nn.Module):
+ def __init__(
+ self,
+ in_channels: int,
+ temb_channels: int,
+ dropout: float = 0.0,
+ num_layers: int = 1,
+ resnet_eps: float = 1e-6,
+ resnet_time_scale_shift: str = "default",
+ resnet_act_fn: str = "swish",
+ resnet_groups: int = 32,
+ resnet_pre_norm: bool = True,
+ attention_head_dim: int = 1,
+ output_scale_factor: float = 1.0,
+ cross_attention_dim: int = 1280,
+ skip_time_act: bool = False,
+ only_cross_attention: bool = False,
+ cross_attention_norm: Optional[str] = None,
+ ):
+ super().__init__()
+
+ self.has_cross_attention = True
+
+ self.attention_head_dim = attention_head_dim
+ resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+
+ self.num_heads = in_channels // self.attention_head_dim
+
+ # there is always at least one resnet
+ resnets = [
+ ResnetBlockFlat(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ ]
+ attentions = []
+
+ for _ in range(num_layers):
+ processor = (
+ AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor()
+ )
+
+ attentions.append(
+ Attention(
+ query_dim=in_channels,
+ cross_attention_dim=in_channels,
+ heads=self.num_heads,
+ dim_head=self.attention_head_dim,
+ added_kv_proj_dim=cross_attention_dim,
+ norm_num_groups=resnet_groups,
+ bias=True,
+ upcast_softmax=True,
+ only_cross_attention=only_cross_attention,
+ cross_attention_norm=cross_attention_norm,
+ processor=processor,
+ )
+ )
+ resnets.append(
+ ResnetBlockFlat(
+ in_channels=in_channels,
+ out_channels=in_channels,
+ temb_channels=temb_channels,
+ eps=resnet_eps,
+ groups=resnet_groups,
+ dropout=dropout,
+ time_embedding_norm=resnet_time_scale_shift,
+ non_linearity=resnet_act_fn,
+ output_scale_factor=output_scale_factor,
+ pre_norm=resnet_pre_norm,
+ skip_time_act=skip_time_act,
+ )
+ )
+
+ self.attentions = nn.ModuleList(attentions)
+ self.resnets = nn.ModuleList(resnets)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ temb: Optional[torch.FloatTensor] = None,
+ encoder_hidden_states: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+ lora_scale = cross_attention_kwargs.get("scale", 1.0)
+
+ if attention_mask is None:
+ # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask.
+ mask = None if encoder_hidden_states is None else encoder_attention_mask
+ else:
+ # when attention_mask is defined: we don't even check for encoder_attention_mask.
+ # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks.
+ # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask.
+ # then we can simplify this whole if/else block to:
+ # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask
+ mask = attention_mask
+
+ hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale)
+ for attn, resnet in zip(self.attentions, self.resnets[1:]):
+ # attn
+ hidden_states = attn(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=mask,
+ **cross_attention_kwargs,
+ )
+
+ # resnet
+ hidden_states = resnet(hidden_states, temb, scale=lora_scale)
+
+ return hidden_states
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion.py b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..4455d20df2132166c3ab94155365b17f373d6a2c
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion.py
@@ -0,0 +1,421 @@
+import inspect
+from typing import Callable, List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModel
+
+from ....models import AutoencoderKL, UNet2DConditionModel
+from ....schedulers import KarrasDiffusionSchedulers
+from ....utils import logging
+from ...pipeline_utils import DiffusionPipeline
+from .pipeline_versatile_diffusion_dual_guided import VersatileDiffusionDualGuidedPipeline
+from .pipeline_versatile_diffusion_image_variation import VersatileDiffusionImageVariationPipeline
+from .pipeline_versatile_diffusion_text_to_image import VersatileDiffusionTextToImagePipeline
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class VersatileDiffusionPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ tokenizer: CLIPTokenizer
+ image_feature_extractor: CLIPImageProcessor
+ text_encoder: CLIPTextModel
+ image_encoder: CLIPVisionModel
+ image_unet: UNet2DConditionModel
+ text_unet: UNet2DConditionModel
+ vae: AutoencoderKL
+ scheduler: KarrasDiffusionSchedulers
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ image_feature_extractor: CLIPImageProcessor,
+ text_encoder: CLIPTextModel,
+ image_encoder: CLIPVisionModel,
+ image_unet: UNet2DConditionModel,
+ text_unet: UNet2DConditionModel,
+ vae: AutoencoderKL,
+ scheduler: KarrasDiffusionSchedulers,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ tokenizer=tokenizer,
+ image_feature_extractor=image_feature_extractor,
+ text_encoder=text_encoder,
+ image_encoder=image_encoder,
+ image_unet=image_unet,
+ text_unet=text_unet,
+ vae=vae,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ @torch.no_grad()
+ def image_variation(
+ self,
+ image: Union[torch.FloatTensor, PIL.Image.Image],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`):
+ The image prompt or prompts to guide the image generation.
+ height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import VersatileDiffusionPipeline
+ >>> import torch
+ >>> import requests
+ >>> from io import BytesIO
+ >>> from PIL import Image
+
+ >>> # let's download an initial image
+ >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg"
+
+ >>> response = requests.get(url)
+ >>> image = Image.open(BytesIO(response.content)).convert("RGB")
+
+ >>> pipe = VersatileDiffusionPipeline.from_pretrained(
+ ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(0)
+ >>> image = pipe.image_variation(image, generator=generator).images[0]
+ >>> image.save("./car_variation.png")
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ expected_components = inspect.signature(VersatileDiffusionImageVariationPipeline.__init__).parameters.keys()
+ components = {name: component for name, component in self.components.items() if name in expected_components}
+ return VersatileDiffusionImageVariationPipeline(**components)(
+ image=image,
+ height=height,
+ width=width,
+ num_inference_steps=num_inference_steps,
+ guidance_scale=guidance_scale,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ eta=eta,
+ generator=generator,
+ latents=latents,
+ output_type=output_type,
+ return_dict=return_dict,
+ callback=callback,
+ callback_steps=callback_steps,
+ )
+
+ @torch.no_grad()
+ def text_to_image(
+ self,
+ prompt: Union[str, List[str]],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation.
+ height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import VersatileDiffusionPipeline
+ >>> import torch
+
+ >>> pipe = VersatileDiffusionPipeline.from_pretrained(
+ ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(0)
+ >>> image = pipe.text_to_image("an astronaut riding on a horse on mars", generator=generator).images[0]
+ >>> image.save("./astronaut.png")
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ expected_components = inspect.signature(VersatileDiffusionTextToImagePipeline.__init__).parameters.keys()
+ components = {name: component for name, component in self.components.items() if name in expected_components}
+ temp_pipeline = VersatileDiffusionTextToImagePipeline(**components)
+ output = temp_pipeline(
+ prompt=prompt,
+ height=height,
+ width=width,
+ num_inference_steps=num_inference_steps,
+ guidance_scale=guidance_scale,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ eta=eta,
+ generator=generator,
+ latents=latents,
+ output_type=output_type,
+ return_dict=return_dict,
+ callback=callback,
+ callback_steps=callback_steps,
+ )
+ # swap the attention blocks back to the original state
+ temp_pipeline._swap_unet_attention_blocks()
+
+ return output
+
+ @torch.no_grad()
+ def dual_guided(
+ self,
+ prompt: Union[PIL.Image.Image, List[PIL.Image.Image]],
+ image: Union[str, List[str]],
+ text_to_image_strength: float = 0.5,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation.
+ height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import VersatileDiffusionPipeline
+ >>> import torch
+ >>> import requests
+ >>> from io import BytesIO
+ >>> from PIL import Image
+
+ >>> # let's download an initial image
+ >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg"
+
+ >>> response = requests.get(url)
+ >>> image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> text = "a red car in the sun"
+
+ >>> pipe = VersatileDiffusionPipeline.from_pretrained(
+ ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(0)
+ >>> text_to_image_strength = 0.75
+
+ >>> image = pipe.dual_guided(
+ ... prompt=text, image=image, text_to_image_strength=text_to_image_strength, generator=generator
+ ... ).images[0]
+ >>> image.save("./car_variation.png")
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+
+ expected_components = inspect.signature(VersatileDiffusionDualGuidedPipeline.__init__).parameters.keys()
+ components = {name: component for name, component in self.components.items() if name in expected_components}
+ temp_pipeline = VersatileDiffusionDualGuidedPipeline(**components)
+ output = temp_pipeline(
+ prompt=prompt,
+ image=image,
+ text_to_image_strength=text_to_image_strength,
+ height=height,
+ width=width,
+ num_inference_steps=num_inference_steps,
+ guidance_scale=guidance_scale,
+ num_images_per_prompt=num_images_per_prompt,
+ eta=eta,
+ generator=generator,
+ latents=latents,
+ output_type=output_type,
+ return_dict=return_dict,
+ callback=callback,
+ callback_steps=callback_steps,
+ )
+ temp_pipeline._revert_dual_attention()
+
+ return output
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py
new file mode 100644
index 0000000000000000000000000000000000000000..8af739bbe42846d3fc77afdb61245ff3222c9821
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_dual_guided.py
@@ -0,0 +1,556 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.utils.checkpoint
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ....image_processor import VaeImageProcessor
+from ....models import AutoencoderKL, DualTransformer2DModel, Transformer2DModel, UNet2DConditionModel
+from ....schedulers import KarrasDiffusionSchedulers
+from ....utils import deprecate, logging
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .modeling_text_unet import UNetFlatConditionModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class VersatileDiffusionDualGuidedPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for image-text dual-guided generation using Versatile Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ vqvae ([`VQModel`]):
+ Vector-quantized (VQ) model to encode and decode images to and from latent representations.
+ bert ([`LDMBertModel`]):
+ Text-encoder model based on [`~transformers.BERT`].
+ tokenizer ([`~transformers.BertTokenizer`]):
+ A `BertTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "bert->unet->vqvae"
+
+ tokenizer: CLIPTokenizer
+ image_feature_extractor: CLIPImageProcessor
+ text_encoder: CLIPTextModelWithProjection
+ image_encoder: CLIPVisionModelWithProjection
+ image_unet: UNet2DConditionModel
+ text_unet: UNetFlatConditionModel
+ vae: AutoencoderKL
+ scheduler: KarrasDiffusionSchedulers
+
+ _optional_components = ["text_unet"]
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ image_feature_extractor: CLIPImageProcessor,
+ text_encoder: CLIPTextModelWithProjection,
+ image_encoder: CLIPVisionModelWithProjection,
+ image_unet: UNet2DConditionModel,
+ text_unet: UNetFlatConditionModel,
+ vae: AutoencoderKL,
+ scheduler: KarrasDiffusionSchedulers,
+ ):
+ super().__init__()
+ self.register_modules(
+ tokenizer=tokenizer,
+ image_feature_extractor=image_feature_extractor,
+ text_encoder=text_encoder,
+ image_encoder=image_encoder,
+ image_unet=image_unet,
+ text_unet=text_unet,
+ vae=vae,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ if self.text_unet is not None and (
+ "dual_cross_attention" not in self.image_unet.config or not self.image_unet.config.dual_cross_attention
+ ):
+ # if loading from a universal checkpoint rather than a saved dual-guided pipeline
+ self._convert_to_dual_attention()
+
+ def remove_unused_weights(self):
+ self.register_modules(text_unet=None)
+
+ def _convert_to_dual_attention(self):
+ """
+ Replace image_unet's `Transformer2DModel` blocks with `DualTransformer2DModel` that contains transformer blocks
+ from both `image_unet` and `text_unet`
+ """
+ for name, module in self.image_unet.named_modules():
+ if isinstance(module, Transformer2DModel):
+ parent_name, index = name.rsplit(".", 1)
+ index = int(index)
+
+ image_transformer = self.image_unet.get_submodule(parent_name)[index]
+ text_transformer = self.text_unet.get_submodule(parent_name)[index]
+
+ config = image_transformer.config
+ dual_transformer = DualTransformer2DModel(
+ num_attention_heads=config.num_attention_heads,
+ attention_head_dim=config.attention_head_dim,
+ in_channels=config.in_channels,
+ num_layers=config.num_layers,
+ dropout=config.dropout,
+ norm_num_groups=config.norm_num_groups,
+ cross_attention_dim=config.cross_attention_dim,
+ attention_bias=config.attention_bias,
+ sample_size=config.sample_size,
+ num_vector_embeds=config.num_vector_embeds,
+ activation_fn=config.activation_fn,
+ num_embeds_ada_norm=config.num_embeds_ada_norm,
+ )
+ dual_transformer.transformers[0] = image_transformer
+ dual_transformer.transformers[1] = text_transformer
+
+ self.image_unet.get_submodule(parent_name)[index] = dual_transformer
+ self.image_unet.register_to_config(dual_cross_attention=True)
+
+ def _revert_dual_attention(self):
+ """
+ Revert the image_unet `DualTransformer2DModel` blocks back to `Transformer2DModel` with image_unet weights Call
+ this function if you reuse `image_unet` in another pipeline, e.g. `VersatileDiffusionPipeline`
+ """
+ for name, module in self.image_unet.named_modules():
+ if isinstance(module, DualTransformer2DModel):
+ parent_name, index = name.rsplit(".", 1)
+ index = int(index)
+ self.image_unet.get_submodule(parent_name)[index] = module.transformers[0]
+
+ self.image_unet.register_to_config(dual_cross_attention=False)
+
+ def _encode_text_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ """
+
+ def normalize_embeddings(encoder_output):
+ embeds = self.text_encoder.text_projection(encoder_output.last_hidden_state)
+ embeds_pooled = encoder_output.text_embeds
+ embeds = embeds / torch.norm(embeds_pooled.unsqueeze(1), dim=-1, keepdim=True)
+ return embeds
+
+ batch_size = len(prompt)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids
+
+ if not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = normalize_embeddings(prompt_embeds)
+
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance:
+ uncond_tokens = [""] * batch_size
+ max_length = text_input_ids.shape[-1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = normalize_embeddings(negative_prompt_embeds)
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ def _encode_image_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ """
+
+ def normalize_embeddings(encoder_output):
+ embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state)
+ embeds = self.image_encoder.visual_projection(embeds)
+ embeds_pooled = embeds[:, 0:1]
+ embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True)
+ return embeds
+
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+ # get prompt text embeddings
+ image_input = self.image_feature_extractor(images=prompt, return_tensors="pt")
+ pixel_values = image_input.pixel_values.to(device).to(self.image_encoder.dtype)
+ image_embeddings = self.image_encoder(pixel_values)
+ image_embeddings = normalize_embeddings(image_embeddings)
+
+ # duplicate image embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = image_embeddings.shape
+ image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)
+ image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance:
+ uncond_images = [np.zeros((512, 512, 3)) + 0.5] * batch_size
+ uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt")
+ pixel_values = uncond_images.pixel_values.to(device).to(self.image_encoder.dtype)
+ negative_prompt_embeds = self.image_encoder(pixel_values)
+ negative_prompt_embeds = normalize_embeddings(negative_prompt_embeds)
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and conditional embeddings into a single batch
+ # to avoid doing two forward passes
+ image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings])
+
+ return image_embeddings
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(self, prompt, image, height, width, callback_steps):
+ if not isinstance(prompt, str) and not isinstance(prompt, PIL.Image.Image) and not isinstance(prompt, list):
+ raise ValueError(f"`prompt` has to be of type `str` `PIL.Image` or `list` but is {type(prompt)}")
+ if not isinstance(image, str) and not isinstance(image, PIL.Image.Image) and not isinstance(image, list):
+ raise ValueError(f"`image` has to be of type `str` `PIL.Image` or `list` but is {type(image)}")
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def set_transformer_params(self, mix_ratio: float = 0.5, condition_types: Tuple = ("text", "image")):
+ for name, module in self.image_unet.named_modules():
+ if isinstance(module, DualTransformer2DModel):
+ module.mix_ratio = mix_ratio
+
+ for i, type in enumerate(condition_types):
+ if type == "text":
+ module.condition_lengths[i] = self.text_encoder.config.max_position_embeddings
+ module.transformer_index_for_condition[i] = 1 # use the second (text) transformer
+ else:
+ module.condition_lengths[i] = 257
+ module.transformer_index_for_condition[i] = 0 # use the first (image) transformer
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[PIL.Image.Image, List[PIL.Image.Image]],
+ image: Union[str, List[str]],
+ text_to_image_strength: float = 0.5,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation.
+ height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import VersatileDiffusionDualGuidedPipeline
+ >>> import torch
+ >>> import requests
+ >>> from io import BytesIO
+ >>> from PIL import Image
+
+ >>> # let's download an initial image
+ >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg"
+
+ >>> response = requests.get(url)
+ >>> image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> text = "a red car in the sun"
+
+ >>> pipe = VersatileDiffusionDualGuidedPipeline.from_pretrained(
+ ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16
+ ... )
+ >>> pipe.remove_unused_weights()
+ >>> pipe = pipe.to("cuda")
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(0)
+ >>> text_to_image_strength = 0.75
+
+ >>> image = pipe(
+ ... prompt=text, image=image, text_to_image_strength=text_to_image_strength, generator=generator
+ ... ).images[0]
+ >>> image.save("./car_variation.png")
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+ # 0. Default height and width to unet
+ height = height or self.image_unet.config.sample_size * self.vae_scale_factor
+ width = width or self.image_unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(prompt, image, height, width, callback_steps)
+
+ # 2. Define call parameters
+ prompt = [prompt] if not isinstance(prompt, list) else prompt
+ image = [image] if not isinstance(image, list) else image
+ batch_size = len(prompt)
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompts
+ prompt_embeds = self._encode_text_prompt(prompt, device, num_images_per_prompt, do_classifier_free_guidance)
+ image_embeddings = self._encode_image_prompt(image, device, num_images_per_prompt, do_classifier_free_guidance)
+ dual_prompt_embeddings = torch.cat([prompt_embeds, image_embeddings], dim=1)
+ prompt_types = ("text", "image")
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.image_unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ dual_prompt_embeddings.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Combine the attention blocks of the image and text UNets
+ self.set_transformer_params(text_to_image_strength, prompt_types)
+
+ # 8. Denoising loop
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=dual_prompt_embeddings).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = latents
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py
new file mode 100644
index 0000000000000000000000000000000000000000..345c15f18d893bfd0eaddc1c7d3578c0ab40ddf2
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_image_variation.py
@@ -0,0 +1,397 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.utils.checkpoint
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+
+from ....image_processor import VaeImageProcessor
+from ....models import AutoencoderKL, UNet2DConditionModel
+from ....schedulers import KarrasDiffusionSchedulers
+from ....utils import deprecate, logging
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class VersatileDiffusionImageVariationPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for image variation using Versatile Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ vqvae ([`VQModel`]):
+ Vector-quantized (VQ) model to encode and decode images to and from latent representations.
+ bert ([`LDMBertModel`]):
+ Text-encoder model based on [`~transformers.BERT`].
+ tokenizer ([`~transformers.BertTokenizer`]):
+ A `BertTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "bert->unet->vqvae"
+
+ image_feature_extractor: CLIPImageProcessor
+ image_encoder: CLIPVisionModelWithProjection
+ image_unet: UNet2DConditionModel
+ vae: AutoencoderKL
+ scheduler: KarrasDiffusionSchedulers
+
+ def __init__(
+ self,
+ image_feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection,
+ image_unet: UNet2DConditionModel,
+ vae: AutoencoderKL,
+ scheduler: KarrasDiffusionSchedulers,
+ ):
+ super().__init__()
+ self.register_modules(
+ image_feature_extractor=image_feature_extractor,
+ image_encoder=image_encoder,
+ image_unet=image_unet,
+ vae=vae,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ """
+
+ def normalize_embeddings(encoder_output):
+ embeds = self.image_encoder.vision_model.post_layernorm(encoder_output.last_hidden_state)
+ embeds = self.image_encoder.visual_projection(embeds)
+ embeds_pooled = embeds[:, 0:1]
+ embeds = embeds / torch.norm(embeds_pooled, dim=-1, keepdim=True)
+ return embeds
+
+ if isinstance(prompt, torch.Tensor) and len(prompt.shape) == 4:
+ prompt = list(prompt)
+
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+ # get prompt text embeddings
+ image_input = self.image_feature_extractor(images=prompt, return_tensors="pt")
+ pixel_values = image_input.pixel_values.to(device).to(self.image_encoder.dtype)
+ image_embeddings = self.image_encoder(pixel_values)
+ image_embeddings = normalize_embeddings(image_embeddings)
+
+ # duplicate image embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = image_embeddings.shape
+ image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)
+ image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance:
+ uncond_images: List[str]
+ if negative_prompt is None:
+ uncond_images = [np.zeros((512, 512, 3)) + 0.5] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, PIL.Image.Image):
+ uncond_images = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_images = negative_prompt
+
+ uncond_images = self.image_feature_extractor(images=uncond_images, return_tensors="pt")
+ pixel_values = uncond_images.pixel_values.to(device).to(self.image_encoder.dtype)
+ negative_prompt_embeds = self.image_encoder(pixel_values)
+ negative_prompt_embeds = normalize_embeddings(negative_prompt_embeds)
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and conditional embeddings into a single batch
+ # to avoid doing two forward passes
+ image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings])
+
+ return image_embeddings
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_image_variation.StableDiffusionImageVariationPipeline.check_inputs
+ def check_inputs(self, image, height, width, callback_steps):
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+ f" {type(image)}"
+ )
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.Tensor],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`PIL.Image.Image`, `List[PIL.Image.Image]` or `torch.Tensor`):
+ The image prompt or prompts to guide the image generation.
+ height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import VersatileDiffusionImageVariationPipeline
+ >>> import torch
+ >>> import requests
+ >>> from io import BytesIO
+ >>> from PIL import Image
+
+ >>> # let's download an initial image
+ >>> url = "https://huggingface.co/datasets/diffusers/images/resolve/main/benz.jpg"
+
+ >>> response = requests.get(url)
+ >>> image = Image.open(BytesIO(response.content)).convert("RGB")
+
+ >>> pipe = VersatileDiffusionImageVariationPipeline.from_pretrained(
+ ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(0)
+ >>> image = pipe(image, generator=generator).images[0]
+ >>> image.save("./car_variation.png")
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images.
+ """
+ # 0. Default height and width to unet
+ height = height or self.image_unet.config.sample_size * self.vae_scale_factor
+ width = width or self.image_unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(image, height, width, callback_steps)
+
+ # 2. Define call parameters
+ batch_size = 1 if isinstance(image, PIL.Image.Image) else len(image)
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ image_embeddings = self._encode_prompt(
+ image, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.image_unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ image_embeddings.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = latents
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py
new file mode 100644
index 0000000000000000000000000000000000000000..0b2518f7e2448f0135bdd524e19400aaa5a84a7d
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/versatile_diffusion/pipeline_versatile_diffusion_text_to_image.py
@@ -0,0 +1,475 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import torch
+import torch.utils.checkpoint
+from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer
+
+from ....image_processor import VaeImageProcessor
+from ....models import AutoencoderKL, Transformer2DModel, UNet2DConditionModel
+from ....schedulers import KarrasDiffusionSchedulers
+from ....utils import deprecate, logging
+from ....utils.torch_utils import randn_tensor
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .modeling_text_unet import UNetFlatConditionModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class VersatileDiffusionTextToImagePipeline(DiffusionPipeline):
+ r"""
+ Pipeline for text-to-image generation using Versatile Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ vqvae ([`VQModel`]):
+ Vector-quantized (VQ) model to encode and decode images to and from latent representations.
+ bert ([`LDMBertModel`]):
+ Text-encoder model based on [`~transformers.BERT`].
+ tokenizer ([`~transformers.BertTokenizer`]):
+ A `BertTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "bert->unet->vqvae"
+
+ tokenizer: CLIPTokenizer
+ image_feature_extractor: CLIPImageProcessor
+ text_encoder: CLIPTextModelWithProjection
+ image_unet: UNet2DConditionModel
+ text_unet: UNetFlatConditionModel
+ vae: AutoencoderKL
+ scheduler: KarrasDiffusionSchedulers
+
+ _optional_components = ["text_unet"]
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModelWithProjection,
+ image_unet: UNet2DConditionModel,
+ text_unet: UNetFlatConditionModel,
+ vae: AutoencoderKL,
+ scheduler: KarrasDiffusionSchedulers,
+ ):
+ super().__init__()
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ image_unet=image_unet,
+ text_unet=text_unet,
+ vae=vae,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ if self.text_unet is not None:
+ self._swap_unet_attention_blocks()
+
+ def _swap_unet_attention_blocks(self):
+ """
+ Swap the `Transformer2DModel` blocks between the image and text UNets
+ """
+ for name, module in self.image_unet.named_modules():
+ if isinstance(module, Transformer2DModel):
+ parent_name, index = name.rsplit(".", 1)
+ index = int(index)
+ self.image_unet.get_submodule(parent_name)[index], self.text_unet.get_submodule(parent_name)[index] = (
+ self.text_unet.get_submodule(parent_name)[index],
+ self.image_unet.get_submodule(parent_name)[index],
+ )
+
+ def remove_unused_weights(self):
+ self.register_modules(text_unet=None)
+
+ def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ """
+
+ def normalize_embeddings(encoder_output):
+ embeds = self.text_encoder.text_projection(encoder_output.last_hidden_state)
+ embeds_pooled = encoder_output.text_embeds
+ embeds = embeds / torch.norm(embeds_pooled.unsqueeze(1), dim=-1, keepdim=True)
+ return embeds
+
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids
+
+ if not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = normalize_embeddings(prompt_embeds)
+
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = text_input_ids.shape[-1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = normalize_embeddings(negative_prompt_embeds)
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation.
+ height (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.image_unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import VersatileDiffusionTextToImagePipeline
+ >>> import torch
+
+ >>> pipe = VersatileDiffusionTextToImagePipeline.from_pretrained(
+ ... "shi-labs/versatile-diffusion", torch_dtype=torch.float16
+ ... )
+ >>> pipe.remove_unused_weights()
+ >>> pipe = pipe.to("cuda")
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(0)
+ >>> image = pipe("an astronaut riding on a horse on mars", generator=generator).images[0]
+ >>> image.save("./astronaut.png")
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images.
+ """
+ # 0. Default height and width to unet
+ height = height or self.image_unet.config.sample_size * self.vae_scale_factor
+ width = width or self.image_unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(prompt, height, width, callback_steps)
+
+ # 2. Define call parameters
+ batch_size = 1 if isinstance(prompt, str) else len(prompt)
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.image_unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.image_unet(latent_model_input, t, encoder_hidden_states=prompt_embeds).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = latents
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/deprecated/vq_diffusion/__init__.py b/model/diffusers_c/pipelines/deprecated/vq_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..070903377c7188415af0417d4839d74a8a34dc01
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/vq_diffusion/__init__.py
@@ -0,0 +1,57 @@
+from typing import TYPE_CHECKING
+
+from ....utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ....utils.dummy_torch_and_transformers_objects import (
+ LearnedClassifierFreeSamplingEmbeddings,
+ VQDiffusionPipeline,
+ )
+
+ _dummy_objects.update(
+ {
+ "LearnedClassifierFreeSamplingEmbeddings": LearnedClassifierFreeSamplingEmbeddings,
+ "VQDiffusionPipeline": VQDiffusionPipeline,
+ }
+ )
+else:
+ _import_structure["pipeline_vq_diffusion"] = ["LearnedClassifierFreeSamplingEmbeddings", "VQDiffusionPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ....utils.dummy_torch_and_transformers_objects import (
+ LearnedClassifierFreeSamplingEmbeddings,
+ VQDiffusionPipeline,
+ )
+ else:
+ from .pipeline_vq_diffusion import LearnedClassifierFreeSamplingEmbeddings, VQDiffusionPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/deprecated/vq_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/vq_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..fa37bdbd5b412d451301f9482130dcfa50256113
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/vq_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/vq_diffusion/__pycache__/pipeline_vq_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/deprecated/vq_diffusion/__pycache__/pipeline_vq_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..62d6da7a6e782a9bb7b2772cc86f0f9d4bdb702b
Binary files /dev/null and b/model/diffusers_c/pipelines/deprecated/vq_diffusion/__pycache__/pipeline_vq_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/deprecated/vq_diffusion/pipeline_vq_diffusion.py b/model/diffusers_c/pipelines/deprecated/vq_diffusion/pipeline_vq_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..0c55d04e67de3f7a5e0cb53fea0baeda191cd268
--- /dev/null
+++ b/model/diffusers_c/pipelines/deprecated/vq_diffusion/pipeline_vq_diffusion.py
@@ -0,0 +1,325 @@
+# Copyright 2024 Microsoft and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ....configuration_utils import ConfigMixin, register_to_config
+from ....models import ModelMixin, Transformer2DModel, VQModel
+from ....schedulers import VQDiffusionScheduler
+from ....utils import logging
+from ...pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class LearnedClassifierFreeSamplingEmbeddings(ModelMixin, ConfigMixin):
+ """
+ Utility class for storing learned text embeddings for classifier free sampling
+ """
+
+ @register_to_config
+ def __init__(self, learnable: bool, hidden_size: Optional[int] = None, length: Optional[int] = None):
+ super().__init__()
+
+ self.learnable = learnable
+
+ if self.learnable:
+ assert hidden_size is not None, "learnable=True requires `hidden_size` to be set"
+ assert length is not None, "learnable=True requires `length` to be set"
+
+ embeddings = torch.zeros(length, hidden_size)
+ else:
+ embeddings = None
+
+ self.embeddings = torch.nn.Parameter(embeddings)
+
+
+class VQDiffusionPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for text-to-image generation using VQ Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vqvae ([`VQModel`]):
+ Vector Quantized Variational Auto-Encoder (VAE) model to encode and decode images to and from latent
+ representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ transformer ([`Transformer2DModel`]):
+ A conditional `Transformer2DModel` to denoise the encoded image latents.
+ scheduler ([`VQDiffusionScheduler`]):
+ A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+ """
+
+ vqvae: VQModel
+ text_encoder: CLIPTextModel
+ tokenizer: CLIPTokenizer
+ transformer: Transformer2DModel
+ learned_classifier_free_sampling_embeddings: LearnedClassifierFreeSamplingEmbeddings
+ scheduler: VQDiffusionScheduler
+
+ def __init__(
+ self,
+ vqvae: VQModel,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ transformer: Transformer2DModel,
+ scheduler: VQDiffusionScheduler,
+ learned_classifier_free_sampling_embeddings: LearnedClassifierFreeSamplingEmbeddings,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vqvae=vqvae,
+ transformer=transformer,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ scheduler=scheduler,
+ learned_classifier_free_sampling_embeddings=learned_classifier_free_sampling_embeddings,
+ )
+
+ def _encode_prompt(self, prompt, num_images_per_prompt, do_classifier_free_guidance):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+
+ if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+ removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+ prompt_embeds = self.text_encoder(text_input_ids.to(self.device))[0]
+
+ # NOTE: This additional step of normalizing the text embeddings is from VQ-Diffusion.
+ # While CLIP does normalize the pooled output of the text transformer when combining
+ # the image and text embeddings, CLIP does not directly normalize the last hidden state.
+ #
+ # CLIP normalizing the pooled output.
+ # https://github.com/huggingface/transformers/blob/d92e22d1f28324f513f3080e5c47c071a3916721/src/transformers/models/clip/modeling_clip.py#L1052-L1053
+ prompt_embeds = prompt_embeds / prompt_embeds.norm(dim=-1, keepdim=True)
+
+ # duplicate text embeddings for each generation per prompt
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ if self.learned_classifier_free_sampling_embeddings.learnable:
+ negative_prompt_embeds = self.learned_classifier_free_sampling_embeddings.embeddings
+ negative_prompt_embeds = negative_prompt_embeds.unsqueeze(0).repeat(batch_size, 1, 1)
+ else:
+ uncond_tokens = [""] * batch_size
+
+ max_length = text_input_ids.shape[-1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ negative_prompt_embeds = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+ # See comment for normalizing text embeddings
+ negative_prompt_embeds = negative_prompt_embeds / negative_prompt_embeds.norm(dim=-1, keepdim=True)
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ num_inference_steps: int = 100,
+ guidance_scale: float = 5.0,
+ truncation_rate: float = 1.0,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ) -> Union[ImagePipelineOutput, Tuple]:
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ truncation_rate (`float`, *optional*, defaults to 1.0 (equivalent to no truncation)):
+ Used to "truncate" the predicted classes for x_0 such that the cumulative probability for a pixel is at
+ most `truncation_rate`. The lowest probabilities that would increase the cumulative probability above
+ `truncation_rate` are set to zero.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor` of shape (batch), *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Must be valid embedding indices.If not provided, a latents tensor will be generated of
+ completely masked latent pixels.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+ if isinstance(prompt, str):
+ batch_size = 1
+ elif isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ batch_size = batch_size * num_images_per_prompt
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ prompt_embeds = self._encode_prompt(prompt, num_images_per_prompt, do_classifier_free_guidance)
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ # get the initial completely masked latents unless the user supplied it
+
+ latents_shape = (batch_size, self.transformer.num_latent_pixels)
+ if latents is None:
+ mask_class = self.transformer.num_vector_embeds - 1
+ latents = torch.full(latents_shape, mask_class).to(self.device)
+ else:
+ if latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+ if (latents < 0).any() or (latents >= self.transformer.num_vector_embeds).any():
+ raise ValueError(
+ "Unexpected latents value(s). All latents be valid embedding indices i.e. in the range 0,"
+ f" {self.transformer.num_vector_embeds - 1} (inclusive)."
+ )
+ latents = latents.to(self.device)
+
+ # set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=self.device)
+
+ timesteps_tensor = self.scheduler.timesteps.to(self.device)
+
+ sample = latents
+
+ for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+ # expand the sample if we are doing classifier free guidance
+ latent_model_input = torch.cat([sample] * 2) if do_classifier_free_guidance else sample
+
+ # predict the un-noised image
+ # model_output == `log_p_x_0`
+ model_output = self.transformer(latent_model_input, encoder_hidden_states=prompt_embeds, timestep=t).sample
+
+ if do_classifier_free_guidance:
+ model_output_uncond, model_output_text = model_output.chunk(2)
+ model_output = model_output_uncond + guidance_scale * (model_output_text - model_output_uncond)
+ model_output -= torch.logsumexp(model_output, dim=1, keepdim=True)
+
+ model_output = self.truncate(model_output, truncation_rate)
+
+ # remove `log(0)`'s (`-inf`s)
+ model_output = model_output.clamp(-70)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ sample = self.scheduler.step(model_output, timestep=t, sample=sample, generator=generator).prev_sample
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, sample)
+
+ embedding_channels = self.vqvae.config.vq_embed_dim
+ embeddings_shape = (batch_size, self.transformer.height, self.transformer.width, embedding_channels)
+ embeddings = self.vqvae.quantize.get_codebook_entry(sample, shape=embeddings_shape)
+ image = self.vqvae.decode(embeddings, force_not_quantize=True).sample
+
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
+
+ def truncate(self, log_p_x_0: torch.FloatTensor, truncation_rate: float) -> torch.FloatTensor:
+ """
+ Truncates `log_p_x_0` such that for each column vector, the total cumulative probability is `truncation_rate`
+ The lowest probabilities that would increase the cumulative probability above `truncation_rate` are set to
+ zero.
+ """
+ sorted_log_p_x_0, indices = torch.sort(log_p_x_0, 1, descending=True)
+ sorted_p_x_0 = torch.exp(sorted_log_p_x_0)
+ keep_mask = sorted_p_x_0.cumsum(dim=1) < truncation_rate
+
+ # Ensure that at least the largest probability is not zeroed out
+ all_true = torch.full_like(keep_mask[:, 0:1, :], True)
+ keep_mask = torch.cat((all_true, keep_mask), dim=1)
+ keep_mask = keep_mask[:, :-1, :]
+
+ keep_mask = keep_mask.gather(1, indices.argsort(1))
+
+ rv = log_p_x_0.clone()
+
+ rv[~keep_mask] = -torch.inf # -inf = log(0)
+
+ return rv
diff --git a/model/diffusers_c/pipelines/dit/__init__.py b/model/diffusers_c/pipelines/dit/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..fe2a94f3cba77d867f97111a41895918842adc27
--- /dev/null
+++ b/model/diffusers_c/pipelines/dit/__init__.py
@@ -0,0 +1,19 @@
+from typing import TYPE_CHECKING
+
+from ...utils import DIFFUSERS_SLOW_IMPORT, _LazyModule
+
+
+_import_structure = {"pipeline_dit": ["DiTPipeline"]}
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from .pipeline_dit import DiTPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
diff --git a/model/diffusers_c/pipelines/dit/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/dit/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..52bc361b4c9beb321060ecd06885ee715af8d961
Binary files /dev/null and b/model/diffusers_c/pipelines/dit/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/dit/__pycache__/pipeline_dit.cpython-39.pyc b/model/diffusers_c/pipelines/dit/__pycache__/pipeline_dit.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0fe391f4b5b41c6a837737c1de42affc62712b55
Binary files /dev/null and b/model/diffusers_c/pipelines/dit/__pycache__/pipeline_dit.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/dit/pipeline_dit.py b/model/diffusers_c/pipelines/dit/pipeline_dit.py
new file mode 100644
index 0000000000000000000000000000000000000000..289ea496028d075c12797bcb921e2fb911c70699
--- /dev/null
+++ b/model/diffusers_c/pipelines/dit/pipeline_dit.py
@@ -0,0 +1,233 @@
+# Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
+# William Peebles and Saining Xie
+#
+# Copyright (c) 2021 OpenAI
+# MIT License
+#
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Dict, List, Optional, Tuple, Union
+
+import torch
+
+from ...models import AutoencoderKL, Transformer2DModel
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+class DiTPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for image generation based on a Transformer backbone instead of a UNet.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ transformer ([`Transformer2DModel`]):
+ A class conditioned `Transformer2DModel` to denoise the encoded image latents.
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ scheduler ([`DDIMScheduler`]):
+ A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+ """
+
+ model_cpu_offload_seq = "transformer->vae"
+
+ def __init__(
+ self,
+ transformer: Transformer2DModel,
+ vae: AutoencoderKL,
+ scheduler: KarrasDiffusionSchedulers,
+ id2label: Optional[Dict[int, str]] = None,
+ ):
+ super().__init__()
+ self.register_modules(transformer=transformer, vae=vae, scheduler=scheduler)
+
+ # create a imagenet -> id dictionary for easier use
+ self.labels = {}
+ if id2label is not None:
+ for key, value in id2label.items():
+ for label in value.split(","):
+ self.labels[label.lstrip().rstrip()] = int(key)
+ self.labels = dict(sorted(self.labels.items()))
+
+ def get_label_ids(self, label: Union[str, List[str]]) -> List[int]:
+ r"""
+
+ Map label strings from ImageNet to corresponding class ids.
+
+ Parameters:
+ label (`str` or `dict` of `str`):
+ Label strings to be mapped to class ids.
+
+ Returns:
+ `list` of `int`:
+ Class ids to be processed by pipeline.
+ """
+
+ if not isinstance(label, list):
+ label = list(label)
+
+ for l in label:
+ if l not in self.labels:
+ raise ValueError(
+ f"{l} does not exist. Please make sure to select one of the following labels: \n {self.labels}."
+ )
+
+ return [self.labels[l] for l in label]
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ class_labels: List[int],
+ guidance_scale: float = 4.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ num_inference_steps: int = 50,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ) -> Union[ImagePipelineOutput, Tuple]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ class_labels (List[int]):
+ List of ImageNet class labels for the images to be generated.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ num_inference_steps (`int`, *optional*, defaults to 250):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import DiTPipeline, DPMSolverMultistepScheduler
+ >>> import torch
+
+ >>> pipe = DiTPipeline.from_pretrained("facebook/DiT-XL-2-256", torch_dtype=torch.float16)
+ >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+ >>> pipe = pipe.to("cuda")
+
+ >>> # pick words from Imagenet class labels
+ >>> pipe.labels # to print all available words
+
+ >>> # pick words that exist in ImageNet
+ >>> words = ["white shark", "umbrella"]
+
+ >>> class_ids = pipe.get_label_ids(words)
+
+ >>> generator = torch.manual_seed(33)
+ >>> output = pipe(class_labels=class_ids, num_inference_steps=25, generator=generator)
+
+ >>> image = output.images[0] # label 'white shark'
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images
+ """
+
+ batch_size = len(class_labels)
+ latent_size = self.transformer.config.sample_size
+ latent_channels = self.transformer.config.in_channels
+
+ latents = randn_tensor(
+ shape=(batch_size, latent_channels, latent_size, latent_size),
+ generator=generator,
+ device=self._execution_device,
+ dtype=self.transformer.dtype,
+ )
+ latent_model_input = torch.cat([latents] * 2) if guidance_scale > 1 else latents
+
+ class_labels = torch.tensor(class_labels, device=self._execution_device).reshape(-1)
+ class_null = torch.tensor([1000] * batch_size, device=self._execution_device)
+ class_labels_input = torch.cat([class_labels, class_null], 0) if guidance_scale > 1 else class_labels
+
+ # set step values
+ self.scheduler.set_timesteps(num_inference_steps)
+ for t in self.progress_bar(self.scheduler.timesteps):
+ if guidance_scale > 1:
+ half = latent_model_input[: len(latent_model_input) // 2]
+ latent_model_input = torch.cat([half, half], dim=0)
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ timesteps = t
+ if not torch.is_tensor(timesteps):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = latent_model_input.device.type == "mps"
+ if isinstance(timesteps, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=latent_model_input.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(latent_model_input.device)
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(latent_model_input.shape[0])
+ # predict noise model_output
+ noise_pred = self.transformer(
+ latent_model_input, timestep=timesteps, class_labels=class_labels_input
+ ).sample
+
+ # perform guidance
+ if guidance_scale > 1:
+ eps, rest = noise_pred[:, :latent_channels], noise_pred[:, latent_channels:]
+ cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
+
+ half_eps = uncond_eps + guidance_scale * (cond_eps - uncond_eps)
+ eps = torch.cat([half_eps, half_eps], dim=0)
+
+ noise_pred = torch.cat([eps, rest], dim=1)
+
+ # learned sigma
+ if self.transformer.config.out_channels // 2 == latent_channels:
+ model_output, _ = torch.split(noise_pred, latent_channels, dim=1)
+ else:
+ model_output = noise_pred
+
+ # compute previous image: x_t -> x_t-1
+ latent_model_input = self.scheduler.step(model_output, t, latent_model_input).prev_sample
+
+ if guidance_scale > 1:
+ latents, _ = latent_model_input.chunk(2, dim=0)
+ else:
+ latents = latent_model_input
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ samples = self.vae.decode(latents).sample
+
+ samples = (samples / 2 + 0.5).clamp(0, 1)
+
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ samples = samples.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ samples = self.numpy_to_pil(samples)
+
+ if not return_dict:
+ return (samples,)
+
+ return ImagePipelineOutput(images=samples)
diff --git a/model/diffusers_c/pipelines/free_init_utils.py b/model/diffusers_c/pipelines/free_init_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..50c28cc69f4469fa88cbf6277c44550714ee73d9
--- /dev/null
+++ b/model/diffusers_c/pipelines/free_init_utils.py
@@ -0,0 +1,184 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Tuple, Union
+
+import torch
+import torch.fft as fft
+
+from ..utils.torch_utils import randn_tensor
+
+
+class FreeInitMixin:
+ r"""Mixin class for FreeInit."""
+
+ def enable_free_init(
+ self,
+ num_iters: int = 3,
+ use_fast_sampling: bool = False,
+ method: str = "butterworth",
+ order: int = 4,
+ spatial_stop_frequency: float = 0.25,
+ temporal_stop_frequency: float = 0.25,
+ ):
+ """Enables the FreeInit mechanism as in https://arxiv.org/abs/2312.07537.
+
+ This implementation has been adapted from the [official repository](https://github.com/TianxingWu/FreeInit).
+
+ Args:
+ num_iters (`int`, *optional*, defaults to `3`):
+ Number of FreeInit noise re-initialization iterations.
+ use_fast_sampling (`bool`, *optional*, defaults to `False`):
+ Whether or not to speedup sampling procedure at the cost of probably lower quality results. Enables
+ the "Coarse-to-Fine Sampling" strategy, as mentioned in the paper, if set to `True`.
+ method (`str`, *optional*, defaults to `butterworth`):
+ Must be one of `butterworth`, `ideal` or `gaussian` to use as the filtering method for the
+ FreeInit low pass filter.
+ order (`int`, *optional*, defaults to `4`):
+ Order of the filter used in `butterworth` method. Larger values lead to `ideal` method behaviour
+ whereas lower values lead to `gaussian` method behaviour.
+ spatial_stop_frequency (`float`, *optional*, defaults to `0.25`):
+ Normalized stop frequency for spatial dimensions. Must be between 0 to 1. Referred to as `d_s` in
+ the original implementation.
+ temporal_stop_frequency (`float`, *optional*, defaults to `0.25`):
+ Normalized stop frequency for temporal dimensions. Must be between 0 to 1. Referred to as `d_t` in
+ the original implementation.
+ """
+ self._free_init_num_iters = num_iters
+ self._free_init_use_fast_sampling = use_fast_sampling
+ self._free_init_method = method
+ self._free_init_order = order
+ self._free_init_spatial_stop_frequency = spatial_stop_frequency
+ self._free_init_temporal_stop_frequency = temporal_stop_frequency
+
+ def disable_free_init(self):
+ """Disables the FreeInit mechanism if enabled."""
+ self._free_init_num_iters = None
+
+ @property
+ def free_init_enabled(self):
+ return hasattr(self, "_free_init_num_iters") and self._free_init_num_iters is not None
+
+ def _get_free_init_freq_filter(
+ self,
+ shape: Tuple[int, ...],
+ device: Union[str, torch.dtype],
+ filter_type: str,
+ order: float,
+ spatial_stop_frequency: float,
+ temporal_stop_frequency: float,
+ ) -> torch.Tensor:
+ r"""Returns the FreeInit filter based on filter type and other input conditions."""
+
+ time, height, width = shape[-3], shape[-2], shape[-1]
+ mask = torch.zeros(shape)
+
+ if spatial_stop_frequency == 0 or temporal_stop_frequency == 0:
+ return mask
+
+ if filter_type == "butterworth":
+
+ def retrieve_mask(x):
+ return 1 / (1 + (x / spatial_stop_frequency**2) ** order)
+ elif filter_type == "gaussian":
+
+ def retrieve_mask(x):
+ return math.exp(-1 / (2 * spatial_stop_frequency**2) * x)
+ elif filter_type == "ideal":
+
+ def retrieve_mask(x):
+ return 1 if x <= spatial_stop_frequency * 2 else 0
+ else:
+ raise NotImplementedError("`filter_type` must be one of gaussian, butterworth or ideal")
+
+ for t in range(time):
+ for h in range(height):
+ for w in range(width):
+ d_square = (
+ ((spatial_stop_frequency / temporal_stop_frequency) * (2 * t / time - 1)) ** 2
+ + (2 * h / height - 1) ** 2
+ + (2 * w / width - 1) ** 2
+ )
+ mask[..., t, h, w] = retrieve_mask(d_square)
+
+ return mask.to(device)
+
+ def _apply_freq_filter(self, x: torch.Tensor, noise: torch.Tensor, low_pass_filter: torch.Tensor) -> torch.Tensor:
+ r"""Noise reinitialization."""
+ # FFT
+ x_freq = fft.fftn(x, dim=(-3, -2, -1))
+ x_freq = fft.fftshift(x_freq, dim=(-3, -2, -1))
+ noise_freq = fft.fftn(noise, dim=(-3, -2, -1))
+ noise_freq = fft.fftshift(noise_freq, dim=(-3, -2, -1))
+
+ # frequency mix
+ high_pass_filter = 1 - low_pass_filter
+ x_freq_low = x_freq * low_pass_filter
+ noise_freq_high = noise_freq * high_pass_filter
+ x_freq_mixed = x_freq_low + noise_freq_high # mix in freq domain
+
+ # IFFT
+ x_freq_mixed = fft.ifftshift(x_freq_mixed, dim=(-3, -2, -1))
+ x_mixed = fft.ifftn(x_freq_mixed, dim=(-3, -2, -1)).real
+
+ return x_mixed
+
+ def _apply_free_init(
+ self,
+ latents: torch.Tensor,
+ free_init_iteration: int,
+ num_inference_steps: int,
+ device: torch.device,
+ dtype: torch.dtype,
+ generator: torch.Generator,
+ ):
+ if free_init_iteration == 0:
+ self._free_init_initial_noise = latents.detach().clone()
+ return latents, self.scheduler.timesteps
+
+ latent_shape = latents.shape
+
+ free_init_filter_shape = (1, *latent_shape[1:])
+ free_init_freq_filter = self._get_free_init_freq_filter(
+ shape=free_init_filter_shape,
+ device=device,
+ filter_type=self._free_init_method,
+ order=self._free_init_order,
+ spatial_stop_frequency=self._free_init_spatial_stop_frequency,
+ temporal_stop_frequency=self._free_init_temporal_stop_frequency,
+ )
+
+ current_diffuse_timestep = self.scheduler.config.num_train_timesteps - 1
+ diffuse_timesteps = torch.full((latent_shape[0],), current_diffuse_timestep).long()
+
+ z_t = self.scheduler.add_noise(
+ original_samples=latents, noise=self._free_init_initial_noise, timesteps=diffuse_timesteps.to(device)
+ ).to(dtype=torch.float32)
+
+ z_rand = randn_tensor(
+ shape=latent_shape,
+ generator=generator,
+ device=device,
+ dtype=torch.float32,
+ )
+ latents = self._apply_freq_filter(z_t, z_rand, low_pass_filter=free_init_freq_filter)
+ latents = latents.to(dtype)
+
+ # Coarse-to-Fine Sampling for faster inference (can lead to lower quality)
+ if self._free_init_use_fast_sampling:
+ num_inference_steps = int(num_inference_steps / self._free_init_num_iters * (free_init_iteration + 1))
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+ return latents, self.scheduler.timesteps
diff --git a/model/diffusers_c/pipelines/i2vgen_xl/__init__.py b/model/diffusers_c/pipelines/i2vgen_xl/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..b24a7e4cee7fb843b9424469a05f511adfa758de
--- /dev/null
+++ b/model/diffusers_c/pipelines/i2vgen_xl/__init__.py
@@ -0,0 +1,46 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_i2vgen_xl"] = ["I2VGenXLPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import * # noqa F403
+ else:
+ from .pipeline_i2vgen_xl import I2VGenXLPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/i2vgen_xl/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/i2vgen_xl/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..14b8ce0c7d4515453c3a971977b2539ea67c07bb
Binary files /dev/null and b/model/diffusers_c/pipelines/i2vgen_xl/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/i2vgen_xl/__pycache__/pipeline_i2vgen_xl.cpython-39.pyc b/model/diffusers_c/pipelines/i2vgen_xl/__pycache__/pipeline_i2vgen_xl.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f823c27b7505bad7c429ca5d63efd2b573111604
Binary files /dev/null and b/model/diffusers_c/pipelines/i2vgen_xl/__pycache__/pipeline_i2vgen_xl.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/i2vgen_xl/pipeline_i2vgen_xl.py b/model/diffusers_c/pipelines/i2vgen_xl/pipeline_i2vgen_xl.py
new file mode 100644
index 0000000000000000000000000000000000000000..cb6f3e30090445670457ce36ed0b415fa2c62c39
--- /dev/null
+++ b/model/diffusers_c/pipelines/i2vgen_xl/pipeline_i2vgen_xl.py
@@ -0,0 +1,798 @@
+# Copyright 2024 Alibaba DAMO-VILAB and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...models import AutoencoderKL
+from ...models.unets.unet_i2vgen_xl import I2VGenXLUNet
+from ...schedulers import DDIMScheduler
+from ...utils import (
+ BaseOutput,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import I2VGenXLPipeline
+ >>> from diffusers.utils import export_to_gif, load_image
+
+ >>> pipeline = I2VGenXLPipeline.from_pretrained("ali-vilab/i2vgen-xl", torch_dtype=torch.float16, variant="fp16")
+ >>> pipeline.enable_model_cpu_offload()
+
+ >>> image_url = "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/i2vgen_xl_images/img_0009.png"
+ >>> image = load_image(image_url).convert("RGB")
+
+ >>> prompt = "Papers were floating in the air on a table in the library"
+ >>> negative_prompt = "Distorted, discontinuous, Ugly, blurry, low resolution, motionless, static, disfigured, disconnected limbs, Ugly faces, incomplete arms"
+ >>> generator = torch.manual_seed(8888)
+
+ >>> frames = pipeline(
+ ... prompt=prompt,
+ ... image=image,
+ ... num_inference_steps=50,
+ ... negative_prompt=negative_prompt,
+ ... guidance_scale=9.0,
+ ... generator=generator
+ ... ).frames[0]
+ >>> video_path = export_to_gif(frames, "i2v.gif")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor: "VaeImageProcessor", output_type: str = "np"):
+ batch_size, channels, num_frames, height, width = video.shape
+ outputs = []
+ for batch_idx in range(batch_size):
+ batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+ batch_output = processor.postprocess(batch_vid, output_type)
+
+ outputs.append(batch_output)
+
+ if output_type == "np":
+ outputs = np.stack(outputs)
+
+ elif output_type == "pt":
+ outputs = torch.stack(outputs)
+
+ elif not output_type == "pil":
+ raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+ return outputs
+
+
+@dataclass
+class I2VGenXLPipelineOutput(BaseOutput):
+ r"""
+ Output class for image-to-video pipeline.
+
+ Args:
+ frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+ List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing denoised
+ PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+ `(batch_size, num_frames, channels, height, width)`
+ """
+
+ frames: Union[torch.Tensor, np.ndarray, List[List[PIL.Image.Image]]]
+
+
+class I2VGenXLPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+):
+ r"""
+ Pipeline for image-to-video generation as proposed in [I2VGenXL](https://i2vgen-xl.github.io/).
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`] to tokenize text.
+ unet ([`I2VGenXLUNet`]):
+ A [`I2VGenXLUNet`] to denoise the encoded video latents.
+ scheduler ([`DDIMScheduler`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ image_encoder: CLIPVisionModelWithProjection,
+ feature_extractor: CLIPImageProcessor,
+ unet: I2VGenXLUNet,
+ scheduler: DDIMScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ image_encoder=image_encoder,
+ feature_extractor=feature_extractor,
+ unet=unet,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ # `do_resize=False` as we do custom resizing.
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_resize=False)
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_videos_per_prompt,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_videos_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if self.do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ # Apply clip_skip to negative prompt embeds
+ if clip_skip is None:
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+ else:
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ negative_prompt_embeds = negative_prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ negative_prompt_embeds = self.text_encoder.text_model.final_layer_norm(negative_prompt_embeds)
+
+ if self.do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_videos_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_videos_per_prompt, seq_len, -1)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def _encode_image(self, image, device, num_videos_per_prompt):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.image_processor.pil_to_numpy(image)
+ image = self.image_processor.numpy_to_pt(image)
+
+ # Normalize the image with CLIP training stats.
+ image = self.feature_extractor(
+ images=image,
+ do_normalize=True,
+ do_center_crop=False,
+ do_resize=False,
+ do_rescale=False,
+ return_tensors="pt",
+ ).pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ image_embeddings = self.image_encoder(image).image_embeds
+ image_embeddings = image_embeddings.unsqueeze(1)
+
+ # duplicate image embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = image_embeddings.shape
+ image_embeddings = image_embeddings.repeat(1, num_videos_per_prompt, 1)
+ image_embeddings = image_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+ if self.do_classifier_free_guidance:
+ negative_image_embeddings = torch.zeros_like(image_embeddings)
+ image_embeddings = torch.cat([negative_image_embeddings, image_embeddings])
+
+ return image_embeddings
+
+ def decode_latents(self, latents, decode_chunk_size=None):
+ latents = 1 / self.vae.config.scaling_factor * latents
+
+ batch_size, channels, num_frames, height, width = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+ if decode_chunk_size is not None:
+ frames = []
+ for i in range(0, latents.shape[0], decode_chunk_size):
+ frame = self.vae.decode(latents[i : i + decode_chunk_size]).sample
+ frames.append(frame)
+ image = torch.cat(frames, dim=0)
+ else:
+ image = self.vae.decode(latents).sample
+
+ decode_shape = (batch_size, num_frames, -1) + image.shape[2:]
+ video = image[None, :].reshape(decode_shape).permute(0, 2, 1, 3, 4)
+
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ video = video.float()
+ return video
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ height,
+ width,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+ f" {type(image)}"
+ )
+
+ def prepare_image_latents(
+ self,
+ image,
+ device,
+ num_frames,
+ num_videos_per_prompt,
+ ):
+ image = image.to(device=device)
+ image_latents = self.vae.encode(image).latent_dist.sample()
+ image_latents = image_latents * self.vae.config.scaling_factor
+
+ # Add frames dimension to image latents
+ image_latents = image_latents.unsqueeze(2)
+
+ # Append a position mask for each subsequent frame
+ # after the intial image latent frame
+ frame_position_mask = []
+ for frame_idx in range(num_frames - 1):
+ scale = (frame_idx + 1) / (num_frames - 1)
+ frame_position_mask.append(torch.ones_like(image_latents[:, :, :1]) * scale)
+ if frame_position_mask:
+ frame_position_mask = torch.cat(frame_position_mask, dim=2)
+ image_latents = torch.cat([image_latents, frame_position_mask], dim=2)
+
+ # duplicate image_latents for each generation per prompt, using mps friendly method
+ image_latents = image_latents.repeat(num_videos_per_prompt, 1, 1, 1, 1)
+
+ if self.do_classifier_free_guidance:
+ image_latents = torch.cat([image_latents] * 2)
+
+ return image_latents
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
+ def prepare_latents(
+ self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+ ):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ num_frames,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ height: Optional[int] = 704,
+ width: Optional[int] = 1280,
+ target_fps: Optional[int] = 16,
+ num_frames: int = 16,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 9.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ eta: float = 0.0,
+ num_videos_per_prompt: Optional[int] = 1,
+ decode_chunk_size: Optional[int] = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = 1,
+ ):
+ r"""
+ The call function to the pipeline for image-to-video generation with [`I2VGenXLPipeline`].
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+ Image or images to guide image generation. If you provide a tensor, it needs to be compatible with
+ [`CLIPImageProcessor`](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json).
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ target_fps (`int`, *optional*):
+ Frames per second. The rate at which the generated images shall be exported to a video after generation. This is also used as a "micro-condition" while generation.
+ num_frames (`int`, *optional*):
+ The number of video frames to generate.
+ num_inference_steps (`int`, *optional*):
+ The number of denoising steps.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ eta (`float`, *optional*):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ num_videos_per_prompt (`int`, *optional*):
+ The number of images to generate per prompt.
+ decode_chunk_size (`int`, *optional*):
+ The number of frames to decode at a time. The higher the chunk size, the higher the temporal consistency
+ between frames, but also the higher the memory consumption. By default, the decoder will decode all frames at once
+ for maximal quality. Reduce `decode_chunk_size` to reduce memory usage.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+
+ Examples:
+
+ Returns:
+ [`pipelines.i2vgen_xl.pipeline_i2vgen_xl.I2VGenXLPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`pipelines.i2vgen_xl.pipeline_i2vgen_xl.I2VGenXLPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(prompt, image, height, width, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ self._guidance_scale = guidance_scale
+
+ # 3.1 Encode input text prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_videos_per_prompt,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 3.2 Encode image prompt
+ # 3.2.1 Image encodings.
+ # https://github.com/ali-vilab/i2vgen-xl/blob/2539c9262ff8a2a22fa9daecbfd13f0a2dbc32d0/tools/inferences/inference_i2vgen_entrance.py#L114
+ cropped_image = _center_crop_wide(image, (width, width))
+ cropped_image = _resize_bilinear(
+ cropped_image, (self.feature_extractor.crop_size["width"], self.feature_extractor.crop_size["height"])
+ )
+ image_embeddings = self._encode_image(cropped_image, device, num_videos_per_prompt)
+
+ # 3.2.2 Image latents.
+ resized_image = _center_crop_wide(image, (width, height))
+ image = self.image_processor.preprocess(resized_image).to(device=device, dtype=image_embeddings.dtype)
+ image_latents = self.prepare_image_latents(
+ image,
+ device=device,
+ num_frames=num_frames,
+ num_videos_per_prompt=num_videos_per_prompt,
+ )
+
+ # 3.3 Prepare additional conditions for the UNet.
+ if self.do_classifier_free_guidance:
+ fps_tensor = torch.tensor([target_fps, target_fps]).to(device)
+ else:
+ fps_tensor = torch.tensor([target_fps]).to(device)
+ fps_tensor = fps_tensor.repeat(batch_size * num_videos_per_prompt, 1).ravel()
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_videos_per_prompt,
+ num_channels_latents,
+ num_frames,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ fps=fps_tensor,
+ image_latents=image_latents,
+ image_embeddings=image_embeddings,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # reshape latents
+ batch_size, channel, frames, width, height = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channel, width, height)
+ noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(batch_size * frames, channel, width, height)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # reshape latents back
+ latents = latents[None, :].reshape(batch_size, frames, channel, width, height).permute(0, 2, 1, 3, 4)
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+
+ # 8. Post processing
+ if output_type == "latent":
+ video = latents
+ else:
+ video_tensor = self.decode_latents(latents, decode_chunk_size=decode_chunk_size)
+ video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)
+
+ # 9. Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (video,)
+
+ return I2VGenXLPipelineOutput(frames=video)
+
+
+# The following utilities are taken and adapted from
+# https://github.com/ali-vilab/i2vgen-xl/blob/main/utils/transforms.py.
+
+
+def _convert_pt_to_pil(image: Union[torch.Tensor, List[torch.Tensor]]):
+ if isinstance(image, list) and isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, 0)
+
+ if isinstance(image, torch.Tensor):
+ if image.ndim == 3:
+ image = image.unsqueeze(0)
+
+ image_numpy = VaeImageProcessor.pt_to_numpy(image)
+ image_pil = VaeImageProcessor.numpy_to_pil(image_numpy)
+ image = image_pil
+
+ return image
+
+
+def _resize_bilinear(
+ image: Union[torch.Tensor, List[torch.Tensor], PIL.Image.Image, List[PIL.Image.Image]], resolution: Tuple[int, int]
+):
+ # First convert the images to PIL in case they are float tensors (only relevant for tests now).
+ image = _convert_pt_to_pil(image)
+
+ if isinstance(image, list):
+ image = [u.resize(resolution, PIL.Image.BILINEAR) for u in image]
+ else:
+ image = image.resize(resolution, PIL.Image.BILINEAR)
+ return image
+
+
+def _center_crop_wide(
+ image: Union[torch.Tensor, List[torch.Tensor], PIL.Image.Image, List[PIL.Image.Image]], resolution: Tuple[int, int]
+):
+ # First convert the images to PIL in case they are float tensors (only relevant for tests now).
+ image = _convert_pt_to_pil(image)
+
+ if isinstance(image, list):
+ scale = min(image[0].size[0] / resolution[0], image[0].size[1] / resolution[1])
+ image = [u.resize((round(u.width // scale), round(u.height // scale)), resample=PIL.Image.BOX) for u in image]
+
+ # center crop
+ x1 = (image[0].width - resolution[0]) // 2
+ y1 = (image[0].height - resolution[1]) // 2
+ image = [u.crop((x1, y1, x1 + resolution[0], y1 + resolution[1])) for u in image]
+ return image
+ else:
+ scale = min(image.size[0] / resolution[0], image.size[1] / resolution[1])
+ image = image.resize((round(image.width // scale), round(image.height // scale)), resample=PIL.Image.BOX)
+ x1 = (image.width - resolution[0]) // 2
+ y1 = (image.height - resolution[1]) // 2
+ image = image.crop((x1, y1, x1 + resolution[0], y1 + resolution[1]))
+ return image
diff --git a/model/diffusers_c/pipelines/kandinsky/__init__.py b/model/diffusers_c/pipelines/kandinsky/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..606f7b378a79489bbcbaa87db2040bd4196bbd8a
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky/__init__.py
@@ -0,0 +1,66 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_kandinsky"] = ["KandinskyPipeline"]
+ _import_structure["pipeline_kandinsky_combined"] = [
+ "KandinskyCombinedPipeline",
+ "KandinskyImg2ImgCombinedPipeline",
+ "KandinskyInpaintCombinedPipeline",
+ ]
+ _import_structure["pipeline_kandinsky_img2img"] = ["KandinskyImg2ImgPipeline"]
+ _import_structure["pipeline_kandinsky_inpaint"] = ["KandinskyInpaintPipeline"]
+ _import_structure["pipeline_kandinsky_prior"] = ["KandinskyPriorPipeline", "KandinskyPriorPipelineOutput"]
+ _import_structure["text_encoder"] = ["MultilingualCLIP"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+
+ else:
+ from .pipeline_kandinsky import KandinskyPipeline
+ from .pipeline_kandinsky_combined import (
+ KandinskyCombinedPipeline,
+ KandinskyImg2ImgCombinedPipeline,
+ KandinskyInpaintCombinedPipeline,
+ )
+ from .pipeline_kandinsky_img2img import KandinskyImg2ImgPipeline
+ from .pipeline_kandinsky_inpaint import KandinskyInpaintPipeline
+ from .pipeline_kandinsky_prior import KandinskyPriorPipeline, KandinskyPriorPipelineOutput
+ from .text_encoder import MultilingualCLIP
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/kandinsky/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..cf6679072c06261b3270ec97a883a78fd6e1d7e5
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c586720031ac1675eb1d9e1fb33b7356607d4e6e
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_combined.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_combined.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b4b77d0eda159e3380481cf4007d05f4d5be72a4
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_combined.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..608c268f0af1e341611b19e0c7d4efea1d250151
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_inpaint.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_inpaint.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ef035904ba86b6def42013e15f008522897908d9
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_inpaint.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_prior.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_prior.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..15bbe54b5146fa956bfea6ea937ce7e08635e651
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky/__pycache__/pipeline_kandinsky_prior.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky/__pycache__/text_encoder.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky/__pycache__/text_encoder.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..fb663146125cb733297ba5ae6ce0ff09cf69f93d
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky/__pycache__/text_encoder.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky.py b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky.py
new file mode 100644
index 0000000000000000000000000000000000000000..34b5a47c2541aca330f65cbd8e31ce722e55fb96
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky.py
@@ -0,0 +1,407 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, List, Optional, Union
+
+import torch
+from transformers import (
+ XLMRobertaTokenizer,
+)
+
+from ...models import UNet2DConditionModel, VQModel
+from ...schedulers import DDIMScheduler, DDPMScheduler
+from ...utils import (
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .text_encoder import MultilingualCLIP
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyPipeline, KandinskyPriorPipeline
+ >>> import torch
+
+ >>> pipe_prior = KandinskyPriorPipeline.from_pretrained("kandinsky-community/Kandinsky-2-1-prior")
+ >>> pipe_prior.to("cuda")
+
+ >>> prompt = "red cat, 4k photo"
+ >>> out = pipe_prior(prompt)
+ >>> image_emb = out.image_embeds
+ >>> negative_image_emb = out.negative_image_embeds
+
+ >>> pipe = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1")
+ >>> pipe.to("cuda")
+
+ >>> image = pipe(
+ ... prompt,
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=negative_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=100,
+ ... ).images
+
+ >>> image[0].save("cat.png")
+ ```
+"""
+
+
+def get_new_h_w(h, w, scale_factor=8):
+ new_h = h // scale_factor**2
+ if h % scale_factor**2 != 0:
+ new_h += 1
+ new_w = w // scale_factor**2
+ if w % scale_factor**2 != 0:
+ new_w += 1
+ return new_h * scale_factor, new_w * scale_factor
+
+
+class KandinskyPipeline(DiffusionPipeline):
+ """
+ Pipeline for text-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ text_encoder ([`MultilingualCLIP`]):
+ Frozen text-encoder.
+ tokenizer ([`XLMRobertaTokenizer`]):
+ Tokenizer of class
+ scheduler (Union[`DDIMScheduler`,`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->movq"
+
+ def __init__(
+ self,
+ text_encoder: MultilingualCLIP,
+ tokenizer: XLMRobertaTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: Union[DDIMScheduler, DDPMScheduler],
+ movq: VQModel,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+ self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ ):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ truncation=True,
+ max_length=77,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ text_input_ids = text_input_ids.to(device)
+ text_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds, text_encoder_hidden_states = self.text_encoder(
+ input_ids=text_input_ids, attention_mask=text_mask
+ )
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=77,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ uncond_text_input_ids = uncond_input.input_ids.to(device)
+ uncond_text_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds, uncond_text_encoder_hidden_states = self.text_encoder(
+ input_ids=uncond_text_input_ids, attention_mask=uncond_text_mask
+ )
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_encoder_hidden_states.shape[1]
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_encoder_hidden_states, text_mask
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ negative_image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for text prompt, that will be used to condition the image generation.
+ negative_image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for negative text prompt, will be used to condition the image generation.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+
+ if isinstance(prompt, str):
+ batch_size = 1
+ elif isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ device = self._execution_device
+
+ batch_size = batch_size * num_images_per_prompt
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ prompt_embeds, text_encoder_hidden_states, _ = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+ )
+
+ if isinstance(image_embeds, list):
+ image_embeds = torch.cat(image_embeds, dim=0)
+ if isinstance(negative_image_embeds, list):
+ negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
+
+ if do_classifier_free_guidance:
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to(
+ dtype=prompt_embeds.dtype, device=device
+ )
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps_tensor = self.scheduler.timesteps
+
+ num_channels_latents = self.unet.config.in_channels
+
+ height, width = get_new_h_w(height, width, self.movq_scale_factor)
+
+ # create initial latent
+ latents = self.prepare_latents(
+ (batch_size, num_channels_latents, height, width),
+ text_encoder_hidden_states.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+
+ for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ added_cond_kwargs = {"text_embeds": prompt_embeds, "image_embeds": image_embeds}
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=text_encoder_hidden_states,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ if do_classifier_free_guidance:
+ noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1)
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ _, variance_pred_text = variance_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1)
+
+ if not (
+ hasattr(self.scheduler.config, "variance_type")
+ and self.scheduler.config.variance_type in ["learned", "learned_range"]
+ ):
+ noise_pred, _ = noise_pred.split(latents.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ ).prev_sample
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # post-processing
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+
+ self.maybe_free_model_hooks()
+
+ if output_type not in ["pt", "np", "pil"]:
+ raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}")
+
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_combined.py b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_combined.py
new file mode 100644
index 0000000000000000000000000000000000000000..da5ff52ed15fd0658dfe9bd547225f30cc835f69
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_combined.py
@@ -0,0 +1,814 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Callable, List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+ XLMRobertaTokenizer,
+)
+
+from ...models import PriorTransformer, UNet2DConditionModel, VQModel
+from ...schedulers import DDIMScheduler, DDPMScheduler, UnCLIPScheduler
+from ...utils import (
+ replace_example_docstring,
+)
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_kandinsky import KandinskyPipeline
+from .pipeline_kandinsky_img2img import KandinskyImg2ImgPipeline
+from .pipeline_kandinsky_inpaint import KandinskyInpaintPipeline
+from .pipeline_kandinsky_prior import KandinskyPriorPipeline
+from .text_encoder import MultilingualCLIP
+
+
+TEXT2IMAGE_EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import AutoPipelineForText2Image
+ import torch
+
+ pipe = AutoPipelineForText2Image.from_pretrained(
+ "kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16
+ )
+ pipe.enable_model_cpu_offload()
+
+ prompt = "A lion in galaxies, spirals, nebulae, stars, smoke, iridescent, intricate detail, octane render, 8k"
+
+ image = pipe(prompt=prompt, num_inference_steps=25).images[0]
+ ```
+"""
+
+IMAGE2IMAGE_EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import AutoPipelineForImage2Image
+ import torch
+ import requests
+ from io import BytesIO
+ from PIL import Image
+ import os
+
+ pipe = AutoPipelineForImage2Image.from_pretrained(
+ "kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16
+ )
+ pipe.enable_model_cpu_offload()
+
+ prompt = "A fantasy landscape, Cinematic lighting"
+ negative_prompt = "low quality, bad quality"
+
+ url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+
+ response = requests.get(url)
+ image = Image.open(BytesIO(response.content)).convert("RGB")
+ image.thumbnail((768, 768))
+
+ image = pipe(prompt=prompt, image=original_image, num_inference_steps=25).images[0]
+ ```
+"""
+
+INPAINT_EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import AutoPipelineForInpainting
+ from diffusers.utils import load_image
+ import torch
+ import numpy as np
+
+ pipe = AutoPipelineForInpainting.from_pretrained(
+ "kandinsky-community/kandinsky-2-1-inpaint", torch_dtype=torch.float16
+ )
+ pipe.enable_model_cpu_offload()
+
+ prompt = "A fantasy landscape, Cinematic lighting"
+ negative_prompt = "low quality, bad quality"
+
+ original_image = load_image(
+ "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" "/kandinsky/cat.png"
+ )
+
+ mask = np.zeros((768, 768), dtype=np.float32)
+ # Let's mask out an area above the cat's head
+ mask[:250, 250:-250] = 1
+
+ image = pipe(prompt=prompt, image=original_image, mask_image=mask, num_inference_steps=25).images[0]
+ ```
+"""
+
+
+class KandinskyCombinedPipeline(DiffusionPipeline):
+ """
+ Combined Pipeline for text-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ text_encoder ([`MultilingualCLIP`]):
+ Frozen text-encoder.
+ tokenizer ([`XLMRobertaTokenizer`]):
+ Tokenizer of class
+ scheduler (Union[`DDIMScheduler`,`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ prior_prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ prior_image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ prior_text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ prior_tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ prior_scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ """
+
+ _load_connected_pipes = True
+ model_cpu_offload_seq = "text_encoder->unet->movq->prior_prior->prior_image_encoder->prior_text_encoder"
+
+ def __init__(
+ self,
+ text_encoder: MultilingualCLIP,
+ tokenizer: XLMRobertaTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: Union[DDIMScheduler, DDPMScheduler],
+ movq: VQModel,
+ prior_prior: PriorTransformer,
+ prior_image_encoder: CLIPVisionModelWithProjection,
+ prior_text_encoder: CLIPTextModelWithProjection,
+ prior_tokenizer: CLIPTokenizer,
+ prior_scheduler: UnCLIPScheduler,
+ prior_image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ prior_prior=prior_prior,
+ prior_image_encoder=prior_image_encoder,
+ prior_text_encoder=prior_text_encoder,
+ prior_tokenizer=prior_tokenizer,
+ prior_scheduler=prior_scheduler,
+ prior_image_processor=prior_image_processor,
+ )
+ self.prior_pipe = KandinskyPriorPipeline(
+ prior=prior_prior,
+ image_encoder=prior_image_encoder,
+ text_encoder=prior_text_encoder,
+ tokenizer=prior_tokenizer,
+ scheduler=prior_scheduler,
+ image_processor=prior_image_processor,
+ )
+ self.decoder_pipe = KandinskyPipeline(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
+
+ def enable_sequential_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗
+ Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a
+ GPU only when their specific submodule's `forward` method is called. Offloading happens on a submodule basis.
+ Memory savings are higher than using `enable_model_cpu_offload`, but performance is lower.
+ """
+ self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+
+ def progress_bar(self, iterable=None, total=None):
+ self.prior_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.enable_model_cpu_offload()
+
+ def set_progress_bar_config(self, **kwargs):
+ self.prior_pipe.set_progress_bar_config(**kwargs)
+ self.decoder_pipe.set_progress_bar_config(**kwargs)
+
+ @torch.no_grad()
+ @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ height: int = 512,
+ width: int = 512,
+ prior_guidance_scale: float = 4.0,
+ prior_num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ prior_num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ prior_outputs = self.prior_pipe(
+ prompt=prompt,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ num_inference_steps=prior_num_inference_steps,
+ generator=generator,
+ latents=latents,
+ guidance_scale=prior_guidance_scale,
+ output_type="pt",
+ return_dict=False,
+ )
+ image_embeds = prior_outputs[0]
+ negative_image_embeds = prior_outputs[1]
+
+ prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
+
+ if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
+ prompt = (image_embeds.shape[0] // len(prompt)) * prompt
+
+ outputs = self.decoder_pipe(
+ prompt=prompt,
+ image_embeds=image_embeds,
+ negative_image_embeds=negative_image_embeds,
+ width=width,
+ height=height,
+ num_inference_steps=num_inference_steps,
+ generator=generator,
+ guidance_scale=guidance_scale,
+ output_type=output_type,
+ callback=callback,
+ callback_steps=callback_steps,
+ return_dict=return_dict,
+ )
+
+ self.maybe_free_model_hooks()
+
+ return outputs
+
+
+class KandinskyImg2ImgCombinedPipeline(DiffusionPipeline):
+ """
+ Combined Pipeline for image-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ text_encoder ([`MultilingualCLIP`]):
+ Frozen text-encoder.
+ tokenizer ([`XLMRobertaTokenizer`]):
+ Tokenizer of class
+ scheduler (Union[`DDIMScheduler`,`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ prior_prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ prior_image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ prior_text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ prior_tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ prior_scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ """
+
+ _load_connected_pipes = True
+ model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->prior_prior->" "text_encoder->unet->movq"
+
+ def __init__(
+ self,
+ text_encoder: MultilingualCLIP,
+ tokenizer: XLMRobertaTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: Union[DDIMScheduler, DDPMScheduler],
+ movq: VQModel,
+ prior_prior: PriorTransformer,
+ prior_image_encoder: CLIPVisionModelWithProjection,
+ prior_text_encoder: CLIPTextModelWithProjection,
+ prior_tokenizer: CLIPTokenizer,
+ prior_scheduler: UnCLIPScheduler,
+ prior_image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ prior_prior=prior_prior,
+ prior_image_encoder=prior_image_encoder,
+ prior_text_encoder=prior_text_encoder,
+ prior_tokenizer=prior_tokenizer,
+ prior_scheduler=prior_scheduler,
+ prior_image_processor=prior_image_processor,
+ )
+ self.prior_pipe = KandinskyPriorPipeline(
+ prior=prior_prior,
+ image_encoder=prior_image_encoder,
+ text_encoder=prior_text_encoder,
+ tokenizer=prior_tokenizer,
+ scheduler=prior_scheduler,
+ image_processor=prior_image_processor,
+ )
+ self.decoder_pipe = KandinskyImg2ImgPipeline(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
+
+ def enable_sequential_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+ text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+ `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+ Note that offloading happens on a submodule basis. Memory savings are higher than with
+ `enable_model_cpu_offload`, but performance is lower.
+ """
+ self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+
+ def progress_bar(self, iterable=None, total=None):
+ self.prior_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.enable_model_cpu_offload()
+
+ def set_progress_bar_config(self, **kwargs):
+ self.prior_pipe.set_progress_bar_config(**kwargs)
+ self.decoder_pipe.set_progress_bar_config(**kwargs)
+
+ @torch.no_grad()
+ @replace_example_docstring(IMAGE2IMAGE_EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ strength: float = 0.3,
+ height: int = 512,
+ width: int = 512,
+ prior_guidance_scale: float = 4.0,
+ prior_num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded
+ again.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ strength (`float`, *optional*, defaults to 0.3):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ prior_num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ prior_outputs = self.prior_pipe(
+ prompt=prompt,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ num_inference_steps=prior_num_inference_steps,
+ generator=generator,
+ latents=latents,
+ guidance_scale=prior_guidance_scale,
+ output_type="pt",
+ return_dict=False,
+ )
+ image_embeds = prior_outputs[0]
+ negative_image_embeds = prior_outputs[1]
+
+ prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
+ image = [image] if isinstance(prompt, PIL.Image.Image) else image
+
+ if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
+ prompt = (image_embeds.shape[0] // len(prompt)) * prompt
+
+ if (
+ isinstance(image, (list, tuple))
+ and len(image) < image_embeds.shape[0]
+ and image_embeds.shape[0] % len(image) == 0
+ ):
+ image = (image_embeds.shape[0] // len(image)) * image
+
+ outputs = self.decoder_pipe(
+ prompt=prompt,
+ image=image,
+ image_embeds=image_embeds,
+ negative_image_embeds=negative_image_embeds,
+ strength=strength,
+ width=width,
+ height=height,
+ num_inference_steps=num_inference_steps,
+ generator=generator,
+ guidance_scale=guidance_scale,
+ output_type=output_type,
+ callback=callback,
+ callback_steps=callback_steps,
+ return_dict=return_dict,
+ )
+
+ self.maybe_free_model_hooks()
+
+ return outputs
+
+
+class KandinskyInpaintCombinedPipeline(DiffusionPipeline):
+ """
+ Combined Pipeline for generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ text_encoder ([`MultilingualCLIP`]):
+ Frozen text-encoder.
+ tokenizer ([`XLMRobertaTokenizer`]):
+ Tokenizer of class
+ scheduler (Union[`DDIMScheduler`,`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ prior_prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ prior_image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ prior_text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ prior_tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ prior_scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ """
+
+ _load_connected_pipes = True
+ model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->prior_prior->text_encoder->unet->movq"
+
+ def __init__(
+ self,
+ text_encoder: MultilingualCLIP,
+ tokenizer: XLMRobertaTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: Union[DDIMScheduler, DDPMScheduler],
+ movq: VQModel,
+ prior_prior: PriorTransformer,
+ prior_image_encoder: CLIPVisionModelWithProjection,
+ prior_text_encoder: CLIPTextModelWithProjection,
+ prior_tokenizer: CLIPTokenizer,
+ prior_scheduler: UnCLIPScheduler,
+ prior_image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ prior_prior=prior_prior,
+ prior_image_encoder=prior_image_encoder,
+ prior_text_encoder=prior_text_encoder,
+ prior_tokenizer=prior_tokenizer,
+ prior_scheduler=prior_scheduler,
+ prior_image_processor=prior_image_processor,
+ )
+ self.prior_pipe = KandinskyPriorPipeline(
+ prior=prior_prior,
+ image_encoder=prior_image_encoder,
+ text_encoder=prior_text_encoder,
+ tokenizer=prior_tokenizer,
+ scheduler=prior_scheduler,
+ image_processor=prior_image_processor,
+ )
+ self.decoder_pipe = KandinskyInpaintPipeline(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
+
+ def enable_sequential_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+ text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+ `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+ Note that offloading happens on a submodule basis. Memory savings are higher than with
+ `enable_model_cpu_offload`, but performance is lower.
+ """
+ self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+
+ def progress_bar(self, iterable=None, total=None):
+ self.prior_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.enable_model_cpu_offload()
+
+ def set_progress_bar_config(self, **kwargs):
+ self.prior_pipe.set_progress_bar_config(**kwargs)
+ self.decoder_pipe.set_progress_bar_config(**kwargs)
+
+ @torch.no_grad()
+ @replace_example_docstring(INPAINT_EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ height: int = 512,
+ width: int = 512,
+ prior_guidance_scale: float = 4.0,
+ prior_num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded
+ again.
+ mask_image (`np.array`):
+ Tensor representing an image batch, to mask `image`. White pixels in the mask will be repainted, while
+ black pixels will be preserved. If `mask_image` is a PIL image, it will be converted to a single
+ channel (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3,
+ so the expected shape would be `(B, H, W, 1)`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ prior_num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ prior_outputs = self.prior_pipe(
+ prompt=prompt,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ num_inference_steps=prior_num_inference_steps,
+ generator=generator,
+ latents=latents,
+ guidance_scale=prior_guidance_scale,
+ output_type="pt",
+ return_dict=False,
+ )
+ image_embeds = prior_outputs[0]
+ negative_image_embeds = prior_outputs[1]
+
+ prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
+ image = [image] if isinstance(prompt, PIL.Image.Image) else image
+ mask_image = [mask_image] if isinstance(mask_image, PIL.Image.Image) else mask_image
+
+ if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
+ prompt = (image_embeds.shape[0] // len(prompt)) * prompt
+
+ if (
+ isinstance(image, (list, tuple))
+ and len(image) < image_embeds.shape[0]
+ and image_embeds.shape[0] % len(image) == 0
+ ):
+ image = (image_embeds.shape[0] // len(image)) * image
+
+ if (
+ isinstance(mask_image, (list, tuple))
+ and len(mask_image) < image_embeds.shape[0]
+ and image_embeds.shape[0] % len(mask_image) == 0
+ ):
+ mask_image = (image_embeds.shape[0] // len(mask_image)) * mask_image
+
+ outputs = self.decoder_pipe(
+ prompt=prompt,
+ image=image,
+ mask_image=mask_image,
+ image_embeds=image_embeds,
+ negative_image_embeds=negative_image_embeds,
+ width=width,
+ height=height,
+ num_inference_steps=num_inference_steps,
+ generator=generator,
+ guidance_scale=guidance_scale,
+ output_type=output_type,
+ callback=callback,
+ callback_steps=callback_steps,
+ return_dict=return_dict,
+ )
+
+ self.maybe_free_model_hooks()
+
+ return outputs
diff --git a/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_img2img.py b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..4d091e7d7a7c180844a719d7acf0db9ab931e84e
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_img2img.py
@@ -0,0 +1,500 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from PIL import Image
+from transformers import (
+ XLMRobertaTokenizer,
+)
+
+from ...models import UNet2DConditionModel, VQModel
+from ...schedulers import DDIMScheduler
+from ...utils import (
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .text_encoder import MultilingualCLIP
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyImg2ImgPipeline, KandinskyPriorPipeline
+ >>> from diffusers.utils import load_image
+ >>> import torch
+
+ >>> pipe_prior = KandinskyPriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior.to("cuda")
+
+ >>> prompt = "A red cartoon frog, 4k"
+ >>> image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False)
+
+ >>> pipe = KandinskyImg2ImgPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16
+ ... )
+ >>> pipe.to("cuda")
+
+ >>> init_image = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/frog.png"
+ ... )
+
+ >>> image = pipe(
+ ... prompt,
+ ... image=init_image,
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=zero_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=100,
+ ... strength=0.2,
+ ... ).images
+
+ >>> image[0].save("red_frog.png")
+ ```
+"""
+
+
+def get_new_h_w(h, w, scale_factor=8):
+ new_h = h // scale_factor**2
+ if h % scale_factor**2 != 0:
+ new_h += 1
+ new_w = w // scale_factor**2
+ if w % scale_factor**2 != 0:
+ new_w += 1
+ return new_h * scale_factor, new_w * scale_factor
+
+
+def prepare_image(pil_image, w=512, h=512):
+ pil_image = pil_image.resize((w, h), resample=Image.BICUBIC, reducing_gap=1)
+ arr = np.array(pil_image.convert("RGB"))
+ arr = arr.astype(np.float32) / 127.5 - 1
+ arr = np.transpose(arr, [2, 0, 1])
+ image = torch.from_numpy(arr).unsqueeze(0)
+ return image
+
+
+class KandinskyImg2ImgPipeline(DiffusionPipeline):
+ """
+ Pipeline for image-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ text_encoder ([`MultilingualCLIP`]):
+ Frozen text-encoder.
+ tokenizer ([`XLMRobertaTokenizer`]):
+ Tokenizer of class
+ scheduler ([`DDIMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ image encoder and decoder
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->movq"
+
+ def __init__(
+ self,
+ text_encoder: MultilingualCLIP,
+ movq: VQModel,
+ tokenizer: XLMRobertaTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: DDIMScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+ self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(self, latents, latent_timestep, shape, dtype, device, generator, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+
+ shape = latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ latents = self.add_noise(latents, noise, latent_timestep)
+ return latents
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ ):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=77,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ text_input_ids = text_input_ids.to(device)
+ text_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds, text_encoder_hidden_states = self.text_encoder(
+ input_ids=text_input_ids, attention_mask=text_mask
+ )
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=77,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ uncond_text_input_ids = uncond_input.input_ids.to(device)
+ uncond_text_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds, uncond_text_encoder_hidden_states = self.text_encoder(
+ input_ids=uncond_text_input_ids, attention_mask=uncond_text_mask
+ )
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_encoder_hidden_states.shape[1]
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_encoder_hidden_states, text_mask
+
+ # add_noise method to overwrite the one in schedule because it use a different beta schedule for adding noise vs sampling
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ betas = torch.linspace(0.0001, 0.02, 1000, dtype=torch.float32)
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_cumprod = alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+
+ return noisy_samples
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ image_embeds: torch.FloatTensor,
+ negative_image_embeds: torch.FloatTensor,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 100,
+ strength: float = 0.3,
+ guidance_scale: float = 7.0,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`torch.FloatTensor`, `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for text prompt, that will be used to condition the image generation.
+ negative_image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for negative text prompt, will be used to condition the image generation.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ strength (`float`, *optional*, defaults to 0.3):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ # 1. Define call parameters
+ if isinstance(prompt, str):
+ batch_size = 1
+ elif isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ device = self._execution_device
+
+ batch_size = batch_size * num_images_per_prompt
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 2. get text and image embeddings
+ prompt_embeds, text_encoder_hidden_states, _ = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+ )
+
+ if isinstance(image_embeds, list):
+ image_embeds = torch.cat(image_embeds, dim=0)
+ if isinstance(negative_image_embeds, list):
+ negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
+
+ if do_classifier_free_guidance:
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to(
+ dtype=prompt_embeds.dtype, device=device
+ )
+
+ # 3. pre-processing initial image
+ if not isinstance(image, list):
+ image = [image]
+ if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image):
+ raise ValueError(
+ f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support PIL image and pytorch tensor"
+ )
+
+ image = torch.cat([prepare_image(i, width, height) for i in image], dim=0)
+ image = image.to(dtype=prompt_embeds.dtype, device=device)
+
+ latents = self.movq.encode(image)["latents"]
+ latents = latents.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # 4. set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+ timesteps_tensor, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+
+ # the formular to calculate timestep for add_noise is taken from the original kandinsky repo
+ latent_timestep = int(self.scheduler.config.num_train_timesteps * strength) - 2
+
+ latent_timestep = torch.tensor([latent_timestep] * batch_size, dtype=timesteps_tensor.dtype, device=device)
+
+ num_channels_latents = self.unet.config.in_channels
+
+ height, width = get_new_h_w(height, width, self.movq_scale_factor)
+
+ # 5. Create initial latent
+ latents = self.prepare_latents(
+ latents,
+ latent_timestep,
+ (batch_size, num_channels_latents, height, width),
+ text_encoder_hidden_states.dtype,
+ device,
+ generator,
+ self.scheduler,
+ )
+
+ # 6. Denoising loop
+ for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ added_cond_kwargs = {"text_embeds": prompt_embeds, "image_embeds": image_embeds}
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=text_encoder_hidden_states,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ if do_classifier_free_guidance:
+ noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1)
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ _, variance_pred_text = variance_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1)
+
+ if not (
+ hasattr(self.scheduler.config, "variance_type")
+ and self.scheduler.config.variance_type in ["learned", "learned_range"]
+ ):
+ noise_pred, _ = noise_pred.split(latents.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ ).prev_sample
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 7. post-processing
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+
+ self.maybe_free_model_hooks()
+
+ if output_type not in ["pt", "np", "pil"]:
+ raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}")
+
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_inpaint.py b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_inpaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..0555cbe754bb5d29e662f642ca011914992626f5
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_inpaint.py
@@ -0,0 +1,635 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from copy import deepcopy
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from packaging import version
+from PIL import Image
+from transformers import (
+ XLMRobertaTokenizer,
+)
+
+from ... import __version__
+from ...models import UNet2DConditionModel, VQModel
+from ...schedulers import DDIMScheduler
+from ...utils import (
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .text_encoder import MultilingualCLIP
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyInpaintPipeline, KandinskyPriorPipeline
+ >>> from diffusers.utils import load_image
+ >>> import torch
+ >>> import numpy as np
+
+ >>> pipe_prior = KandinskyPriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior.to("cuda")
+
+ >>> prompt = "a hat"
+ >>> image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False)
+
+ >>> pipe = KandinskyInpaintPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-1-inpaint", torch_dtype=torch.float16
+ ... )
+ >>> pipe.to("cuda")
+
+ >>> init_image = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... )
+
+ >>> mask = np.zeros((768, 768), dtype=np.float32)
+ >>> mask[:250, 250:-250] = 1
+
+ >>> out = pipe(
+ ... prompt,
+ ... image=init_image,
+ ... mask_image=mask,
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=zero_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=50,
+ ... )
+
+ >>> image = out.images[0]
+ >>> image.save("cat_with_hat.png")
+ ```
+"""
+
+
+def get_new_h_w(h, w, scale_factor=8):
+ new_h = h // scale_factor**2
+ if h % scale_factor**2 != 0:
+ new_h += 1
+ new_w = w // scale_factor**2
+ if w % scale_factor**2 != 0:
+ new_w += 1
+ return new_h * scale_factor, new_w * scale_factor
+
+
+def prepare_mask(masks):
+ prepared_masks = []
+ for mask in masks:
+ old_mask = deepcopy(mask)
+ for i in range(mask.shape[1]):
+ for j in range(mask.shape[2]):
+ if old_mask[0][i][j] == 1:
+ continue
+ if i != 0:
+ mask[:, i - 1, j] = 0
+ if j != 0:
+ mask[:, i, j - 1] = 0
+ if i != 0 and j != 0:
+ mask[:, i - 1, j - 1] = 0
+ if i != mask.shape[1] - 1:
+ mask[:, i + 1, j] = 0
+ if j != mask.shape[2] - 1:
+ mask[:, i, j + 1] = 0
+ if i != mask.shape[1] - 1 and j != mask.shape[2] - 1:
+ mask[:, i + 1, j + 1] = 0
+ prepared_masks.append(mask)
+ return torch.stack(prepared_masks, dim=0)
+
+
+def prepare_mask_and_masked_image(image, mask, height, width):
+ r"""
+ Prepares a pair (mask, image) to be consumed by the Kandinsky inpaint pipeline. This means that those inputs will
+ be converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for
+ the ``image`` and ``1`` for the ``mask``.
+
+ The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+ binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+ Args:
+ image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+ ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+ mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+ ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+
+
+ Raises:
+ ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+ should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+ TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+ (ot the other way around).
+
+ Returns:
+ tuple[torch.Tensor]: The pair (mask, image) as ``torch.Tensor`` with 4
+ dimensions: ``batch x channels x height x width``.
+ """
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ if mask is None:
+ raise ValueError("`mask_image` input cannot be undefined.")
+
+ if isinstance(image, torch.Tensor):
+ if not isinstance(mask, torch.Tensor):
+ raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
+
+ # Batch single image
+ if image.ndim == 3:
+ assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
+ image = image.unsqueeze(0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Single batched mask, no channel dim or single mask not batched but channel dim
+ if mask.shape[0] == 1:
+ mask = mask.unsqueeze(0)
+
+ # Batched masks no channel dim
+ else:
+ mask = mask.unsqueeze(1)
+
+ assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+ assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+ assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+
+ # Check image is in [-1, 1]
+ if image.min() < -1 or image.max() > 1:
+ raise ValueError("Image should be in [-1, 1] range")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("Mask should be in [0, 1] range")
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ # Image as float32
+ image = image.to(dtype=torch.float32)
+ elif isinstance(mask, torch.Tensor):
+ raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+ else:
+ # preprocess image
+ if isinstance(image, (PIL.Image.Image, np.ndarray)):
+ image = [image]
+
+ if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+ # resize all images w.r.t passed height an width
+ image = [i.resize((width, height), resample=Image.BICUBIC, reducing_gap=1) for i in image]
+ image = [np.array(i.convert("RGB"))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+ image = np.concatenate([i[None, :] for i in image], axis=0)
+
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ # preprocess mask
+ if isinstance(mask, (PIL.Image.Image, np.ndarray)):
+ mask = [mask]
+
+ if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
+ mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
+ mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+ mask = mask.astype(np.float32) / 255.0
+ elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+ mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+ mask = torch.from_numpy(mask)
+
+ mask = 1 - mask
+
+ return mask, image
+
+
+class KandinskyInpaintPipeline(DiffusionPipeline):
+ """
+ Pipeline for text-guided image inpainting using Kandinsky2.1
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ text_encoder ([`MultilingualCLIP`]):
+ Frozen text-encoder.
+ tokenizer ([`XLMRobertaTokenizer`]):
+ Tokenizer of class
+ scheduler ([`DDIMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ image encoder and decoder
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->movq"
+
+ def __init__(
+ self,
+ text_encoder: MultilingualCLIP,
+ movq: VQModel,
+ tokenizer: XLMRobertaTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: DDIMScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ text_encoder=text_encoder,
+ movq=movq,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ )
+ self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+ self._warn_has_been_called = False
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ ):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=77,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ text_input_ids = text_input_ids.to(device)
+ text_mask = text_inputs.attention_mask.to(device)
+
+ prompt_embeds, text_encoder_hidden_states = self.text_encoder(
+ input_ids=text_input_ids, attention_mask=text_mask
+ )
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=77,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ uncond_text_input_ids = uncond_input.input_ids.to(device)
+ uncond_text_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds, uncond_text_encoder_hidden_states = self.text_encoder(
+ input_ids=uncond_text_input_ids, attention_mask=uncond_text_mask
+ )
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_encoder_hidden_states.shape[1]
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_encoder_hidden_states, text_mask
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[torch.FloatTensor, PIL.Image.Image],
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
+ image_embeds: torch.FloatTensor,
+ negative_image_embeds: torch.FloatTensor,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`torch.FloatTensor`, `PIL.Image.Image` or `np.ndarray`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ mask_image (`PIL.Image.Image`,`torch.FloatTensor` or `np.ndarray`):
+ `Image`, or a tensor representing an image batch, to mask `image`. White pixels in the mask will be
+ repainted, while black pixels will be preserved. You can pass a pytorch tensor as mask only if the
+ image you passed is a pytorch tensor, and it should contain one color channel (L) instead of 3, so the
+ expected shape would be either `(B, 1, H, W,)`, `(B, H, W)`, `(1, H, W)` or `(H, W)` If image is an PIL
+ image or numpy array, mask should also be a either PIL image or numpy array. If it is a PIL image, it
+ will be converted to a single channel (luminance) before use. If it is a nummpy array, the expected
+ shape is `(H, W)`.
+ image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for text prompt, that will be used to condition the image generation.
+ negative_image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for negative text prompt, will be used to condition the image generation.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ if not self._warn_has_been_called and version.parse(version.parse(__version__).base_version) < version.parse(
+ "0.23.0.dev0"
+ ):
+ logger.warn(
+ "Please note that the expected format of `mask_image` has recently been changed. "
+ "Before diffusers == 0.19.0, Kandinsky Inpainting pipelines repainted black pixels and preserved black pixels. "
+ "As of diffusers==0.19.0 this behavior has been inverted. Now white pixels are repainted and black pixels are preserved. "
+ "This way, Kandinsky's masking behavior is aligned with Stable Diffusion. "
+ "THIS means that you HAVE to invert the input mask to have the same behavior as before as explained in https://github.com/huggingface/diffusers/pull/4207. "
+ "This warning will be surpressed after the first inference call and will be removed in diffusers>0.23.0"
+ )
+ self._warn_has_been_called = True
+
+ # Define call parameters
+ if isinstance(prompt, str):
+ batch_size = 1
+ elif isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ device = self._execution_device
+
+ batch_size = batch_size * num_images_per_prompt
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ prompt_embeds, text_encoder_hidden_states, _ = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+ )
+
+ if isinstance(image_embeds, list):
+ image_embeds = torch.cat(image_embeds, dim=0)
+ if isinstance(negative_image_embeds, list):
+ negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
+
+ if do_classifier_free_guidance:
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to(
+ dtype=prompt_embeds.dtype, device=device
+ )
+
+ # preprocess image and mask
+ mask_image, image = prepare_mask_and_masked_image(image, mask_image, height, width)
+
+ image = image.to(dtype=prompt_embeds.dtype, device=device)
+ image = self.movq.encode(image)["latents"]
+
+ mask_image = mask_image.to(dtype=prompt_embeds.dtype, device=device)
+
+ image_shape = tuple(image.shape[-2:])
+ mask_image = F.interpolate(
+ mask_image,
+ image_shape,
+ mode="nearest",
+ )
+ mask_image = prepare_mask(mask_image)
+ masked_image = image * mask_image
+
+ mask_image = mask_image.repeat_interleave(num_images_per_prompt, dim=0)
+ masked_image = masked_image.repeat_interleave(num_images_per_prompt, dim=0)
+ if do_classifier_free_guidance:
+ mask_image = mask_image.repeat(2, 1, 1, 1)
+ masked_image = masked_image.repeat(2, 1, 1, 1)
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps_tensor = self.scheduler.timesteps
+
+ num_channels_latents = self.movq.config.latent_channels
+
+ # get h, w for latents
+ sample_height, sample_width = get_new_h_w(height, width, self.movq_scale_factor)
+
+ # create initial latent
+ latents = self.prepare_latents(
+ (batch_size, num_channels_latents, sample_height, sample_width),
+ text_encoder_hidden_states.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+
+ # Check that sizes of mask, masked image and latents match with expected
+ num_channels_mask = mask_image.shape[1]
+ num_channels_masked_image = masked_image.shape[1]
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+
+ for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = torch.cat([latent_model_input, masked_image, mask_image], dim=1)
+
+ added_cond_kwargs = {"text_embeds": prompt_embeds, "image_embeds": image_embeds}
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=text_encoder_hidden_states,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ if do_classifier_free_guidance:
+ noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1)
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ _, variance_pred_text = variance_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1)
+
+ if not (
+ hasattr(self.scheduler.config, "variance_type")
+ and self.scheduler.config.variance_type in ["learned", "learned_range"]
+ ):
+ noise_pred, _ = noise_pred.split(latents.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ ).prev_sample
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # post-processing
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+
+ self.maybe_free_model_hooks()
+
+ if output_type not in ["pt", "np", "pil"]:
+ raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}")
+
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_prior.py b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_prior.py
new file mode 100644
index 0000000000000000000000000000000000000000..0d9f54349660d28960410e33b87770388d363c99
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky/pipeline_kandinsky_prior.py
@@ -0,0 +1,547 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...models import PriorTransformer
+from ...schedulers import UnCLIPScheduler
+from ...utils import (
+ BaseOutput,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyPipeline, KandinskyPriorPipeline
+ >>> import torch
+
+ >>> pipe_prior = KandinskyPriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-1-prior")
+ >>> pipe_prior.to("cuda")
+
+ >>> prompt = "red cat, 4k photo"
+ >>> out = pipe_prior(prompt)
+ >>> image_emb = out.image_embeds
+ >>> negative_image_emb = out.negative_image_embeds
+
+ >>> pipe = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1")
+ >>> pipe.to("cuda")
+
+ >>> image = pipe(
+ ... prompt,
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=negative_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=100,
+ ... ).images
+
+ >>> image[0].save("cat.png")
+ ```
+"""
+
+EXAMPLE_INTERPOLATE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyPriorPipeline, KandinskyPipeline
+ >>> from diffusers.utils import load_image
+ >>> import PIL
+
+ >>> import torch
+ >>> from torchvision import transforms
+
+ >>> pipe_prior = KandinskyPriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-1-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior.to("cuda")
+
+ >>> img1 = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... )
+
+ >>> img2 = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/starry_night.jpeg"
+ ... )
+
+ >>> images_texts = ["a cat", img1, img2]
+ >>> weights = [0.3, 0.3, 0.4]
+ >>> image_emb, zero_image_emb = pipe_prior.interpolate(images_texts, weights)
+
+ >>> pipe = KandinskyPipeline.from_pretrained("kandinsky-community/kandinsky-2-1", torch_dtype=torch.float16)
+ >>> pipe.to("cuda")
+
+ >>> image = pipe(
+ ... "",
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=zero_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=150,
+ ... ).images[0]
+
+ >>> image.save("starry_cat.png")
+ ```
+"""
+
+
+@dataclass
+class KandinskyPriorPipelineOutput(BaseOutput):
+ """
+ Output class for KandinskyPriorPipeline.
+
+ Args:
+ image_embeds (`torch.FloatTensor`)
+ clip image embeddings for text prompt
+ negative_image_embeds (`List[PIL.Image.Image]` or `np.ndarray`)
+ clip image embeddings for unconditional tokens
+ """
+
+ image_embeds: Union[torch.FloatTensor, np.ndarray]
+ negative_image_embeds: Union[torch.FloatTensor, np.ndarray]
+
+
+class KandinskyPriorPipeline(DiffusionPipeline):
+ """
+ Pipeline for generating image prior for Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ """
+
+ _exclude_from_cpu_offload = ["prior"]
+ model_cpu_offload_seq = "text_encoder->prior"
+
+ def __init__(
+ self,
+ prior: PriorTransformer,
+ image_encoder: CLIPVisionModelWithProjection,
+ text_encoder: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ scheduler: UnCLIPScheduler,
+ image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ prior=prior,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ scheduler=scheduler,
+ image_encoder=image_encoder,
+ image_processor=image_processor,
+ )
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_INTERPOLATE_DOC_STRING)
+ def interpolate(
+ self,
+ images_and_prompts: List[Union[str, PIL.Image.Image, torch.FloatTensor]],
+ weights: List[float],
+ num_images_per_prompt: int = 1,
+ num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ negative_prior_prompt: Optional[str] = None,
+ negative_prompt: str = "",
+ guidance_scale: float = 4.0,
+ device=None,
+ ):
+ """
+ Function invoked when using the prior pipeline for interpolation.
+
+ Args:
+ images_and_prompts (`List[Union[str, PIL.Image.Image, torch.FloatTensor]]`):
+ list of prompts and images to guide the image generation.
+ weights: (`List[float]`):
+ list of weights for each condition in `images_and_prompts`
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ negative_prior_prompt (`str`, *optional*):
+ The prompt not to guide the prior diffusion process. Ignored when not using guidance (i.e., ignored if
+ `guidance_scale` is less than `1`).
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt not to guide the image generation. Ignored when not using guidance (i.e., ignored if
+ `guidance_scale` is less than `1`).
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+
+ Examples:
+
+ Returns:
+ [`KandinskyPriorPipelineOutput`] or `tuple`
+ """
+
+ device = device or self.device
+
+ if len(images_and_prompts) != len(weights):
+ raise ValueError(
+ f"`images_and_prompts` contains {len(images_and_prompts)} items and `weights` contains {len(weights)} items - they should be lists of same length"
+ )
+
+ image_embeddings = []
+ for cond, weight in zip(images_and_prompts, weights):
+ if isinstance(cond, str):
+ image_emb = self(
+ cond,
+ num_inference_steps=num_inference_steps,
+ num_images_per_prompt=num_images_per_prompt,
+ generator=generator,
+ latents=latents,
+ negative_prompt=negative_prior_prompt,
+ guidance_scale=guidance_scale,
+ ).image_embeds
+
+ elif isinstance(cond, (PIL.Image.Image, torch.Tensor)):
+ if isinstance(cond, PIL.Image.Image):
+ cond = (
+ self.image_processor(cond, return_tensors="pt")
+ .pixel_values[0]
+ .unsqueeze(0)
+ .to(dtype=self.image_encoder.dtype, device=device)
+ )
+
+ image_emb = self.image_encoder(cond)["image_embeds"]
+
+ else:
+ raise ValueError(
+ f"`images_and_prompts` can only contains elements to be of type `str`, `PIL.Image.Image` or `torch.Tensor` but is {type(cond)}"
+ )
+
+ image_embeddings.append(image_emb * weight)
+
+ image_emb = torch.cat(image_embeddings).sum(dim=0, keepdim=True)
+
+ out_zero = self(
+ negative_prompt,
+ num_inference_steps=num_inference_steps,
+ num_images_per_prompt=num_images_per_prompt,
+ generator=generator,
+ latents=latents,
+ negative_prompt=negative_prior_prompt,
+ guidance_scale=guidance_scale,
+ )
+ zero_image_emb = out_zero.negative_image_embeds if negative_prompt == "" else out_zero.image_embeds
+
+ return KandinskyPriorPipelineOutput(image_embeds=image_emb, negative_image_embeds=zero_image_emb)
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def get_zero_embed(self, batch_size=1, device=None):
+ device = device or self.device
+ zero_img = torch.zeros(1, 3, self.image_encoder.config.image_size, self.image_encoder.config.image_size).to(
+ device=device, dtype=self.image_encoder.dtype
+ )
+ zero_image_emb = self.image_encoder(zero_img)["image_embeds"]
+ zero_image_emb = zero_image_emb.repeat(batch_size, 1)
+ return zero_image_emb
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ ):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ text_mask = text_inputs.attention_mask.bool().to(device)
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+
+ text_encoder_output = self.text_encoder(text_input_ids.to(device))
+
+ prompt_embeds = text_encoder_output.text_embeds
+ text_encoder_hidden_states = text_encoder_output.last_hidden_state
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_text_mask = uncond_input.attention_mask.bool().to(device)
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device))
+
+ negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds
+ uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_encoder_hidden_states.shape[1]
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_encoder_hidden_states, text_mask
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: int = 1,
+ num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ guidance_scale: float = 4.0,
+ output_type: Optional[str] = "pt",
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ output_type (`str`, *optional*, defaults to `"pt"`):
+ The output format of the generate image. Choose between: `"np"` (`np.array`) or `"pt"`
+ (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`KandinskyPriorPipelineOutput`] or `tuple`
+ """
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ elif not isinstance(prompt, list):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+ elif not isinstance(negative_prompt, list) and negative_prompt is not None:
+ raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+ # if the negative prompt is defined we double the batch size to
+ # directly retrieve the negative prompt embedding
+ if negative_prompt is not None:
+ prompt = prompt + negative_prompt
+ negative_prompt = 2 * negative_prompt
+
+ device = self._execution_device
+
+ batch_size = len(prompt)
+ batch_size = batch_size * num_images_per_prompt
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+ prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+ )
+
+ # prior
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ prior_timesteps_tensor = self.scheduler.timesteps
+
+ embedding_dim = self.prior.config.embedding_dim
+
+ latents = self.prepare_latents(
+ (batch_size, embedding_dim),
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+
+ for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ predicted_image_embedding = self.prior(
+ latent_model_input,
+ timestep=t,
+ proj_embedding=prompt_embeds,
+ encoder_hidden_states=text_encoder_hidden_states,
+ attention_mask=text_mask,
+ ).predicted_image_embedding
+
+ if do_classifier_free_guidance:
+ predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2)
+ predicted_image_embedding = predicted_image_embedding_uncond + guidance_scale * (
+ predicted_image_embedding_text - predicted_image_embedding_uncond
+ )
+
+ if i + 1 == prior_timesteps_tensor.shape[0]:
+ prev_timestep = None
+ else:
+ prev_timestep = prior_timesteps_tensor[i + 1]
+
+ latents = self.scheduler.step(
+ predicted_image_embedding,
+ timestep=t,
+ sample=latents,
+ generator=generator,
+ prev_timestep=prev_timestep,
+ ).prev_sample
+
+ latents = self.prior.post_process_latents(latents)
+
+ image_embeddings = latents
+
+ # if negative prompt has been defined, we retrieve split the image embedding into two
+ if negative_prompt is None:
+ zero_embeds = self.get_zero_embed(latents.shape[0], device=latents.device)
+
+ self.maybe_free_model_hooks()
+ else:
+ image_embeddings, zero_embeds = image_embeddings.chunk(2)
+
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.prior_hook.offload()
+
+ if output_type not in ["pt", "np"]:
+ raise ValueError(f"Only the output types `pt` and `np` are supported not output_type={output_type}")
+
+ if output_type == "np":
+ image_embeddings = image_embeddings.cpu().numpy()
+ zero_embeds = zero_embeds.cpu().numpy()
+
+ if not return_dict:
+ return (image_embeddings, zero_embeds)
+
+ return KandinskyPriorPipelineOutput(image_embeds=image_embeddings, negative_image_embeds=zero_embeds)
diff --git a/model/diffusers_c/pipelines/kandinsky/text_encoder.py b/model/diffusers_c/pipelines/kandinsky/text_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..caa0029f00ca22818819d5b76b57ec489c6da1d6
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky/text_encoder.py
@@ -0,0 +1,27 @@
+import torch
+from transformers import PreTrainedModel, XLMRobertaConfig, XLMRobertaModel
+
+
+class MCLIPConfig(XLMRobertaConfig):
+ model_type = "M-CLIP"
+
+ def __init__(self, transformerDimSize=1024, imageDimSize=768, **kwargs):
+ self.transformerDimensions = transformerDimSize
+ self.numDims = imageDimSize
+ super().__init__(**kwargs)
+
+
+class MultilingualCLIP(PreTrainedModel):
+ config_class = MCLIPConfig
+
+ def __init__(self, config, *args, **kwargs):
+ super().__init__(config, *args, **kwargs)
+ self.transformer = XLMRobertaModel(config)
+ self.LinearTransformation = torch.nn.Linear(
+ in_features=config.transformerDimensions, out_features=config.numDims
+ )
+
+ def forward(self, input_ids, attention_mask):
+ embs = self.transformer(input_ids=input_ids, attention_mask=attention_mask)[0]
+ embs2 = (embs * attention_mask.unsqueeze(2)).sum(dim=1) / attention_mask.sum(dim=1)[:, None]
+ return self.LinearTransformation(embs2), embs
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__init__.py b/model/diffusers_c/pipelines/kandinsky2_2/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..67e97f161173ac8981dadf757fd8d6438307c973
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/__init__.py
@@ -0,0 +1,70 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_kandinsky2_2"] = ["KandinskyV22Pipeline"]
+ _import_structure["pipeline_kandinsky2_2_combined"] = [
+ "KandinskyV22CombinedPipeline",
+ "KandinskyV22Img2ImgCombinedPipeline",
+ "KandinskyV22InpaintCombinedPipeline",
+ ]
+ _import_structure["pipeline_kandinsky2_2_controlnet"] = ["KandinskyV22ControlnetPipeline"]
+ _import_structure["pipeline_kandinsky2_2_controlnet_img2img"] = ["KandinskyV22ControlnetImg2ImgPipeline"]
+ _import_structure["pipeline_kandinsky2_2_img2img"] = ["KandinskyV22Img2ImgPipeline"]
+ _import_structure["pipeline_kandinsky2_2_inpainting"] = ["KandinskyV22InpaintPipeline"]
+ _import_structure["pipeline_kandinsky2_2_prior"] = ["KandinskyV22PriorPipeline"]
+ _import_structure["pipeline_kandinsky2_2_prior_emb2emb"] = ["KandinskyV22PriorEmb2EmbPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_kandinsky2_2 import KandinskyV22Pipeline
+ from .pipeline_kandinsky2_2_combined import (
+ KandinskyV22CombinedPipeline,
+ KandinskyV22Img2ImgCombinedPipeline,
+ KandinskyV22InpaintCombinedPipeline,
+ )
+ from .pipeline_kandinsky2_2_controlnet import KandinskyV22ControlnetPipeline
+ from .pipeline_kandinsky2_2_controlnet_img2img import KandinskyV22ControlnetImg2ImgPipeline
+ from .pipeline_kandinsky2_2_img2img import KandinskyV22Img2ImgPipeline
+ from .pipeline_kandinsky2_2_inpainting import KandinskyV22InpaintPipeline
+ from .pipeline_kandinsky2_2_prior import KandinskyV22PriorPipeline
+ from .pipeline_kandinsky2_2_prior_emb2emb import KandinskyV22PriorEmb2EmbPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a65ce28c84e3adbcfebbf70376463a56718ee814
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2d5fc3f5abba98b27f99b2c0180f46681f887945
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_combined.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_combined.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6d80a24b9e41a3e0781dbbe065825e555b3168c0
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_combined.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_controlnet.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_controlnet.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..379f413072a67ffea6bfb8bf8910eabd4f26eb7c
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_controlnet.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_controlnet_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_controlnet_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..19c0dbf5d6d75544802ba289f525fc8ba81a8233
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_controlnet_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3c2c891b5909057c431bd2ffa799f1010bd76c40
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_inpainting.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_inpainting.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..89dcc7bc1edb127d6d8a64e959b111661c88c289
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_inpainting.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_prior.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_prior.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..400252a6da2c6bef788c531ac3ac1a46d28d5852
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_prior.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_prior_emb2emb.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_prior_emb2emb.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..fc0ce094be986161adecd1b0f6a8af8140ba6119
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky2_2/__pycache__/pipeline_kandinsky2_2_prior_emb2emb.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py
new file mode 100644
index 0000000000000000000000000000000000000000..4b977af0d6f792a31ef4bdaa9609134eac4b3430
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2.py
@@ -0,0 +1,320 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Dict, List, Optional, Union
+
+import torch
+
+from ...models import UNet2DConditionModel, VQModel
+from ...schedulers import DDPMScheduler
+from ...utils import deprecate, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyV22Pipeline, KandinskyV22PriorPipeline
+ >>> import torch
+
+ >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-prior")
+ >>> pipe_prior.to("cuda")
+ >>> prompt = "red cat, 4k photo"
+ >>> out = pipe_prior(prompt)
+ >>> image_emb = out.image_embeds
+ >>> zero_image_emb = out.negative_image_embeds
+ >>> pipe = KandinskyV22Pipeline.from_pretrained("kandinsky-community/kandinsky-2-2-decoder")
+ >>> pipe.to("cuda")
+ >>> image = pipe(
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=zero_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=50,
+ ... ).images
+ >>> image[0].save("cat.png")
+ ```
+"""
+
+
+def downscale_height_and_width(height, width, scale_factor=8):
+ new_height = height // scale_factor**2
+ if height % scale_factor**2 != 0:
+ new_height += 1
+ new_width = width // scale_factor**2
+ if width % scale_factor**2 != 0:
+ new_width += 1
+ return new_height * scale_factor, new_width * scale_factor
+
+
+class KandinskyV22Pipeline(DiffusionPipeline):
+ """
+ Pipeline for text-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ scheduler (Union[`DDIMScheduler`,`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ """
+
+ model_cpu_offload_seq = "unet->movq"
+ _callback_tensor_inputs = ["latents", "image_embeds", "negative_image_embeds"]
+
+ def __init__(
+ self,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+ self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ negative_image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for text prompt, that will be used to condition the image generation.
+ negative_image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for negative text prompt, will be used to condition the image generation.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ device = self._execution_device
+
+ self._guidance_scale = guidance_scale
+
+ if isinstance(image_embeds, list):
+ image_embeds = torch.cat(image_embeds, dim=0)
+ batch_size = image_embeds.shape[0] * num_images_per_prompt
+ if isinstance(negative_image_embeds, list):
+ negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
+
+ if self.do_classifier_free_guidance:
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to(
+ dtype=self.unet.dtype, device=device
+ )
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ num_channels_latents = self.unet.config.in_channels
+
+ height, width = downscale_height_and_width(height, width, self.movq_scale_factor)
+
+ # create initial latent
+ latents = self.prepare_latents(
+ (batch_size, num_channels_latents, height, width),
+ image_embeds.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+
+ self._num_timesteps = len(timesteps)
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ added_cond_kwargs = {"image_embeds": image_embeds}
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=None,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ if self.do_classifier_free_guidance:
+ noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1)
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ _, variance_pred_text = variance_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1)
+
+ if not (
+ hasattr(self.scheduler.config, "variance_type")
+ and self.scheduler.config.variance_type in ["learned", "learned_range"]
+ ):
+ noise_pred, _ = noise_pred.split(latents.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ )[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ image_embeds = callback_outputs.pop("image_embeds", image_embeds)
+ negative_image_embeds = callback_outputs.pop("negative_image_embeds", negative_image_embeds)
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if output_type not in ["pt", "np", "pil", "latent"]:
+ raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}")
+
+ if not output_type == "latent":
+ # post-processing
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+ else:
+ image = latents
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py
new file mode 100644
index 0000000000000000000000000000000000000000..65ba22cd6e26035f5ca6777af75a2cdc537288c5
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_combined.py
@@ -0,0 +1,851 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Dict, List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...models import PriorTransformer, UNet2DConditionModel, VQModel
+from ...schedulers import DDPMScheduler, UnCLIPScheduler
+from ...utils import deprecate, logging, replace_example_docstring
+from ..pipeline_utils import DiffusionPipeline
+from .pipeline_kandinsky2_2 import KandinskyV22Pipeline
+from .pipeline_kandinsky2_2_img2img import KandinskyV22Img2ImgPipeline
+from .pipeline_kandinsky2_2_inpainting import KandinskyV22InpaintPipeline
+from .pipeline_kandinsky2_2_prior import KandinskyV22PriorPipeline
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+TEXT2IMAGE_EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import AutoPipelineForText2Image
+ import torch
+
+ pipe = AutoPipelineForText2Image.from_pretrained(
+ "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16
+ )
+ pipe.enable_model_cpu_offload()
+
+ prompt = "A lion in galaxies, spirals, nebulae, stars, smoke, iridescent, intricate detail, octane render, 8k"
+
+ image = pipe(prompt=prompt, num_inference_steps=25).images[0]
+ ```
+"""
+
+IMAGE2IMAGE_EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import AutoPipelineForImage2Image
+ import torch
+ import requests
+ from io import BytesIO
+ from PIL import Image
+ import os
+
+ pipe = AutoPipelineForImage2Image.from_pretrained(
+ "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16
+ )
+ pipe.enable_model_cpu_offload()
+
+ prompt = "A fantasy landscape, Cinematic lighting"
+ negative_prompt = "low quality, bad quality"
+
+ url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+
+ response = requests.get(url)
+ image = Image.open(BytesIO(response.content)).convert("RGB")
+ image.thumbnail((768, 768))
+
+ image = pipe(prompt=prompt, image=original_image, num_inference_steps=25).images[0]
+ ```
+"""
+
+INPAINT_EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import AutoPipelineForInpainting
+ from diffusers.utils import load_image
+ import torch
+ import numpy as np
+
+ pipe = AutoPipelineForInpainting.from_pretrained(
+ "kandinsky-community/kandinsky-2-2-decoder-inpaint", torch_dtype=torch.float16
+ )
+ pipe.enable_model_cpu_offload()
+
+ prompt = "A fantasy landscape, Cinematic lighting"
+ negative_prompt = "low quality, bad quality"
+
+ original_image = load_image(
+ "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main" "/kandinsky/cat.png"
+ )
+
+ mask = np.zeros((768, 768), dtype=np.float32)
+ # Let's mask out an area above the cat's head
+ mask[:250, 250:-250] = 1
+
+ image = pipe(prompt=prompt, image=original_image, mask_image=mask, num_inference_steps=25).images[0]
+ ```
+"""
+
+
+class KandinskyV22CombinedPipeline(DiffusionPipeline):
+ """
+ Combined Pipeline for text-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ scheduler (Union[`DDIMScheduler`,`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ prior_prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ prior_image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ prior_text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ prior_tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ prior_scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ prior_image_processor ([`CLIPImageProcessor`]):
+ A image_processor to be used to preprocess image from clip.
+ """
+
+ model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->unet->movq"
+ _load_connected_pipes = True
+
+ def __init__(
+ self,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ prior_prior: PriorTransformer,
+ prior_image_encoder: CLIPVisionModelWithProjection,
+ prior_text_encoder: CLIPTextModelWithProjection,
+ prior_tokenizer: CLIPTokenizer,
+ prior_scheduler: UnCLIPScheduler,
+ prior_image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ prior_prior=prior_prior,
+ prior_image_encoder=prior_image_encoder,
+ prior_text_encoder=prior_text_encoder,
+ prior_tokenizer=prior_tokenizer,
+ prior_scheduler=prior_scheduler,
+ prior_image_processor=prior_image_processor,
+ )
+ self.prior_pipe = KandinskyV22PriorPipeline(
+ prior=prior_prior,
+ image_encoder=prior_image_encoder,
+ text_encoder=prior_text_encoder,
+ tokenizer=prior_tokenizer,
+ scheduler=prior_scheduler,
+ image_processor=prior_image_processor,
+ )
+ self.decoder_pipe = KandinskyV22Pipeline(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
+
+ def enable_sequential_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+ text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+ `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+ Note that offloading happens on a submodule basis. Memory savings are higher than with
+ `enable_model_cpu_offload`, but performance is lower.
+ """
+ self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+
+ def progress_bar(self, iterable=None, total=None):
+ self.prior_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.enable_model_cpu_offload()
+
+ def set_progress_bar_config(self, **kwargs):
+ self.prior_pipe.set_progress_bar_config(**kwargs)
+ self.decoder_pipe.set_progress_bar_config(**kwargs)
+
+ @torch.no_grad()
+ @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ height: int = 512,
+ width: int = 512,
+ prior_guidance_scale: float = 4.0,
+ prior_num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ prior_callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ prior_callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ prior_num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ prior_callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference of the prior pipeline.
+ The function is called with the following arguments: `prior_callback_on_step_end(self:
+ DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`.
+ prior_callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the
+ list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in
+ the `._callback_tensor_inputs` attribute of your prior pipeline class.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference of the decoder pipeline.
+ The function is called with the following arguments: `callback_on_step_end(self: DiffusionPipeline,
+ step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors
+ as specified by `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ prior_outputs = self.prior_pipe(
+ prompt=prompt,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ num_inference_steps=prior_num_inference_steps,
+ generator=generator,
+ latents=latents,
+ guidance_scale=prior_guidance_scale,
+ output_type="pt",
+ return_dict=False,
+ callback_on_step_end=prior_callback_on_step_end,
+ callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs,
+ )
+ image_embeds = prior_outputs[0]
+ negative_image_embeds = prior_outputs[1]
+
+ prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
+
+ if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
+ prompt = (image_embeds.shape[0] // len(prompt)) * prompt
+
+ outputs = self.decoder_pipe(
+ image_embeds=image_embeds,
+ negative_image_embeds=negative_image_embeds,
+ width=width,
+ height=height,
+ num_inference_steps=num_inference_steps,
+ generator=generator,
+ guidance_scale=guidance_scale,
+ output_type=output_type,
+ callback=callback,
+ callback_steps=callback_steps,
+ return_dict=return_dict,
+ callback_on_step_end=callback_on_step_end,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ )
+ self.maybe_free_model_hooks()
+
+ return outputs
+
+
+class KandinskyV22Img2ImgCombinedPipeline(DiffusionPipeline):
+ """
+ Combined Pipeline for image-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ scheduler (Union[`DDIMScheduler`,`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ prior_prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ prior_image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ prior_text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ prior_tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ prior_scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ prior_image_processor ([`CLIPImageProcessor`]):
+ A image_processor to be used to preprocess image from clip.
+ """
+
+ model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->unet->movq"
+ _load_connected_pipes = True
+
+ def __init__(
+ self,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ prior_prior: PriorTransformer,
+ prior_image_encoder: CLIPVisionModelWithProjection,
+ prior_text_encoder: CLIPTextModelWithProjection,
+ prior_tokenizer: CLIPTokenizer,
+ prior_scheduler: UnCLIPScheduler,
+ prior_image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ prior_prior=prior_prior,
+ prior_image_encoder=prior_image_encoder,
+ prior_text_encoder=prior_text_encoder,
+ prior_tokenizer=prior_tokenizer,
+ prior_scheduler=prior_scheduler,
+ prior_image_processor=prior_image_processor,
+ )
+ self.prior_pipe = KandinskyV22PriorPipeline(
+ prior=prior_prior,
+ image_encoder=prior_image_encoder,
+ text_encoder=prior_text_encoder,
+ tokenizer=prior_tokenizer,
+ scheduler=prior_scheduler,
+ image_processor=prior_image_processor,
+ )
+ self.decoder_pipe = KandinskyV22Img2ImgPipeline(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
+
+ def enable_model_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+ """
+ self.prior_pipe.enable_model_cpu_offload()
+ self.decoder_pipe.enable_model_cpu_offload()
+
+ def enable_sequential_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+ text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+ `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+ Note that offloading happens on a submodule basis. Memory savings are higher than with
+ `enable_model_cpu_offload`, but performance is lower.
+ """
+ self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+
+ def progress_bar(self, iterable=None, total=None):
+ self.prior_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.enable_model_cpu_offload()
+
+ def set_progress_bar_config(self, **kwargs):
+ self.prior_pipe.set_progress_bar_config(**kwargs)
+ self.decoder_pipe.set_progress_bar_config(**kwargs)
+
+ @torch.no_grad()
+ @replace_example_docstring(IMAGE2IMAGE_EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ strength: float = 0.3,
+ num_images_per_prompt: int = 1,
+ height: int = 512,
+ width: int = 512,
+ prior_guidance_scale: float = 4.0,
+ prior_num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ prior_callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ prior_callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded
+ again.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ strength (`float`, *optional*, defaults to 0.3):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ prior_num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ prior_outputs = self.prior_pipe(
+ prompt=prompt,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ num_inference_steps=prior_num_inference_steps,
+ generator=generator,
+ latents=latents,
+ guidance_scale=prior_guidance_scale,
+ output_type="pt",
+ return_dict=False,
+ callback_on_step_end=prior_callback_on_step_end,
+ callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs,
+ )
+ image_embeds = prior_outputs[0]
+ negative_image_embeds = prior_outputs[1]
+
+ prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
+ image = [image] if isinstance(prompt, PIL.Image.Image) else image
+
+ if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
+ prompt = (image_embeds.shape[0] // len(prompt)) * prompt
+
+ if (
+ isinstance(image, (list, tuple))
+ and len(image) < image_embeds.shape[0]
+ and image_embeds.shape[0] % len(image) == 0
+ ):
+ image = (image_embeds.shape[0] // len(image)) * image
+
+ outputs = self.decoder_pipe(
+ image=image,
+ image_embeds=image_embeds,
+ negative_image_embeds=negative_image_embeds,
+ width=width,
+ height=height,
+ strength=strength,
+ num_inference_steps=num_inference_steps,
+ generator=generator,
+ guidance_scale=guidance_scale,
+ output_type=output_type,
+ callback=callback,
+ callback_steps=callback_steps,
+ return_dict=return_dict,
+ callback_on_step_end=callback_on_step_end,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ )
+
+ self.maybe_free_model_hooks()
+ return outputs
+
+
+class KandinskyV22InpaintCombinedPipeline(DiffusionPipeline):
+ """
+ Combined Pipeline for inpainting generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ scheduler (Union[`DDIMScheduler`,`DDPMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ prior_prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ prior_image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ prior_text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ prior_tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ prior_scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ prior_image_processor ([`CLIPImageProcessor`]):
+ A image_processor to be used to preprocess image from clip.
+ """
+
+ model_cpu_offload_seq = "prior_text_encoder->prior_image_encoder->unet->movq"
+ _load_connected_pipes = True
+
+ def __init__(
+ self,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ prior_prior: PriorTransformer,
+ prior_image_encoder: CLIPVisionModelWithProjection,
+ prior_text_encoder: CLIPTextModelWithProjection,
+ prior_tokenizer: CLIPTokenizer,
+ prior_scheduler: UnCLIPScheduler,
+ prior_image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ prior_prior=prior_prior,
+ prior_image_encoder=prior_image_encoder,
+ prior_text_encoder=prior_text_encoder,
+ prior_tokenizer=prior_tokenizer,
+ prior_scheduler=prior_scheduler,
+ prior_image_processor=prior_image_processor,
+ )
+ self.prior_pipe = KandinskyV22PriorPipeline(
+ prior=prior_prior,
+ image_encoder=prior_image_encoder,
+ text_encoder=prior_text_encoder,
+ tokenizer=prior_tokenizer,
+ scheduler=prior_scheduler,
+ image_processor=prior_image_processor,
+ )
+ self.decoder_pipe = KandinskyV22InpaintPipeline(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
+
+ def enable_sequential_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet,
+ text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a
+ `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called.
+ Note that offloading happens on a submodule basis. Memory savings are higher than with
+ `enable_model_cpu_offload`, but performance is lower.
+ """
+ self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+
+ def progress_bar(self, iterable=None, total=None):
+ self.prior_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.enable_model_cpu_offload()
+
+ def set_progress_bar_config(self, **kwargs):
+ self.prior_pipe.set_progress_bar_config(**kwargs)
+ self.decoder_pipe.set_progress_bar_config(**kwargs)
+
+ @torch.no_grad()
+ @replace_example_docstring(INPAINT_EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ height: int = 512,
+ width: int = 512,
+ prior_guidance_scale: float = 4.0,
+ prior_num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ prior_callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ prior_callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded
+ again.
+ mask_image (`np.array`):
+ Tensor representing an image batch, to mask `image`. White pixels in the mask will be repainted, while
+ black pixels will be preserved. If `mask_image` is a PIL image, it will be converted to a single
+ channel (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3,
+ so the expected shape would be `(B, H, W, 1)`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ prior_num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ prior_callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `prior_callback_on_step_end(self: DiffusionPipeline, step: int, timestep:
+ int, callback_kwargs: Dict)`.
+ prior_callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the
+ list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in
+ the `._callback_tensor_inputs` attribute of your pipeline class.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ prior_kwargs = {}
+ if kwargs.get("prior_callback", None) is not None:
+ prior_kwargs["callback"] = kwargs.pop("prior_callback")
+ deprecate(
+ "prior_callback",
+ "1.0.0",
+ "Passing `prior_callback` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`",
+ )
+ if kwargs.get("prior_callback_steps", None) is not None:
+ deprecate(
+ "prior_callback_steps",
+ "1.0.0",
+ "Passing `prior_callback_steps` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`",
+ )
+ prior_kwargs["callback_steps"] = kwargs.pop("prior_callback_steps")
+
+ prior_outputs = self.prior_pipe(
+ prompt=prompt,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ num_inference_steps=prior_num_inference_steps,
+ generator=generator,
+ latents=latents,
+ guidance_scale=prior_guidance_scale,
+ output_type="pt",
+ return_dict=False,
+ callback_on_step_end=prior_callback_on_step_end,
+ callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs,
+ **prior_kwargs,
+ )
+ image_embeds = prior_outputs[0]
+ negative_image_embeds = prior_outputs[1]
+
+ prompt = [prompt] if not isinstance(prompt, (list, tuple)) else prompt
+ image = [image] if isinstance(prompt, PIL.Image.Image) else image
+ mask_image = [mask_image] if isinstance(mask_image, PIL.Image.Image) else mask_image
+
+ if len(prompt) < image_embeds.shape[0] and image_embeds.shape[0] % len(prompt) == 0:
+ prompt = (image_embeds.shape[0] // len(prompt)) * prompt
+
+ if (
+ isinstance(image, (list, tuple))
+ and len(image) < image_embeds.shape[0]
+ and image_embeds.shape[0] % len(image) == 0
+ ):
+ image = (image_embeds.shape[0] // len(image)) * image
+
+ if (
+ isinstance(mask_image, (list, tuple))
+ and len(mask_image) < image_embeds.shape[0]
+ and image_embeds.shape[0] % len(mask_image) == 0
+ ):
+ mask_image = (image_embeds.shape[0] // len(mask_image)) * mask_image
+
+ outputs = self.decoder_pipe(
+ image=image,
+ mask_image=mask_image,
+ image_embeds=image_embeds,
+ negative_image_embeds=negative_image_embeds,
+ width=width,
+ height=height,
+ num_inference_steps=num_inference_steps,
+ generator=generator,
+ guidance_scale=guidance_scale,
+ output_type=output_type,
+ return_dict=return_dict,
+ callback_on_step_end=callback_on_step_end,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ **kwargs,
+ )
+ self.maybe_free_model_hooks()
+
+ return outputs
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet.py b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..de87dd3c345df456f79386a40fbe5ed6638afda2
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet.py
@@ -0,0 +1,320 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, List, Optional, Union
+
+import torch
+
+from ...models import UNet2DConditionModel, VQModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ logging,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> import numpy as np
+
+ >>> from diffusers import KandinskyV22PriorPipeline, KandinskyV22ControlnetPipeline
+ >>> from transformers import pipeline
+ >>> from diffusers.utils import load_image
+
+
+ >>> def make_hint(image, depth_estimator):
+ ... image = depth_estimator(image)["depth"]
+ ... image = np.array(image)
+ ... image = image[:, :, None]
+ ... image = np.concatenate([image, image, image], axis=2)
+ ... detected_map = torch.from_numpy(image).float() / 255.0
+ ... hint = detected_map.permute(2, 0, 1)
+ ... return hint
+
+
+ >>> depth_estimator = pipeline("depth-estimation")
+
+ >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior = pipe_prior.to("cuda")
+
+ >>> pipe = KandinskyV22ControlnetPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-controlnet-depth", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+
+ >>> img = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... ).resize((768, 768))
+
+ >>> hint = make_hint(img, depth_estimator).unsqueeze(0).half().to("cuda")
+
+ >>> prompt = "A robot, 4k photo"
+ >>> negative_prior_prompt = "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature"
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(43)
+
+ >>> image_emb, zero_image_emb = pipe_prior(
+ ... prompt=prompt, negative_prompt=negative_prior_prompt, generator=generator
+ ... ).to_tuple()
+
+ >>> images = pipe(
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=zero_image_emb,
+ ... hint=hint,
+ ... num_inference_steps=50,
+ ... generator=generator,
+ ... height=768,
+ ... width=768,
+ ... ).images
+
+ >>> images[0].save("robot_cat.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.downscale_height_and_width
+def downscale_height_and_width(height, width, scale_factor=8):
+ new_height = height // scale_factor**2
+ if height % scale_factor**2 != 0:
+ new_height += 1
+ new_width = width // scale_factor**2
+ if width % scale_factor**2 != 0:
+ new_width += 1
+ return new_height * scale_factor, new_width * scale_factor
+
+
+class KandinskyV22ControlnetPipeline(DiffusionPipeline):
+ """
+ Pipeline for text-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ scheduler ([`DDIMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ """
+
+ model_cpu_offload_seq = "unet->movq"
+
+ def __init__(
+ self,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+ self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ negative_image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ hint: torch.FloatTensor,
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ hint (`torch.FloatTensor`):
+ The controlnet condition.
+ image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for text prompt, that will be used to condition the image generation.
+ negative_image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for negative text prompt, will be used to condition the image generation.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ device = self._execution_device
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ if isinstance(image_embeds, list):
+ image_embeds = torch.cat(image_embeds, dim=0)
+ if isinstance(negative_image_embeds, list):
+ negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
+ if isinstance(hint, list):
+ hint = torch.cat(hint, dim=0)
+
+ batch_size = image_embeds.shape[0] * num_images_per_prompt
+
+ if do_classifier_free_guidance:
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ hint = hint.repeat_interleave(num_images_per_prompt, dim=0)
+
+ image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to(
+ dtype=self.unet.dtype, device=device
+ )
+ hint = torch.cat([hint, hint], dim=0).to(dtype=self.unet.dtype, device=device)
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps_tensor = self.scheduler.timesteps
+
+ num_channels_latents = self.movq.config.latent_channels
+
+ height, width = downscale_height_and_width(height, width, self.movq_scale_factor)
+
+ # create initial latent
+ latents = self.prepare_latents(
+ (batch_size, num_channels_latents, height, width),
+ image_embeds.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+
+ for i, t in enumerate(self.progress_bar(timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ added_cond_kwargs = {"image_embeds": image_embeds, "hint": hint}
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=None,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ if do_classifier_free_guidance:
+ noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1)
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ _, variance_pred_text = variance_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1)
+
+ if not (
+ hasattr(self.scheduler.config, "variance_type")
+ and self.scheduler.config.variance_type in ["learned", "learned_range"]
+ ):
+ noise_pred, _ = noise_pred.split(latents.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ )[0]
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+ # post-processing
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if output_type not in ["pt", "np", "pil"]:
+ raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}")
+
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet_img2img.py b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..c3ac7bcf6099aa33be2785224b17fdb5618fcc6b
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_controlnet_img2img.py
@@ -0,0 +1,381 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from PIL import Image
+
+from ...models import UNet2DConditionModel, VQModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ logging,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> import numpy as np
+
+ >>> from diffusers import KandinskyV22PriorEmb2EmbPipeline, KandinskyV22ControlnetImg2ImgPipeline
+ >>> from transformers import pipeline
+ >>> from diffusers.utils import load_image
+
+
+ >>> def make_hint(image, depth_estimator):
+ ... image = depth_estimator(image)["depth"]
+ ... image = np.array(image)
+ ... image = image[:, :, None]
+ ... image = np.concatenate([image, image, image], axis=2)
+ ... detected_map = torch.from_numpy(image).float() / 255.0
+ ... hint = detected_map.permute(2, 0, 1)
+ ... return hint
+
+
+ >>> depth_estimator = pipeline("depth-estimation")
+
+ >>> pipe_prior = KandinskyV22PriorEmb2EmbPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior = pipe_prior.to("cuda")
+
+ >>> pipe = KandinskyV22ControlnetImg2ImgPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-controlnet-depth", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> img = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... ).resize((768, 768))
+
+
+ >>> hint = make_hint(img, depth_estimator).unsqueeze(0).half().to("cuda")
+
+ >>> prompt = "A robot, 4k photo"
+ >>> negative_prior_prompt = "lowres, text, error, cropped, worst quality, low quality, jpeg artifacts, ugly, duplicate, morbid, mutilated, out of frame, extra fingers, mutated hands, poorly drawn hands, poorly drawn face, mutation, deformed, blurry, dehydrated, bad anatomy, bad proportions, extra limbs, cloned face, disfigured, gross proportions, malformed limbs, missing arms, missing legs, extra arms, extra legs, fused fingers, too many fingers, long neck, username, watermark, signature"
+
+ >>> generator = torch.Generator(device="cuda").manual_seed(43)
+
+ >>> img_emb = pipe_prior(prompt=prompt, image=img, strength=0.85, generator=generator)
+ >>> negative_emb = pipe_prior(prompt=negative_prior_prompt, image=img, strength=1, generator=generator)
+
+ >>> images = pipe(
+ ... image=img,
+ ... strength=0.5,
+ ... image_embeds=img_emb.image_embeds,
+ ... negative_image_embeds=negative_emb.image_embeds,
+ ... hint=hint,
+ ... num_inference_steps=50,
+ ... generator=generator,
+ ... height=768,
+ ... width=768,
+ ... ).images
+
+ >>> images[0].save("robot_cat.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.downscale_height_and_width
+def downscale_height_and_width(height, width, scale_factor=8):
+ new_height = height // scale_factor**2
+ if height % scale_factor**2 != 0:
+ new_height += 1
+ new_width = width // scale_factor**2
+ if width % scale_factor**2 != 0:
+ new_width += 1
+ return new_height * scale_factor, new_width * scale_factor
+
+
+# Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_img2img.prepare_image
+def prepare_image(pil_image, w=512, h=512):
+ pil_image = pil_image.resize((w, h), resample=Image.BICUBIC, reducing_gap=1)
+ arr = np.array(pil_image.convert("RGB"))
+ arr = arr.astype(np.float32) / 127.5 - 1
+ arr = np.transpose(arr, [2, 0, 1])
+ image = torch.from_numpy(arr).unsqueeze(0)
+ return image
+
+
+class KandinskyV22ControlnetImg2ImgPipeline(DiffusionPipeline):
+ """
+ Pipeline for image-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ scheduler ([`DDIMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ """
+
+ model_cpu_offload_seq = "unet->movq"
+
+ def __init__(
+ self,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+ self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+
+ # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_img2img.KandinskyImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2_img2img.KandinskyV22Img2ImgPipeline.prepare_latents
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ self.movq.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = self.movq.encode(image).latent_dist.sample(generator)
+
+ init_latents = self.movq.config.scaling_factor * init_latents
+
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+
+ latents = init_latents
+
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ negative_image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ hint: torch.FloatTensor,
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ strength: float = 0.3,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "pil",
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for text prompt, that will be used to condition the image generation.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded
+ again.
+ strength (`float`, *optional*, defaults to 0.8):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ hint (`torch.FloatTensor`):
+ The controlnet condition.
+ negative_image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for negative text prompt, will be used to condition the image generation.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ device = self._execution_device
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ if isinstance(image_embeds, list):
+ image_embeds = torch.cat(image_embeds, dim=0)
+ if isinstance(negative_image_embeds, list):
+ negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
+ if isinstance(hint, list):
+ hint = torch.cat(hint, dim=0)
+
+ batch_size = image_embeds.shape[0]
+
+ if do_classifier_free_guidance:
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ hint = hint.repeat_interleave(num_images_per_prompt, dim=0)
+
+ image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to(
+ dtype=self.unet.dtype, device=device
+ )
+ hint = torch.cat([hint, hint], dim=0).to(dtype=self.unet.dtype, device=device)
+
+ if not isinstance(image, list):
+ image = [image]
+ if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image):
+ raise ValueError(
+ f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support PIL image and pytorch tensor"
+ )
+
+ image = torch.cat([prepare_image(i, width, height) for i in image], dim=0)
+ image = image.to(dtype=image_embeds.dtype, device=device)
+
+ latents = self.movq.encode(image)["latents"]
+ latents = latents.repeat_interleave(num_images_per_prompt, dim=0)
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ height, width = downscale_height_and_width(height, width, self.movq_scale_factor)
+ latents = self.prepare_latents(
+ latents, latent_timestep, batch_size, num_images_per_prompt, image_embeds.dtype, device, generator
+ )
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ added_cond_kwargs = {"image_embeds": image_embeds, "hint": hint}
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=None,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ if do_classifier_free_guidance:
+ noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1)
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ _, variance_pred_text = variance_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1)
+
+ if not (
+ hasattr(self.scheduler.config, "variance_type")
+ and self.scheduler.config.variance_type in ["learned", "learned_range"]
+ ):
+ noise_pred, _ = noise_pred.split(latents.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ )[0]
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # post-processing
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if output_type not in ["pt", "np", "pil"]:
+ raise ValueError(f"Only the output types `pt`, `pil` and `np` are supported not output_type={output_type}")
+
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..3fdae934adfe29d468e8ff2705955821cdd1126e
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_img2img.py
@@ -0,0 +1,399 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from PIL import Image
+
+from ...models import UNet2DConditionModel, VQModel
+from ...schedulers import DDPMScheduler
+from ...utils import deprecate, logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyV22Img2ImgPipeline, KandinskyV22PriorPipeline
+ >>> from diffusers.utils import load_image
+ >>> import torch
+
+ >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior.to("cuda")
+
+ >>> prompt = "A red cartoon frog, 4k"
+ >>> image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False)
+
+ >>> pipe = KandinskyV22Img2ImgPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16
+ ... )
+ >>> pipe.to("cuda")
+
+ >>> init_image = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/frog.png"
+ ... )
+
+ >>> image = pipe(
+ ... image=init_image,
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=zero_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=100,
+ ... strength=0.2,
+ ... ).images
+
+ >>> image[0].save("red_frog.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.downscale_height_and_width
+def downscale_height_and_width(height, width, scale_factor=8):
+ new_height = height // scale_factor**2
+ if height % scale_factor**2 != 0:
+ new_height += 1
+ new_width = width // scale_factor**2
+ if width % scale_factor**2 != 0:
+ new_width += 1
+ return new_height * scale_factor, new_width * scale_factor
+
+
+# Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_img2img.prepare_image
+def prepare_image(pil_image, w=512, h=512):
+ pil_image = pil_image.resize((w, h), resample=Image.BICUBIC, reducing_gap=1)
+ arr = np.array(pil_image.convert("RGB"))
+ arr = arr.astype(np.float32) / 127.5 - 1
+ arr = np.transpose(arr, [2, 0, 1])
+ image = torch.from_numpy(arr).unsqueeze(0)
+ return image
+
+
+class KandinskyV22Img2ImgPipeline(DiffusionPipeline):
+ """
+ Pipeline for image-to-image generation using Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ scheduler ([`DDIMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ """
+
+ model_cpu_offload_seq = "unet->movq"
+ _callback_tensor_inputs = ["latents", "image_embeds", "negative_image_embeds"]
+
+ def __init__(
+ self,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+ self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+
+ # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_img2img.KandinskyImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ self.movq.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = self.movq.encode(image).latent_dist.sample(generator)
+
+ init_latents = self.movq.config.scaling_factor * init_latents
+
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+
+ latents = init_latents
+
+ return latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]],
+ negative_image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ strength: float = 0.3,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for text prompt, that will be used to condition the image generation.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process. Can also accept image latents as `image`, if passing latents directly, it will not be encoded
+ again.
+ strength (`float`, *optional*, defaults to 0.8):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ negative_image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for negative text prompt, will be used to condition the image generation.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ device = self._execution_device
+
+ self._guidance_scale = guidance_scale
+
+ if isinstance(image_embeds, list):
+ image_embeds = torch.cat(image_embeds, dim=0)
+ batch_size = image_embeds.shape[0]
+ if isinstance(negative_image_embeds, list):
+ negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
+
+ if self.do_classifier_free_guidance:
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to(
+ dtype=self.unet.dtype, device=device
+ )
+
+ if not isinstance(image, list):
+ image = [image]
+ if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image):
+ raise ValueError(
+ f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support PIL image and pytorch tensor"
+ )
+
+ image = torch.cat([prepare_image(i, width, height) for i in image], dim=0)
+ image = image.to(dtype=image_embeds.dtype, device=device)
+
+ latents = self.movq.encode(image)["latents"]
+ latents = latents.repeat_interleave(num_images_per_prompt, dim=0)
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ height, width = downscale_height_and_width(height, width, self.movq_scale_factor)
+ latents = self.prepare_latents(
+ latents, latent_timestep, batch_size, num_images_per_prompt, image_embeds.dtype, device, generator
+ )
+ self._num_timesteps = len(timesteps)
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ added_cond_kwargs = {"image_embeds": image_embeds}
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=None,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ if self.do_classifier_free_guidance:
+ noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1)
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ _, variance_pred_text = variance_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1)
+
+ if not (
+ hasattr(self.scheduler.config, "variance_type")
+ and self.scheduler.config.variance_type in ["learned", "learned_range"]
+ ):
+ noise_pred, _ = noise_pred.split(latents.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ )[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ image_embeds = callback_outputs.pop("image_embeds", image_embeds)
+ negative_image_embeds = callback_outputs.pop("negative_image_embeds", negative_image_embeds)
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if output_type not in ["pt", "np", "pil", "latent"]:
+ raise ValueError(
+ f"Only the output types `pt`, `pil` ,`np` and `latent` are supported not output_type={output_type}"
+ )
+
+ if not output_type == "latent":
+ # post-processing
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+ else:
+ image = latents
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpainting.py b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpainting.py
new file mode 100644
index 0000000000000000000000000000000000000000..4983ed7218d6ff1917fecec1ce1c70c3d7c42ae4
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_inpainting.py
@@ -0,0 +1,556 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from copy import deepcopy
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from packaging import version
+from PIL import Image
+
+from ... import __version__
+from ...models import UNet2DConditionModel, VQModel
+from ...schedulers import DDPMScheduler
+from ...utils import deprecate, logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyV22InpaintPipeline, KandinskyV22PriorPipeline
+ >>> from diffusers.utils import load_image
+ >>> import torch
+ >>> import numpy as np
+
+ >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior.to("cuda")
+
+ >>> prompt = "a hat"
+ >>> image_emb, zero_image_emb = pipe_prior(prompt, return_dict=False)
+
+ >>> pipe = KandinskyV22InpaintPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-decoder-inpaint", torch_dtype=torch.float16
+ ... )
+ >>> pipe.to("cuda")
+
+ >>> init_image = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... )
+
+ >>> mask = np.zeros((768, 768), dtype=np.float32)
+ >>> mask[:250, 250:-250] = 1
+
+ >>> out = pipe(
+ ... image=init_image,
+ ... mask_image=mask,
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=zero_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=50,
+ ... )
+
+ >>> image = out.images[0]
+ >>> image.save("cat_with_hat.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.kandinsky2_2.pipeline_kandinsky2_2.downscale_height_and_width
+def downscale_height_and_width(height, width, scale_factor=8):
+ new_height = height // scale_factor**2
+ if height % scale_factor**2 != 0:
+ new_height += 1
+ new_width = width // scale_factor**2
+ if width % scale_factor**2 != 0:
+ new_width += 1
+ return new_height * scale_factor, new_width * scale_factor
+
+
+# Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_inpaint.prepare_mask
+def prepare_mask(masks):
+ prepared_masks = []
+ for mask in masks:
+ old_mask = deepcopy(mask)
+ for i in range(mask.shape[1]):
+ for j in range(mask.shape[2]):
+ if old_mask[0][i][j] == 1:
+ continue
+ if i != 0:
+ mask[:, i - 1, j] = 0
+ if j != 0:
+ mask[:, i, j - 1] = 0
+ if i != 0 and j != 0:
+ mask[:, i - 1, j - 1] = 0
+ if i != mask.shape[1] - 1:
+ mask[:, i + 1, j] = 0
+ if j != mask.shape[2] - 1:
+ mask[:, i, j + 1] = 0
+ if i != mask.shape[1] - 1 and j != mask.shape[2] - 1:
+ mask[:, i + 1, j + 1] = 0
+ prepared_masks.append(mask)
+ return torch.stack(prepared_masks, dim=0)
+
+
+# Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_inpaint.prepare_mask_and_masked_image
+def prepare_mask_and_masked_image(image, mask, height, width):
+ r"""
+ Prepares a pair (mask, image) to be consumed by the Kandinsky inpaint pipeline. This means that those inputs will
+ be converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for
+ the ``image`` and ``1`` for the ``mask``.
+
+ The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+ binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+ Args:
+ image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+ ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+ mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+ ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+
+
+ Raises:
+ ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+ should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+ TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+ (ot the other way around).
+
+ Returns:
+ tuple[torch.Tensor]: The pair (mask, image) as ``torch.Tensor`` with 4
+ dimensions: ``batch x channels x height x width``.
+ """
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ if mask is None:
+ raise ValueError("`mask_image` input cannot be undefined.")
+
+ if isinstance(image, torch.Tensor):
+ if not isinstance(mask, torch.Tensor):
+ raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
+
+ # Batch single image
+ if image.ndim == 3:
+ assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
+ image = image.unsqueeze(0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Single batched mask, no channel dim or single mask not batched but channel dim
+ if mask.shape[0] == 1:
+ mask = mask.unsqueeze(0)
+
+ # Batched masks no channel dim
+ else:
+ mask = mask.unsqueeze(1)
+
+ assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+ assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+ assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+
+ # Check image is in [-1, 1]
+ if image.min() < -1 or image.max() > 1:
+ raise ValueError("Image should be in [-1, 1] range")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("Mask should be in [0, 1] range")
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ # Image as float32
+ image = image.to(dtype=torch.float32)
+ elif isinstance(mask, torch.Tensor):
+ raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+ else:
+ # preprocess image
+ if isinstance(image, (PIL.Image.Image, np.ndarray)):
+ image = [image]
+
+ if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+ # resize all images w.r.t passed height an width
+ image = [i.resize((width, height), resample=Image.BICUBIC, reducing_gap=1) for i in image]
+ image = [np.array(i.convert("RGB"))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+ image = np.concatenate([i[None, :] for i in image], axis=0)
+
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ # preprocess mask
+ if isinstance(mask, (PIL.Image.Image, np.ndarray)):
+ mask = [mask]
+
+ if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
+ mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
+ mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+ mask = mask.astype(np.float32) / 255.0
+ elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+ mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+ mask = torch.from_numpy(mask)
+
+ mask = 1 - mask
+
+ return mask, image
+
+
+class KandinskyV22InpaintPipeline(DiffusionPipeline):
+ """
+ Pipeline for text-guided image inpainting using Kandinsky2.1
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ scheduler ([`DDIMScheduler`]):
+ A scheduler to be used in combination with `unet` to generate image latents.
+ unet ([`UNet2DConditionModel`]):
+ Conditional U-Net architecture to denoise the image embedding.
+ movq ([`VQModel`]):
+ MoVQ Decoder to generate the image from the latents.
+ """
+
+ model_cpu_offload_seq = "unet->movq"
+ _callback_tensor_inputs = ["latents", "image_embeds", "negative_image_embeds", "masked_image", "mask_image"]
+
+ def __init__(
+ self,
+ unet: UNet2DConditionModel,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ unet=unet,
+ scheduler=scheduler,
+ movq=movq,
+ )
+ self.movq_scale_factor = 2 ** (len(self.movq.config.block_out_channels) - 1)
+ self._warn_has_been_called = False
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ image: Union[torch.FloatTensor, PIL.Image.Image],
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image, np.ndarray],
+ negative_image_embeds: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ height: int = 512,
+ width: int = 512,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 4.0,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for text prompt, that will be used to condition the image generation.
+ image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
+ be masked out with `mask_image` and repainted according to `prompt`.
+ mask_image (`np.array`):
+ Tensor representing an image batch, to mask `image`. White pixels in the mask will be repainted, while
+ black pixels will be preserved. If `mask_image` is a PIL image, it will be converted to a single
+ channel (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3,
+ so the expected shape would be `(B, H, W, 1)`.
+ negative_image_embeds (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ The clip image embeddings for negative text prompt, will be used to condition the image generation.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+ """
+ if not self._warn_has_been_called and version.parse(version.parse(__version__).base_version) < version.parse(
+ "0.23.0.dev0"
+ ):
+ logger.warn(
+ "Please note that the expected format of `mask_image` has recently been changed. "
+ "Before diffusers == 0.19.0, Kandinsky Inpainting pipelines repainted black pixels and preserved black pixels. "
+ "As of diffusers==0.19.0 this behavior has been inverted. Now white pixels are repainted and black pixels are preserved. "
+ "This way, Kandinsky's masking behavior is aligned with Stable Diffusion. "
+ "THIS means that you HAVE to invert the input mask to have the same behavior as before as explained in https://github.com/huggingface/diffusers/pull/4207. "
+ "This warning will be surpressed after the first inference call and will be removed in diffusers>0.23.0"
+ )
+ self._warn_has_been_called = True
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ self._guidance_scale = guidance_scale
+
+ device = self._execution_device
+
+ if isinstance(image_embeds, list):
+ image_embeds = torch.cat(image_embeds, dim=0)
+ batch_size = image_embeds.shape[0] * num_images_per_prompt
+ if isinstance(negative_image_embeds, list):
+ negative_image_embeds = torch.cat(negative_image_embeds, dim=0)
+
+ if self.do_classifier_free_guidance:
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ negative_image_embeds = negative_image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ image_embeds = torch.cat([negative_image_embeds, image_embeds], dim=0).to(
+ dtype=self.unet.dtype, device=device
+ )
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # preprocess image and mask
+ mask_image, image = prepare_mask_and_masked_image(image, mask_image, height, width)
+
+ image = image.to(dtype=image_embeds.dtype, device=device)
+ image = self.movq.encode(image)["latents"]
+
+ mask_image = mask_image.to(dtype=image_embeds.dtype, device=device)
+
+ image_shape = tuple(image.shape[-2:])
+ mask_image = F.interpolate(
+ mask_image,
+ image_shape,
+ mode="nearest",
+ )
+ mask_image = prepare_mask(mask_image)
+ masked_image = image * mask_image
+
+ mask_image = mask_image.repeat_interleave(num_images_per_prompt, dim=0)
+ masked_image = masked_image.repeat_interleave(num_images_per_prompt, dim=0)
+ if self.do_classifier_free_guidance:
+ mask_image = mask_image.repeat(2, 1, 1, 1)
+ masked_image = masked_image.repeat(2, 1, 1, 1)
+
+ num_channels_latents = self.movq.config.latent_channels
+
+ height, width = downscale_height_and_width(height, width, self.movq_scale_factor)
+
+ # create initial latent
+ latents = self.prepare_latents(
+ (batch_size, num_channels_latents, height, width),
+ image_embeds.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+ noise = torch.clone(latents)
+
+ self._num_timesteps = len(timesteps)
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = torch.cat([latent_model_input, masked_image, mask_image], dim=1)
+
+ added_cond_kwargs = {"image_embeds": image_embeds}
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=None,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ if self.do_classifier_free_guidance:
+ noise_pred, variance_pred = noise_pred.split(latents.shape[1], dim=1)
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ _, variance_pred_text = variance_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, variance_pred_text], dim=1)
+
+ if not (
+ hasattr(self.scheduler.config, "variance_type")
+ and self.scheduler.config.variance_type in ["learned", "learned_range"]
+ ):
+ noise_pred, _ = noise_pred.split(latents.shape[1], dim=1)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ )[0]
+ init_latents_proper = image[:1]
+ init_mask = mask_image[:1]
+
+ if i < len(timesteps) - 1:
+ noise_timestep = timesteps[i + 1]
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_proper, noise, torch.tensor([noise_timestep])
+ )
+
+ latents = init_mask * init_latents_proper + (1 - init_mask) * latents
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ image_embeds = callback_outputs.pop("image_embeds", image_embeds)
+ negative_image_embeds = callback_outputs.pop("negative_image_embeds", negative_image_embeds)
+ masked_image = callback_outputs.pop("masked_image", masked_image)
+ mask_image = callback_outputs.pop("mask_image", mask_image)
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # post-processing
+ latents = mask_image[:1] * image[:1] + (1 - mask_image[:1]) * latents
+
+ if output_type not in ["pt", "np", "pil", "latent"]:
+ raise ValueError(
+ f"Only the output types `pt`, `pil`, `np` and `latent` are supported not output_type={output_type}"
+ )
+
+ if not output_type == "latent":
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+ else:
+ image = latents
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py
new file mode 100644
index 0000000000000000000000000000000000000000..83427c68f20892a3503dd11034310252dc667870
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior.py
@@ -0,0 +1,549 @@
+from typing import Callable, Dict, List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...models import PriorTransformer
+from ...schedulers import UnCLIPScheduler
+from ...utils import (
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..kandinsky import KandinskyPriorPipelineOutput
+from ..pipeline_utils import DiffusionPipeline
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyV22Pipeline, KandinskyV22PriorPipeline
+ >>> import torch
+
+ >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained("kandinsky-community/kandinsky-2-2-prior")
+ >>> pipe_prior.to("cuda")
+ >>> prompt = "red cat, 4k photo"
+ >>> image_emb, negative_image_emb = pipe_prior(prompt).to_tuple()
+
+ >>> pipe = KandinskyV22Pipeline.from_pretrained("kandinsky-community/kandinsky-2-2-decoder")
+ >>> pipe.to("cuda")
+ >>> image = pipe(
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=negative_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=50,
+ ... ).images
+ >>> image[0].save("cat.png")
+ ```
+"""
+
+EXAMPLE_INTERPOLATE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyV22PriorPipeline, KandinskyV22Pipeline
+ >>> from diffusers.utils import load_image
+ >>> import PIL
+ >>> import torch
+ >>> from torchvision import transforms
+
+ >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior.to("cuda")
+ >>> img1 = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... )
+ >>> img2 = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/starry_night.jpeg"
+ ... )
+ >>> images_texts = ["a cat", img1, img2]
+ >>> weights = [0.3, 0.3, 0.4]
+ >>> out = pipe_prior.interpolate(images_texts, weights)
+ >>> pipe = KandinskyV22Pipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16
+ ... )
+ >>> pipe.to("cuda")
+ >>> image = pipe(
+ ... image_embeds=out.image_embeds,
+ ... negative_image_embeds=out.negative_image_embeds,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=50,
+ ... ).images[0]
+ >>> image.save("starry_cat.png")
+ ```
+"""
+
+
+class KandinskyV22PriorPipeline(DiffusionPipeline):
+ """
+ Pipeline for generating image prior for Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ image_processor ([`CLIPImageProcessor`]):
+ A image_processor to be used to preprocess image from clip.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->prior"
+ _exclude_from_cpu_offload = ["prior"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "text_encoder_hidden_states", "text_mask"]
+
+ def __init__(
+ self,
+ prior: PriorTransformer,
+ image_encoder: CLIPVisionModelWithProjection,
+ text_encoder: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ scheduler: UnCLIPScheduler,
+ image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ prior=prior,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ scheduler=scheduler,
+ image_encoder=image_encoder,
+ image_processor=image_processor,
+ )
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_INTERPOLATE_DOC_STRING)
+ def interpolate(
+ self,
+ images_and_prompts: List[Union[str, PIL.Image.Image, torch.FloatTensor]],
+ weights: List[float],
+ num_images_per_prompt: int = 1,
+ num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ negative_prior_prompt: Optional[str] = None,
+ negative_prompt: str = "",
+ guidance_scale: float = 4.0,
+ device=None,
+ ):
+ """
+ Function invoked when using the prior pipeline for interpolation.
+
+ Args:
+ images_and_prompts (`List[Union[str, PIL.Image.Image, torch.FloatTensor]]`):
+ list of prompts and images to guide the image generation.
+ weights: (`List[float]`):
+ list of weights for each condition in `images_and_prompts`
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ negative_prior_prompt (`str`, *optional*):
+ The prompt not to guide the prior diffusion process. Ignored when not using guidance (i.e., ignored if
+ `guidance_scale` is less than `1`).
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt not to guide the image generation. Ignored when not using guidance (i.e., ignored if
+ `guidance_scale` is less than `1`).
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+
+ Examples:
+
+ Returns:
+ [`KandinskyPriorPipelineOutput`] or `tuple`
+ """
+
+ device = device or self.device
+
+ if len(images_and_prompts) != len(weights):
+ raise ValueError(
+ f"`images_and_prompts` contains {len(images_and_prompts)} items and `weights` contains {len(weights)} items - they should be lists of same length"
+ )
+
+ image_embeddings = []
+ for cond, weight in zip(images_and_prompts, weights):
+ if isinstance(cond, str):
+ image_emb = self(
+ cond,
+ num_inference_steps=num_inference_steps,
+ num_images_per_prompt=num_images_per_prompt,
+ generator=generator,
+ latents=latents,
+ negative_prompt=negative_prior_prompt,
+ guidance_scale=guidance_scale,
+ ).image_embeds.unsqueeze(0)
+
+ elif isinstance(cond, (PIL.Image.Image, torch.Tensor)):
+ if isinstance(cond, PIL.Image.Image):
+ cond = (
+ self.image_processor(cond, return_tensors="pt")
+ .pixel_values[0]
+ .unsqueeze(0)
+ .to(dtype=self.image_encoder.dtype, device=device)
+ )
+
+ image_emb = self.image_encoder(cond)["image_embeds"].repeat(num_images_per_prompt, 1).unsqueeze(0)
+
+ else:
+ raise ValueError(
+ f"`images_and_prompts` can only contains elements to be of type `str`, `PIL.Image.Image` or `torch.Tensor` but is {type(cond)}"
+ )
+
+ image_embeddings.append(image_emb * weight)
+
+ image_emb = torch.cat(image_embeddings).sum(dim=0)
+
+ out_zero = self(
+ negative_prompt,
+ num_inference_steps=num_inference_steps,
+ num_images_per_prompt=num_images_per_prompt,
+ generator=generator,
+ latents=latents,
+ negative_prompt=negative_prior_prompt,
+ guidance_scale=guidance_scale,
+ )
+ zero_image_emb = out_zero.negative_image_embeds if negative_prompt == "" else out_zero.image_embeds
+
+ return KandinskyPriorPipelineOutput(image_embeds=image_emb, negative_image_embeds=zero_image_emb)
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline.get_zero_embed
+ def get_zero_embed(self, batch_size=1, device=None):
+ device = device or self.device
+ zero_img = torch.zeros(1, 3, self.image_encoder.config.image_size, self.image_encoder.config.image_size).to(
+ device=device, dtype=self.image_encoder.dtype
+ )
+ zero_image_emb = self.image_encoder(zero_img)["image_embeds"]
+ zero_image_emb = zero_image_emb.repeat(batch_size, 1)
+ return zero_image_emb
+
+ # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ ):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ text_mask = text_inputs.attention_mask.bool().to(device)
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+
+ text_encoder_output = self.text_encoder(text_input_ids.to(device))
+
+ prompt_embeds = text_encoder_output.text_embeds
+ text_encoder_hidden_states = text_encoder_output.last_hidden_state
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_text_mask = uncond_input.attention_mask.bool().to(device)
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device))
+
+ negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds
+ uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_encoder_hidden_states.shape[1]
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_encoder_hidden_states, text_mask
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: int = 1,
+ num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ guidance_scale: float = 4.0,
+ output_type: Optional[str] = "pt", # pt only
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ output_type (`str`, *optional*, defaults to `"pt"`):
+ The output format of the generate image. Choose between: `"np"` (`np.array`) or `"pt"`
+ (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`KandinskyPriorPipelineOutput`] or `tuple`
+ """
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ elif not isinstance(prompt, list):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+ elif not isinstance(negative_prompt, list) and negative_prompt is not None:
+ raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+ # if the negative prompt is defined we double the batch size to
+ # directly retrieve the negative prompt embedding
+ if negative_prompt is not None:
+ prompt = prompt + negative_prompt
+ negative_prompt = 2 * negative_prompt
+
+ device = self._execution_device
+
+ batch_size = len(prompt)
+ batch_size = batch_size * num_images_per_prompt
+
+ self._guidance_scale = guidance_scale
+
+ prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt(
+ prompt, device, num_images_per_prompt, self.do_classifier_free_guidance, negative_prompt
+ )
+
+ # prior
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ embedding_dim = self.prior.config.embedding_dim
+
+ latents = self.prepare_latents(
+ (batch_size, embedding_dim),
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+ self._num_timesteps = len(timesteps)
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ predicted_image_embedding = self.prior(
+ latent_model_input,
+ timestep=t,
+ proj_embedding=prompt_embeds,
+ encoder_hidden_states=text_encoder_hidden_states,
+ attention_mask=text_mask,
+ ).predicted_image_embedding
+
+ if self.do_classifier_free_guidance:
+ predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2)
+ predicted_image_embedding = predicted_image_embedding_uncond + self.guidance_scale * (
+ predicted_image_embedding_text - predicted_image_embedding_uncond
+ )
+
+ if i + 1 == timesteps.shape[0]:
+ prev_timestep = None
+ else:
+ prev_timestep = timesteps[i + 1]
+
+ latents = self.scheduler.step(
+ predicted_image_embedding,
+ timestep=t,
+ sample=latents,
+ generator=generator,
+ prev_timestep=prev_timestep,
+ ).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ text_encoder_hidden_states = callback_outputs.pop(
+ "text_encoder_hidden_states", text_encoder_hidden_states
+ )
+ text_mask = callback_outputs.pop("text_mask", text_mask)
+
+ latents = self.prior.post_process_latents(latents)
+
+ image_embeddings = latents
+
+ # if negative prompt has been defined, we retrieve split the image embedding into two
+ if negative_prompt is None:
+ zero_embeds = self.get_zero_embed(latents.shape[0], device=latents.device)
+ else:
+ image_embeddings, zero_embeds = image_embeddings.chunk(2)
+
+ self.maybe_free_model_hooks()
+
+ if output_type not in ["pt", "np"]:
+ raise ValueError(f"Only the output types `pt` and `np` are supported not output_type={output_type}")
+
+ if output_type == "np":
+ image_embeddings = image_embeddings.cpu().numpy()
+ zero_embeds = zero_embeds.cpu().numpy()
+
+ if not return_dict:
+ return (image_embeddings, zero_embeds)
+
+ return KandinskyPriorPipelineOutput(image_embeds=image_embeddings, negative_image_embeds=zero_embeds)
diff --git a/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior_emb2emb.py b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior_emb2emb.py
new file mode 100644
index 0000000000000000000000000000000000000000..bef70821c60530b5099c3d4aaa4f81df62190293
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky2_2/pipeline_kandinsky2_2_prior_emb2emb.py
@@ -0,0 +1,563 @@
+from typing import List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...models import PriorTransformer
+from ...schedulers import UnCLIPScheduler
+from ...utils import (
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..kandinsky import KandinskyPriorPipelineOutput
+from ..pipeline_utils import DiffusionPipeline
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyV22Pipeline, KandinskyV22PriorEmb2EmbPipeline
+ >>> import torch
+
+ >>> pipe_prior = KandinskyPriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior.to("cuda")
+
+ >>> prompt = "red cat, 4k photo"
+ >>> img = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... )
+ >>> image_emb, nagative_image_emb = pipe_prior(prompt, image=img, strength=0.2).to_tuple()
+
+ >>> pipe = KandinskyPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-decoder, torch_dtype=torch.float16"
+ ... )
+ >>> pipe.to("cuda")
+
+ >>> image = pipe(
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=negative_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=100,
+ ... ).images
+
+ >>> image[0].save("cat.png")
+ ```
+"""
+
+EXAMPLE_INTERPOLATE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import KandinskyV22PriorEmb2EmbPipeline, KandinskyV22Pipeline
+ >>> from diffusers.utils import load_image
+ >>> import PIL
+
+ >>> import torch
+ >>> from torchvision import transforms
+
+ >>> pipe_prior = KandinskyV22PriorPipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-prior", torch_dtype=torch.float16
+ ... )
+ >>> pipe_prior.to("cuda")
+
+ >>> img1 = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/cat.png"
+ ... )
+
+ >>> img2 = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main"
+ ... "/kandinsky/starry_night.jpeg"
+ ... )
+
+ >>> images_texts = ["a cat", img1, img2]
+ >>> weights = [0.3, 0.3, 0.4]
+ >>> image_emb, zero_image_emb = pipe_prior.interpolate(images_texts, weights)
+
+ >>> pipe = KandinskyV22Pipeline.from_pretrained(
+ ... "kandinsky-community/kandinsky-2-2-decoder", torch_dtype=torch.float16
+ ... )
+ >>> pipe.to("cuda")
+
+ >>> image = pipe(
+ ... image_embeds=image_emb,
+ ... negative_image_embeds=zero_image_emb,
+ ... height=768,
+ ... width=768,
+ ... num_inference_steps=150,
+ ... ).images[0]
+
+ >>> image.save("starry_cat.png")
+ ```
+"""
+
+
+class KandinskyV22PriorEmb2EmbPipeline(DiffusionPipeline):
+ """
+ Pipeline for generating image prior for Kandinsky
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen image-encoder.
+ text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ scheduler ([`UnCLIPScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->prior"
+ _exclude_from_cpu_offload = ["prior"]
+
+ def __init__(
+ self,
+ prior: PriorTransformer,
+ image_encoder: CLIPVisionModelWithProjection,
+ text_encoder: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ scheduler: UnCLIPScheduler,
+ image_processor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ prior=prior,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ scheduler=scheduler,
+ image_encoder=image_encoder,
+ image_processor=image_processor,
+ )
+
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_INTERPOLATE_DOC_STRING)
+ def interpolate(
+ self,
+ images_and_prompts: List[Union[str, PIL.Image.Image, torch.FloatTensor]],
+ weights: List[float],
+ num_images_per_prompt: int = 1,
+ num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ negative_prior_prompt: Optional[str] = None,
+ negative_prompt: str = "",
+ guidance_scale: float = 4.0,
+ device=None,
+ ):
+ """
+ Function invoked when using the prior pipeline for interpolation.
+
+ Args:
+ images_and_prompts (`List[Union[str, PIL.Image.Image, torch.FloatTensor]]`):
+ list of prompts and images to guide the image generation.
+ weights: (`List[float]`):
+ list of weights for each condition in `images_and_prompts`
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ negative_prior_prompt (`str`, *optional*):
+ The prompt not to guide the prior diffusion process. Ignored when not using guidance (i.e., ignored if
+ `guidance_scale` is less than `1`).
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt not to guide the image generation. Ignored when not using guidance (i.e., ignored if
+ `guidance_scale` is less than `1`).
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+
+ Examples:
+
+ Returns:
+ [`KandinskyPriorPipelineOutput`] or `tuple`
+ """
+
+ device = device or self.device
+
+ if len(images_and_prompts) != len(weights):
+ raise ValueError(
+ f"`images_and_prompts` contains {len(images_and_prompts)} items and `weights` contains {len(weights)} items - they should be lists of same length"
+ )
+
+ image_embeddings = []
+ for cond, weight in zip(images_and_prompts, weights):
+ if isinstance(cond, str):
+ image_emb = self(
+ cond,
+ num_inference_steps=num_inference_steps,
+ num_images_per_prompt=num_images_per_prompt,
+ generator=generator,
+ latents=latents,
+ negative_prompt=negative_prior_prompt,
+ guidance_scale=guidance_scale,
+ ).image_embeds.unsqueeze(0)
+
+ elif isinstance(cond, (PIL.Image.Image, torch.Tensor)):
+ image_emb = self._encode_image(
+ cond, device=device, num_images_per_prompt=num_images_per_prompt
+ ).unsqueeze(0)
+
+ else:
+ raise ValueError(
+ f"`images_and_prompts` can only contains elements to be of type `str`, `PIL.Image.Image` or `torch.Tensor` but is {type(cond)}"
+ )
+
+ image_embeddings.append(image_emb * weight)
+
+ image_emb = torch.cat(image_embeddings).sum(dim=0)
+
+ return KandinskyPriorPipelineOutput(image_embeds=image_emb, negative_image_embeds=torch.randn_like(image_emb))
+
+ def _encode_image(
+ self,
+ image: Union[torch.Tensor, List[PIL.Image.Image]],
+ device,
+ num_images_per_prompt,
+ ):
+ if not isinstance(image, torch.Tensor):
+ image = self.image_processor(image, return_tensors="pt").pixel_values.to(
+ dtype=self.image_encoder.dtype, device=device
+ )
+
+ image_emb = self.image_encoder(image)["image_embeds"] # B, D
+ image_emb = image_emb.repeat_interleave(num_images_per_prompt, dim=0)
+ image_emb.to(device=device)
+
+ return image_emb
+
+ def prepare_latents(self, emb, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ emb = emb.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ init_latents = emb
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+
+ return latents
+
+ # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline.get_zero_embed
+ def get_zero_embed(self, batch_size=1, device=None):
+ device = device or self.device
+ zero_img = torch.zeros(1, 3, self.image_encoder.config.image_size, self.image_encoder.config.image_size).to(
+ device=device, dtype=self.image_encoder.dtype
+ )
+ zero_image_emb = self.image_encoder(zero_img)["image_embeds"]
+ zero_image_emb = zero_image_emb.repeat(batch_size, 1)
+ return zero_image_emb
+
+ # Copied from diffusers.pipelines.kandinsky.pipeline_kandinsky_prior.KandinskyPriorPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ ):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ text_mask = text_inputs.attention_mask.bool().to(device)
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+
+ text_encoder_output = self.text_encoder(text_input_ids.to(device))
+
+ prompt_embeds = text_encoder_output.text_embeds
+ text_encoder_hidden_states = text_encoder_output.last_hidden_state
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_text_mask = uncond_input.attention_mask.bool().to(device)
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device))
+
+ negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds
+ uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_encoder_hidden_states.shape[1]
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_encoder_hidden_states, text_mask
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[torch.Tensor, List[torch.Tensor], PIL.Image.Image, List[PIL.Image.Image]],
+ strength: float = 0.3,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: int = 1,
+ num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ guidance_scale: float = 4.0,
+ output_type: Optional[str] = "pt", # pt only
+ return_dict: bool = True,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ strength (`float`, *optional*, defaults to 0.8):
+ Conceptually, indicates how much to transform the reference `emb`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added.
+ emb (`torch.FloatTensor`):
+ The image embedding.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ output_type (`str`, *optional*, defaults to `"pt"`):
+ The output format of the generate image. Choose between: `"np"` (`np.array`) or `"pt"`
+ (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`KandinskyPriorPipelineOutput`] or `tuple`
+ """
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ elif not isinstance(prompt, list):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+ elif not isinstance(negative_prompt, list) and negative_prompt is not None:
+ raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}")
+
+ # if the negative prompt is defined we double the batch size to
+ # directly retrieve the negative prompt embedding
+ if negative_prompt is not None:
+ prompt = prompt + negative_prompt
+ negative_prompt = 2 * negative_prompt
+
+ device = self._execution_device
+
+ batch_size = len(prompt)
+ batch_size = batch_size * num_images_per_prompt
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+ prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
+ )
+
+ if not isinstance(image, List):
+ image = [image]
+
+ if isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+
+ if isinstance(image, torch.Tensor) and image.ndim == 2:
+ # allow user to pass image_embeds directly
+ image_embeds = image.repeat_interleave(num_images_per_prompt, dim=0)
+ elif isinstance(image, torch.Tensor) and image.ndim != 4:
+ raise ValueError(
+ f" if pass `image` as pytorch tensor, or a list of pytorch tensor, please make sure each tensor has shape [batch_size, channels, height, width], currently {image[0].unsqueeze(0).shape}"
+ )
+ else:
+ image_embeds = self._encode_image(image, device, num_images_per_prompt)
+
+ # prior
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+
+ latents = image_embeds
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size)
+ latents = self.prepare_latents(
+ latents,
+ latent_timestep,
+ batch_size // num_images_per_prompt,
+ num_images_per_prompt,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ )
+
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ predicted_image_embedding = self.prior(
+ latent_model_input,
+ timestep=t,
+ proj_embedding=prompt_embeds,
+ encoder_hidden_states=text_encoder_hidden_states,
+ attention_mask=text_mask,
+ ).predicted_image_embedding
+
+ if do_classifier_free_guidance:
+ predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2)
+ predicted_image_embedding = predicted_image_embedding_uncond + guidance_scale * (
+ predicted_image_embedding_text - predicted_image_embedding_uncond
+ )
+
+ if i + 1 == timesteps.shape[0]:
+ prev_timestep = None
+ else:
+ prev_timestep = timesteps[i + 1]
+
+ latents = self.scheduler.step(
+ predicted_image_embedding,
+ timestep=t,
+ sample=latents,
+ generator=generator,
+ prev_timestep=prev_timestep,
+ ).prev_sample
+
+ latents = self.prior.post_process_latents(latents)
+
+ image_embeddings = latents
+
+ # if negative prompt has been defined, we retrieve split the image embedding into two
+ if negative_prompt is None:
+ zero_embeds = self.get_zero_embed(latents.shape[0], device=latents.device)
+ else:
+ image_embeddings, zero_embeds = image_embeddings.chunk(2)
+
+ self.maybe_free_model_hooks()
+
+ if output_type not in ["pt", "np"]:
+ raise ValueError(f"Only the output types `pt` and `np` are supported not output_type={output_type}")
+
+ if output_type == "np":
+ image_embeddings = image_embeddings.cpu().numpy()
+ zero_embeds = zero_embeds.cpu().numpy()
+
+ if not return_dict:
+ return (image_embeddings, zero_embeds)
+
+ return KandinskyPriorPipelineOutput(image_embeds=image_embeddings, negative_image_embeds=zero_embeds)
diff --git a/model/diffusers_c/pipelines/kandinsky3/__init__.py b/model/diffusers_c/pipelines/kandinsky3/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e8a3063141b5e62682ce3ad8e8e0d02473402665
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky3/__init__.py
@@ -0,0 +1,49 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_kandinsky3"] = ["Kandinsky3Pipeline"]
+ _import_structure["pipeline_kandinsky3_img2img"] = ["Kandinsky3Img2ImgPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_kandinsky3 import Kandinsky3Pipeline
+ from .pipeline_kandinsky3_img2img import Kandinsky3Img2ImgPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/kandinsky3/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky3/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2765069d8c9c3b0c34519b94705cdf28a51ae72c
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky3/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky3/__pycache__/pipeline_kandinsky3.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky3/__pycache__/pipeline_kandinsky3.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4c173a223427e892b31cd7b050d668009e265794
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky3/__pycache__/pipeline_kandinsky3.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky3/__pycache__/pipeline_kandinsky3_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/kandinsky3/__pycache__/pipeline_kandinsky3_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b336c555d6462f961a531467f427a9be875c5c0a
Binary files /dev/null and b/model/diffusers_c/pipelines/kandinsky3/__pycache__/pipeline_kandinsky3_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/kandinsky3/convert_kandinsky3_unet.py b/model/diffusers_c/pipelines/kandinsky3/convert_kandinsky3_unet.py
new file mode 100644
index 0000000000000000000000000000000000000000..4fe8c54eb7fc17150a6d927aeefd5ca21dce9636
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky3/convert_kandinsky3_unet.py
@@ -0,0 +1,98 @@
+#!/usr/bin/env python3
+import argparse
+import fnmatch
+
+from safetensors.torch import load_file
+
+from diffusers import Kandinsky3UNet
+
+
+MAPPING = {
+ "to_time_embed.1": "time_embedding.linear_1",
+ "to_time_embed.3": "time_embedding.linear_2",
+ "in_layer": "conv_in",
+ "out_layer.0": "conv_norm_out",
+ "out_layer.2": "conv_out",
+ "down_samples": "down_blocks",
+ "up_samples": "up_blocks",
+ "projection_lin": "encoder_hid_proj.projection_linear",
+ "projection_ln": "encoder_hid_proj.projection_norm",
+ "feature_pooling": "add_time_condition",
+ "to_query": "to_q",
+ "to_key": "to_k",
+ "to_value": "to_v",
+ "output_layer": "to_out.0",
+ "self_attention_block": "attentions.0",
+}
+
+DYNAMIC_MAP = {
+ "resnet_attn_blocks.*.0": "resnets_in.*",
+ "resnet_attn_blocks.*.1": ("attentions.*", 1),
+ "resnet_attn_blocks.*.2": "resnets_out.*",
+}
+# MAPPING = {}
+
+
+def convert_state_dict(unet_state_dict):
+ """
+ Convert the state dict of a U-Net model to match the key format expected by Kandinsky3UNet model.
+ Args:
+ unet_model (torch.nn.Module): The original U-Net model.
+ unet_kandi3_model (torch.nn.Module): The Kandinsky3UNet model to match keys with.
+
+ Returns:
+ OrderedDict: The converted state dictionary.
+ """
+ # Example of renaming logic (this will vary based on your model's architecture)
+ converted_state_dict = {}
+ for key in unet_state_dict:
+ new_key = key
+ for pattern, new_pattern in MAPPING.items():
+ new_key = new_key.replace(pattern, new_pattern)
+
+ for dyn_pattern, dyn_new_pattern in DYNAMIC_MAP.items():
+ has_matched = False
+ if fnmatch.fnmatch(new_key, f"*.{dyn_pattern}.*") and not has_matched:
+ star = int(new_key.split(dyn_pattern.split(".")[0])[-1].split(".")[1])
+
+ if isinstance(dyn_new_pattern, tuple):
+ new_star = star + dyn_new_pattern[-1]
+ dyn_new_pattern = dyn_new_pattern[0]
+ else:
+ new_star = star
+
+ pattern = dyn_pattern.replace("*", str(star))
+ new_pattern = dyn_new_pattern.replace("*", str(new_star))
+
+ new_key = new_key.replace(pattern, new_pattern)
+ has_matched = True
+
+ converted_state_dict[new_key] = unet_state_dict[key]
+
+ return converted_state_dict
+
+
+def main(model_path, output_path):
+ # Load your original U-Net model
+ unet_state_dict = load_file(model_path)
+
+ # Initialize your Kandinsky3UNet model
+ config = {}
+
+ # Convert the state dict
+ converted_state_dict = convert_state_dict(unet_state_dict)
+
+ unet = Kandinsky3UNet(config)
+ unet.load_state_dict(converted_state_dict)
+
+ unet.save_pretrained(output_path)
+ print(f"Converted model saved to {output_path}")
+
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser(description="Convert U-Net PyTorch model to Kandinsky3UNet format")
+ parser.add_argument("--model_path", type=str, required=True, help="Path to the original U-Net PyTorch model")
+ parser.add_argument("--output_path", type=str, required=True, help="Path to save the converted model")
+
+ args = parser.parse_args()
+ main(args.model_path, args.output_path)
diff --git a/model/diffusers_c/pipelines/kandinsky3/pipeline_kandinsky3.py b/model/diffusers_c/pipelines/kandinsky3/pipeline_kandinsky3.py
new file mode 100644
index 0000000000000000000000000000000000000000..fcf7ddcb9966ebdaabee4761d7c7015109deb865
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky3/pipeline_kandinsky3.py
@@ -0,0 +1,589 @@
+from typing import Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...models import Kandinsky3UNet, VQModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ deprecate,
+ is_accelerate_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import AutoPipelineForText2Image
+ >>> import torch
+
+ >>> pipe = AutoPipelineForText2Image.from_pretrained("kandinsky-community/kandinsky-3", variant="fp16", torch_dtype=torch.float16)
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = "A photograph of the inside of a subway train. There are raccoons sitting on the seats. One of them is reading a newspaper. The window shows the city in the background."
+
+ >>> generator = torch.Generator(device="cpu").manual_seed(0)
+ >>> image = pipe(prompt, num_inference_steps=25, generator=generator).images[0]
+ ```
+
+"""
+
+
+def downscale_height_and_width(height, width, scale_factor=8):
+ new_height = height // scale_factor**2
+ if height % scale_factor**2 != 0:
+ new_height += 1
+ new_width = width // scale_factor**2
+ if width % scale_factor**2 != 0:
+ new_width += 1
+ return new_height * scale_factor, new_width * scale_factor
+
+
+class Kandinsky3Pipeline(DiffusionPipeline, LoraLoaderMixin):
+ model_cpu_offload_seq = "text_encoder->unet->movq"
+ _callback_tensor_inputs = [
+ "latents",
+ "prompt_embeds",
+ "negative_prompt_embeds",
+ "negative_attention_mask",
+ "attention_mask",
+ ]
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ unet: Kandinsky3UNet,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ tokenizer=tokenizer, text_encoder=text_encoder, unet=unet, scheduler=scheduler, movq=movq
+ )
+
+ def remove_all_hooks(self):
+ if is_accelerate_available():
+ from accelerate.hooks import remove_hook_from_module
+ else:
+ raise ImportError("Please install accelerate via `pip install accelerate`")
+
+ for model in [self.text_encoder, self.unet, self.movq]:
+ if model is not None:
+ remove_hook_from_module(model, recurse=True)
+
+ self.unet_offload_hook = None
+ self.text_encoder_offload_hook = None
+ self.final_offload_hook = None
+
+ def process_embeds(self, embeddings, attention_mask, cut_context):
+ if cut_context:
+ embeddings[attention_mask == 0] = torch.zeros_like(embeddings[attention_mask == 0])
+ max_seq_length = attention_mask.sum(-1).max() + 1
+ embeddings = embeddings[:, :max_seq_length]
+ attention_mask = attention_mask[:, :max_seq_length]
+ return embeddings, attention_mask
+
+ @torch.no_grad()
+ def encode_prompt(
+ self,
+ prompt,
+ do_classifier_free_guidance=True,
+ num_images_per_prompt=1,
+ device=None,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ _cut_context=False,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ negative_attention_mask: Optional[torch.FloatTensor] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated attention mask. Must provide if passing `prompt_embeds` directly.
+ negative_attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated negative attention mask. Must provide if passing `negative_prompt_embeds` directly.
+ """
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ max_length = 128
+
+ if prompt_embeds is None:
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids.to(device)
+ attention_mask = text_inputs.attention_mask.to(device)
+ prompt_embeds = self.text_encoder(
+ text_input_ids,
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+ prompt_embeds, attention_mask = self.process_embeds(prompt_embeds, attention_mask, _cut_context)
+ prompt_embeds = prompt_embeds * attention_mask.unsqueeze(2)
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+ attention_mask = attention_mask.repeat(num_images_per_prompt, 1)
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+ if negative_prompt is not None:
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=128,
+ truncation=True,
+ return_attention_mask=True,
+ return_tensors="pt",
+ )
+ text_input_ids = uncond_input.input_ids.to(device)
+ negative_attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ text_input_ids,
+ attention_mask=negative_attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds[:, : prompt_embeds.shape[1]]
+ negative_attention_mask = negative_attention_mask[:, : prompt_embeds.shape[1]]
+ negative_prompt_embeds = negative_prompt_embeds * negative_attention_mask.unsqueeze(2)
+
+ else:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_attention_mask = torch.zeros_like(attention_mask)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+ if negative_prompt_embeds.shape != prompt_embeds.shape:
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+ negative_attention_mask = negative_attention_mask.repeat(num_images_per_prompt, 1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ else:
+ negative_prompt_embeds = None
+ negative_attention_mask = None
+ return prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask
+
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def check_inputs(
+ self,
+ prompt,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ attention_mask=None,
+ negative_attention_mask=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+ if negative_prompt_embeds is not None and negative_attention_mask is None:
+ raise ValueError("Please provide `negative_attention_mask` along with `negative_prompt_embeds`")
+
+ if negative_prompt_embeds is not None and negative_attention_mask is not None:
+ if negative_prompt_embeds.shape[:2] != negative_attention_mask.shape:
+ raise ValueError(
+ "`negative_prompt_embeds` and `negative_attention_mask` must have the same batch_size and token length when passed directly, but"
+ f" got: `negative_prompt_embeds` {negative_prompt_embeds.shape[:2]} != `negative_attention_mask`"
+ f" {negative_attention_mask.shape}."
+ )
+
+ if prompt_embeds is not None and attention_mask is None:
+ raise ValueError("Please provide `attention_mask` along with `prompt_embeds`")
+
+ if prompt_embeds is not None and attention_mask is not None:
+ if prompt_embeds.shape[:2] != attention_mask.shape:
+ raise ValueError(
+ "`prompt_embeds` and `attention_mask` must have the same batch_size and token length when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape[:2]} != `attention_mask`"
+ f" {attention_mask.shape}."
+ )
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ num_inference_steps: int = 25,
+ guidance_scale: float = 3.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ height: Optional[int] = 1024,
+ width: Optional[int] = 1024,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ negative_attention_mask: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ latents=None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 3.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size):
+ The width in pixels of the generated image.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated attention mask. Must provide if passing `prompt_embeds` directly.
+ negative_attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated negative attention mask. Must provide if passing `negative_prompt_embeds` directly.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ clean_caption (`bool`, *optional*, defaults to `True`):
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
+ prompt.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ cut_context = True
+ device = self._execution_device
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ callback_on_step_end_tensor_inputs,
+ attention_mask,
+ negative_attention_mask,
+ )
+
+ self._guidance_scale = guidance_scale
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask = self.encode_prompt(
+ prompt,
+ self.do_classifier_free_guidance,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ _cut_context=cut_context,
+ attention_mask=attention_mask,
+ negative_attention_mask=negative_attention_mask,
+ )
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ attention_mask = torch.cat([negative_attention_mask, attention_mask]).bool()
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latents
+ height, width = downscale_height_and_width(height, width, 8)
+
+ latents = self.prepare_latents(
+ (batch_size * num_images_per_prompt, 4, height, width),
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
+ self.text_encoder_offload_hook.offload()
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ encoder_attention_mask=attention_mask,
+ return_dict=False,
+ )[0]
+
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+
+ noise_pred = (guidance_scale + 1.0) * noise_pred_text - guidance_scale * noise_pred_uncond
+ # noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ ).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+ attention_mask = callback_outputs.pop("attention_mask", attention_mask)
+ negative_attention_mask = callback_outputs.pop("negative_attention_mask", negative_attention_mask)
+
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # post-processing
+ if output_type not in ["pt", "np", "pil", "latent"]:
+ raise ValueError(
+ f"Only the output types `pt`, `pil`, `np` and `latent` are supported not output_type={output_type}"
+ )
+
+ if not output_type == "latent":
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+ else:
+ image = latents
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/kandinsky3/pipeline_kandinsky3_img2img.py b/model/diffusers_c/pipelines/kandinsky3/pipeline_kandinsky3_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..7f4164a04d1edb6dcc97821e877908a1892fae3b
--- /dev/null
+++ b/model/diffusers_c/pipelines/kandinsky3/pipeline_kandinsky3_img2img.py
@@ -0,0 +1,654 @@
+import inspect
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL
+import PIL.Image
+import torch
+from transformers import T5EncoderModel, T5Tokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...models import Kandinsky3UNet, VQModel
+from ...schedulers import DDPMScheduler
+from ...utils import (
+ deprecate,
+ is_accelerate_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import AutoPipelineForImage2Image
+ >>> from diffusers.utils import load_image
+ >>> import torch
+
+ >>> pipe = AutoPipelineForImage2Image.from_pretrained("kandinsky-community/kandinsky-3", variant="fp16", torch_dtype=torch.float16)
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = "A painting of the inside of a subway train with tiny raccoons."
+ >>> image = load_image("https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/kandinsky3/t2i.png")
+
+ >>> generator = torch.Generator(device="cpu").manual_seed(0)
+ >>> image = pipe(prompt, image=image, strength=0.75, num_inference_steps=25, generator=generator).images[0]
+ ```
+"""
+
+
+def downscale_height_and_width(height, width, scale_factor=8):
+ new_height = height // scale_factor**2
+ if height % scale_factor**2 != 0:
+ new_height += 1
+ new_width = width // scale_factor**2
+ if width % scale_factor**2 != 0:
+ new_width += 1
+ return new_height * scale_factor, new_width * scale_factor
+
+
+def prepare_image(pil_image):
+ arr = np.array(pil_image.convert("RGB"))
+ arr = arr.astype(np.float32) / 127.5 - 1
+ arr = np.transpose(arr, [2, 0, 1])
+ image = torch.from_numpy(arr).unsqueeze(0)
+ return image
+
+
+class Kandinsky3Img2ImgPipeline(DiffusionPipeline, LoraLoaderMixin):
+ model_cpu_offload_seq = "text_encoder->movq->unet->movq"
+ _callback_tensor_inputs = [
+ "latents",
+ "prompt_embeds",
+ "negative_prompt_embeds",
+ "negative_attention_mask",
+ "attention_mask",
+ ]
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ unet: Kandinsky3UNet,
+ scheduler: DDPMScheduler,
+ movq: VQModel,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ tokenizer=tokenizer, text_encoder=text_encoder, unet=unet, scheduler=scheduler, movq=movq
+ )
+
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start:]
+
+ return timesteps, num_inference_steps - t_start
+
+ def remove_all_hooks(self):
+ if is_accelerate_available():
+ from accelerate.hooks import remove_hook_from_module
+ else:
+ raise ImportError("Please install accelerate via `pip install accelerate`")
+
+ for model in [self.text_encoder, self.unet]:
+ if model is not None:
+ remove_hook_from_module(model, recurse=True)
+
+ self.unet_offload_hook = None
+ self.text_encoder_offload_hook = None
+ self.final_offload_hook = None
+
+ def _process_embeds(self, embeddings, attention_mask, cut_context):
+ # return embeddings, attention_mask
+ if cut_context:
+ embeddings[attention_mask == 0] = torch.zeros_like(embeddings[attention_mask == 0])
+ max_seq_length = attention_mask.sum(-1).max() + 1
+ embeddings = embeddings[:, :max_seq_length]
+ attention_mask = attention_mask[:, :max_seq_length]
+ return embeddings, attention_mask
+
+ @torch.no_grad()
+ def encode_prompt(
+ self,
+ prompt,
+ do_classifier_free_guidance=True,
+ num_images_per_prompt=1,
+ device=None,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ _cut_context=False,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ negative_attention_mask: Optional[torch.FloatTensor] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated attention mask. Must provide if passing `prompt_embeds` directly.
+ negative_attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated negative attention mask. Must provide if passing `negative_prompt_embeds` directly.
+ """
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ max_length = 128
+
+ if prompt_embeds is None:
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids.to(device)
+ attention_mask = text_inputs.attention_mask.to(device)
+ prompt_embeds = self.text_encoder(
+ text_input_ids,
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+ prompt_embeds, attention_mask = self._process_embeds(prompt_embeds, attention_mask, _cut_context)
+ prompt_embeds = prompt_embeds * attention_mask.unsqueeze(2)
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+ attention_mask = attention_mask.repeat(num_images_per_prompt, 1)
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+ if negative_prompt is not None:
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=128,
+ truncation=True,
+ return_attention_mask=True,
+ return_tensors="pt",
+ )
+ text_input_ids = uncond_input.input_ids.to(device)
+ negative_attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ text_input_ids,
+ attention_mask=negative_attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds[:, : prompt_embeds.shape[1]]
+ negative_attention_mask = negative_attention_mask[:, : prompt_embeds.shape[1]]
+ negative_prompt_embeds = negative_prompt_embeds * negative_attention_mask.unsqueeze(2)
+
+ else:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_attention_mask = torch.zeros_like(attention_mask)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+ if negative_prompt_embeds.shape != prompt_embeds.shape:
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+ negative_attention_mask = negative_attention_mask.repeat(num_images_per_prompt, 1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ else:
+ negative_prompt_embeds = None
+ negative_attention_mask = None
+ return prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask
+
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ self.movq.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = self.movq.encode(image).latent_dist.sample(generator)
+
+ init_latents = self.movq.config.scaling_factor * init_latents
+
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+
+ latents = init_latents
+
+ return latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ attention_mask=None,
+ negative_attention_mask=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if negative_prompt_embeds is not None and negative_attention_mask is None:
+ raise ValueError("Please provide `negative_attention_mask` along with `negative_prompt_embeds`")
+
+ if negative_prompt_embeds is not None and negative_attention_mask is not None:
+ if negative_prompt_embeds.shape[:2] != negative_attention_mask.shape:
+ raise ValueError(
+ "`negative_prompt_embeds` and `negative_attention_mask` must have the same batch_size and token length when passed directly, but"
+ f" got: `negative_prompt_embeds` {negative_prompt_embeds.shape[:2]} != `negative_attention_mask`"
+ f" {negative_attention_mask.shape}."
+ )
+
+ if prompt_embeds is not None and attention_mask is None:
+ raise ValueError("Please provide `attention_mask` along with `prompt_embeds`")
+
+ if prompt_embeds is not None and attention_mask is not None:
+ if prompt_embeds.shape[:2] != attention_mask.shape:
+ raise ValueError(
+ "`prompt_embeds` and `attention_mask` must have the same batch_size and token length when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape[:2]} != `attention_mask`"
+ f" {attention_mask.shape}."
+ )
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]] = None,
+ strength: float = 0.3,
+ num_inference_steps: int = 25,
+ guidance_scale: float = 3.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ attention_mask: Optional[torch.FloatTensor] = None,
+ negative_attention_mask: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 3.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated attention mask. Must provide if passing `prompt_embeds` directly.
+ negative_attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated negative attention mask. Must provide if passing `negative_prompt_embeds` directly.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`
+
+ """
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ cut_context = True
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ callback_on_step_end_tensor_inputs,
+ attention_mask,
+ negative_attention_mask,
+ )
+
+ self._guidance_scale = guidance_scale
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds, attention_mask, negative_attention_mask = self.encode_prompt(
+ prompt,
+ self.do_classifier_free_guidance,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ _cut_context=cut_context,
+ attention_mask=attention_mask,
+ negative_attention_mask=negative_attention_mask,
+ )
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ attention_mask = torch.cat([negative_attention_mask, attention_mask]).bool()
+ if not isinstance(image, list):
+ image = [image]
+ if not all(isinstance(i, (PIL.Image.Image, torch.Tensor)) for i in image):
+ raise ValueError(
+ f"Input is in incorrect format: {[type(i) for i in image]}. Currently, we only support PIL image and pytorch tensor"
+ )
+
+ image = torch.cat([prepare_image(i) for i in image], dim=0)
+ image = image.to(dtype=prompt_embeds.dtype, device=device)
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ # 5. Prepare latents
+ latents = self.movq.encode(image)["latents"]
+ latents = latents.repeat_interleave(num_images_per_prompt, dim=0)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ latents = self.prepare_latents(
+ latents, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
+ )
+ if hasattr(self, "text_encoder_offload_hook") and self.text_encoder_offload_hook is not None:
+ self.text_encoder_offload_hook.offload()
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ encoder_attention_mask=attention_mask,
+ )[0]
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+
+ noise_pred = (guidance_scale + 1.0) * noise_pred_text - guidance_scale * noise_pred_uncond
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ noise_pred,
+ t,
+ latents,
+ generator=generator,
+ ).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+ attention_mask = callback_outputs.pop("attention_mask", attention_mask)
+ negative_attention_mask = callback_outputs.pop("negative_attention_mask", negative_attention_mask)
+
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # post-processing
+ if output_type not in ["pt", "np", "pil", "latent"]:
+ raise ValueError(
+ f"Only the output types `pt`, `pil`, `np` and `latent` are supported not output_type={output_type}"
+ )
+ if not output_type == "latent":
+ image = self.movq.decode(latents, force_not_quantize=True)["sample"]
+
+ if output_type in ["np", "pil"]:
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+ else:
+ image = latents
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/latent_consistency_models/__init__.py b/model/diffusers_c/pipelines/latent_consistency_models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8f79d3c4773f393ed689a949041d36ad77e20968
--- /dev/null
+++ b/model/diffusers_c/pipelines/latent_consistency_models/__init__.py
@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_latent_consistency_img2img"] = ["LatentConsistencyModelImg2ImgPipeline"]
+ _import_structure["pipeline_latent_consistency_text2img"] = ["LatentConsistencyModelPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_latent_consistency_img2img import LatentConsistencyModelImg2ImgPipeline
+ from .pipeline_latent_consistency_text2img import LatentConsistencyModelPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4e6e857c6cd21fc0ec7206329e6158b203f4d2fa
Binary files /dev/null and b/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/pipeline_latent_consistency_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/pipeline_latent_consistency_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9e8a4305de70a5c4f0a3955ce23524b213f0826a
Binary files /dev/null and b/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/pipeline_latent_consistency_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/pipeline_latent_consistency_text2img.cpython-39.pyc b/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/pipeline_latent_consistency_text2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..392c1fb38f56380615bae76e575dba81e4b08322
Binary files /dev/null and b/model/diffusers_c/pipelines/latent_consistency_models/__pycache__/pipeline_latent_consistency_text2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py b/model/diffusers_c/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..f64854ea982ba9dfdbf9f87558ec17bdccdb1e2e
--- /dev/null
+++ b/model/diffusers_c/pipelines/latent_consistency_models/pipeline_latent_consistency_img2img.py
@@ -0,0 +1,956 @@
+# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import LCMScheduler
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import AutoPipelineForImage2Image
+ >>> import torch
+ >>> import PIL
+
+ >>> pipe = AutoPipelineForImage2Image.from_pretrained("SimianLuo/LCM_Dreamshaper_v7")
+ >>> # To save GPU memory, torch.float16 can be used, but it may compromise image quality.
+ >>> pipe.to(torch_device="cuda", torch_dtype=torch.float32)
+
+ >>> prompt = "High altitude snowy mountains"
+ >>> image = PIL.Image.open("./snowy_mountains.png")
+
+ >>> # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
+ >>> num_inference_steps = 4
+ >>> images = pipe(
+ ... prompt=prompt, image=image, num_inference_steps=num_inference_steps, guidance_scale=8.0
+ ... ).images
+
+ >>> images[0].save("image.png")
+ ```
+
+"""
+
+
+class LatentConsistencyModelImg2ImgPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for image-to-image generation using a latent consistency model.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Currently only
+ supports [`LCMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ requires_safety_checker (`bool`, *optional*, defaults to `True`):
+ Whether the pipeline requires a safety checker component.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "denoised", "prompt_embeds", "w_embedding"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: LCMScheduler,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.prepare_latents
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+
+ return latents
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ def check_inputs(
+ self,
+ prompt: Union[str, List[str]],
+ strength: float,
+ callback_steps: int,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ @property
+ def do_classifier_free_guidance(self):
+ return False
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ num_inference_steps: int = 4,
+ strength: float = 0.8,
+ original_inference_steps: int = None,
+ timesteps: List[int] = None,
+ guidance_scale: float = 8.5,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ original_inference_steps (`int`, *optional*):
+ The original number of inference steps use to generate a linearly-spaced timestep schedule, from which
+ we will draw `num_inference_steps` evenly spaced timesteps from as our final timestep schedule,
+ following the Skipping-Step method in the paper (see Section 4.3). If not set this will default to the
+ scheduler's `original_inference_steps` attribute.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps on the original LCM training/distillation timestep schedule are used. Must be in descending
+ order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ Note that the original latent consistency models paper uses a different CFG formulation where the
+ guidance scales are decreased by 1 (so in the paper formulation CFG is enabled when `guidance_scale >
+ 0`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ strength,
+ callback_steps,
+ prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 3. Encode input prompt
+ lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ # NOTE: when a LCM is distilled from an LDM via latent consistency distillation (Algorithm 1) with guided
+ # distillation, the forward pass of the LCM learns to approximate sampling from the LDM using CFG with the
+ # unconditional prompt "" (the empty string). Due to this, LCMs currently do not support negative prompts.
+ prompt_embeds, _ = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=None,
+ lora_scale=lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 4. Encode image
+ image = self.image_processor.preprocess(image)
+
+ # 5. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(
+ self.scheduler,
+ num_inference_steps,
+ device,
+ timesteps,
+ original_inference_steps=original_inference_steps,
+ strength=strength,
+ )
+
+ # 6. Prepare latent variables
+ original_inference_steps = (
+ original_inference_steps
+ if original_inference_steps is not None
+ else self.scheduler.config.original_inference_steps
+ )
+ latent_timestep = timesteps[:1]
+ latents = self.prepare_latents(
+ image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
+ )
+ bs = batch_size * num_images_per_prompt
+
+ # 6. Get Guidance Scale Embedding
+ # NOTE: We use the Imagen CFG formulation that StableDiffusionPipeline uses rather than the original LCM paper
+ # CFG formulation, so we need to subtract 1 from the input guidance_scale.
+ # LCM CFG formulation: cfg_noise = noise_cond + cfg_scale * (noise_cond - noise_uncond), (cfg_scale > 0.0 using CFG)
+ w = torch.tensor(self.guidance_scale - 1).repeat(bs)
+ w_embedding = self.get_guidance_scale_embedding(w, embedding_dim=self.unet.config.time_cond_proj_dim).to(
+ device=device, dtype=latents.dtype
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, None)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 8. LCM Multistep Sampling Loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ latents = latents.to(prompt_embeds.dtype)
+
+ # model prediction (v-prediction, eps, x)
+ model_pred = self.unet(
+ latents,
+ t,
+ timestep_cond=w_embedding,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents, denoised = self.scheduler.step(model_pred, t, latents, **extra_step_kwargs, return_dict=False)
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ w_embedding = callback_outputs.pop("w_embedding", w_embedding)
+ denoised = callback_outputs.pop("denoised", denoised)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ denoised = denoised.to(prompt_embeds.dtype)
+ if not output_type == "latent":
+ image = self.vae.decode(denoised / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = denoised
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/latent_consistency_models/pipeline_latent_consistency_text2img.py b/model/diffusers_c/pipelines/latent_consistency_models/pipeline_latent_consistency_text2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..e9bacaa89ba55e06e8c8b18834f682989c3ffc62
--- /dev/null
+++ b/model/diffusers_c/pipelines/latent_consistency_models/pipeline_latent_consistency_text2img.py
@@ -0,0 +1,888 @@
+# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import LCMScheduler
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import DiffusionPipeline
+ >>> import torch
+
+ >>> pipe = DiffusionPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7")
+ >>> # To save GPU memory, torch.float16 can be used, but it may compromise image quality.
+ >>> pipe.to(torch_device="cuda", torch_dtype=torch.float32)
+
+ >>> prompt = "Self-portrait oil painting, a beautiful cyborg with golden hair, 8k"
+
+ >>> # Can be set to 1~50 steps. LCM support fast inference even <= 4 steps. Recommend: 1~8 steps.
+ >>> num_inference_steps = 4
+ >>> images = pipe(prompt=prompt, num_inference_steps=num_inference_steps, guidance_scale=8.0).images
+ >>> images[0].save("image.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class LatentConsistencyModelPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using a latent consistency model.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Currently only
+ supports [`LCMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ requires_safety_checker (`bool`, *optional*, defaults to `True`):
+ Whether the pipeline requires a safety checker component.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "denoised", "prompt_embeds", "w_embedding"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: LCMScheduler,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Currently StableDiffusionPipeline.check_inputs with negative prompt stuff removed
+ def check_inputs(
+ self,
+ prompt: Union[str, List[str]],
+ height: int,
+ width: int,
+ callback_steps: int,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ @property
+ def do_classifier_free_guidance(self):
+ return False
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 4,
+ original_inference_steps: int = None,
+ timesteps: List[int] = None,
+ guidance_scale: float = 8.5,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ original_inference_steps (`int`, *optional*):
+ The original number of inference steps use to generate a linearly-spaced timestep schedule, from which
+ we will draw `num_inference_steps` evenly spaced timesteps from as our final timestep schedule,
+ following the Skipping-Step method in the paper (see Section 4.3). If not set this will default to the
+ scheduler's `original_inference_steps` attribute.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps on the original LCM training/distillation timestep schedule are used. Must be in descending
+ order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ Note that the original latent consistency models paper uses a different CFG formulation where the
+ guidance scales are decreased by 1 (so in the paper formulation CFG is enabled when `guidance_scale >
+ 0`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ callback_steps,
+ prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 3. Encode input prompt
+ lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ # NOTE: when a LCM is distilled from an LDM via latent consistency distillation (Algorithm 1) with guided
+ # distillation, the forward pass of the LCM learns to approximate sampling from the LDM using CFG with the
+ # unconditional prompt "" (the empty string). Due to this, LCMs currently do not support negative prompts.
+ prompt_embeds, _ = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=None,
+ lora_scale=lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(
+ self.scheduler, num_inference_steps, device, timesteps, original_inference_steps=original_inference_steps
+ )
+
+ # 5. Prepare latent variable
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+ bs = batch_size * num_images_per_prompt
+
+ # 6. Get Guidance Scale Embedding
+ # NOTE: We use the Imagen CFG formulation that StableDiffusionPipeline uses rather than the original LCM paper
+ # CFG formulation, so we need to subtract 1 from the input guidance_scale.
+ # LCM CFG formulation: cfg_noise = noise_cond + cfg_scale * (noise_cond - noise_uncond), (cfg_scale > 0.0 using CFG)
+ w = torch.tensor(self.guidance_scale - 1).repeat(bs)
+ w_embedding = self.get_guidance_scale_embedding(w, embedding_dim=self.unet.config.time_cond_proj_dim).to(
+ device=device, dtype=latents.dtype
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, None)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 8. LCM MultiStep Sampling Loop:
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ latents = latents.to(prompt_embeds.dtype)
+
+ # model prediction (v-prediction, eps, x)
+ model_pred = self.unet(
+ latents,
+ t,
+ timestep_cond=w_embedding,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents, denoised = self.scheduler.step(model_pred, t, latents, **extra_step_kwargs, return_dict=False)
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ w_embedding = callback_outputs.pop("w_embedding", w_embedding)
+ denoised = callback_outputs.pop("denoised", denoised)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ denoised = denoised.to(prompt_embeds.dtype)
+ if not output_type == "latent":
+ image = self.vae.decode(denoised / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = denoised
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/latent_diffusion/__init__.py b/model/diffusers_c/pipelines/latent_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..561f96fc71dc7b4404e09571e0b7eaa4ee02fde8
--- /dev/null
+++ b/model/diffusers_c/pipelines/latent_diffusion/__init__.py
@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_latent_diffusion"] = ["LDMBertModel", "LDMTextToImagePipeline"]
+ _import_structure["pipeline_latent_diffusion_superresolution"] = ["LDMSuperResolutionPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_latent_diffusion import LDMBertModel, LDMTextToImagePipeline
+ from .pipeline_latent_diffusion_superresolution import LDMSuperResolutionPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/latent_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/latent_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d18674d492afea0a78cb17d217d8056dada1013b
Binary files /dev/null and b/model/diffusers_c/pipelines/latent_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/latent_diffusion/__pycache__/pipeline_latent_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/latent_diffusion/__pycache__/pipeline_latent_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e6241fa6e5e84f3c6a9124e2dd19a6e92b21f2f6
Binary files /dev/null and b/model/diffusers_c/pipelines/latent_diffusion/__pycache__/pipeline_latent_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/latent_diffusion/__pycache__/pipeline_latent_diffusion_superresolution.cpython-39.pyc b/model/diffusers_c/pipelines/latent_diffusion/__pycache__/pipeline_latent_diffusion_superresolution.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..61e893136fdd5ea94e388f37b33c5bfce49225fb
Binary files /dev/null and b/model/diffusers_c/pipelines/latent_diffusion/__pycache__/pipeline_latent_diffusion_superresolution.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/latent_diffusion/pipeline_latent_diffusion.py b/model/diffusers_c/pipelines/latent_diffusion/pipeline_latent_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..f39cbc839652c32d938b39fdb8b6802e93742230
--- /dev/null
+++ b/model/diffusers_c/pipelines/latent_diffusion/pipeline_latent_diffusion.py
@@ -0,0 +1,746 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from transformers import PretrainedConfig, PreTrainedModel, PreTrainedTokenizer
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import BaseModelOutput
+from transformers.utils import logging
+
+from ...models import AutoencoderKL, UNet2DConditionModel, UNet2DModel, VQModel
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+class LDMTextToImagePipeline(DiffusionPipeline):
+ r"""
+ Pipeline for text-to-image generation using latent diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ vqvae ([`VQModel`]):
+ Vector-quantized (VQ) model to encode and decode images to and from latent representations.
+ bert ([`LDMBertModel`]):
+ Text-encoder model based on [`~transformers.BERT`].
+ tokenizer ([`~transformers.BertTokenizer`]):
+ A `BertTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "bert->unet->vqvae"
+
+ def __init__(
+ self,
+ vqvae: Union[VQModel, AutoencoderKL],
+ bert: PreTrainedModel,
+ tokenizer: PreTrainedTokenizer,
+ unet: Union[UNet2DModel, UNet2DConditionModel],
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+ ):
+ super().__init__()
+ self.register_modules(vqvae=vqvae, bert=bert, tokenizer=tokenizer, unet=unet, scheduler=scheduler)
+ self.vae_scale_factor = 2 ** (len(self.vqvae.config.block_out_channels) - 1)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: Optional[int] = 50,
+ guidance_scale: Optional[float] = 1.0,
+ eta: Optional[float] = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ **kwargs,
+ ) -> Union[Tuple, ImagePipelineOutput]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 1.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ ```py
+ >>> from diffusers import DiffusionPipeline
+
+ >>> # load model and scheduler
+ >>> ldm = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256")
+
+ >>> # run pipeline in inference (sample random noise and denoise)
+ >>> prompt = "A painting of a squirrel eating a burger"
+ >>> images = ldm([prompt], num_inference_steps=50, eta=0.3, guidance_scale=6).images
+
+ >>> # save images
+ >>> for idx, image in enumerate(images):
+ ... image.save(f"squirrel-{idx}.png")
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ if isinstance(prompt, str):
+ batch_size = 1
+ elif isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ # get unconditional embeddings for classifier free guidance
+ if guidance_scale != 1.0:
+ uncond_input = self.tokenizer(
+ [""] * batch_size, padding="max_length", max_length=77, truncation=True, return_tensors="pt"
+ )
+ negative_prompt_embeds = self.bert(uncond_input.input_ids.to(self._execution_device))[0]
+
+ # get prompt text embeddings
+ text_input = self.tokenizer(prompt, padding="max_length", max_length=77, truncation=True, return_tensors="pt")
+ prompt_embeds = self.bert(text_input.input_ids.to(self._execution_device))[0]
+
+ # get the initial random noise unless the user supplied it
+ latents_shape = (batch_size, self.unet.config.in_channels, height // 8, width // 8)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(
+ latents_shape, generator=generator, device=self._execution_device, dtype=prompt_embeds.dtype
+ )
+ else:
+ if latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+ latents = latents.to(self._execution_device)
+
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+
+ extra_kwargs = {}
+ if accepts_eta:
+ extra_kwargs["eta"] = eta
+
+ for t in self.progress_bar(self.scheduler.timesteps):
+ if guidance_scale == 1.0:
+ # guidance_scale of 1 means no guidance
+ latents_input = latents
+ context = prompt_embeds
+ else:
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ latents_input = torch.cat([latents] * 2)
+ context = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # predict the noise residual
+ noise_pred = self.unet(latents_input, t, encoder_hidden_states=context).sample
+ # perform guidance
+ if guidance_scale != 1.0:
+ noise_pred_uncond, noise_prediction_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_kwargs).prev_sample
+
+ # scale and decode the image latents with vae
+ latents = 1 / self.vqvae.config.scaling_factor * latents
+ image = self.vqvae.decode(latents).sample
+
+ image = (image / 2 + 0.5).clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
+
+
+################################################################################
+# Code for the text transformer model
+################################################################################
+""" PyTorch LDMBERT model."""
+
+
+logger = logging.get_logger(__name__)
+
+LDMBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
+ "ldm-bert",
+ # See all LDMBert models at https://huggingface.co/models?filter=ldmbert
+]
+
+
+LDMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+ "ldm-bert": "https://huggingface.co/valhalla/ldm-bert/blob/main/config.json",
+}
+
+
+""" LDMBERT model configuration"""
+
+
+class LDMBertConfig(PretrainedConfig):
+ model_type = "ldmbert"
+ keys_to_ignore_at_inference = ["past_key_values"]
+ attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
+
+ def __init__(
+ self,
+ vocab_size=30522,
+ max_position_embeddings=77,
+ encoder_layers=32,
+ encoder_ffn_dim=5120,
+ encoder_attention_heads=8,
+ head_dim=64,
+ encoder_layerdrop=0.0,
+ activation_function="gelu",
+ d_model=1280,
+ dropout=0.1,
+ attention_dropout=0.0,
+ activation_dropout=0.0,
+ init_std=0.02,
+ classifier_dropout=0.0,
+ scale_embedding=False,
+ use_cache=True,
+ pad_token_id=0,
+ **kwargs,
+ ):
+ self.vocab_size = vocab_size
+ self.max_position_embeddings = max_position_embeddings
+ self.d_model = d_model
+ self.encoder_ffn_dim = encoder_ffn_dim
+ self.encoder_layers = encoder_layers
+ self.encoder_attention_heads = encoder_attention_heads
+ self.head_dim = head_dim
+ self.dropout = dropout
+ self.attention_dropout = attention_dropout
+ self.activation_dropout = activation_dropout
+ self.activation_function = activation_function
+ self.init_std = init_std
+ self.encoder_layerdrop = encoder_layerdrop
+ self.classifier_dropout = classifier_dropout
+ self.use_cache = use_cache
+ self.num_hidden_layers = encoder_layers
+ self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True
+
+ super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+ """
+ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+ """
+ bsz, src_len = mask.size()
+ tgt_len = tgt_len if tgt_len is not None else src_len
+
+ expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)
+
+ inverted_mask = 1.0 - expanded_mask
+
+ return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
+
+
+# Copied from transformers.models.bart.modeling_bart.BartAttention with Bart->LDMBert
+class LDMBertAttention(nn.Module):
+ """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+ def __init__(
+ self,
+ embed_dim: int,
+ num_heads: int,
+ head_dim: int,
+ dropout: float = 0.0,
+ is_decoder: bool = False,
+ bias: bool = False,
+ ):
+ super().__init__()
+ self.embed_dim = embed_dim
+ self.num_heads = num_heads
+ self.dropout = dropout
+ self.head_dim = head_dim
+ self.inner_dim = head_dim * num_heads
+
+ self.scaling = self.head_dim**-0.5
+ self.is_decoder = is_decoder
+
+ self.k_proj = nn.Linear(embed_dim, self.inner_dim, bias=bias)
+ self.v_proj = nn.Linear(embed_dim, self.inner_dim, bias=bias)
+ self.q_proj = nn.Linear(embed_dim, self.inner_dim, bias=bias)
+ self.out_proj = nn.Linear(self.inner_dim, embed_dim)
+
+ def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+ return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ key_value_states: Optional[torch.Tensor] = None,
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ layer_head_mask: Optional[torch.Tensor] = None,
+ output_attentions: bool = False,
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+ """Input shape: Batch x Time x Channel"""
+
+ # if key_value_states are provided this layer is used as a cross-attention layer
+ # for the decoder
+ is_cross_attention = key_value_states is not None
+
+ bsz, tgt_len, _ = hidden_states.size()
+
+ # get query proj
+ query_states = self.q_proj(hidden_states) * self.scaling
+ # get key, value proj
+ if is_cross_attention and past_key_value is not None:
+ # reuse k,v, cross_attentions
+ key_states = past_key_value[0]
+ value_states = past_key_value[1]
+ elif is_cross_attention:
+ # cross_attentions
+ key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+ value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+ elif past_key_value is not None:
+ # reuse k, v, self_attention
+ key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+ value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
+ else:
+ # self_attention
+ key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+ value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+ if self.is_decoder:
+ # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+ # Further calls to cross_attention layer can then reuse all cross-attention
+ # key/value_states (first "if" case)
+ # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+ # all previous decoder key/value_states. Further calls to uni-directional self-attention
+ # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+ # if encoder bi-directional self-attention `past_key_value` is always `None`
+ past_key_value = (key_states, value_states)
+
+ proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+ query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)
+ key_states = key_states.view(*proj_shape)
+ value_states = value_states.view(*proj_shape)
+
+ src_len = key_states.size(1)
+ attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))
+
+ if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):
+ raise ValueError(
+ f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+ f" {attn_weights.size()}"
+ )
+
+ if attention_mask is not None:
+ if attention_mask.size() != (bsz, 1, tgt_len, src_len):
+ raise ValueError(
+ f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}"
+ )
+ attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + attention_mask
+ attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+ if layer_head_mask is not None:
+ if layer_head_mask.size() != (self.num_heads,):
+ raise ValueError(
+ f"Head mask for a single layer should be of size {(self.num_heads,)}, but is"
+ f" {layer_head_mask.size()}"
+ )
+ attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+ attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
+
+ if output_attentions:
+ # this operation is a bit awkward, but it's required to
+ # make sure that attn_weights keeps its gradient.
+ # In order to do so, attn_weights have to be reshaped
+ # twice and have to be reused in the following
+ attn_weights_reshaped = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)
+ attn_weights = attn_weights_reshaped.view(bsz * self.num_heads, tgt_len, src_len)
+ else:
+ attn_weights_reshaped = None
+
+ attn_probs = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+
+ attn_output = torch.bmm(attn_probs, value_states)
+
+ if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):
+ raise ValueError(
+ f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+ f" {attn_output.size()}"
+ )
+
+ attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)
+ attn_output = attn_output.transpose(1, 2)
+
+ # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be
+ # partitioned across GPUs when using tensor-parallelism.
+ attn_output = attn_output.reshape(bsz, tgt_len, self.inner_dim)
+
+ attn_output = self.out_proj(attn_output)
+
+ return attn_output, attn_weights_reshaped, past_key_value
+
+
+class LDMBertEncoderLayer(nn.Module):
+ def __init__(self, config: LDMBertConfig):
+ super().__init__()
+ self.embed_dim = config.d_model
+ self.self_attn = LDMBertAttention(
+ embed_dim=self.embed_dim,
+ num_heads=config.encoder_attention_heads,
+ head_dim=config.head_dim,
+ dropout=config.attention_dropout,
+ )
+ self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)
+ self.dropout = config.dropout
+ self.activation_fn = ACT2FN[config.activation_function]
+ self.activation_dropout = config.activation_dropout
+ self.fc1 = nn.Linear(self.embed_dim, config.encoder_ffn_dim)
+ self.fc2 = nn.Linear(config.encoder_ffn_dim, self.embed_dim)
+ self.final_layer_norm = nn.LayerNorm(self.embed_dim)
+
+ def forward(
+ self,
+ hidden_states: torch.FloatTensor,
+ attention_mask: torch.FloatTensor,
+ layer_head_mask: torch.FloatTensor,
+ output_attentions: Optional[bool] = False,
+ ) -> Tuple[torch.FloatTensor, Optional[torch.FloatTensor]]:
+ """
+ Args:
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+ attention_mask (`torch.FloatTensor`): attention mask of size
+ `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+ layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size
+ `(encoder_attention_heads,)`.
+ output_attentions (`bool`, *optional*):
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+ returned tensors for more detail.
+ """
+ residual = hidden_states
+ hidden_states = self.self_attn_layer_norm(hidden_states)
+ hidden_states, attn_weights, _ = self.self_attn(
+ hidden_states=hidden_states,
+ attention_mask=attention_mask,
+ layer_head_mask=layer_head_mask,
+ output_attentions=output_attentions,
+ )
+ hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+ hidden_states = residual + hidden_states
+
+ residual = hidden_states
+ hidden_states = self.final_layer_norm(hidden_states)
+ hidden_states = self.activation_fn(self.fc1(hidden_states))
+ hidden_states = nn.functional.dropout(hidden_states, p=self.activation_dropout, training=self.training)
+ hidden_states = self.fc2(hidden_states)
+ hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+ hidden_states = residual + hidden_states
+
+ if hidden_states.dtype == torch.float16 and (
+ torch.isinf(hidden_states).any() or torch.isnan(hidden_states).any()
+ ):
+ clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+ hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+ outputs = (hidden_states,)
+
+ if output_attentions:
+ outputs += (attn_weights,)
+
+ return outputs
+
+
+# Copied from transformers.models.bart.modeling_bart.BartPretrainedModel with Bart->LDMBert
+class LDMBertPreTrainedModel(PreTrainedModel):
+ config_class = LDMBertConfig
+ base_model_prefix = "model"
+ _supports_gradient_checkpointing = True
+ _keys_to_ignore_on_load_unexpected = [r"encoder\.version", r"decoder\.version"]
+
+ def _init_weights(self, module):
+ std = self.config.init_std
+ if isinstance(module, nn.Linear):
+ module.weight.data.normal_(mean=0.0, std=std)
+ if module.bias is not None:
+ module.bias.data.zero_()
+ elif isinstance(module, nn.Embedding):
+ module.weight.data.normal_(mean=0.0, std=std)
+ if module.padding_idx is not None:
+ module.weight.data[module.padding_idx].zero_()
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ if isinstance(module, (LDMBertEncoder,)):
+ module.gradient_checkpointing = value
+
+ @property
+ def dummy_inputs(self):
+ pad_token = self.config.pad_token_id
+ input_ids = torch.tensor([[0, 6, 10, 4, 2], [0, 8, 12, 2, pad_token]], device=self.device)
+ dummy_inputs = {
+ "attention_mask": input_ids.ne(pad_token),
+ "input_ids": input_ids,
+ }
+ return dummy_inputs
+
+
+class LDMBertEncoder(LDMBertPreTrainedModel):
+ """
+ Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+ [`LDMBertEncoderLayer`].
+
+ Args:
+ config: LDMBertConfig
+ embed_tokens (nn.Embedding): output embedding
+ """
+
+ def __init__(self, config: LDMBertConfig):
+ super().__init__(config)
+
+ self.dropout = config.dropout
+
+ embed_dim = config.d_model
+ self.padding_idx = config.pad_token_id
+ self.max_source_positions = config.max_position_embeddings
+
+ self.embed_tokens = nn.Embedding(config.vocab_size, embed_dim)
+ self.embed_positions = nn.Embedding(config.max_position_embeddings, embed_dim)
+ self.layers = nn.ModuleList([LDMBertEncoderLayer(config) for _ in range(config.encoder_layers)])
+ self.layer_norm = nn.LayerNorm(embed_dim)
+
+ self.gradient_checkpointing = False
+ # Initialize weights and apply final processing
+ self.post_init()
+
+ def get_input_embeddings(self):
+ return self.embed_tokens
+
+ def set_input_embeddings(self, value):
+ self.embed_tokens = value
+
+ def forward(
+ self,
+ input_ids: torch.LongTensor = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.LongTensor] = None,
+ head_mask: Optional[torch.Tensor] = None,
+ inputs_embeds: Optional[torch.FloatTensor] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ ) -> Union[Tuple, BaseModelOutput]:
+ r"""
+ Args:
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+ provide it.
+
+ Indices can be obtained using [`BartTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+ [`PreTrainedTokenizer.__call__`] for details.
+
+ [What are input IDs?](../glossary#input-ids)
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+ - 1 for tokens that are **not masked**,
+ - 0 for tokens that are **masked**.
+
+ [What are attention masks?](../glossary#attention-mask)
+ head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+ Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+ - 1 indicates the head is **not masked**,
+ - 0 indicates the head is **masked**.
+
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+ This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+ than the model's internal embedding lookup matrix.
+ output_attentions (`bool`, *optional*):
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+ returned tensors for more detail.
+ output_hidden_states (`bool`, *optional*):
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+ for more detail.
+ return_dict (`bool`, *optional*):
+ Whether or not to return a [`~utils.BaseModelOutput`] instead of a plain tuple.
+ """
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+ output_hidden_states = (
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+ )
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+ # retrieve input_ids and inputs_embeds
+ if input_ids is not None and inputs_embeds is not None:
+ raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+ elif input_ids is not None:
+ input_shape = input_ids.size()
+ input_ids = input_ids.view(-1, input_shape[-1])
+ elif inputs_embeds is not None:
+ input_shape = inputs_embeds.size()[:-1]
+ else:
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+ if inputs_embeds is None:
+ inputs_embeds = self.embed_tokens(input_ids)
+
+ seq_len = input_shape[1]
+ if position_ids is None:
+ position_ids = torch.arange(seq_len, dtype=torch.long, device=inputs_embeds.device).expand((1, -1))
+ embed_pos = self.embed_positions(position_ids)
+
+ hidden_states = inputs_embeds + embed_pos
+ hidden_states = nn.functional.dropout(hidden_states, p=self.dropout, training=self.training)
+
+ # expand attention_mask
+ if attention_mask is not None:
+ # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+ attention_mask = _expand_mask(attention_mask, inputs_embeds.dtype)
+
+ encoder_states = () if output_hidden_states else None
+ all_attentions = () if output_attentions else None
+
+ # check if head_mask has a correct number of layers specified if desired
+ if head_mask is not None:
+ if head_mask.size()[0] != (len(self.layers)):
+ raise ValueError(
+ f"The head_mask should be specified for {len(self.layers)} layers, but it is for"
+ f" {head_mask.size()[0]}."
+ )
+
+ for idx, encoder_layer in enumerate(self.layers):
+ if output_hidden_states:
+ encoder_states = encoder_states + (hidden_states,)
+ if self.gradient_checkpointing and self.training:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs, output_attentions)
+
+ return custom_forward
+
+ layer_outputs = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(encoder_layer),
+ hidden_states,
+ attention_mask,
+ (head_mask[idx] if head_mask is not None else None),
+ )
+ else:
+ layer_outputs = encoder_layer(
+ hidden_states,
+ attention_mask,
+ layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+ output_attentions=output_attentions,
+ )
+
+ hidden_states = layer_outputs[0]
+
+ if output_attentions:
+ all_attentions = all_attentions + (layer_outputs[1],)
+
+ hidden_states = self.layer_norm(hidden_states)
+
+ if output_hidden_states:
+ encoder_states = encoder_states + (hidden_states,)
+
+ if not return_dict:
+ return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+ return BaseModelOutput(
+ last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+ )
+
+
+class LDMBertModel(LDMBertPreTrainedModel):
+ _no_split_modules = []
+
+ def __init__(self, config: LDMBertConfig):
+ super().__init__(config)
+ self.model = LDMBertEncoder(config)
+ self.to_logits = nn.Linear(config.hidden_size, config.vocab_size)
+
+ def forward(
+ self,
+ input_ids=None,
+ attention_mask=None,
+ position_ids=None,
+ head_mask=None,
+ inputs_embeds=None,
+ output_attentions=None,
+ output_hidden_states=None,
+ return_dict=None,
+ ):
+ outputs = self.model(
+ input_ids,
+ attention_mask=attention_mask,
+ position_ids=position_ids,
+ head_mask=head_mask,
+ inputs_embeds=inputs_embeds,
+ output_attentions=output_attentions,
+ output_hidden_states=output_hidden_states,
+ return_dict=return_dict,
+ )
+ return outputs
diff --git a/model/diffusers_c/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py b/model/diffusers_c/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py
new file mode 100644
index 0000000000000000000000000000000000000000..bb72b4d4eb8e387d596b22cca65c82aef0ab9e75
--- /dev/null
+++ b/model/diffusers_c/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py
@@ -0,0 +1,189 @@
+import inspect
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.utils.checkpoint
+
+from ...models import UNet2DModel, VQModel
+from ...schedulers import (
+ DDIMScheduler,
+ DPMSolverMultistepScheduler,
+ EulerAncestralDiscreteScheduler,
+ EulerDiscreteScheduler,
+ LMSDiscreteScheduler,
+ PNDMScheduler,
+)
+from ...utils import PIL_INTERPOLATION
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+def preprocess(image):
+ w, h = image.size
+ w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32
+ image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image[None].transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image)
+ return 2.0 * image - 1.0
+
+
+class LDMSuperResolutionPipeline(DiffusionPipeline):
+ r"""
+ A pipeline for image super-resolution using latent diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Parameters:
+ vqvae ([`VQModel`]):
+ Vector-quantized (VQ) model to encode and decode images to and from latent representations.
+ unet ([`UNet2DModel`]):
+ A `UNet2DModel` to denoise the encoded image.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], [`EulerDiscreteScheduler`],
+ [`EulerAncestralDiscreteScheduler`], [`DPMSolverMultistepScheduler`], or [`PNDMScheduler`].
+ """
+
+ def __init__(
+ self,
+ vqvae: VQModel,
+ unet: UNet2DModel,
+ scheduler: Union[
+ DDIMScheduler,
+ PNDMScheduler,
+ LMSDiscreteScheduler,
+ EulerDiscreteScheduler,
+ EulerAncestralDiscreteScheduler,
+ DPMSolverMultistepScheduler,
+ ],
+ ):
+ super().__init__()
+ self.register_modules(vqvae=vqvae, unet=unet, scheduler=scheduler)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image: Union[torch.Tensor, PIL.Image.Image] = None,
+ batch_size: Optional[int] = 1,
+ num_inference_steps: Optional[int] = 100,
+ eta: Optional[float] = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ) -> Union[Tuple, ImagePipelineOutput]:
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`torch.Tensor` or `PIL.Image.Image`):
+ `Image` or tensor representing an image batch to be used as the starting point for the process.
+ batch_size (`int`, *optional*, defaults to 1):
+ Number of images to generate.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
+
+ Example:
+
+ ```py
+ >>> import requests
+ >>> from PIL import Image
+ >>> from io import BytesIO
+ >>> from diffusers import LDMSuperResolutionPipeline
+ >>> import torch
+
+ >>> # load model and scheduler
+ >>> pipeline = LDMSuperResolutionPipeline.from_pretrained("CompVis/ldm-super-resolution-4x-openimages")
+ >>> pipeline = pipeline.to("cuda")
+
+ >>> # let's download an image
+ >>> url = (
+ ... "https://user-images.githubusercontent.com/38061659/199705896-b48e17b8-b231-47cd-a270-4ffa5a93fa3e.png"
+ ... )
+ >>> response = requests.get(url)
+ >>> low_res_img = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> low_res_img = low_res_img.resize((128, 128))
+
+ >>> # run pipeline in inference (sample random noise and denoise)
+ >>> upscaled_image = pipeline(low_res_img, num_inference_steps=100, eta=1).images[0]
+ >>> # save image
+ >>> upscaled_image.save("ldm_generated_image.png")
+ ```
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images
+ """
+ if isinstance(image, PIL.Image.Image):
+ batch_size = 1
+ elif isinstance(image, torch.Tensor):
+ batch_size = image.shape[0]
+ else:
+ raise ValueError(f"`image` has to be of type `PIL.Image.Image` or `torch.Tensor` but is {type(image)}")
+
+ if isinstance(image, PIL.Image.Image):
+ image = preprocess(image)
+
+ height, width = image.shape[-2:]
+
+ # in_channels should be 6: 3 for latents, 3 for low resolution image
+ latents_shape = (batch_size, self.unet.config.in_channels // 2, height, width)
+ latents_dtype = next(self.unet.parameters()).dtype
+
+ latents = randn_tensor(latents_shape, generator=generator, device=self.device, dtype=latents_dtype)
+
+ image = image.to(device=self.device, dtype=latents_dtype)
+
+ # set timesteps and move to the correct device
+ self.scheduler.set_timesteps(num_inference_steps, device=self.device)
+ timesteps_tensor = self.scheduler.timesteps
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature.
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_kwargs = {}
+ if accepts_eta:
+ extra_kwargs["eta"] = eta
+
+ for t in self.progress_bar(timesteps_tensor):
+ # concat latents and low resolution image in the channel dimension.
+ latents_input = torch.cat([latents, image], dim=1)
+ latents_input = self.scheduler.scale_model_input(latents_input, t)
+ # predict the noise residual
+ noise_pred = self.unet(latents_input, t).sample
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_kwargs).prev_sample
+
+ # decode the image latents with the VQVAE
+ image = self.vqvae.decode(latents).sample
+ image = torch.clamp(image, -1.0, 1.0)
+ image = image / 2 + 0.5
+ image = image.cpu().permute(0, 2, 3, 1).numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/musicldm/__init__.py b/model/diffusers_c/pipelines/musicldm/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..ed71eeb1d99b28f20f7cd94776c0303208620653
--- /dev/null
+++ b/model/diffusers_c/pipelines/musicldm/__init__.py
@@ -0,0 +1,49 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+ is_transformers_version,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_musicldm"] = ["MusicLDMPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.27.0")):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_musicldm import MusicLDMPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/musicldm/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/musicldm/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8e92a83a24bbe2cf1e0b04d86570dd0c8e5ceadb
Binary files /dev/null and b/model/diffusers_c/pipelines/musicldm/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/musicldm/__pycache__/pipeline_musicldm.cpython-39.pyc b/model/diffusers_c/pipelines/musicldm/__pycache__/pipeline_musicldm.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..07704a6ce70efc204ef866b871eb1b1bb183d364
Binary files /dev/null and b/model/diffusers_c/pipelines/musicldm/__pycache__/pipeline_musicldm.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/musicldm/pipeline_musicldm.py b/model/diffusers_c/pipelines/musicldm/pipeline_musicldm.py
new file mode 100644
index 0000000000000000000000000000000000000000..5fde3450b9a0e788b43f676b683c9e87abef70ff
--- /dev/null
+++ b/model/diffusers_c/pipelines/musicldm/pipeline_musicldm.py
@@ -0,0 +1,635 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import (
+ ClapFeatureExtractor,
+ ClapModel,
+ ClapTextModelWithProjection,
+ RobertaTokenizer,
+ RobertaTokenizerFast,
+ SpeechT5HifiGan,
+)
+
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ is_accelerate_available,
+ is_accelerate_version,
+ is_librosa_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import AudioPipelineOutput, DiffusionPipeline, StableDiffusionMixin
+
+
+if is_librosa_available():
+ import librosa
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import MusicLDMPipeline
+ >>> import torch
+ >>> import scipy
+
+ >>> repo_id = "ucsd-reach/musicldm"
+ >>> pipe = MusicLDMPipeline.from_pretrained(repo_id, torch_dtype=torch.float16)
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "Techno music with a strong, upbeat tempo and high melodic riffs"
+ >>> audio = pipe(prompt, num_inference_steps=10, audio_length_in_s=5.0).audios[0]
+
+ >>> # save the audio sample as a .wav file
+ >>> scipy.io.wavfile.write("techno.wav", rate=16000, data=audio)
+ ```
+"""
+
+
+class MusicLDMPipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+ Pipeline for text-to-audio generation using MusicLDM.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.ClapModel`]):
+ Frozen text-audio embedding model (`ClapTextModel`), specifically the
+ [laion/clap-htsat-unfused](https://huggingface.co/laion/clap-htsat-unfused) variant.
+ tokenizer ([`PreTrainedTokenizer`]):
+ A [`~transformers.RobertaTokenizer`] to tokenize text.
+ feature_extractor ([`~transformers.ClapFeatureExtractor`]):
+ Feature extractor to compute mel-spectrograms from audio waveforms.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded audio latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded audio latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ vocoder ([`~transformers.SpeechT5HifiGan`]):
+ Vocoder of class `SpeechT5HifiGan`.
+ """
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: Union[ClapTextModelWithProjection, ClapModel],
+ tokenizer: Union[RobertaTokenizer, RobertaTokenizerFast],
+ feature_extractor: Optional[ClapFeatureExtractor],
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ vocoder: SpeechT5HifiGan,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ feature_extractor=feature_extractor,
+ unet=unet,
+ scheduler=scheduler,
+ vocoder=vocoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_waveforms_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device (`torch.device`):
+ torch device
+ num_waveforms_per_prompt (`int`):
+ number of waveforms that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the audio generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ attention_mask = text_inputs.attention_mask
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLAP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = self.text_encoder.get_text_features(
+ text_input_ids.to(device),
+ attention_mask=attention_mask.to(device),
+ )
+
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.text_model.dtype, device=device)
+
+ (
+ bs_embed,
+ seq_len,
+ ) = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_waveforms_per_prompt)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_waveforms_per_prompt, seq_len)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ uncond_input_ids = uncond_input.input_ids.to(device)
+ attention_mask = uncond_input.attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder.get_text_features(
+ uncond_input_ids,
+ attention_mask=attention_mask,
+ )
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.text_model.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_waveforms_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_waveforms_per_prompt, seq_len)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.audioldm.pipeline_audioldm.AudioLDMPipeline.mel_spectrogram_to_waveform
+ def mel_spectrogram_to_waveform(self, mel_spectrogram):
+ if mel_spectrogram.dim() == 4:
+ mel_spectrogram = mel_spectrogram.squeeze(1)
+
+ waveform = self.vocoder(mel_spectrogram)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ waveform = waveform.cpu().float()
+ return waveform
+
+ # Copied from diffusers.pipelines.audioldm2.pipeline_audioldm2.AudioLDM2Pipeline.score_waveforms
+ def score_waveforms(self, text, audio, num_waveforms_per_prompt, device, dtype):
+ if not is_librosa_available():
+ logger.info(
+ "Automatic scoring of the generated audio waveforms against the input prompt text requires the "
+ "`librosa` package to resample the generated waveforms. Returning the audios in the order they were "
+ "generated. To enable automatic scoring, install `librosa` with: `pip install librosa`."
+ )
+ return audio
+ inputs = self.tokenizer(text, return_tensors="pt", padding=True)
+ resampled_audio = librosa.resample(
+ audio.numpy(), orig_sr=self.vocoder.config.sampling_rate, target_sr=self.feature_extractor.sampling_rate
+ )
+ inputs["input_features"] = self.feature_extractor(
+ list(resampled_audio), return_tensors="pt", sampling_rate=self.feature_extractor.sampling_rate
+ ).input_features.type(dtype)
+ inputs = inputs.to(device)
+
+ # compute the audio-text similarity score using the CLAP model
+ logits_per_text = self.text_encoder(**inputs).logits_per_text
+ # sort by the highest matching generations per prompt
+ indices = torch.argsort(logits_per_text, dim=1, descending=True)[:, :num_waveforms_per_prompt]
+ audio = torch.index_select(audio, 0, indices.reshape(-1).cpu())
+ return audio
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.audioldm.pipeline_audioldm.AudioLDMPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ audio_length_in_s,
+ vocoder_upsample_factor,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ min_audio_length_in_s = vocoder_upsample_factor * self.vae_scale_factor
+ if audio_length_in_s < min_audio_length_in_s:
+ raise ValueError(
+ f"`audio_length_in_s` has to be a positive value greater than or equal to {min_audio_length_in_s}, but "
+ f"is {audio_length_in_s}."
+ )
+
+ if self.vocoder.config.model_in_dim % self.vae_scale_factor != 0:
+ raise ValueError(
+ f"The number of frequency bins in the vocoder's log-mel spectrogram has to be divisible by the "
+ f"VAE scale factor, but got {self.vocoder.config.model_in_dim} bins and a scale factor of "
+ f"{self.vae_scale_factor}."
+ )
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.audioldm.pipeline_audioldm.AudioLDMPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, dtype, device, generator, latents=None):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ height // self.vae_scale_factor,
+ self.vocoder.config.model_in_dim // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def enable_model_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+ """
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+ from accelerate import cpu_offload_with_hook
+ else:
+ raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+ device = torch.device(f"cuda:{gpu_id}")
+
+ if self.device.type != "cpu":
+ self.to("cpu", silence_dtype_warnings=True)
+ torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
+
+ model_sequence = [
+ self.text_encoder.text_model,
+ self.text_encoder.text_projection,
+ self.unet,
+ self.vae,
+ self.vocoder,
+ self.text_encoder,
+ ]
+
+ hook = None
+ for cpu_offloaded_model in model_sequence:
+ _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+ # We'll offload the last model manually.
+ self.final_offload_hook = hook
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ audio_length_in_s: Optional[float] = None,
+ num_inference_steps: int = 200,
+ guidance_scale: float = 2.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_waveforms_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ output_type: Optional[str] = "np",
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide audio generation. If not defined, you need to pass `prompt_embeds`.
+ audio_length_in_s (`int`, *optional*, defaults to 10.24):
+ The length of the generated audio sample in seconds.
+ num_inference_steps (`int`, *optional*, defaults to 200):
+ The number of denoising steps. More denoising steps usually lead to a higher quality audio at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 2.0):
+ A higher guidance scale value encourages the model to generate audio that is closely linked to the text
+ `prompt` at the expense of lower sound quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in audio generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_waveforms_per_prompt (`int`, *optional*, defaults to 1):
+ The number of waveforms to generate per prompt. If `num_waveforms_per_prompt > 1`, the text encoding
+ model is a joint text-audio model ([`~transformers.ClapModel`]), and the tokenizer is a
+ `[~transformers.ClapProcessor]`, then automatic scoring will be performed between the generated outputs
+ and the input text. This scoring ranks the generated waveforms based on their cosine similarity to text
+ input in the joint text-audio embedding space.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.AudioPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ output_type (`str`, *optional*, defaults to `"np"`):
+ The output format of the generated audio. Choose between `"np"` to return a NumPy `np.ndarray` or
+ `"pt"` to return a PyTorch `torch.Tensor` object. Set to `"latent"` to return the latent diffusion
+ model (LDM) output.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.AudioPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.AudioPipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated audio.
+ """
+ # 0. Convert audio input length from seconds to spectrogram height
+ vocoder_upsample_factor = np.prod(self.vocoder.config.upsample_rates) / self.vocoder.config.sampling_rate
+
+ if audio_length_in_s is None:
+ audio_length_in_s = self.unet.config.sample_size * self.vae_scale_factor * vocoder_upsample_factor
+
+ height = int(audio_length_in_s / vocoder_upsample_factor)
+
+ original_waveform_length = int(audio_length_in_s * self.vocoder.config.sampling_rate)
+ if height % self.vae_scale_factor != 0:
+ height = int(np.ceil(height / self.vae_scale_factor)) * self.vae_scale_factor
+ logger.info(
+ f"Audio length in seconds {audio_length_in_s} is increased to {height * vocoder_upsample_factor} "
+ f"so that it can be handled by the model. It will be cut to {audio_length_in_s} after the "
+ f"denoising process."
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ audio_length_in_s,
+ vocoder_upsample_factor,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds = self._encode_prompt(
+ prompt,
+ device,
+ num_waveforms_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_waveforms_per_prompt,
+ num_channels_latents,
+ height,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=None,
+ class_labels=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ self.maybe_free_model_hooks()
+
+ # 8. Post-processing
+ if not output_type == "latent":
+ latents = 1 / self.vae.config.scaling_factor * latents
+ mel_spectrogram = self.vae.decode(latents).sample
+ else:
+ return AudioPipelineOutput(audios=latents)
+
+ audio = self.mel_spectrogram_to_waveform(mel_spectrogram)
+
+ audio = audio[:, :original_waveform_length]
+
+ # 9. Automatic scoring
+ if num_waveforms_per_prompt > 1 and prompt is not None:
+ audio = self.score_waveforms(
+ text=prompt,
+ audio=audio,
+ num_waveforms_per_prompt=num_waveforms_per_prompt,
+ device=device,
+ dtype=prompt_embeds.dtype,
+ )
+
+ if output_type == "np":
+ audio = audio.numpy()
+
+ if not return_dict:
+ return (audio,)
+
+ return AudioPipelineOutput(audios=audio)
diff --git a/model/diffusers_c/pipelines/onnx_utils.py b/model/diffusers_c/pipelines/onnx_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..e1a368a22d1717e8cae49b1f13e35b62a28eee79
--- /dev/null
+++ b/model/diffusers_c/pipelines/onnx_utils.py
@@ -0,0 +1,214 @@
+# Copyright 2024 The HuggingFace Inc. team.
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import os
+import shutil
+from pathlib import Path
+from typing import Optional, Union
+
+import numpy as np
+from huggingface_hub import hf_hub_download
+from huggingface_hub.utils import validate_hf_hub_args
+
+from ..utils import ONNX_EXTERNAL_WEIGHTS_NAME, ONNX_WEIGHTS_NAME, is_onnx_available, logging
+
+
+if is_onnx_available():
+ import onnxruntime as ort
+
+
+logger = logging.get_logger(__name__)
+
+ORT_TO_NP_TYPE = {
+ "tensor(bool)": np.bool_,
+ "tensor(int8)": np.int8,
+ "tensor(uint8)": np.uint8,
+ "tensor(int16)": np.int16,
+ "tensor(uint16)": np.uint16,
+ "tensor(int32)": np.int32,
+ "tensor(uint32)": np.uint32,
+ "tensor(int64)": np.int64,
+ "tensor(uint64)": np.uint64,
+ "tensor(float16)": np.float16,
+ "tensor(float)": np.float32,
+ "tensor(double)": np.float64,
+}
+
+
+class OnnxRuntimeModel:
+ def __init__(self, model=None, **kwargs):
+ logger.info("`diffusers.OnnxRuntimeModel` is experimental and might change in the future.")
+ self.model = model
+ self.model_save_dir = kwargs.get("model_save_dir", None)
+ self.latest_model_name = kwargs.get("latest_model_name", ONNX_WEIGHTS_NAME)
+
+ def __call__(self, **kwargs):
+ inputs = {k: np.array(v) for k, v in kwargs.items()}
+ return self.model.run(None, inputs)
+
+ @staticmethod
+ def load_model(path: Union[str, Path], provider=None, sess_options=None):
+ """
+ Loads an ONNX Inference session with an ExecutionProvider. Default provider is `CPUExecutionProvider`
+
+ Arguments:
+ path (`str` or `Path`):
+ Directory from which to load
+ provider(`str`, *optional*):
+ Onnxruntime execution provider to use for loading the model, defaults to `CPUExecutionProvider`
+ """
+ if provider is None:
+ logger.info("No onnxruntime provider specified, using CPUExecutionProvider")
+ provider = "CPUExecutionProvider"
+
+ return ort.InferenceSession(path, providers=[provider], sess_options=sess_options)
+
+ def _save_pretrained(self, save_directory: Union[str, Path], file_name: Optional[str] = None, **kwargs):
+ """
+ Save a model and its configuration file to a directory, so that it can be re-loaded using the
+ [`~optimum.onnxruntime.modeling_ort.ORTModel.from_pretrained`] class method. It will always save the
+ latest_model_name.
+
+ Arguments:
+ save_directory (`str` or `Path`):
+ Directory where to save the model file.
+ file_name(`str`, *optional*):
+ Overwrites the default model file name from `"model.onnx"` to `file_name`. This allows you to save the
+ model with a different name.
+ """
+ model_file_name = file_name if file_name is not None else ONNX_WEIGHTS_NAME
+
+ src_path = self.model_save_dir.joinpath(self.latest_model_name)
+ dst_path = Path(save_directory).joinpath(model_file_name)
+ try:
+ shutil.copyfile(src_path, dst_path)
+ except shutil.SameFileError:
+ pass
+
+ # copy external weights (for models >2GB)
+ src_path = self.model_save_dir.joinpath(ONNX_EXTERNAL_WEIGHTS_NAME)
+ if src_path.exists():
+ dst_path = Path(save_directory).joinpath(ONNX_EXTERNAL_WEIGHTS_NAME)
+ try:
+ shutil.copyfile(src_path, dst_path)
+ except shutil.SameFileError:
+ pass
+
+ def save_pretrained(
+ self,
+ save_directory: Union[str, os.PathLike],
+ **kwargs,
+ ):
+ """
+ Save a model to a directory, so that it can be re-loaded using the [`~OnnxModel.from_pretrained`] class
+ method.:
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to which to save. Will be created if it doesn't exist.
+ """
+ if os.path.isfile(save_directory):
+ logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
+ return
+
+ os.makedirs(save_directory, exist_ok=True)
+
+ # saving model weights/files
+ self._save_pretrained(save_directory, **kwargs)
+
+ @classmethod
+ @validate_hf_hub_args
+ def _from_pretrained(
+ cls,
+ model_id: Union[str, Path],
+ token: Optional[Union[bool, str, None]] = None,
+ revision: Optional[Union[str, None]] = None,
+ force_download: bool = False,
+ cache_dir: Optional[str] = None,
+ file_name: Optional[str] = None,
+ provider: Optional[str] = None,
+ sess_options: Optional["ort.SessionOptions"] = None,
+ **kwargs,
+ ):
+ """
+ Load a model from a directory or the HF Hub.
+
+ Arguments:
+ model_id (`str` or `Path`):
+ Directory from which to load
+ token (`str` or `bool`):
+ Is needed to load models from a private or gated repository
+ revision (`str`):
+ Revision is the specific model version to use. It can be a branch name, a tag name, or a commit id
+ cache_dir (`Union[str, Path]`, *optional*):
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the
+ standard cache should not be used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ file_name(`str`):
+ Overwrites the default model file name from `"model.onnx"` to `file_name`. This allows you to load
+ different model files from the same repository or directory.
+ provider(`str`):
+ The ONNX runtime provider, e.g. `CPUExecutionProvider` or `CUDAExecutionProvider`.
+ kwargs (`Dict`, *optional*):
+ kwargs will be passed to the model during initialization
+ """
+ model_file_name = file_name if file_name is not None else ONNX_WEIGHTS_NAME
+ # load model from local directory
+ if os.path.isdir(model_id):
+ model = OnnxRuntimeModel.load_model(
+ Path(model_id, model_file_name).as_posix(), provider=provider, sess_options=sess_options
+ )
+ kwargs["model_save_dir"] = Path(model_id)
+ # load model from hub
+ else:
+ # download model
+ model_cache_path = hf_hub_download(
+ repo_id=model_id,
+ filename=model_file_name,
+ token=token,
+ revision=revision,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ )
+ kwargs["model_save_dir"] = Path(model_cache_path).parent
+ kwargs["latest_model_name"] = Path(model_cache_path).name
+ model = OnnxRuntimeModel.load_model(model_cache_path, provider=provider, sess_options=sess_options)
+ return cls(model=model, **kwargs)
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(
+ cls,
+ model_id: Union[str, Path],
+ force_download: bool = True,
+ token: Optional[str] = None,
+ cache_dir: Optional[str] = None,
+ **model_kwargs,
+ ):
+ revision = None
+ if len(str(model_id).split("@")) == 2:
+ model_id, revision = model_id.split("@")
+
+ return cls._from_pretrained(
+ model_id=model_id,
+ revision=revision,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ token=token,
+ **model_kwargs,
+ )
diff --git a/model/diffusers_c/pipelines/paint_by_example/__init__.py b/model/diffusers_c/pipelines/paint_by_example/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..aaa775f690c3d290074662c029f242df3c61e003
--- /dev/null
+++ b/model/diffusers_c/pipelines/paint_by_example/__init__.py
@@ -0,0 +1,55 @@
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, List, Optional, Union
+
+import numpy as np
+import PIL
+from PIL import Image
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["image_encoder"] = ["PaintByExampleImageEncoder"]
+ _import_structure["pipeline_paint_by_example"] = ["PaintByExamplePipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .image_encoder import PaintByExampleImageEncoder
+ from .pipeline_paint_by_example import PaintByExamplePipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/paint_by_example/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/paint_by_example/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..458a6c620905f1229c06d2818a9a49e710d29b2a
Binary files /dev/null and b/model/diffusers_c/pipelines/paint_by_example/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/paint_by_example/__pycache__/image_encoder.cpython-39.pyc b/model/diffusers_c/pipelines/paint_by_example/__pycache__/image_encoder.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a94433b3ffc35c8252a06d0544f3dd5c62298354
Binary files /dev/null and b/model/diffusers_c/pipelines/paint_by_example/__pycache__/image_encoder.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/paint_by_example/__pycache__/pipeline_paint_by_example.cpython-39.pyc b/model/diffusers_c/pipelines/paint_by_example/__pycache__/pipeline_paint_by_example.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..60276976249fafbd4053d34311b640d9742689d5
Binary files /dev/null and b/model/diffusers_c/pipelines/paint_by_example/__pycache__/pipeline_paint_by_example.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/paint_by_example/image_encoder.py b/model/diffusers_c/pipelines/paint_by_example/image_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..2fd0338b1f9190e63053c086e4bbd74da9d98c54
--- /dev/null
+++ b/model/diffusers_c/pipelines/paint_by_example/image_encoder.py
@@ -0,0 +1,67 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+from torch import nn
+from transformers import CLIPPreTrainedModel, CLIPVisionModel
+
+from ...models.attention import BasicTransformerBlock
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class PaintByExampleImageEncoder(CLIPPreTrainedModel):
+ def __init__(self, config, proj_size=None):
+ super().__init__(config)
+ self.proj_size = proj_size or getattr(config, "projection_dim", 768)
+
+ self.model = CLIPVisionModel(config)
+ self.mapper = PaintByExampleMapper(config)
+ self.final_layer_norm = nn.LayerNorm(config.hidden_size)
+ self.proj_out = nn.Linear(config.hidden_size, self.proj_size)
+
+ # uncondition for scaling
+ self.uncond_vector = nn.Parameter(torch.randn((1, 1, self.proj_size)))
+
+ def forward(self, pixel_values, return_uncond_vector=False):
+ clip_output = self.model(pixel_values=pixel_values)
+ latent_states = clip_output.pooler_output
+ latent_states = self.mapper(latent_states[:, None])
+ latent_states = self.final_layer_norm(latent_states)
+ latent_states = self.proj_out(latent_states)
+ if return_uncond_vector:
+ return latent_states, self.uncond_vector
+
+ return latent_states
+
+
+class PaintByExampleMapper(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ num_layers = (config.num_hidden_layers + 1) // 5
+ hid_size = config.hidden_size
+ num_heads = 1
+ self.blocks = nn.ModuleList(
+ [
+ BasicTransformerBlock(hid_size, num_heads, hid_size, activation_fn="gelu", attention_bias=True)
+ for _ in range(num_layers)
+ ]
+ )
+
+ def forward(self, hidden_states):
+ for block in self.blocks:
+ hidden_states = block(hidden_states)
+
+ return hidden_states
diff --git a/model/diffusers_c/pipelines/paint_by_example/pipeline_paint_by_example.py b/model/diffusers_c/pipelines/paint_by_example/pipeline_paint_by_example.py
new file mode 100644
index 0000000000000000000000000000000000000000..8a24f134e7937757664060839c25205516d75f01
--- /dev/null
+++ b/model/diffusers_c/pipelines/paint_by_example/pipeline_paint_by_example.py
@@ -0,0 +1,621 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor
+
+from ...image_processor import VaeImageProcessor
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import deprecate, logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from .image_encoder import PaintByExampleImageEncoder
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+def prepare_mask_and_masked_image(image, mask):
+ """
+ Prepares a pair (image, mask) to be consumed by the Paint by Example pipeline. This means that those inputs will be
+ converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
+ ``image`` and ``1`` for the ``mask``.
+
+ The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+ binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+ Args:
+ image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+ ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+ mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+ ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+
+
+ Raises:
+ ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+ should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+ TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+ (ot the other way around).
+
+ Returns:
+ tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
+ dimensions: ``batch x channels x height x width``.
+ """
+ if isinstance(image, torch.Tensor):
+ if not isinstance(mask, torch.Tensor):
+ raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
+
+ # Batch single image
+ if image.ndim == 3:
+ assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
+ image = image.unsqueeze(0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Batched mask
+ if mask.shape[0] == image.shape[0]:
+ mask = mask.unsqueeze(1)
+ else:
+ mask = mask.unsqueeze(0)
+
+ assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+ assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+ assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+ assert mask.shape[1] == 1, "Mask image must have a single channel"
+
+ # Check image is in [-1, 1]
+ if image.min() < -1 or image.max() > 1:
+ raise ValueError("Image should be in [-1, 1] range")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("Mask should be in [0, 1] range")
+
+ # paint-by-example inverses the mask
+ mask = 1 - mask
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ # Image as float32
+ image = image.to(dtype=torch.float32)
+ elif isinstance(mask, torch.Tensor):
+ raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+ else:
+ if isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ image = np.concatenate([np.array(i.convert("RGB"))[None, :] for i in image], axis=0)
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ # preprocess mask
+ if isinstance(mask, PIL.Image.Image):
+ mask = [mask]
+
+ mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+ mask = mask.astype(np.float32) / 255.0
+
+ # paint-by-example inverses the mask
+ mask = 1 - mask
+
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+ mask = torch.from_numpy(mask)
+
+ masked_image = image * mask
+
+ return mask, masked_image
+
+
+class PaintByExamplePipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+
+
+ 🧪 This is an experimental feature!
+
+
+
+ Pipeline for image-guided image inpainting using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ image_encoder ([`PaintByExampleImageEncoder`]):
+ Encodes the example input image. The `unet` is conditioned on the example image instead of a text prompt.
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+
+ """
+
+ # TODO: feature_extractor is required to encode initial images (if they are in PIL format),
+ # we should give a descriptive message if the pipeline doesn't have one.
+
+ model_cpu_offload_seq = "unet->vae"
+ _exclude_from_cpu_offload = ["image_encoder"]
+ _optional_components = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ image_encoder: PaintByExampleImageEncoder,
+ unet: UNet2DConditionModel,
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = False,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ image_encoder=image_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_image_variation.StableDiffusionImageVariationPipeline.check_inputs
+ def check_inputs(self, image, height, width, callback_steps):
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+ f" {type(image)}"
+ )
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_mask_latents
+ def prepare_mask_latents(
+ self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+ ):
+ # resize the mask to latents shape as we concatenate the mask to the latents
+ # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+ # and half precision
+ mask = torch.nn.functional.interpolate(
+ mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+ )
+ mask = mask.to(device=device, dtype=dtype)
+
+ masked_image = masked_image.to(device=device, dtype=dtype)
+
+ if masked_image.shape[1] == 4:
+ masked_image_latents = masked_image
+ else:
+ masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+
+ # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+ if mask.shape[0] < batch_size:
+ if not batch_size % mask.shape[0] == 0:
+ raise ValueError(
+ "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+ f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+ " of masks that you pass is divisible by the total requested batch size."
+ )
+ mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+ if masked_image_latents.shape[0] < batch_size:
+ if not batch_size % masked_image_latents.shape[0] == 0:
+ raise ValueError(
+ "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+ f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+ " Make sure the number of images that you pass is divisible by the total requested batch size."
+ )
+ masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)
+
+ mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+ masked_image_latents = (
+ torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+ )
+
+ # aligning device to prevent device errors when concating it with the latent model input
+ masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+ return mask, masked_image_latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline._encode_vae_image
+ def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+ if isinstance(generator, list):
+ image_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(image.shape[0])
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ image_latents = self.vae.config.scaling_factor * image_latents
+
+ return image_latents
+
+ def _encode_image(self, image, device, num_images_per_prompt, do_classifier_free_guidance):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(images=image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ image_embeddings, negative_prompt_embeds = self.image_encoder(image, return_uncond_vector=True)
+
+ # duplicate image embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = image_embeddings.shape
+ image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)
+ image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, image_embeddings.shape[0], 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(bs_embed * num_images_per_prompt, 1, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings])
+
+ return image_embeddings
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ example_image: Union[torch.FloatTensor, PIL.Image.Image],
+ image: Union[torch.FloatTensor, PIL.Image.Image],
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ example_image (`torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]`):
+ An example image to guide image generation.
+ image (`torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]`):
+ `Image` or tensor representing an image batch to be inpainted (parts of the image are masked out with
+ `mask_image` and repainted according to `prompt`).
+ mask_image (`torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]`):
+ `Image` or tensor representing an image batch to mask `image`. White pixels in the mask are repainted,
+ while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a single channel
+ (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, so the
+ expected shape would be `(B, H, W, 1)`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Example:
+
+ ```py
+ >>> import PIL
+ >>> import requests
+ >>> import torch
+ >>> from io import BytesIO
+ >>> from diffusers import PaintByExamplePipeline
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = (
+ ... "https://raw.githubusercontent.com/Fantasy-Studio/Paint-by-Example/main/examples/image/example_1.png"
+ ... )
+ >>> mask_url = (
+ ... "https://raw.githubusercontent.com/Fantasy-Studio/Paint-by-Example/main/examples/mask/example_1.png"
+ ... )
+ >>> example_url = "https://raw.githubusercontent.com/Fantasy-Studio/Paint-by-Example/main/examples/reference/example_1.jpg"
+
+ >>> init_image = download_image(img_url).resize((512, 512))
+ >>> mask_image = download_image(mask_url).resize((512, 512))
+ >>> example_image = download_image(example_url).resize((512, 512))
+
+ >>> pipe = PaintByExamplePipeline.from_pretrained(
+ ... "Fantasy-Studio/Paint-by-Example",
+ ... torch_dtype=torch.float16,
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> image = pipe(image=init_image, mask_image=mask_image, example_image=example_image).images[0]
+ >>> image
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 1. Define call parameters
+ if isinstance(image, PIL.Image.Image):
+ batch_size = 1
+ elif isinstance(image, list):
+ batch_size = len(image)
+ else:
+ batch_size = image.shape[0]
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 2. Preprocess mask and image
+ mask, masked_image = prepare_mask_and_masked_image(image, mask_image)
+ height, width = masked_image.shape[-2:]
+
+ # 3. Check inputs
+ self.check_inputs(example_image, height, width, callback_steps)
+
+ # 4. Encode input image
+ image_embeddings = self._encode_image(
+ example_image, device, num_images_per_prompt, do_classifier_free_guidance
+ )
+
+ # 5. set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ image_embeddings.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 7. Prepare mask latent variables
+ mask, masked_image_latents = self.prepare_mask_latents(
+ mask,
+ masked_image,
+ batch_size * num_images_per_prompt,
+ height,
+ width,
+ image_embeddings.dtype,
+ device,
+ generator,
+ do_classifier_free_guidance,
+ )
+
+ # 8. Check that sizes of mask, masked image and latents match
+ num_channels_mask = mask.shape[1]
+ num_channels_masked_image = masked_image_latents.shape[1]
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+
+ # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 10. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+ latent_model_input = torch.cat([latent_model_input, masked_image_latents, mask], dim=1)
+
+ # predict the noise residual
+ noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ self.maybe_free_model_hooks()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, image_embeddings.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/pia/__init__.py b/model/diffusers_c/pipelines/pia/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..16e8004966e58387f7dba2f6ff3175575fe0abee
--- /dev/null
+++ b/model/diffusers_c/pipelines/pia/__init__.py
@@ -0,0 +1,46 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_pia"] = ["PIAPipeline", "PIAPipelineOutput"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+
+ else:
+ from .pipeline_pia import PIAPipeline, PIAPipelineOutput
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/pia/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/pia/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0c24fccd94589fe91fa3b992042e64693a45887a
Binary files /dev/null and b/model/diffusers_c/pipelines/pia/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/pia/__pycache__/pipeline_pia.cpython-39.pyc b/model/diffusers_c/pipelines/pia/__pycache__/pipeline_pia.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8fa3376d97cb8d21111c855bcd4fc8a07c0be361
Binary files /dev/null and b/model/diffusers_c/pipelines/pia/__pycache__/pipeline_pia.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/pia/pipeline_pia.py b/model/diffusers_c/pipelines/pia/pipeline_pia.py
new file mode 100644
index 0000000000000000000000000000000000000000..507088991a5e07361f210237089e25fd263ef205
--- /dev/null
+++ b/model/diffusers_c/pipelines/pia/pipeline_pia.py
@@ -0,0 +1,1034 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL
+import torch
+import torch.fft as fft
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...models.unets.unet_motion_model import MotionAdapter
+from ...schedulers import (
+ DDIMScheduler,
+ DPMSolverMultistepScheduler,
+ EulerAncestralDiscreteScheduler,
+ EulerDiscreteScheduler,
+ LMSDiscreteScheduler,
+ PNDMScheduler,
+)
+from ...utils import (
+ USE_PEFT_BACKEND,
+ BaseOutput,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..free_init_utils import FreeInitMixin
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import (
+ ... EulerDiscreteScheduler,
+ ... MotionAdapter,
+ ... PIAPipeline,
+ ... )
+ >>> from diffusers.utils import export_to_gif, load_image
+ >>> adapter = MotionAdapter.from_pretrained("../checkpoints/pia-diffusers")
+ >>> pipe = PIAPipeline.from_pretrained("SG161222/Realistic_Vision_V6.0_B1_noVAE", motion_adapter=adapter)
+ >>> pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config)
+ >>> image = load_image(
+ ... "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/pix2pix/cat_6.png?download=true"
+ ... )
+ >>> image = image.resize((512, 512))
+ >>> prompt = "cat in a hat"
+ >>> negative_prompt = "wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg"
+ >>> generator = torch.Generator("cpu").manual_seed(0)
+ >>> output = pipe(image=image, prompt=prompt, negative_prompt=negative_prompt, generator=generator)
+ >>> frames = output.frames[0]
+ >>> export_to_gif(frames, "pia-animation.gif")
+ ```
+"""
+
+RANGE_LIST = [
+ [1.0, 0.9, 0.85, 0.85, 0.85, 0.8], # 0 Small Motion
+ [1.0, 0.8, 0.8, 0.8, 0.79, 0.78, 0.75], # Moderate Motion
+ [1.0, 0.8, 0.7, 0.7, 0.7, 0.7, 0.7, 0.7, 0.7, 0.7, 0.6, 0.5, 0.5], # Large Motion
+ [1.0, 0.9, 0.85, 0.85, 0.85, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.8, 0.85, 0.85, 0.9, 1.0], # Loop
+ [1.0, 0.8, 0.8, 0.8, 0.79, 0.78, 0.75, 0.75, 0.75, 0.75, 0.75, 0.78, 0.79, 0.8, 0.8, 1.0], # Loop
+ [1.0, 0.8, 0.7, 0.7, 0.7, 0.7, 0.6, 0.5, 0.5, 0.6, 0.7, 0.7, 0.7, 0.7, 0.8, 1.0], # Loop
+ [0.5, 0.4, 0.4, 0.4, 0.35, 0.3], # Style Transfer Candidate Small Motion
+ [0.5, 0.4, 0.4, 0.4, 0.35, 0.35, 0.3, 0.25, 0.2], # Style Transfer Moderate Motion
+ [0.5, 0.2], # Style Transfer Large Motion
+]
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor: "VaeImageProcessor", output_type: str = "np"):
+ batch_size, channels, num_frames, height, width = video.shape
+ outputs = []
+ for batch_idx in range(batch_size):
+ batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+ batch_output = processor.postprocess(batch_vid, output_type)
+
+ outputs.append(batch_output)
+
+ if output_type == "np":
+ outputs = np.stack(outputs)
+
+ elif output_type == "pt":
+ outputs = torch.stack(outputs)
+
+ elif not output_type == "pil":
+ raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+ return outputs
+
+
+def prepare_mask_coef_by_statistics(num_frames: int, cond_frame: int, motion_scale: int):
+ assert num_frames > 0, "video_length should be greater than 0"
+
+ assert num_frames > cond_frame, "video_length should be greater than cond_frame"
+
+ range_list = RANGE_LIST
+
+ assert motion_scale < len(range_list), f"motion_scale type{motion_scale} not implemented"
+
+ coef = range_list[motion_scale]
+ coef = coef + ([coef[-1]] * (num_frames - len(coef)))
+
+ order = [abs(i - cond_frame) for i in range(num_frames)]
+ coef = [coef[order[i]] for i in range(num_frames)]
+
+ return coef
+
+
+def _get_freeinit_freq_filter(
+ shape: Tuple[int, ...],
+ device: Union[str, torch.dtype],
+ filter_type: str,
+ order: float,
+ spatial_stop_frequency: float,
+ temporal_stop_frequency: float,
+) -> torch.Tensor:
+ r"""Returns the FreeInit filter based on filter type and other input conditions."""
+
+ time, height, width = shape[-3], shape[-2], shape[-1]
+ mask = torch.zeros(shape)
+
+ if spatial_stop_frequency == 0 or temporal_stop_frequency == 0:
+ return mask
+
+ if filter_type == "butterworth":
+
+ def retrieve_mask(x):
+ return 1 / (1 + (x / spatial_stop_frequency**2) ** order)
+ elif filter_type == "gaussian":
+
+ def retrieve_mask(x):
+ return math.exp(-1 / (2 * spatial_stop_frequency**2) * x)
+ elif filter_type == "ideal":
+
+ def retrieve_mask(x):
+ return 1 if x <= spatial_stop_frequency * 2 else 0
+ else:
+ raise NotImplementedError("`filter_type` must be one of gaussian, butterworth or ideal")
+
+ for t in range(time):
+ for h in range(height):
+ for w in range(width):
+ d_square = (
+ ((spatial_stop_frequency / temporal_stop_frequency) * (2 * t / time - 1)) ** 2
+ + (2 * h / height - 1) ** 2
+ + (2 * w / width - 1) ** 2
+ )
+ mask[..., t, h, w] = retrieve_mask(d_square)
+
+ return mask.to(device)
+
+
+def _freq_mix_3d(x: torch.Tensor, noise: torch.Tensor, LPF: torch.Tensor) -> torch.Tensor:
+ r"""Noise reinitialization."""
+ # FFT
+ x_freq = fft.fftn(x, dim=(-3, -2, -1))
+ x_freq = fft.fftshift(x_freq, dim=(-3, -2, -1))
+ noise_freq = fft.fftn(noise, dim=(-3, -2, -1))
+ noise_freq = fft.fftshift(noise_freq, dim=(-3, -2, -1))
+
+ # frequency mix
+ HPF = 1 - LPF
+ x_freq_low = x_freq * LPF
+ noise_freq_high = noise_freq * HPF
+ x_freq_mixed = x_freq_low + noise_freq_high # mix in freq domain
+
+ # IFFT
+ x_freq_mixed = fft.ifftshift(x_freq_mixed, dim=(-3, -2, -1))
+ x_mixed = fft.ifftn(x_freq_mixed, dim=(-3, -2, -1)).real
+
+ return x_mixed
+
+
+@dataclass
+class PIAPipelineOutput(BaseOutput):
+ r"""
+ Output class for PIAPipeline.
+
+ Args:
+ frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+ Nested list of length `batch_size` with denoised PIL image sequences of length `num_frames`,
+ NumPy array of shape `(batch_size, num_frames, channels, height, width,
+ Torch tensor of shape `(batch_size, num_frames, channels, height, width)`.
+ """
+
+ frames: Union[torch.Tensor, np.ndarray, List[List[PIL.Image.Image]]]
+
+
+class PIAPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FromSingleFileMixin,
+ FreeInitMixin,
+):
+ r"""
+ Pipeline for text-to-video generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`] to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A [`UNet2DConditionModel`] used to create a UNetMotionModel to denoise the encoded video latents.
+ motion_adapter ([`MotionAdapter`]):
+ A [`MotionAdapter`] to be used in combination with `unet` to denoise the encoded video latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["feature_extractor", "image_encoder", "motion_adapter"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: Union[UNet2DConditionModel, UNetMotionModel],
+ scheduler: Union[
+ DDIMScheduler,
+ PNDMScheduler,
+ LMSDiscreteScheduler,
+ EulerDiscreteScheduler,
+ EulerAncestralDiscreteScheduler,
+ DPMSolverMultistepScheduler,
+ ],
+ motion_adapter: Optional[MotionAdapter] = None,
+ feature_extractor: CLIPImageProcessor = None,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ ):
+ super().__init__()
+ if isinstance(unet, UNet2DConditionModel):
+ unet = UNetMotionModel.from_unet2d(unet, motion_adapter)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ motion_adapter=motion_adapter,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with num_images_per_prompt -> num_videos_per_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis/pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
+ def decode_latents(self, latents):
+ latents = 1 / self.vae.config.scaling_factor * latents
+
+ batch_size, channels, num_frames, height, width = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+ image = self.vae.decode(latents).sample
+ video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ video = video.float()
+ return video
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
+ def prepare_latents(
+ self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+ ):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ num_frames,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def prepare_masked_condition(
+ self,
+ image,
+ batch_size,
+ num_channels_latents,
+ num_frames,
+ height,
+ width,
+ dtype,
+ device,
+ generator,
+ motion_scale=0,
+ ):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ num_frames,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ _, _, _, scaled_height, scaled_width = shape
+
+ image = self.image_processor.preprocess(image)
+ image = image.to(device, dtype)
+
+ if isinstance(generator, list):
+ image_latent = [
+ self.vae.encode(image[k : k + 1]).latent_dist.sample(generator[k]) for k in range(batch_size)
+ ]
+ image_latent = torch.cat(image_latent, dim=0)
+ else:
+ image_latent = self.vae.encode(image).latent_dist.sample(generator)
+
+ image_latent = image_latent.to(device=device, dtype=dtype)
+ image_latent = torch.nn.functional.interpolate(image_latent, size=[scaled_height, scaled_width])
+ image_latent_padding = image_latent.clone() * self.vae.config.scaling_factor
+
+ mask = torch.zeros((batch_size, 1, num_frames, scaled_height, scaled_width)).to(device=device, dtype=dtype)
+ mask_coef = prepare_mask_coef_by_statistics(num_frames, 0, motion_scale)
+ masked_image = torch.zeros(batch_size, 4, num_frames, scaled_height, scaled_width).to(
+ device=device, dtype=self.unet.dtype
+ )
+ for f in range(num_frames):
+ mask[:, :, f, :, :] = mask_coef[f]
+ masked_image[:, :, f, :, :] = image_latent_padding.clone()
+
+ mask = torch.cat([mask] * 2) if self.do_classifier_free_guidance else mask
+ masked_image = torch.cat([masked_image] * 2) if self.do_classifier_free_guidance else masked_image
+
+ return mask, masked_image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image: PipelineImageInput,
+ prompt: Union[str, List[str]] = None,
+ strength: float = 1.0,
+ num_frames: Optional[int] = 16,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_videos_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ motion_scale: int = 0,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`PipelineImageInput`):
+ The input image to be used for video generation.
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ strength (`float`, *optional*, defaults to 1.0): Indicates extent to transform the reference `image`. Must be between 0 and 1.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated video.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated video.
+ num_frames (`int`, *optional*, defaults to 16):
+ The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds
+ amounts to 2 seconds of video.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`. Latents should be of shape
+ `(batch_size, num_channel, num_frames, height, width)`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ motion_scale: (`int`, *optional*, defaults to 0):
+ Parameter that controls the amount and type of motion that is added to the image. Increasing the value increases the amount of motion, while specific
+ ranges of values control the type of motion that is added. Must be between 0 and 8.
+ Set between 0-2 to only increase the amount of motion.
+ Set between 3-5 to create looping motion.
+ Set between 6-8 to perform motion with image style transfer.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated video. Choose between `torch.FloatTensor`, `PIL.Image` or
+ `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead
+ of a plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.pia.pipeline_pia.PIAPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.pia.pipeline_pia.PIAPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ num_videos_per_prompt = 1
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_videos_per_prompt)
+ self._num_timesteps = len(timesteps)
+
+ # 5. Prepare latent variables
+ latents = self.prepare_latents(
+ batch_size * num_videos_per_prompt,
+ 4,
+ num_frames,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents=latents,
+ )
+ mask, masked_image = self.prepare_masked_condition(
+ image,
+ batch_size * num_videos_per_prompt,
+ 4,
+ num_frames=num_frames,
+ height=height,
+ width=width,
+ dtype=self.unet.dtype,
+ device=device,
+ generator=generator,
+ motion_scale=motion_scale,
+ )
+ if strength < 1.0:
+ noise = randn_tensor(latents.shape, generator=generator, device=device, dtype=latents.dtype)
+ latents = self.scheduler.add_noise(masked_image[0], noise, latent_timestep)
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 8. Denoising loop
+ num_free_init_iters = self._free_init_num_iters if self.free_init_enabled else 1
+ for free_init_iter in range(num_free_init_iters):
+ if self.free_init_enabled:
+ latents, timesteps = self._apply_free_init(
+ latents, free_init_iter, num_inference_steps, device, latents.dtype, generator
+ )
+
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+ latent_model_input = torch.cat([latent_model_input, mask, masked_image], dim=1)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ ).sample
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+
+ # 9. Post processing
+ if output_type == "latent":
+ video = latents
+ else:
+ video_tensor = self.decode_latents(latents)
+ video = tensor2vid(video_tensor, self.image_processor, output_type=output_type)
+
+ # 10. Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (video,)
+
+ return PIAPipelineOutput(frames=video)
diff --git a/model/diffusers_c/pipelines/pipeline_flax_utils.py b/model/diffusers_c/pipelines/pipeline_flax_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..97ad6988aa8e84a38cb4574e952b63ba8e9c3cc7
--- /dev/null
+++ b/model/diffusers_c/pipelines/pipeline_flax_utils.py
@@ -0,0 +1,615 @@
+# Copyright 2024 The HuggingFace Inc. team.
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib
+import inspect
+import os
+from typing import Any, Dict, List, Optional, Union
+
+import flax
+import numpy as np
+import PIL.Image
+from flax.core.frozen_dict import FrozenDict
+from huggingface_hub import create_repo, snapshot_download
+from huggingface_hub.utils import validate_hf_hub_args
+from PIL import Image
+from tqdm.auto import tqdm
+
+from ..configuration_utils import ConfigMixin
+from ..models.modeling_flax_utils import FLAX_WEIGHTS_NAME, FlaxModelMixin
+from ..schedulers.scheduling_utils_flax import SCHEDULER_CONFIG_NAME, FlaxSchedulerMixin
+from ..utils import (
+ CONFIG_NAME,
+ BaseOutput,
+ PushToHubMixin,
+ http_user_agent,
+ is_transformers_available,
+ logging,
+)
+
+
+if is_transformers_available():
+ from transformers import FlaxPreTrainedModel
+
+INDEX_FILE = "diffusion_flax_model.bin"
+
+
+logger = logging.get_logger(__name__)
+
+
+LOADABLE_CLASSES = {
+ "diffusers": {
+ "FlaxModelMixin": ["save_pretrained", "from_pretrained"],
+ "FlaxSchedulerMixin": ["save_pretrained", "from_pretrained"],
+ "FlaxDiffusionPipeline": ["save_pretrained", "from_pretrained"],
+ },
+ "transformers": {
+ "PreTrainedTokenizer": ["save_pretrained", "from_pretrained"],
+ "PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"],
+ "FlaxPreTrainedModel": ["save_pretrained", "from_pretrained"],
+ "FeatureExtractionMixin": ["save_pretrained", "from_pretrained"],
+ "ProcessorMixin": ["save_pretrained", "from_pretrained"],
+ "ImageProcessingMixin": ["save_pretrained", "from_pretrained"],
+ },
+}
+
+ALL_IMPORTABLE_CLASSES = {}
+for library in LOADABLE_CLASSES:
+ ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])
+
+
+def import_flax_or_no_model(module, class_name):
+ try:
+ # 1. First make sure that if a Flax object is present, import this one
+ class_obj = getattr(module, "Flax" + class_name)
+ except AttributeError:
+ # 2. If this doesn't work, it's not a model and we don't append "Flax"
+ class_obj = getattr(module, class_name)
+ except AttributeError:
+ raise ValueError(f"Neither Flax{class_name} nor {class_name} exist in {module}")
+
+ return class_obj
+
+
+@flax.struct.dataclass
+class FlaxImagePipelineOutput(BaseOutput):
+ """
+ Output class for image pipelines.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+class FlaxDiffusionPipeline(ConfigMixin, PushToHubMixin):
+ r"""
+ Base class for Flax-based pipelines.
+
+ [`FlaxDiffusionPipeline`] stores all components (models, schedulers, and processors) for diffusion pipelines and
+ provides methods for loading, downloading and saving models. It also includes methods to:
+
+ - enable/disable the progress bar for the denoising iteration
+
+ Class attributes:
+
+ - **config_name** ([`str`]) -- The configuration filename that stores the class and module names of all the
+ diffusion pipeline's components.
+ """
+
+ config_name = "model_index.json"
+
+ def register_modules(self, **kwargs):
+ # import it here to avoid circular import
+ from diffusers import pipelines
+
+ for name, module in kwargs.items():
+ if module is None:
+ register_dict = {name: (None, None)}
+ else:
+ # retrieve library
+ library = module.__module__.split(".")[0]
+
+ # check if the module is a pipeline module
+ pipeline_dir = module.__module__.split(".")[-2]
+ path = module.__module__.split(".")
+ is_pipeline_module = pipeline_dir in path and hasattr(pipelines, pipeline_dir)
+
+ # if library is not in LOADABLE_CLASSES, then it is a custom module.
+ # Or if it's a pipeline module, then the module is inside the pipeline
+ # folder so we set the library to module name.
+ if library not in LOADABLE_CLASSES or is_pipeline_module:
+ library = pipeline_dir
+
+ # retrieve class_name
+ class_name = module.__class__.__name__
+
+ register_dict = {name: (library, class_name)}
+
+ # save model index config
+ self.register_to_config(**register_dict)
+
+ # set models
+ setattr(self, name, module)
+
+ def save_pretrained(
+ self,
+ save_directory: Union[str, os.PathLike],
+ params: Union[Dict, FrozenDict],
+ push_to_hub: bool = False,
+ **kwargs,
+ ):
+ # TODO: handle inference_state
+ """
+ Save all saveable variables of the pipeline to a directory. A pipeline variable can be saved and loaded if its
+ class implements both a save and loading method. The pipeline is easily reloaded using the
+ [`~FlaxDiffusionPipeline.from_pretrained`] class method.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to which to save. Will be created if it doesn't exist.
+ push_to_hub (`bool`, *optional*, defaults to `False`):
+ Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the
+ repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+ namespace).
+ kwargs (`Dict[str, Any]`, *optional*):
+ Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+ """
+ self.save_config(save_directory)
+
+ model_index_dict = dict(self.config)
+ model_index_dict.pop("_class_name")
+ model_index_dict.pop("_diffusers_version")
+ model_index_dict.pop("_module", None)
+
+ if push_to_hub:
+ commit_message = kwargs.pop("commit_message", None)
+ private = kwargs.pop("private", False)
+ create_pr = kwargs.pop("create_pr", False)
+ token = kwargs.pop("token", None)
+ repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
+ repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id
+
+ for pipeline_component_name in model_index_dict.keys():
+ sub_model = getattr(self, pipeline_component_name)
+ if sub_model is None:
+ # edge case for saving a pipeline with safety_checker=None
+ continue
+
+ model_cls = sub_model.__class__
+
+ save_method_name = None
+ # search for the model's base class in LOADABLE_CLASSES
+ for library_name, library_classes in LOADABLE_CLASSES.items():
+ library = importlib.import_module(library_name)
+ for base_class, save_load_methods in library_classes.items():
+ class_candidate = getattr(library, base_class, None)
+ if class_candidate is not None and issubclass(model_cls, class_candidate):
+ # if we found a suitable base class in LOADABLE_CLASSES then grab its save method
+ save_method_name = save_load_methods[0]
+ break
+ if save_method_name is not None:
+ break
+
+ save_method = getattr(sub_model, save_method_name)
+ expects_params = "params" in set(inspect.signature(save_method).parameters.keys())
+
+ if expects_params:
+ save_method(
+ os.path.join(save_directory, pipeline_component_name), params=params[pipeline_component_name]
+ )
+ else:
+ save_method(os.path.join(save_directory, pipeline_component_name))
+
+ if push_to_hub:
+ self._upload_folder(
+ save_directory,
+ repo_id,
+ token=token,
+ commit_message=commit_message,
+ create_pr=create_pr,
+ )
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
+ r"""
+ Instantiate a Flax-based diffusion pipeline from pretrained pipeline weights.
+
+ The pipeline is set in evaluation mode (`model.eval()) by default and dropout modules are deactivated.
+
+ If you get the error message below, you need to finetune the weights for your downstream task:
+
+ ```
+ Some weights of FlaxUNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+ ```
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *repo id* (for example `runwayml/stable-diffusion-v1-5`) of a pretrained pipeline
+ hosted on the Hub.
+ - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+ using [`~FlaxDiffusionPipeline.save_pretrained`].
+ dtype (`str` or `jnp.dtype`, *optional*):
+ Override the default `jnp.dtype` and load the model under this dtype. If `"auto"`, the dtype is
+ automatically derived from the model's weights.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to overwrite load and saveable variables (the pipeline components) of the specific pipeline
+ class. The overwritten components are passed directly to the pipelines `__init__` method.
+
+
+
+ To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
+ `huggingface-cli login`.
+
+
+
+ Examples:
+
+ ```py
+ >>> from diffusers import FlaxDiffusionPipeline
+
+ >>> # Download pipeline from huggingface.co and cache.
+ >>> # Requires to be logged in to Hugging Face hub,
+ >>> # see more in [the documentation](https://huggingface.co/docs/hub/security-tokens)
+ >>> pipeline, params = FlaxDiffusionPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5",
+ ... revision="bf16",
+ ... dtype=jnp.bfloat16,
+ ... )
+
+ >>> # Download pipeline, but use a different scheduler
+ >>> from diffusers import FlaxDPMSolverMultistepScheduler
+
+ >>> model_id = "runwayml/stable-diffusion-v1-5"
+ >>> dpmpp, dpmpp_state = FlaxDPMSolverMultistepScheduler.from_pretrained(
+ ... model_id,
+ ... subfolder="scheduler",
+ ... )
+
+ >>> dpm_pipe, dpm_params = FlaxStableDiffusionPipeline.from_pretrained(
+ ... model_id, revision="bf16", dtype=jnp.bfloat16, scheduler=dpmpp
+ ... )
+ >>> dpm_params["scheduler"] = dpmpp_state
+ ```
+ """
+ cache_dir = kwargs.pop("cache_dir", None)
+ resume_download = kwargs.pop("resume_download", False)
+ proxies = kwargs.pop("proxies", None)
+ local_files_only = kwargs.pop("local_files_only", False)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ from_pt = kwargs.pop("from_pt", False)
+ use_memory_efficient_attention = kwargs.pop("use_memory_efficient_attention", False)
+ split_head_dim = kwargs.pop("split_head_dim", False)
+ dtype = kwargs.pop("dtype", None)
+
+ # 1. Download the checkpoints and configs
+ # use snapshot download here to get it working from from_pretrained
+ if not os.path.isdir(pretrained_model_name_or_path):
+ config_dict = cls.load_config(
+ pretrained_model_name_or_path,
+ cache_dir=cache_dir,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ )
+ # make sure we only download sub-folders and `diffusers` filenames
+ folder_names = [k for k in config_dict.keys() if not k.startswith("_")]
+ allow_patterns = [os.path.join(k, "*") for k in folder_names]
+ allow_patterns += [FLAX_WEIGHTS_NAME, SCHEDULER_CONFIG_NAME, CONFIG_NAME, cls.config_name]
+
+ ignore_patterns = ["*.bin", "*.safetensors"] if not from_pt else []
+ ignore_patterns += ["*.onnx", "*.onnx_data", "*.xml", "*.pb"]
+
+ if cls != FlaxDiffusionPipeline:
+ requested_pipeline_class = cls.__name__
+ else:
+ requested_pipeline_class = config_dict.get("_class_name", cls.__name__)
+ requested_pipeline_class = (
+ requested_pipeline_class
+ if requested_pipeline_class.startswith("Flax")
+ else "Flax" + requested_pipeline_class
+ )
+
+ user_agent = {"pipeline_class": requested_pipeline_class}
+ user_agent = http_user_agent(user_agent)
+
+ # download all allow_patterns
+ cached_folder = snapshot_download(
+ pretrained_model_name_or_path,
+ cache_dir=cache_dir,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ allow_patterns=allow_patterns,
+ ignore_patterns=ignore_patterns,
+ user_agent=user_agent,
+ )
+ else:
+ cached_folder = pretrained_model_name_or_path
+
+ config_dict = cls.load_config(cached_folder)
+
+ # 2. Load the pipeline class, if using custom module then load it from the hub
+ # if we load from explicit class, let's use it
+ if cls != FlaxDiffusionPipeline:
+ pipeline_class = cls
+ else:
+ diffusers_module = importlib.import_module(cls.__module__.split(".")[0])
+ class_name = (
+ config_dict["_class_name"]
+ if config_dict["_class_name"].startswith("Flax")
+ else "Flax" + config_dict["_class_name"]
+ )
+ pipeline_class = getattr(diffusers_module, class_name)
+
+ # some modules can be passed directly to the init
+ # in this case they are already instantiated in `kwargs`
+ # extract them here
+ expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class)
+ passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
+ passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
+
+ init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs)
+
+ # define init kwargs
+ init_kwargs = {k: init_dict.pop(k) for k in optional_kwargs if k in init_dict}
+ init_kwargs = {**init_kwargs, **passed_pipe_kwargs}
+
+ # remove `null` components
+ def load_module(name, value):
+ if value[0] is None:
+ return False
+ if name in passed_class_obj and passed_class_obj[name] is None:
+ return False
+ return True
+
+ init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)}
+
+ # Throw nice warnings / errors for fast accelerate loading
+ if len(unused_kwargs) > 0:
+ logger.warning(
+ f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored."
+ )
+
+ # inference_params
+ params = {}
+
+ # import it here to avoid circular import
+ from diffusers import pipelines
+
+ # 3. Load each module in the pipeline
+ for name, (library_name, class_name) in init_dict.items():
+ if class_name is None:
+ # edge case for when the pipeline was saved with safety_checker=None
+ init_kwargs[name] = None
+ continue
+
+ is_pipeline_module = hasattr(pipelines, library_name)
+ loaded_sub_model = None
+ sub_model_should_be_defined = True
+
+ # if the model is in a pipeline module, then we load it from the pipeline
+ if name in passed_class_obj:
+ # 1. check that passed_class_obj has correct parent class
+ if not is_pipeline_module:
+ library = importlib.import_module(library_name)
+ class_obj = getattr(library, class_name)
+ importable_classes = LOADABLE_CLASSES[library_name]
+ class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()}
+
+ expected_class_obj = None
+ for class_name, class_candidate in class_candidates.items():
+ if class_candidate is not None and issubclass(class_obj, class_candidate):
+ expected_class_obj = class_candidate
+
+ if not issubclass(passed_class_obj[name].__class__, expected_class_obj):
+ raise ValueError(
+ f"{passed_class_obj[name]} is of type: {type(passed_class_obj[name])}, but should be"
+ f" {expected_class_obj}"
+ )
+ elif passed_class_obj[name] is None:
+ logger.warning(
+ f"You have passed `None` for {name} to disable its functionality in {pipeline_class}. Note"
+ f" that this might lead to problems when using {pipeline_class} and is not recommended."
+ )
+ sub_model_should_be_defined = False
+ else:
+ logger.warning(
+ f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it"
+ " has the correct type"
+ )
+
+ # set passed class object
+ loaded_sub_model = passed_class_obj[name]
+ elif is_pipeline_module:
+ pipeline_module = getattr(pipelines, library_name)
+ class_obj = import_flax_or_no_model(pipeline_module, class_name)
+
+ importable_classes = ALL_IMPORTABLE_CLASSES
+ class_candidates = {c: class_obj for c in importable_classes.keys()}
+ else:
+ # else we just import it from the library.
+ library = importlib.import_module(library_name)
+ class_obj = import_flax_or_no_model(library, class_name)
+
+ importable_classes = LOADABLE_CLASSES[library_name]
+ class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()}
+
+ if loaded_sub_model is None and sub_model_should_be_defined:
+ load_method_name = None
+ for class_name, class_candidate in class_candidates.items():
+ if class_candidate is not None and issubclass(class_obj, class_candidate):
+ load_method_name = importable_classes[class_name][1]
+
+ load_method = getattr(class_obj, load_method_name)
+
+ # check if the module is in a subdirectory
+ if os.path.isdir(os.path.join(cached_folder, name)):
+ loadable_folder = os.path.join(cached_folder, name)
+ else:
+ loaded_sub_model = cached_folder
+
+ if issubclass(class_obj, FlaxModelMixin):
+ loaded_sub_model, loaded_params = load_method(
+ loadable_folder,
+ from_pt=from_pt,
+ use_memory_efficient_attention=use_memory_efficient_attention,
+ split_head_dim=split_head_dim,
+ dtype=dtype,
+ )
+ params[name] = loaded_params
+ elif is_transformers_available() and issubclass(class_obj, FlaxPreTrainedModel):
+ if from_pt:
+ # TODO(Suraj): Fix this in Transformers. We should be able to use `_do_init=False` here
+ loaded_sub_model = load_method(loadable_folder, from_pt=from_pt)
+ loaded_params = loaded_sub_model.params
+ del loaded_sub_model._params
+ else:
+ loaded_sub_model, loaded_params = load_method(loadable_folder, _do_init=False)
+ params[name] = loaded_params
+ elif issubclass(class_obj, FlaxSchedulerMixin):
+ loaded_sub_model, scheduler_state = load_method(loadable_folder)
+ params[name] = scheduler_state
+ else:
+ loaded_sub_model = load_method(loadable_folder)
+
+ init_kwargs[name] = loaded_sub_model # UNet(...), # DiffusionSchedule(...)
+
+ # 4. Potentially add passed objects if expected
+ missing_modules = set(expected_modules) - set(init_kwargs.keys())
+ passed_modules = list(passed_class_obj.keys())
+
+ if len(missing_modules) > 0 and missing_modules <= set(passed_modules):
+ for module in missing_modules:
+ init_kwargs[module] = passed_class_obj.get(module, None)
+ elif len(missing_modules) > 0:
+ passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - optional_kwargs
+ raise ValueError(
+ f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed."
+ )
+
+ model = pipeline_class(**init_kwargs, dtype=dtype)
+ return model, params
+
+ @classmethod
+ def _get_signature_keys(cls, obj):
+ parameters = inspect.signature(obj.__init__).parameters
+ required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty}
+ optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty})
+ expected_modules = set(required_parameters.keys()) - {"self"}
+
+ return expected_modules, optional_parameters
+
+ @property
+ def components(self) -> Dict[str, Any]:
+ r"""
+
+ The `self.components` property can be useful to run different pipelines with the same weights and
+ configurations to not have to re-allocate memory.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import (
+ ... FlaxStableDiffusionPipeline,
+ ... FlaxStableDiffusionImg2ImgPipeline,
+ ... )
+
+ >>> text2img = FlaxStableDiffusionPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", revision="bf16", dtype=jnp.bfloat16
+ ... )
+ >>> img2img = FlaxStableDiffusionImg2ImgPipeline(**text2img.components)
+ ```
+
+ Returns:
+ A dictionary containing all the modules needed to initialize the pipeline.
+ """
+ expected_modules, optional_parameters = self._get_signature_keys(self)
+ components = {
+ k: getattr(self, k) for k in self.config.keys() if not k.startswith("_") and k not in optional_parameters
+ }
+
+ if set(components.keys()) != expected_modules:
+ raise ValueError(
+ f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected"
+ f" {expected_modules} to be defined, but {components} are defined."
+ )
+
+ return components
+
+ @staticmethod
+ def numpy_to_pil(images):
+ """
+ Convert a NumPy image or a batch of images to a PIL image.
+ """
+ if images.ndim == 3:
+ images = images[None, ...]
+ images = (images * 255).round().astype("uint8")
+ if images.shape[-1] == 1:
+ # special case for grayscale (single channel) images
+ pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
+ else:
+ pil_images = [Image.fromarray(image) for image in images]
+
+ return pil_images
+
+ # TODO: make it compatible with jax.lax
+ def progress_bar(self, iterable):
+ if not hasattr(self, "_progress_bar_config"):
+ self._progress_bar_config = {}
+ elif not isinstance(self._progress_bar_config, dict):
+ raise ValueError(
+ f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}."
+ )
+
+ return tqdm(iterable, **self._progress_bar_config)
+
+ def set_progress_bar_config(self, **kwargs):
+ self._progress_bar_config = kwargs
diff --git a/model/diffusers_c/pipelines/pipeline_loading_utils.py b/model/diffusers_c/pipelines/pipeline_loading_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..0c20f09a264e9d9cb0a88ce6475e6910c6d48a79
--- /dev/null
+++ b/model/diffusers_c/pipelines/pipeline_loading_utils.py
@@ -0,0 +1,507 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import importlib
+import os
+import re
+import warnings
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Union
+
+import torch
+from huggingface_hub import (
+ model_info,
+)
+from packaging import version
+
+from ..utils import (
+ SAFETENSORS_WEIGHTS_NAME,
+ WEIGHTS_NAME,
+ get_class_from_dynamic_module,
+ is_peft_available,
+ is_transformers_available,
+ logging,
+)
+from ..utils.torch_utils import is_compiled_module
+
+
+if is_transformers_available():
+ import transformers
+ from transformers import PreTrainedModel
+ from transformers.utils import FLAX_WEIGHTS_NAME as TRANSFORMERS_FLAX_WEIGHTS_NAME
+ from transformers.utils import SAFE_WEIGHTS_NAME as TRANSFORMERS_SAFE_WEIGHTS_NAME
+ from transformers.utils import WEIGHTS_NAME as TRANSFORMERS_WEIGHTS_NAME
+from huggingface_hub.utils import validate_hf_hub_args
+
+from ..utils import FLAX_WEIGHTS_NAME, ONNX_EXTERNAL_WEIGHTS_NAME, ONNX_WEIGHTS_NAME
+
+
+INDEX_FILE = "diffusion_pytorch_model.bin"
+CUSTOM_PIPELINE_FILE_NAME = "pipeline.py"
+DUMMY_MODULES_FOLDER = "diffusers.utils"
+TRANSFORMERS_DUMMY_MODULES_FOLDER = "transformers.utils"
+CONNECTED_PIPES_KEYS = ["prior"]
+
+logger = logging.get_logger(__name__)
+
+LOADABLE_CLASSES = {
+ "diffusers": {
+ "ModelMixin": ["save_pretrained", "from_pretrained"],
+ "SchedulerMixin": ["save_pretrained", "from_pretrained"],
+ "DiffusionPipeline": ["save_pretrained", "from_pretrained"],
+ "OnnxRuntimeModel": ["save_pretrained", "from_pretrained"],
+ },
+ "transformers": {
+ "PreTrainedTokenizer": ["save_pretrained", "from_pretrained"],
+ "PreTrainedTokenizerFast": ["save_pretrained", "from_pretrained"],
+ "PreTrainedModel": ["save_pretrained", "from_pretrained"],
+ "FeatureExtractionMixin": ["save_pretrained", "from_pretrained"],
+ "ProcessorMixin": ["save_pretrained", "from_pretrained"],
+ "ImageProcessingMixin": ["save_pretrained", "from_pretrained"],
+ },
+ "onnxruntime.training": {
+ "ORTModule": ["save_pretrained", "from_pretrained"],
+ },
+}
+
+ALL_IMPORTABLE_CLASSES = {}
+for library in LOADABLE_CLASSES:
+ ALL_IMPORTABLE_CLASSES.update(LOADABLE_CLASSES[library])
+
+
+def is_safetensors_compatible(filenames, variant=None, passed_components=None) -> bool:
+ """
+ Checking for safetensors compatibility:
+ - By default, all models are saved with the default pytorch serialization, so we use the list of default pytorch
+ files to know which safetensors files are needed.
+ - The model is safetensors compatible only if there is a matching safetensors file for every default pytorch file.
+
+ Converting default pytorch serialized filenames to safetensors serialized filenames:
+ - For models from the diffusers library, just replace the ".bin" extension with ".safetensors"
+ - For models from the transformers library, the filename changes from "pytorch_model" to "model", and the ".bin"
+ extension is replaced with ".safetensors"
+ """
+ pt_filenames = []
+
+ sf_filenames = set()
+
+ passed_components = passed_components or []
+
+ for filename in filenames:
+ _, extension = os.path.splitext(filename)
+
+ if len(filename.split("/")) == 2 and filename.split("/")[0] in passed_components:
+ continue
+
+ if extension == ".bin":
+ pt_filenames.append(os.path.normpath(filename))
+ elif extension == ".safetensors":
+ sf_filenames.add(os.path.normpath(filename))
+
+ for filename in pt_filenames:
+ # filename = 'foo/bar/baz.bam' -> path = 'foo/bar', filename = 'baz', extension = '.bam'
+ path, filename = os.path.split(filename)
+ filename, extension = os.path.splitext(filename)
+
+ if filename.startswith("pytorch_model"):
+ filename = filename.replace("pytorch_model", "model")
+ else:
+ filename = filename
+
+ expected_sf_filename = os.path.normpath(os.path.join(path, filename))
+ expected_sf_filename = f"{expected_sf_filename}.safetensors"
+ if expected_sf_filename not in sf_filenames:
+ logger.warning(f"{expected_sf_filename} not found")
+ return False
+
+ return True
+
+
+def variant_compatible_siblings(filenames, variant=None) -> Union[List[os.PathLike], str]:
+ weight_names = [
+ WEIGHTS_NAME,
+ SAFETENSORS_WEIGHTS_NAME,
+ FLAX_WEIGHTS_NAME,
+ ONNX_WEIGHTS_NAME,
+ ONNX_EXTERNAL_WEIGHTS_NAME,
+ ]
+
+ if is_transformers_available():
+ weight_names += [TRANSFORMERS_WEIGHTS_NAME, TRANSFORMERS_SAFE_WEIGHTS_NAME, TRANSFORMERS_FLAX_WEIGHTS_NAME]
+
+ # model_pytorch, diffusion_model_pytorch, ...
+ weight_prefixes = [w.split(".")[0] for w in weight_names]
+ # .bin, .safetensors, ...
+ weight_suffixs = [w.split(".")[-1] for w in weight_names]
+ # -00001-of-00002
+ transformers_index_format = r"\d{5}-of-\d{5}"
+
+ if variant is not None:
+ # `diffusion_pytorch_model.fp16.bin` as well as `model.fp16-00001-of-00002.safetensors`
+ variant_file_re = re.compile(
+ rf"({'|'.join(weight_prefixes)})\.({variant}|{variant}-{transformers_index_format})\.({'|'.join(weight_suffixs)})$"
+ )
+ # `text_encoder/pytorch_model.bin.index.fp16.json`
+ variant_index_re = re.compile(
+ rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.{variant}\.json$"
+ )
+
+ # `diffusion_pytorch_model.bin` as well as `model-00001-of-00002.safetensors`
+ non_variant_file_re = re.compile(
+ rf"({'|'.join(weight_prefixes)})(-{transformers_index_format})?\.({'|'.join(weight_suffixs)})$"
+ )
+ # `text_encoder/pytorch_model.bin.index.json`
+ non_variant_index_re = re.compile(rf"({'|'.join(weight_prefixes)})\.({'|'.join(weight_suffixs)})\.index\.json")
+
+ if variant is not None:
+ variant_weights = {f for f in filenames if variant_file_re.match(f.split("/")[-1]) is not None}
+ variant_indexes = {f for f in filenames if variant_index_re.match(f.split("/")[-1]) is not None}
+ variant_filenames = variant_weights | variant_indexes
+ else:
+ variant_filenames = set()
+
+ non_variant_weights = {f for f in filenames if non_variant_file_re.match(f.split("/")[-1]) is not None}
+ non_variant_indexes = {f for f in filenames if non_variant_index_re.match(f.split("/")[-1]) is not None}
+ non_variant_filenames = non_variant_weights | non_variant_indexes
+
+ # all variant filenames will be used by default
+ usable_filenames = set(variant_filenames)
+
+ def convert_to_variant(filename):
+ if "index" in filename:
+ variant_filename = filename.replace("index", f"index.{variant}")
+ elif re.compile(f"^(.*?){transformers_index_format}").match(filename) is not None:
+ variant_filename = f"{filename.split('-')[0]}.{variant}-{'-'.join(filename.split('-')[1:])}"
+ else:
+ variant_filename = f"{filename.split('.')[0]}.{variant}.{filename.split('.')[1]}"
+ return variant_filename
+
+ for f in non_variant_filenames:
+ variant_filename = convert_to_variant(f)
+ if variant_filename not in usable_filenames:
+ usable_filenames.add(f)
+
+ return usable_filenames, variant_filenames
+
+
+@validate_hf_hub_args
+def warn_deprecated_model_variant(pretrained_model_name_or_path, token, variant, revision, model_filenames):
+ info = model_info(
+ pretrained_model_name_or_path,
+ token=token,
+ revision=None,
+ )
+ filenames = {sibling.rfilename for sibling in info.siblings}
+ comp_model_filenames, _ = variant_compatible_siblings(filenames, variant=revision)
+ comp_model_filenames = [".".join(f.split(".")[:1] + f.split(".")[2:]) for f in comp_model_filenames]
+
+ if set(model_filenames).issubset(set(comp_model_filenames)):
+ warnings.warn(
+ f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'` even though you can load it via `variant=`{revision}`. Loading model variants via `revision='{revision}'` is deprecated and will be removed in diffusers v1. Please use `variant='{revision}'` instead.",
+ FutureWarning,
+ )
+ else:
+ warnings.warn(
+ f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'`. This behavior is deprecated and will be removed in diffusers v1. One should use `variant='{revision}'` instead. However, it appears that {pretrained_model_name_or_path} currently does not have the required variant filenames in the 'main' branch. \n The Diffusers team and community would be very grateful if you could open an issue: https://github.com/huggingface/diffusers/issues/new with the title '{pretrained_model_name_or_path} is missing {revision} files' so that the correct variant file can be added.",
+ FutureWarning,
+ )
+
+
+def _unwrap_model(model):
+ """Unwraps a model."""
+ if is_compiled_module(model):
+ model = model._orig_mod
+
+ if is_peft_available():
+ from peft import PeftModel
+
+ if isinstance(model, PeftModel):
+ model = model.base_model.model
+
+ return model
+
+
+def maybe_raise_or_warn(
+ library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module
+):
+ """Simple helper method to raise or warn in case incorrect module has been passed"""
+ if not is_pipeline_module:
+ library = importlib.import_module(library_name)
+ class_obj = getattr(library, class_name)
+ class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()}
+
+ expected_class_obj = None
+ for class_name, class_candidate in class_candidates.items():
+ if class_candidate is not None and issubclass(class_obj, class_candidate):
+ expected_class_obj = class_candidate
+
+ # Dynamo wraps the original model in a private class.
+ # I didn't find a public API to get the original class.
+ sub_model = passed_class_obj[name]
+ unwrapped_sub_model = _unwrap_model(sub_model)
+ model_cls = unwrapped_sub_model.__class__
+
+ if not issubclass(model_cls, expected_class_obj):
+ raise ValueError(
+ f"{passed_class_obj[name]} is of type: {model_cls}, but should be" f" {expected_class_obj}"
+ )
+ else:
+ logger.warning(
+ f"You have passed a non-standard module {passed_class_obj[name]}. We cannot verify whether it"
+ " has the correct type"
+ )
+
+
+def get_class_obj_and_candidates(
+ library_name, class_name, importable_classes, pipelines, is_pipeline_module, component_name=None, cache_dir=None
+):
+ """Simple helper method to retrieve class object of module as well as potential parent class objects"""
+ component_folder = os.path.join(cache_dir, component_name)
+
+ if is_pipeline_module:
+ pipeline_module = getattr(pipelines, library_name)
+
+ class_obj = getattr(pipeline_module, class_name)
+ class_candidates = {c: class_obj for c in importable_classes.keys()}
+ elif os.path.isfile(os.path.join(component_folder, library_name + ".py")):
+ # load custom component
+ class_obj = get_class_from_dynamic_module(
+ component_folder, module_file=library_name + ".py", class_name=class_name
+ )
+ class_candidates = {c: class_obj for c in importable_classes.keys()}
+ else:
+ # else we just import it from the library.
+ library = importlib.import_module(library_name)
+
+ class_obj = getattr(library, class_name)
+ class_candidates = {c: getattr(library, c, None) for c in importable_classes.keys()}
+
+ return class_obj, class_candidates
+
+
+def _get_pipeline_class(
+ class_obj,
+ config=None,
+ load_connected_pipeline=False,
+ custom_pipeline=None,
+ repo_id=None,
+ hub_revision=None,
+ class_name=None,
+ cache_dir=None,
+ revision=None,
+):
+ if custom_pipeline is not None:
+ if custom_pipeline.endswith(".py"):
+ path = Path(custom_pipeline)
+ # decompose into folder & file
+ file_name = path.name
+ custom_pipeline = path.parent.absolute()
+ elif repo_id is not None:
+ file_name = f"{custom_pipeline}.py"
+ custom_pipeline = repo_id
+ else:
+ file_name = CUSTOM_PIPELINE_FILE_NAME
+
+ if repo_id is not None and hub_revision is not None:
+ # if we load the pipeline code from the Hub
+ # make sure to overwrite the `revision`
+ revision = hub_revision
+
+ return get_class_from_dynamic_module(
+ custom_pipeline,
+ module_file=file_name,
+ class_name=class_name,
+ cache_dir=cache_dir,
+ revision=revision,
+ )
+
+ if class_obj.__name__ != "DiffusionPipeline":
+ return class_obj
+
+ diffusers_module = importlib.import_module(class_obj.__module__.split(".")[0])
+ class_name = class_name or config["_class_name"]
+ if not class_name:
+ raise ValueError(
+ "The class name could not be found in the configuration file. Please make sure to pass the correct `class_name`."
+ )
+
+ class_name = class_name[4:] if class_name.startswith("Flax") else class_name
+
+ pipeline_cls = getattr(diffusers_module, class_name)
+
+ if load_connected_pipeline:
+ from .auto_pipeline import _get_connected_pipeline
+
+ connected_pipeline_cls = _get_connected_pipeline(pipeline_cls)
+ if connected_pipeline_cls is not None:
+ logger.info(
+ f"Loading connected pipeline {connected_pipeline_cls.__name__} instead of {pipeline_cls.__name__} as specified via `load_connected_pipeline=True`"
+ )
+ else:
+ logger.info(f"{pipeline_cls.__name__} has no connected pipeline class. Loading {pipeline_cls.__name__}.")
+
+ pipeline_cls = connected_pipeline_cls or pipeline_cls
+
+ return pipeline_cls
+
+
+def load_sub_model(
+ library_name: str,
+ class_name: str,
+ importable_classes: List[Any],
+ pipelines: Any,
+ is_pipeline_module: bool,
+ pipeline_class: Any,
+ torch_dtype: torch.dtype,
+ provider: Any,
+ sess_options: Any,
+ device_map: Optional[Union[Dict[str, torch.device], str]],
+ max_memory: Optional[Dict[Union[int, str], Union[int, str]]],
+ offload_folder: Optional[Union[str, os.PathLike]],
+ offload_state_dict: bool,
+ model_variants: Dict[str, str],
+ name: str,
+ from_flax: bool,
+ variant: str,
+ low_cpu_mem_usage: bool,
+ cached_folder: Union[str, os.PathLike],
+):
+ """Helper method to load the module `name` from `library_name` and `class_name`"""
+ # retrieve class candidates
+ class_obj, class_candidates = get_class_obj_and_candidates(
+ library_name,
+ class_name,
+ importable_classes,
+ pipelines,
+ is_pipeline_module,
+ component_name=name,
+ cache_dir=cached_folder,
+ )
+
+ load_method_name = None
+ # retrieve load method name
+ for class_name, class_candidate in class_candidates.items():
+ if class_candidate is not None and issubclass(class_obj, class_candidate):
+ load_method_name = importable_classes[class_name][1]
+
+ # if load method name is None, then we have a dummy module -> raise Error
+ if load_method_name is None:
+ none_module = class_obj.__module__
+ is_dummy_path = none_module.startswith(DUMMY_MODULES_FOLDER) or none_module.startswith(
+ TRANSFORMERS_DUMMY_MODULES_FOLDER
+ )
+ if is_dummy_path and "dummy" in none_module:
+ # call class_obj for nice error message of missing requirements
+ class_obj()
+
+ raise ValueError(
+ f"The component {class_obj} of {pipeline_class} cannot be loaded as it does not seem to have"
+ f" any of the loading methods defined in {ALL_IMPORTABLE_CLASSES}."
+ )
+
+ load_method = getattr(class_obj, load_method_name)
+
+ # add kwargs to loading method
+ diffusers_module = importlib.import_module(__name__.split(".")[0] + "." + __name__.split(".")[1])
+ loading_kwargs = {}
+ if issubclass(class_obj, torch.nn.Module):
+ loading_kwargs["torch_dtype"] = torch_dtype
+ if issubclass(class_obj, diffusers_module.OnnxRuntimeModel):
+ loading_kwargs["provider"] = provider
+ loading_kwargs["sess_options"] = sess_options
+
+ is_diffusers_model = issubclass(class_obj, diffusers_module.ModelMixin)
+
+ if is_transformers_available():
+ transformers_version = version.parse(version.parse(transformers.__version__).base_version)
+ else:
+ transformers_version = "N/A"
+
+ is_transformers_model = (
+ is_transformers_available()
+ and issubclass(class_obj, PreTrainedModel)
+ and transformers_version >= version.parse("4.20.0")
+ )
+
+ # When loading a transformers model, if the device_map is None, the weights will be initialized as opposed to diffusers.
+ # To make default loading faster we set the `low_cpu_mem_usage=low_cpu_mem_usage` flag which is `True` by default.
+ # This makes sure that the weights won't be initialized which significantly speeds up loading.
+ if is_diffusers_model or is_transformers_model:
+ loading_kwargs["device_map"] = device_map
+ loading_kwargs["max_memory"] = max_memory
+ loading_kwargs["offload_folder"] = offload_folder
+ loading_kwargs["offload_state_dict"] = offload_state_dict
+ loading_kwargs["variant"] = model_variants.pop(name, None)
+
+ if from_flax:
+ loading_kwargs["from_flax"] = True
+
+ # the following can be deleted once the minimum required `transformers` version
+ # is higher than 4.27
+ if (
+ is_transformers_model
+ and loading_kwargs["variant"] is not None
+ and transformers_version < version.parse("4.27.0")
+ ):
+ raise ImportError(
+ f"When passing `variant='{variant}'`, please make sure to upgrade your `transformers` version to at least 4.27.0.dev0"
+ )
+ elif is_transformers_model and loading_kwargs["variant"] is None:
+ loading_kwargs.pop("variant")
+
+ # if `from_flax` and model is transformer model, can currently not load with `low_cpu_mem_usage`
+ if not (from_flax and is_transformers_model):
+ loading_kwargs["low_cpu_mem_usage"] = low_cpu_mem_usage
+ else:
+ loading_kwargs["low_cpu_mem_usage"] = False
+
+ # check if the module is in a subdirectory
+ if os.path.isdir(os.path.join(cached_folder, name)):
+ loaded_sub_model = load_method(os.path.join(cached_folder, name), **loading_kwargs)
+ else:
+ # else load from the root directory
+ loaded_sub_model = load_method(cached_folder, **loading_kwargs)
+
+ return loaded_sub_model
+
+
+def _fetch_class_library_tuple(module):
+ # import it here to avoid circular import
+ diffusers_module = importlib.import_module(__name__.split(".")[0] + "." + __name__.split(".")[1])
+ pipelines = getattr(diffusers_module, "pipelines")
+
+ # register the config from the original module, not the dynamo compiled one
+ not_compiled_module = _unwrap_model(module)
+ library = not_compiled_module.__module__.split(".")[0]
+
+ # check if the module is a pipeline module
+ module_path_items = not_compiled_module.__module__.split(".")
+ pipeline_dir = module_path_items[-2] if len(module_path_items) > 2 else None
+
+ path = not_compiled_module.__module__.split(".")
+ is_pipeline_module = pipeline_dir in path and hasattr(pipelines, pipeline_dir)
+
+ # if library is not in LOADABLE_CLASSES, then it is a custom module.
+ # Or if it's a pipeline module, then the module is inside the pipeline
+ # folder so we set the library to module name.
+ if is_pipeline_module:
+ library = pipeline_dir
+ elif library not in LOADABLE_CLASSES:
+ library = not_compiled_module.__module__
+
+ # retrieve class_name
+ class_name = not_compiled_module.__class__.__name__
+
+ return (library, class_name)
diff --git a/model/diffusers_c/pipelines/pipeline_utils.py b/model/diffusers_c/pipelines/pipeline_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..ee7029198a0585fee4a1b112343afeae5cff60a2
--- /dev/null
+++ b/model/diffusers_c/pipelines/pipeline_utils.py
@@ -0,0 +1,1768 @@
+# Copyright 2024 The HuggingFace Inc. team.
+# Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import fnmatch
+import importlib
+import inspect
+import os
+import re
+import sys
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import requests
+import torch
+from huggingface_hub import (
+ ModelCard,
+ create_repo,
+ hf_hub_download,
+ model_info,
+ snapshot_download,
+)
+from huggingface_hub.utils import OfflineModeIsEnabled, validate_hf_hub_args
+from packaging import version
+from requests.exceptions import HTTPError
+from tqdm.auto import tqdm
+
+from .. import __version__
+from ..configuration_utils import ConfigMixin
+from ..models import AutoencoderKL
+from ..models.attention_processor import FusedAttnProcessor2_0
+from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT
+from ..schedulers.scheduling_utils import SCHEDULER_CONFIG_NAME
+from ..utils import (
+ CONFIG_NAME,
+ DEPRECATED_REVISION_ARGS,
+ BaseOutput,
+ PushToHubMixin,
+ deprecate,
+ is_accelerate_available,
+ is_accelerate_version,
+ is_torch_npu_available,
+ is_torch_version,
+ logging,
+ numpy_to_pil,
+)
+from ..utils.hub_utils import load_or_create_model_card, populate_model_card
+from ..utils.torch_utils import is_compiled_module
+
+
+if is_torch_npu_available():
+ import torch_npu # noqa: F401
+
+
+from .pipeline_loading_utils import (
+ ALL_IMPORTABLE_CLASSES,
+ CONNECTED_PIPES_KEYS,
+ CUSTOM_PIPELINE_FILE_NAME,
+ LOADABLE_CLASSES,
+ _fetch_class_library_tuple,
+ _get_pipeline_class,
+ _unwrap_model,
+ is_safetensors_compatible,
+ load_sub_model,
+ maybe_raise_or_warn,
+ variant_compatible_siblings,
+ warn_deprecated_model_variant,
+)
+
+
+if is_accelerate_available():
+ import accelerate
+
+
+LIBRARIES = []
+for library in LOADABLE_CLASSES:
+ LIBRARIES.append(library)
+
+logger = logging.get_logger(__name__)
+
+
+@dataclass
+class ImagePipelineOutput(BaseOutput):
+ """
+ Output class for image pipelines.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+@dataclass
+class AudioPipelineOutput(BaseOutput):
+ """
+ Output class for audio pipelines.
+
+ Args:
+ audios (`np.ndarray`)
+ List of denoised audio samples of a NumPy array of shape `(batch_size, num_channels, sample_rate)`.
+ """
+
+ audios: np.ndarray
+
+
+class DiffusionPipeline(ConfigMixin, PushToHubMixin):
+ r"""
+ Base class for all pipelines.
+
+ [`DiffusionPipeline`] stores all components (models, schedulers, and processors) for diffusion pipelines and
+ provides methods for loading, downloading and saving models. It also includes methods to:
+
+ - move all PyTorch modules to the device of your choice
+ - enable/disable the progress bar for the denoising iteration
+
+ Class attributes:
+
+ - **config_name** (`str`) -- The configuration filename that stores the class and module names of all the
+ diffusion pipeline's components.
+ - **_optional_components** (`List[str]`) -- List of all optional components that don't have to be passed to the
+ pipeline to function (should be overridden by subclasses).
+ """
+
+ config_name = "model_index.json"
+ model_cpu_offload_seq = None
+ _optional_components = []
+ _exclude_from_cpu_offload = []
+ _load_connected_pipes = False
+ _is_onnx = False
+
+ def register_modules(self, **kwargs):
+ for name, module in kwargs.items():
+ # retrieve library
+ if module is None or isinstance(module, (tuple, list)) and module[0] is None:
+ register_dict = {name: (None, None)}
+ else:
+ library, class_name = _fetch_class_library_tuple(module)
+ register_dict = {name: (library, class_name)}
+
+ # save model index config
+ self.register_to_config(**register_dict)
+
+ # set models
+ setattr(self, name, module)
+
+ def __setattr__(self, name: str, value: Any):
+ if name in self.__dict__ and hasattr(self.config, name):
+ # We need to overwrite the config if name exists in config
+ if isinstance(getattr(self.config, name), (tuple, list)):
+ if value is not None and self.config[name][0] is not None:
+ class_library_tuple = _fetch_class_library_tuple(value)
+ else:
+ class_library_tuple = (None, None)
+
+ self.register_to_config(**{name: class_library_tuple})
+ else:
+ self.register_to_config(**{name: value})
+
+ super().__setattr__(name, value)
+
+ def save_pretrained(
+ self,
+ save_directory: Union[str, os.PathLike],
+ safe_serialization: bool = True,
+ variant: Optional[str] = None,
+ push_to_hub: bool = False,
+ **kwargs,
+ ):
+ """
+ Save all saveable variables of the pipeline to a directory. A pipeline variable can be saved and loaded if its
+ class implements both a save and loading method. The pipeline is easily reloaded using the
+ [`~DiffusionPipeline.from_pretrained`] class method.
+
+ Arguments:
+ save_directory (`str` or `os.PathLike`):
+ Directory to save a pipeline to. Will be created if it doesn't exist.
+ safe_serialization (`bool`, *optional*, defaults to `True`):
+ Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+ variant (`str`, *optional*):
+ If specified, weights are saved in the format `pytorch_model..bin`.
+ push_to_hub (`bool`, *optional*, defaults to `False`):
+ Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the
+ repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+ namespace).
+ kwargs (`Dict[str, Any]`, *optional*):
+ Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+ """
+ model_index_dict = dict(self.config)
+ model_index_dict.pop("_class_name", None)
+ model_index_dict.pop("_diffusers_version", None)
+ model_index_dict.pop("_module", None)
+ model_index_dict.pop("_name_or_path", None)
+
+ if push_to_hub:
+ commit_message = kwargs.pop("commit_message", None)
+ private = kwargs.pop("private", False)
+ create_pr = kwargs.pop("create_pr", False)
+ token = kwargs.pop("token", None)
+ repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
+ repo_id = create_repo(repo_id, exist_ok=True, private=private, token=token).repo_id
+
+ expected_modules, optional_kwargs = self._get_signature_keys(self)
+
+ def is_saveable_module(name, value):
+ if name not in expected_modules:
+ return False
+ if name in self._optional_components and value[0] is None:
+ return False
+ return True
+
+ model_index_dict = {k: v for k, v in model_index_dict.items() if is_saveable_module(k, v)}
+ for pipeline_component_name in model_index_dict.keys():
+ sub_model = getattr(self, pipeline_component_name)
+ model_cls = sub_model.__class__
+
+ # Dynamo wraps the original model in a private class.
+ # I didn't find a public API to get the original class.
+ if is_compiled_module(sub_model):
+ sub_model = _unwrap_model(sub_model)
+ model_cls = sub_model.__class__
+
+ save_method_name = None
+ # search for the model's base class in LOADABLE_CLASSES
+ for library_name, library_classes in LOADABLE_CLASSES.items():
+ if library_name in sys.modules:
+ library = importlib.import_module(library_name)
+ else:
+ logger.info(
+ f"{library_name} is not installed. Cannot save {pipeline_component_name} as {library_classes} from {library_name}"
+ )
+
+ for base_class, save_load_methods in library_classes.items():
+ class_candidate = getattr(library, base_class, None)
+ if class_candidate is not None and issubclass(model_cls, class_candidate):
+ # if we found a suitable base class in LOADABLE_CLASSES then grab its save method
+ save_method_name = save_load_methods[0]
+ break
+ if save_method_name is not None:
+ break
+
+ if save_method_name is None:
+ logger.warn(f"self.{pipeline_component_name}={sub_model} of type {type(sub_model)} cannot be saved.")
+ # make sure that unsaveable components are not tried to be loaded afterward
+ self.register_to_config(**{pipeline_component_name: (None, None)})
+ continue
+
+ save_method = getattr(sub_model, save_method_name)
+
+ # Call the save method with the argument safe_serialization only if it's supported
+ save_method_signature = inspect.signature(save_method)
+ save_method_accept_safe = "safe_serialization" in save_method_signature.parameters
+ save_method_accept_variant = "variant" in save_method_signature.parameters
+
+ save_kwargs = {}
+ if save_method_accept_safe:
+ save_kwargs["safe_serialization"] = safe_serialization
+ if save_method_accept_variant:
+ save_kwargs["variant"] = variant
+
+ save_method(os.path.join(save_directory, pipeline_component_name), **save_kwargs)
+
+ # finally save the config
+ self.save_config(save_directory)
+
+ if push_to_hub:
+ # Create a new empty model card and eventually tag it
+ model_card = load_or_create_model_card(repo_id, token=token, is_pipeline=True)
+ model_card = populate_model_card(model_card)
+ model_card.save(os.path.join(save_directory, "README.md"))
+
+ self._upload_folder(
+ save_directory,
+ repo_id,
+ token=token,
+ commit_message=commit_message,
+ create_pr=create_pr,
+ )
+
+ def to(self, *args, **kwargs):
+ r"""
+ Performs Pipeline dtype and/or device conversion. A torch.dtype and torch.device are inferred from the
+ arguments of `self.to(*args, **kwargs).`
+
+
+
+ If the pipeline already has the correct torch.dtype and torch.device, then it is returned as is. Otherwise,
+ the returned pipeline is a copy of self with the desired torch.dtype and torch.device.
+
+
+
+
+ Here are the ways to call `to`:
+
+ - `to(dtype, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified
+ [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype)
+ - `to(device, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the specified
+ [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device)
+ - `to(device=None, dtype=None, silence_dtype_warnings=False) → DiffusionPipeline` to return a pipeline with the
+ specified [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device) and
+ [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype)
+
+ Arguments:
+ dtype (`torch.dtype`, *optional*):
+ Returns a pipeline with the specified
+ [`dtype`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.dtype)
+ device (`torch.Device`, *optional*):
+ Returns a pipeline with the specified
+ [`device`](https://pytorch.org/docs/stable/tensor_attributes.html#torch.device)
+ silence_dtype_warnings (`str`, *optional*, defaults to `False`):
+ Whether to omit warnings if the target `dtype` is not compatible with the target `device`.
+
+ Returns:
+ [`DiffusionPipeline`]: The pipeline converted to specified `dtype` and/or `dtype`.
+ """
+ dtype = kwargs.pop("dtype", None)
+ device = kwargs.pop("device", None)
+ silence_dtype_warnings = kwargs.pop("silence_dtype_warnings", False)
+
+ dtype_arg = None
+ device_arg = None
+ if len(args) == 1:
+ if isinstance(args[0], torch.dtype):
+ dtype_arg = args[0]
+ else:
+ device_arg = torch.device(args[0]) if args[0] is not None else None
+ elif len(args) == 2:
+ if isinstance(args[0], torch.dtype):
+ raise ValueError(
+ "When passing two arguments, make sure the first corresponds to `device` and the second to `dtype`."
+ )
+ device_arg = torch.device(args[0]) if args[0] is not None else None
+ dtype_arg = args[1]
+ elif len(args) > 2:
+ raise ValueError("Please make sure to pass at most two arguments (`device` and `dtype`) `.to(...)`")
+
+ if dtype is not None and dtype_arg is not None:
+ raise ValueError(
+ "You have passed `dtype` both as an argument and as a keyword argument. Please only pass one of the two."
+ )
+
+ dtype = dtype or dtype_arg
+
+ if device is not None and device_arg is not None:
+ raise ValueError(
+ "You have passed `device` both as an argument and as a keyword argument. Please only pass one of the two."
+ )
+
+ device = device or device_arg
+
+ # throw warning if pipeline is in "offloaded"-mode but user tries to manually set to GPU.
+ def module_is_sequentially_offloaded(module):
+ if not is_accelerate_available() or is_accelerate_version("<", "0.14.0"):
+ return False
+
+ return hasattr(module, "_hf_hook") and not isinstance(
+ module._hf_hook, (accelerate.hooks.CpuOffload, accelerate.hooks.AlignDevicesHook)
+ )
+
+ def module_is_offloaded(module):
+ if not is_accelerate_available() or is_accelerate_version("<", "0.17.0.dev0"):
+ return False
+
+ return hasattr(module, "_hf_hook") and isinstance(module._hf_hook, accelerate.hooks.CpuOffload)
+
+ # .to("cuda") would raise an error if the pipeline is sequentially offloaded, so we raise our own to make it clearer
+ pipeline_is_sequentially_offloaded = any(
+ module_is_sequentially_offloaded(module) for _, module in self.components.items()
+ )
+ if pipeline_is_sequentially_offloaded and device and torch.device(device).type == "cuda":
+ raise ValueError(
+ "It seems like you have activated sequential model offloading by calling `enable_sequential_cpu_offload`, but are now attempting to move the pipeline to GPU. This is not compatible with offloading. Please, move your pipeline `.to('cpu')` or consider removing the move altogether if you use sequential offloading."
+ )
+
+ # Display a warning in this case (the operation succeeds but the benefits are lost)
+ pipeline_is_offloaded = any(module_is_offloaded(module) for _, module in self.components.items())
+ if pipeline_is_offloaded and device and torch.device(device).type == "cuda":
+ logger.warning(
+ f"It seems like you have activated model offloading by calling `enable_model_cpu_offload`, but are now manually moving the pipeline to GPU. It is strongly recommended against doing so as memory gains from offloading are likely to be lost. Offloading automatically takes care of moving the individual components {', '.join(self.components.keys())} to GPU when needed. To make sure offloading works as expected, you should consider moving the pipeline back to CPU: `pipeline.to('cpu')` or removing the move altogether if you use offloading."
+ )
+
+ module_names, _ = self._get_signature_keys(self)
+ modules = [getattr(self, n, None) for n in module_names]
+ modules = [m for m in modules if isinstance(m, torch.nn.Module)]
+
+ is_offloaded = pipeline_is_offloaded or pipeline_is_sequentially_offloaded
+ for module in modules:
+ is_loaded_in_8bit = hasattr(module, "is_loaded_in_8bit") and module.is_loaded_in_8bit
+
+ if is_loaded_in_8bit and dtype is not None:
+ logger.warning(
+ f"The module '{module.__class__.__name__}' has been loaded in 8bit and conversion to {dtype} is not yet supported. Module is still in 8bit precision."
+ )
+
+ if is_loaded_in_8bit and device is not None:
+ logger.warning(
+ f"The module '{module.__class__.__name__}' has been loaded in 8bit and moving it to {dtype} via `.to()` is not yet supported. Module is still on {module.device}."
+ )
+ else:
+ module.to(device, dtype)
+
+ if (
+ module.dtype == torch.float16
+ and str(device) in ["cpu"]
+ and not silence_dtype_warnings
+ and not is_offloaded
+ ):
+ logger.warning(
+ "Pipelines loaded with `dtype=torch.float16` cannot run with `cpu` device. It"
+ " is not recommended to move them to `cpu` as running them will fail. Please make"
+ " sure to use an accelerator to run the pipeline in inference, due to the lack of"
+ " support for`float16` operations on this device in PyTorch. Please, remove the"
+ " `torch_dtype=torch.float16` argument, or use another device for inference."
+ )
+ return self
+
+ @property
+ def device(self) -> torch.device:
+ r"""
+ Returns:
+ `torch.device`: The torch device on which the pipeline is located.
+ """
+ module_names, _ = self._get_signature_keys(self)
+ modules = [getattr(self, n, None) for n in module_names]
+ modules = [m for m in modules if isinstance(m, torch.nn.Module)]
+
+ for module in modules:
+ return module.device
+
+ return torch.device("cpu")
+
+ @property
+ def dtype(self) -> torch.dtype:
+ r"""
+ Returns:
+ `torch.dtype`: The torch dtype on which the pipeline is located.
+ """
+ module_names, _ = self._get_signature_keys(self)
+ modules = [getattr(self, n, None) for n in module_names]
+ modules = [m for m in modules if isinstance(m, torch.nn.Module)]
+
+ for module in modules:
+ return module.dtype
+
+ return torch.float32
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]], **kwargs):
+ r"""
+ Instantiate a PyTorch diffusion pipeline from pretrained pipeline weights.
+
+ The pipeline is set in evaluation mode (`model.eval()`) by default.
+
+ If you get the error message below, you need to finetune the weights for your downstream task:
+
+ ```
+ Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+ - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
+ You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+ ```
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *repo id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
+ hosted on the Hub.
+ - A path to a *directory* (for example `./my_pipeline_directory/`) containing pipeline weights
+ saved using
+ [`~DiffusionPipeline.save_pretrained`].
+ torch_dtype (`str` or `torch.dtype`, *optional*):
+ Override the default `torch.dtype` and load the model with another dtype. If "auto" is passed, the
+ dtype is automatically derived from the model's weights.
+ custom_pipeline (`str`, *optional*):
+
+
+
+ 🧪 This is an experimental feature and may change in the future.
+
+
+
+ Can be either:
+
+ - A string, the *repo id* (for example `hf-internal-testing/diffusers-dummy-pipeline`) of a custom
+ pipeline hosted on the Hub. The repository must contain a file called pipeline.py that defines
+ the custom pipeline.
+ - A string, the *file name* of a community pipeline hosted on GitHub under
+ [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file
+ names must match the file name and not the pipeline script (`clip_guided_stable_diffusion`
+ instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the
+ current main branch of GitHub.
+ - A path to a directory (`./my_pipeline_directory/`) containing a custom pipeline. The directory
+ must contain a file called `pipeline.py` that defines the custom pipeline.
+
+ For more information on how to load and create custom pipelines, please have a look at [Loading and
+ Adding Custom
+ Pipelines](https://huggingface.co/docs/diffusers/using-diffusers/custom_pipeline_overview)
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ custom_revision (`str`, *optional*):
+ The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
+ `revision` when loading a custom pipeline from the Hub. Defaults to the latest stable 🤗 Diffusers version.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you’re downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+ device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+ A map that specifies where each submodule should go. It doesn’t need to be defined for each
+ parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
+ same device.
+
+ Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
+ more information about each option see [designing a device
+ map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+ max_memory (`Dict`, *optional*):
+ A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
+ each GPU and the available CPU RAM if unset.
+ offload_folder (`str` or `os.PathLike`, *optional*):
+ The path to offload weights if device_map contains the value `"disk"`.
+ offload_state_dict (`bool`, *optional*):
+ If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
+ the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
+ when there is some disk offload.
+ low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+ Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+ tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+ Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+ argument to `True` will raise an error.
+ use_safetensors (`bool`, *optional*, defaults to `None`):
+ If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+ safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+ weights. If set to `False`, safetensors weights are not loaded.
+ use_onnx (`bool`, *optional*, defaults to `None`):
+ If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights
+ will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is
+ `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending
+ with `.onnx` and `.pb`.
+ kwargs (remaining dictionary of keyword arguments, *optional*):
+ Can be used to overwrite load and saveable variables (the pipeline components of the specific pipeline
+ class). The overwritten components are passed directly to the pipelines `__init__` method. See example
+ below for more information.
+ variant (`str`, *optional*):
+ Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
+ loading `from_flax`.
+
+
+
+ To use private or [gated](https://huggingface.co/docs/hub/models-gated#gated-models) models, log-in with
+ `huggingface-cli login`.
+
+
+
+ Examples:
+
+ ```py
+ >>> from diffusers import DiffusionPipeline
+
+ >>> # Download pipeline from huggingface.co and cache.
+ >>> pipeline = DiffusionPipeline.from_pretrained("CompVis/ldm-text2im-large-256")
+
+ >>> # Download pipeline that requires an authorization token
+ >>> # For more information on access tokens, please refer to this section
+ >>> # of the documentation](https://huggingface.co/docs/hub/security-tokens)
+ >>> pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
+
+ >>> # Use a different scheduler
+ >>> from diffusers import LMSDiscreteScheduler
+
+ >>> scheduler = LMSDiscreteScheduler.from_config(pipeline.scheduler.config)
+ >>> pipeline.scheduler = scheduler
+ ```
+ """
+ cache_dir = kwargs.pop("cache_dir", None)
+ resume_download = kwargs.pop("resume_download", False)
+ force_download = kwargs.pop("force_download", False)
+ proxies = kwargs.pop("proxies", None)
+ local_files_only = kwargs.pop("local_files_only", None)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ from_flax = kwargs.pop("from_flax", False)
+ torch_dtype = kwargs.pop("torch_dtype", None)
+ custom_pipeline = kwargs.pop("custom_pipeline", None)
+ custom_revision = kwargs.pop("custom_revision", None)
+ provider = kwargs.pop("provider", None)
+ sess_options = kwargs.pop("sess_options", None)
+ device_map = kwargs.pop("device_map", None)
+ max_memory = kwargs.pop("max_memory", None)
+ offload_folder = kwargs.pop("offload_folder", None)
+ offload_state_dict = kwargs.pop("offload_state_dict", False)
+ low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT)
+ variant = kwargs.pop("variant", None)
+ use_safetensors = kwargs.pop("use_safetensors", None)
+ use_onnx = kwargs.pop("use_onnx", None)
+ load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
+
+ if low_cpu_mem_usage and not is_accelerate_available():
+ low_cpu_mem_usage = False
+ logger.warning(
+ "Cannot initialize model with low cpu memory usage because `accelerate` was not found in the"
+ " environment. Defaulting to `low_cpu_mem_usage=False`. It is strongly recommended to install"
+ " `accelerate` for faster and less memory-intense model loading. You can do so with: \n```\npip"
+ " install accelerate\n```\n."
+ )
+
+ if device_map is not None and not is_torch_version(">=", "1.9.0"):
+ raise NotImplementedError(
+ "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
+ " `device_map=None`."
+ )
+
+ if low_cpu_mem_usage is True and not is_torch_version(">=", "1.9.0"):
+ raise NotImplementedError(
+ "Low memory initialization requires torch >= 1.9.0. Please either update your PyTorch version or set"
+ " `low_cpu_mem_usage=False`."
+ )
+
+ if low_cpu_mem_usage is False and device_map is not None:
+ raise ValueError(
+ f"You cannot set `low_cpu_mem_usage` to False while using device_map={device_map} for loading and"
+ " dispatching. Please make sure to set `low_cpu_mem_usage=True`."
+ )
+
+ # 1. Download the checkpoints and configs
+ # use snapshot download here to get it working from from_pretrained
+ if not os.path.isdir(pretrained_model_name_or_path):
+ if pretrained_model_name_or_path.count("/") > 1:
+ raise ValueError(
+ f'The provided pretrained_model_name_or_path "{pretrained_model_name_or_path}"'
+ " is neither a valid local path nor a valid repo id. Please check the parameter."
+ )
+ cached_folder = cls.download(
+ pretrained_model_name_or_path,
+ cache_dir=cache_dir,
+ resume_download=resume_download,
+ force_download=force_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ from_flax=from_flax,
+ use_safetensors=use_safetensors,
+ use_onnx=use_onnx,
+ custom_pipeline=custom_pipeline,
+ custom_revision=custom_revision,
+ variant=variant,
+ load_connected_pipeline=load_connected_pipeline,
+ **kwargs,
+ )
+ else:
+ cached_folder = pretrained_model_name_or_path
+
+ config_dict = cls.load_config(cached_folder)
+
+ # pop out "_ignore_files" as it is only needed for download
+ config_dict.pop("_ignore_files", None)
+
+ # 2. Define which model components should load variants
+ # We retrieve the information by matching whether variant
+ # model checkpoints exist in the subfolders
+ model_variants = {}
+ if variant is not None:
+ for folder in os.listdir(cached_folder):
+ folder_path = os.path.join(cached_folder, folder)
+ is_folder = os.path.isdir(folder_path) and folder in config_dict
+ variant_exists = is_folder and any(
+ p.split(".")[1].startswith(variant) for p in os.listdir(folder_path)
+ )
+ if variant_exists:
+ model_variants[folder] = variant
+
+ # 3. Load the pipeline class, if using custom module then load it from the hub
+ # if we load from explicit class, let's use it
+ custom_class_name = None
+ if os.path.isfile(os.path.join(cached_folder, f"{custom_pipeline}.py")):
+ custom_pipeline = os.path.join(cached_folder, f"{custom_pipeline}.py")
+ elif isinstance(config_dict["_class_name"], (list, tuple)) and os.path.isfile(
+ os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py")
+ ):
+ custom_pipeline = os.path.join(cached_folder, f"{config_dict['_class_name'][0]}.py")
+ custom_class_name = config_dict["_class_name"][1]
+
+ pipeline_class = _get_pipeline_class(
+ cls,
+ config_dict,
+ load_connected_pipeline=load_connected_pipeline,
+ custom_pipeline=custom_pipeline,
+ class_name=custom_class_name,
+ cache_dir=cache_dir,
+ revision=custom_revision,
+ )
+
+ # DEPRECATED: To be removed in 1.0.0
+ if pipeline_class.__name__ == "StableDiffusionInpaintPipeline" and version.parse(
+ version.parse(config_dict["_diffusers_version"]).base_version
+ ) <= version.parse("0.5.1"):
+ from diffusers import StableDiffusionInpaintPipeline, StableDiffusionInpaintPipelineLegacy
+
+ pipeline_class = StableDiffusionInpaintPipelineLegacy
+
+ deprecation_message = (
+ "You are using a legacy checkpoint for inpainting with Stable Diffusion, therefore we are loading the"
+ f" {StableDiffusionInpaintPipelineLegacy} class instead of {StableDiffusionInpaintPipeline}. For"
+ " better inpainting results, we strongly suggest using Stable Diffusion's official inpainting"
+ " checkpoint: https://huggingface.co/runwayml/stable-diffusion-inpainting instead or adapting your"
+ f" checkpoint {pretrained_model_name_or_path} to the format of"
+ " https://huggingface.co/runwayml/stable-diffusion-inpainting. Note that we do not actively maintain"
+ " the {StableDiffusionInpaintPipelineLegacy} class and will likely remove it in version 1.0.0."
+ )
+ deprecate("StableDiffusionInpaintPipelineLegacy", "1.0.0", deprecation_message, standard_warn=False)
+
+ # 4. Define expected modules given pipeline signature
+ # and define non-None initialized modules (=`init_kwargs`)
+
+ # some modules can be passed directly to the init
+ # in this case they are already instantiated in `kwargs`
+ # extract them here
+ expected_modules, optional_kwargs = cls._get_signature_keys(pipeline_class)
+ passed_class_obj = {k: kwargs.pop(k) for k in expected_modules if k in kwargs}
+ passed_pipe_kwargs = {k: kwargs.pop(k) for k in optional_kwargs if k in kwargs}
+
+ init_dict, unused_kwargs, _ = pipeline_class.extract_init_dict(config_dict, **kwargs)
+
+ # define init kwargs and make sure that optional component modules are filtered out
+ init_kwargs = {
+ k: init_dict.pop(k)
+ for k in optional_kwargs
+ if k in init_dict and k not in pipeline_class._optional_components
+ }
+ init_kwargs = {**init_kwargs, **passed_pipe_kwargs}
+
+ # remove `null` components
+ def load_module(name, value):
+ if value[0] is None:
+ return False
+ if name in passed_class_obj and passed_class_obj[name] is None:
+ return False
+ return True
+
+ init_dict = {k: v for k, v in init_dict.items() if load_module(k, v)}
+
+ # Special case: safety_checker must be loaded separately when using `from_flax`
+ if from_flax and "safety_checker" in init_dict and "safety_checker" not in passed_class_obj:
+ raise NotImplementedError(
+ "The safety checker cannot be automatically loaded when loading weights `from_flax`."
+ " Please, pass `safety_checker=None` to `from_pretrained`, and load the safety checker"
+ " separately if you need it."
+ )
+
+ # 5. Throw nice warnings / errors for fast accelerate loading
+ if len(unused_kwargs) > 0:
+ logger.warning(
+ f"Keyword arguments {unused_kwargs} are not expected by {pipeline_class.__name__} and will be ignored."
+ )
+
+ # import it here to avoid circular import
+ from diffusers import pipelines
+
+ # 6. Load each module in the pipeline
+ for name, (library_name, class_name) in logging.tqdm(init_dict.items(), desc="Loading pipeline components..."):
+ # 6.1 - now that JAX/Flax is an official framework of the library, we might load from Flax names
+ class_name = class_name[4:] if class_name.startswith("Flax") else class_name
+
+ # 6.2 Define all importable classes
+ is_pipeline_module = hasattr(pipelines, library_name)
+ importable_classes = ALL_IMPORTABLE_CLASSES
+ loaded_sub_model = None
+
+ # 6.3 Use passed sub model or load class_name from library_name
+ if name in passed_class_obj:
+ # if the model is in a pipeline module, then we load it from the pipeline
+ # check that passed_class_obj has correct parent class
+ maybe_raise_or_warn(
+ library_name, library, class_name, importable_classes, passed_class_obj, name, is_pipeline_module
+ )
+
+ loaded_sub_model = passed_class_obj[name]
+ else:
+ # load sub model
+ loaded_sub_model = load_sub_model(
+ library_name=library_name,
+ class_name=class_name,
+ importable_classes=importable_classes,
+ pipelines=pipelines,
+ is_pipeline_module=is_pipeline_module,
+ pipeline_class=pipeline_class,
+ torch_dtype=torch_dtype,
+ provider=provider,
+ sess_options=sess_options,
+ device_map=device_map,
+ max_memory=max_memory,
+ offload_folder=offload_folder,
+ offload_state_dict=offload_state_dict,
+ model_variants=model_variants,
+ name=name,
+ from_flax=from_flax,
+ variant=variant,
+ low_cpu_mem_usage=low_cpu_mem_usage,
+ cached_folder=cached_folder,
+ )
+ logger.info(
+ f"Loaded {name} as {class_name} from `{name}` subfolder of {pretrained_model_name_or_path}."
+ )
+
+ init_kwargs[name] = loaded_sub_model # UNet(...), # DiffusionSchedule(...)
+
+ if pipeline_class._load_connected_pipes and os.path.isfile(os.path.join(cached_folder, "README.md")):
+ modelcard = ModelCard.load(os.path.join(cached_folder, "README.md"))
+ connected_pipes = {prefix: getattr(modelcard.data, prefix, [None])[0] for prefix in CONNECTED_PIPES_KEYS}
+ load_kwargs = {
+ "cache_dir": cache_dir,
+ "resume_download": resume_download,
+ "force_download": force_download,
+ "proxies": proxies,
+ "local_files_only": local_files_only,
+ "token": token,
+ "revision": revision,
+ "torch_dtype": torch_dtype,
+ "custom_pipeline": custom_pipeline,
+ "custom_revision": custom_revision,
+ "provider": provider,
+ "sess_options": sess_options,
+ "device_map": device_map,
+ "max_memory": max_memory,
+ "offload_folder": offload_folder,
+ "offload_state_dict": offload_state_dict,
+ "low_cpu_mem_usage": low_cpu_mem_usage,
+ "variant": variant,
+ "use_safetensors": use_safetensors,
+ }
+
+ def get_connected_passed_kwargs(prefix):
+ connected_passed_class_obj = {
+ k.replace(f"{prefix}_", ""): w for k, w in passed_class_obj.items() if k.split("_")[0] == prefix
+ }
+ connected_passed_pipe_kwargs = {
+ k.replace(f"{prefix}_", ""): w for k, w in passed_pipe_kwargs.items() if k.split("_")[0] == prefix
+ }
+
+ connected_passed_kwargs = {**connected_passed_class_obj, **connected_passed_pipe_kwargs}
+ return connected_passed_kwargs
+
+ connected_pipes = {
+ prefix: DiffusionPipeline.from_pretrained(
+ repo_id, **load_kwargs.copy(), **get_connected_passed_kwargs(prefix)
+ )
+ for prefix, repo_id in connected_pipes.items()
+ if repo_id is not None
+ }
+
+ for prefix, connected_pipe in connected_pipes.items():
+ # add connected pipes to `init_kwargs` with _, e.g. "prior_text_encoder"
+ init_kwargs.update(
+ {"_".join([prefix, name]): component for name, component in connected_pipe.components.items()}
+ )
+
+ # 7. Potentially add passed objects if expected
+ missing_modules = set(expected_modules) - set(init_kwargs.keys())
+ passed_modules = list(passed_class_obj.keys())
+ optional_modules = pipeline_class._optional_components
+ if len(missing_modules) > 0 and missing_modules <= set(passed_modules + optional_modules):
+ for module in missing_modules:
+ init_kwargs[module] = passed_class_obj.get(module, None)
+ elif len(missing_modules) > 0:
+ passed_modules = set(list(init_kwargs.keys()) + list(passed_class_obj.keys())) - optional_kwargs
+ raise ValueError(
+ f"Pipeline {pipeline_class} expected {expected_modules}, but only {passed_modules} were passed."
+ )
+
+ # 8. Instantiate the pipeline
+ model = pipeline_class(**init_kwargs)
+
+ # 9. Save where the model was instantiated from
+ model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+ return model
+
+ @property
+ def name_or_path(self) -> str:
+ return getattr(self.config, "_name_or_path", None)
+
+ @property
+ def _execution_device(self):
+ r"""
+ Returns the device on which the pipeline's models will be executed. After calling
+ [`~DiffusionPipeline.enable_sequential_cpu_offload`] the execution device can only be inferred from
+ Accelerate's module hooks.
+ """
+ for name, model in self.components.items():
+ if not isinstance(model, torch.nn.Module) or name in self._exclude_from_cpu_offload:
+ continue
+
+ if not hasattr(model, "_hf_hook"):
+ return self.device
+ for module in model.modules():
+ if (
+ hasattr(module, "_hf_hook")
+ and hasattr(module._hf_hook, "execution_device")
+ and module._hf_hook.execution_device is not None
+ ):
+ return torch.device(module._hf_hook.execution_device)
+ return self.device
+
+ def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+ r"""
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+
+ Arguments:
+ gpu_id (`int`, *optional*):
+ The ID of the accelerator that shall be used in inference. If not specified, it will default to 0.
+ device (`torch.Device` or `str`, *optional*, defaults to "cuda"):
+ The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will
+ default to "cuda".
+ """
+ if self.model_cpu_offload_seq is None:
+ raise ValueError(
+ "Model CPU offload cannot be enabled because no `model_cpu_offload_seq` class attribute is set."
+ )
+
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+ from accelerate import cpu_offload_with_hook
+ else:
+ raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+ torch_device = torch.device(device)
+ device_index = torch_device.index
+
+ if gpu_id is not None and device_index is not None:
+ raise ValueError(
+ f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}"
+ f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}"
+ )
+
+ # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0
+ self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0)
+
+ device_type = torch_device.type
+ device = torch.device(f"{device_type}:{self._offload_gpu_id}")
+ self._offload_device = device
+
+ if self.device.type != "cpu":
+ self.to("cpu", silence_dtype_warnings=True)
+ device_mod = getattr(torch, self.device.type, None)
+ if hasattr(device_mod, "empty_cache") and device_mod.is_available():
+ device_mod.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
+
+ all_model_components = {k: v for k, v in self.components.items() if isinstance(v, torch.nn.Module)}
+
+ self._all_hooks = []
+ hook = None
+ for model_str in self.model_cpu_offload_seq.split("->"):
+ model = all_model_components.pop(model_str, None)
+ if not isinstance(model, torch.nn.Module):
+ continue
+
+ _, hook = cpu_offload_with_hook(model, device, prev_module_hook=hook)
+ self._all_hooks.append(hook)
+
+ # CPU offload models that are not in the seq chain unless they are explicitly excluded
+ # these models will stay on CPU until maybe_free_model_hooks is called
+ # some models cannot be in the seq chain because they are iteratively called, such as controlnet
+ for name, model in all_model_components.items():
+ if not isinstance(model, torch.nn.Module):
+ continue
+
+ if name in self._exclude_from_cpu_offload:
+ model.to(device)
+ else:
+ _, hook = cpu_offload_with_hook(model, device)
+ self._all_hooks.append(hook)
+
+ def maybe_free_model_hooks(self):
+ r"""
+ Function that offloads all components, removes all model hooks that were added when using
+ `enable_model_cpu_offload` and then applies them again. In case the model has not been offloaded this function
+ is a no-op. Make sure to add this function to the end of the `__call__` function of your pipeline so that it
+ functions correctly when applying enable_model_cpu_offload.
+ """
+ if not hasattr(self, "_all_hooks") or len(self._all_hooks) == 0:
+ # `enable_model_cpu_offload` has not be called, so silently do nothing
+ return
+
+ for hook in self._all_hooks:
+ # offload model and remove hook from model
+ hook.offload()
+ hook.remove()
+
+ # make sure the model is in the same state as before calling it
+ self.enable_model_cpu_offload(device=getattr(self, "_offload_device", "cuda"))
+
+ def enable_sequential_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
+ r"""
+ Offloads all models to CPU using 🤗 Accelerate, significantly reducing memory usage. When called, the state
+ dicts of all `torch.nn.Module` components (except those in `self._exclude_from_cpu_offload`) are saved to CPU
+ and then moved to `torch.device('meta')` and loaded to GPU only when their specific submodule has its `forward`
+ method called. Offloading happens on a submodule basis. Memory savings are higher than with
+ `enable_model_cpu_offload`, but performance is lower.
+
+ Arguments:
+ gpu_id (`int`, *optional*):
+ The ID of the accelerator that shall be used in inference. If not specified, it will default to 0.
+ device (`torch.Device` or `str`, *optional*, defaults to "cuda"):
+ The PyTorch device type of the accelerator that shall be used in inference. If not specified, it will
+ default to "cuda".
+ """
+ if is_accelerate_available() and is_accelerate_version(">=", "0.14.0"):
+ from accelerate import cpu_offload
+ else:
+ raise ImportError("`enable_sequential_cpu_offload` requires `accelerate v0.14.0` or higher")
+
+ torch_device = torch.device(device)
+ device_index = torch_device.index
+
+ if gpu_id is not None and device_index is not None:
+ raise ValueError(
+ f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}"
+ f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}"
+ )
+
+ # _offload_gpu_id should be set to passed gpu_id (or id in passed `device`) or default to previously set id or default to 0
+ self._offload_gpu_id = gpu_id or torch_device.index or getattr(self, "_offload_gpu_id", 0)
+
+ device_type = torch_device.type
+ device = torch.device(f"{device_type}:{self._offload_gpu_id}")
+ self._offload_device = device
+
+ if self.device.type != "cpu":
+ self.to("cpu", silence_dtype_warnings=True)
+ device_mod = getattr(torch, self.device.type, None)
+ if hasattr(device_mod, "empty_cache") and device_mod.is_available():
+ device_mod.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
+
+ for name, model in self.components.items():
+ if not isinstance(model, torch.nn.Module):
+ continue
+
+ if name in self._exclude_from_cpu_offload:
+ model.to(device)
+ else:
+ # make sure to offload buffers if not all high level weights
+ # are of type nn.Module
+ offload_buffers = len(model._parameters) > 0
+ cpu_offload(model, device, offload_buffers=offload_buffers)
+
+ @classmethod
+ @validate_hf_hub_args
+ def download(cls, pretrained_model_name, **kwargs) -> Union[str, os.PathLike]:
+ r"""
+ Download and cache a PyTorch diffusion pipeline from pretrained pipeline weights.
+
+ Parameters:
+ pretrained_model_name (`str` or `os.PathLike`, *optional*):
+ A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained pipeline
+ hosted on the Hub.
+ custom_pipeline (`str`, *optional*):
+ Can be either:
+
+ - A string, the *repository id* (for example `CompVis/ldm-text2im-large-256`) of a pretrained
+ pipeline hosted on the Hub. The repository must contain a file called `pipeline.py` that defines
+ the custom pipeline.
+
+ - A string, the *file name* of a community pipeline hosted on GitHub under
+ [Community](https://github.com/huggingface/diffusers/tree/main/examples/community). Valid file
+ names must match the file name and not the pipeline script (`clip_guided_stable_diffusion`
+ instead of `clip_guided_stable_diffusion.py`). Community pipelines are always loaded from the
+ current `main` branch of GitHub.
+
+ - A path to a *directory* (`./my_pipeline_directory/`) containing a custom pipeline. The directory
+ must contain a file called `pipeline.py` that defines the custom pipeline.
+
+
+
+ 🧪 This is an experimental feature and may change in the future.
+
+
+
+ For more information on how to load and create custom pipelines, take a look at [How to contribute a
+ community pipeline](https://huggingface.co/docs/diffusers/main/en/using-diffusers/contribute_pipeline).
+
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+ custom_revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, or a commit id similar to
+ `revision` when loading a custom pipeline from the Hub. It can be a 🤗 Diffusers version when loading a
+ custom pipeline from GitHub, otherwise it defaults to `"main"` when loading from the Hub.
+ mirror (`str`, *optional*):
+ Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+ guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+ information.
+ variant (`str`, *optional*):
+ Load weights from a specified variant filename such as `"fp16"` or `"ema"`. This is ignored when
+ loading `from_flax`.
+ use_safetensors (`bool`, *optional*, defaults to `None`):
+ If set to `None`, the safetensors weights are downloaded if they're available **and** if the
+ safetensors library is installed. If set to `True`, the model is forcibly loaded from safetensors
+ weights. If set to `False`, safetensors weights are not loaded.
+ use_onnx (`bool`, *optional*, defaults to `False`):
+ If set to `True`, ONNX weights will always be downloaded if present. If set to `False`, ONNX weights
+ will never be downloaded. By default `use_onnx` defaults to the `_is_onnx` class attribute which is
+ `False` for non-ONNX pipelines and `True` for ONNX pipelines. ONNX weights include both files ending
+ with `.onnx` and `.pb`.
+ trust_remote_code (`bool`, *optional*, defaults to `False`):
+ Whether or not to allow for custom pipelines and components defined on the Hub in their own files. This
+ option should only be set to `True` for repositories you trust and in which you have read the code, as
+ it will execute code present on the Hub on your local machine.
+
+ Returns:
+ `os.PathLike`:
+ A path to the downloaded pipeline.
+
+
+
+ To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
+ `huggingface-cli login`.
+
+
+
+ """
+ cache_dir = kwargs.pop("cache_dir", None)
+ resume_download = kwargs.pop("resume_download", False)
+ force_download = kwargs.pop("force_download", False)
+ proxies = kwargs.pop("proxies", None)
+ local_files_only = kwargs.pop("local_files_only", None)
+ token = kwargs.pop("token", None)
+ revision = kwargs.pop("revision", None)
+ from_flax = kwargs.pop("from_flax", False)
+ custom_pipeline = kwargs.pop("custom_pipeline", None)
+ custom_revision = kwargs.pop("custom_revision", None)
+ variant = kwargs.pop("variant", None)
+ use_safetensors = kwargs.pop("use_safetensors", None)
+ use_onnx = kwargs.pop("use_onnx", None)
+ load_connected_pipeline = kwargs.pop("load_connected_pipeline", False)
+ trust_remote_code = kwargs.pop("trust_remote_code", False)
+
+ allow_pickle = False
+ if use_safetensors is None:
+ use_safetensors = True
+ allow_pickle = True
+
+ allow_patterns = None
+ ignore_patterns = None
+
+ model_info_call_error: Optional[Exception] = None
+ if not local_files_only:
+ try:
+ info = model_info(pretrained_model_name, token=token, revision=revision)
+ except (HTTPError, OfflineModeIsEnabled, requests.ConnectionError) as e:
+ logger.warn(f"Couldn't connect to the Hub: {e}.\nWill try to load from local cache.")
+ local_files_only = True
+ model_info_call_error = e # save error to reraise it if model is not cached locally
+
+ if not local_files_only:
+ config_file = hf_hub_download(
+ pretrained_model_name,
+ cls.config_name,
+ cache_dir=cache_dir,
+ revision=revision,
+ proxies=proxies,
+ force_download=force_download,
+ resume_download=resume_download,
+ token=token,
+ )
+
+ config_dict = cls._dict_from_json_file(config_file)
+ ignore_filenames = config_dict.pop("_ignore_files", [])
+
+ # retrieve all folder_names that contain relevant files
+ folder_names = [k for k, v in config_dict.items() if isinstance(v, list) and k != "_class_name"]
+
+ filenames = {sibling.rfilename for sibling in info.siblings}
+ model_filenames, variant_filenames = variant_compatible_siblings(filenames, variant=variant)
+
+ diffusers_module = importlib.import_module(__name__.split(".")[0])
+ pipelines = getattr(diffusers_module, "pipelines")
+
+ # optionally create a custom component <> custom file mapping
+ custom_components = {}
+ for component in folder_names:
+ module_candidate = config_dict[component][0]
+
+ if module_candidate is None or not isinstance(module_candidate, str):
+ continue
+
+ # We compute candidate file path on the Hub. Do not use `os.path.join`.
+ candidate_file = f"{component}/{module_candidate}.py"
+
+ if candidate_file in filenames:
+ custom_components[component] = module_candidate
+ elif module_candidate not in LOADABLE_CLASSES and not hasattr(pipelines, module_candidate):
+ raise ValueError(
+ f"{candidate_file} as defined in `model_index.json` does not exist in {pretrained_model_name} and is not a module in 'diffusers/pipelines'."
+ )
+
+ if len(variant_filenames) == 0 and variant is not None:
+ deprecation_message = (
+ f"You are trying to load the model files of the `variant={variant}`, but no such modeling files are available."
+ f"The default model files: {model_filenames} will be loaded instead. Make sure to not load from `variant={variant}`"
+ "if such variant modeling files are not available. Doing so will lead to an error in v0.24.0 as defaulting to non-variant"
+ "modeling files is deprecated."
+ )
+ deprecate("no variant default", "0.24.0", deprecation_message, standard_warn=False)
+
+ # remove ignored filenames
+ model_filenames = set(model_filenames) - set(ignore_filenames)
+ variant_filenames = set(variant_filenames) - set(ignore_filenames)
+
+ # if the whole pipeline is cached we don't have to ping the Hub
+ if revision in DEPRECATED_REVISION_ARGS and version.parse(
+ version.parse(__version__).base_version
+ ) >= version.parse("0.22.0"):
+ warn_deprecated_model_variant(pretrained_model_name, token, variant, revision, model_filenames)
+
+ model_folder_names = {os.path.split(f)[0] for f in model_filenames if os.path.split(f)[0] in folder_names}
+
+ custom_class_name = None
+ if custom_pipeline is None and isinstance(config_dict["_class_name"], (list, tuple)):
+ custom_pipeline = config_dict["_class_name"][0]
+ custom_class_name = config_dict["_class_name"][1]
+
+ # all filenames compatible with variant will be added
+ allow_patterns = list(model_filenames)
+
+ # allow all patterns from non-model folders
+ # this enables downloading schedulers, tokenizers, ...
+ allow_patterns += [f"{k}/*" for k in folder_names if k not in model_folder_names]
+ # add custom component files
+ allow_patterns += [f"{k}/{f}.py" for k, f in custom_components.items()]
+ # add custom pipeline file
+ allow_patterns += [f"{custom_pipeline}.py"] if f"{custom_pipeline}.py" in filenames else []
+ # also allow downloading config.json files with the model
+ allow_patterns += [os.path.join(k, "config.json") for k in model_folder_names]
+
+ allow_patterns += [
+ SCHEDULER_CONFIG_NAME,
+ CONFIG_NAME,
+ cls.config_name,
+ CUSTOM_PIPELINE_FILE_NAME,
+ ]
+
+ load_pipe_from_hub = custom_pipeline is not None and f"{custom_pipeline}.py" in filenames
+ load_components_from_hub = len(custom_components) > 0
+
+ if load_pipe_from_hub and not trust_remote_code:
+ raise ValueError(
+ f"The repository for {pretrained_model_name} contains custom code in {custom_pipeline}.py which must be executed to correctly "
+ f"load the model. You can inspect the repository content at https://hf.co/{pretrained_model_name}/blob/main/{custom_pipeline}.py.\n"
+ f"Please pass the argument `trust_remote_code=True` to allow custom code to be run."
+ )
+
+ if load_components_from_hub and not trust_remote_code:
+ raise ValueError(
+ f"The repository for {pretrained_model_name} contains custom code in {'.py, '.join([os.path.join(k, v) for k,v in custom_components.items()])} which must be executed to correctly "
+ f"load the model. You can inspect the repository content at {', '.join([f'https://hf.co/{pretrained_model_name}/{k}/{v}.py' for k,v in custom_components.items()])}.\n"
+ f"Please pass the argument `trust_remote_code=True` to allow custom code to be run."
+ )
+
+ # retrieve passed components that should not be downloaded
+ pipeline_class = _get_pipeline_class(
+ cls,
+ config_dict,
+ load_connected_pipeline=load_connected_pipeline,
+ custom_pipeline=custom_pipeline,
+ repo_id=pretrained_model_name if load_pipe_from_hub else None,
+ hub_revision=revision,
+ class_name=custom_class_name,
+ cache_dir=cache_dir,
+ revision=custom_revision,
+ )
+ expected_components, _ = cls._get_signature_keys(pipeline_class)
+ passed_components = [k for k in expected_components if k in kwargs]
+
+ if (
+ use_safetensors
+ and not allow_pickle
+ and not is_safetensors_compatible(
+ model_filenames, variant=variant, passed_components=passed_components
+ )
+ ):
+ raise EnvironmentError(
+ f"Could not find the necessary `safetensors` weights in {model_filenames} (variant={variant})"
+ )
+ if from_flax:
+ ignore_patterns = ["*.bin", "*.safetensors", "*.onnx", "*.pb"]
+ elif use_safetensors and is_safetensors_compatible(
+ model_filenames, variant=variant, passed_components=passed_components
+ ):
+ ignore_patterns = ["*.bin", "*.msgpack"]
+
+ use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx
+ if not use_onnx:
+ ignore_patterns += ["*.onnx", "*.pb"]
+
+ safetensors_variant_filenames = {f for f in variant_filenames if f.endswith(".safetensors")}
+ safetensors_model_filenames = {f for f in model_filenames if f.endswith(".safetensors")}
+ if (
+ len(safetensors_variant_filenames) > 0
+ and safetensors_model_filenames != safetensors_variant_filenames
+ ):
+ logger.warn(
+ f"\nA mixture of {variant} and non-{variant} filenames will be loaded.\nLoaded {variant} filenames:\n[{', '.join(safetensors_variant_filenames)}]\nLoaded non-{variant} filenames:\n[{', '.join(safetensors_model_filenames - safetensors_variant_filenames)}\nIf this behavior is not expected, please check your folder structure."
+ )
+ else:
+ ignore_patterns = ["*.safetensors", "*.msgpack"]
+
+ use_onnx = use_onnx if use_onnx is not None else pipeline_class._is_onnx
+ if not use_onnx:
+ ignore_patterns += ["*.onnx", "*.pb"]
+
+ bin_variant_filenames = {f for f in variant_filenames if f.endswith(".bin")}
+ bin_model_filenames = {f for f in model_filenames if f.endswith(".bin")}
+ if len(bin_variant_filenames) > 0 and bin_model_filenames != bin_variant_filenames:
+ logger.warn(
+ f"\nA mixture of {variant} and non-{variant} filenames will be loaded.\nLoaded {variant} filenames:\n[{', '.join(bin_variant_filenames)}]\nLoaded non-{variant} filenames:\n[{', '.join(bin_model_filenames - bin_variant_filenames)}\nIf this behavior is not expected, please check your folder structure."
+ )
+
+ # Don't download any objects that are passed
+ allow_patterns = [
+ p for p in allow_patterns if not (len(p.split("/")) == 2 and p.split("/")[0] in passed_components)
+ ]
+
+ if pipeline_class._load_connected_pipes:
+ allow_patterns.append("README.md")
+
+ # Don't download index files of forbidden patterns either
+ ignore_patterns = ignore_patterns + [f"{i}.index.*json" for i in ignore_patterns]
+
+ re_ignore_pattern = [re.compile(fnmatch.translate(p)) for p in ignore_patterns]
+ re_allow_pattern = [re.compile(fnmatch.translate(p)) for p in allow_patterns]
+
+ expected_files = [f for f in filenames if not any(p.match(f) for p in re_ignore_pattern)]
+ expected_files = [f for f in expected_files if any(p.match(f) for p in re_allow_pattern)]
+
+ snapshot_folder = Path(config_file).parent
+ pipeline_is_cached = all((snapshot_folder / f).is_file() for f in expected_files)
+
+ if pipeline_is_cached and not force_download:
+ # if the pipeline is cached, we can directly return it
+ # else call snapshot_download
+ return snapshot_folder
+
+ user_agent = {"pipeline_class": cls.__name__}
+ if custom_pipeline is not None and not custom_pipeline.endswith(".py"):
+ user_agent["custom_pipeline"] = custom_pipeline
+
+ # download all allow_patterns - ignore_patterns
+ try:
+ cached_folder = snapshot_download(
+ pretrained_model_name,
+ cache_dir=cache_dir,
+ resume_download=resume_download,
+ proxies=proxies,
+ local_files_only=local_files_only,
+ token=token,
+ revision=revision,
+ allow_patterns=allow_patterns,
+ ignore_patterns=ignore_patterns,
+ user_agent=user_agent,
+ )
+
+ # retrieve pipeline class from local file
+ cls_name = cls.load_config(os.path.join(cached_folder, "model_index.json")).get("_class_name", None)
+ cls_name = cls_name[4:] if isinstance(cls_name, str) and cls_name.startswith("Flax") else cls_name
+
+ diffusers_module = importlib.import_module(__name__.split(".")[0])
+ pipeline_class = getattr(diffusers_module, cls_name, None) if isinstance(cls_name, str) else None
+
+ if pipeline_class is not None and pipeline_class._load_connected_pipes:
+ modelcard = ModelCard.load(os.path.join(cached_folder, "README.md"))
+ connected_pipes = sum([getattr(modelcard.data, k, []) for k in CONNECTED_PIPES_KEYS], [])
+ for connected_pipe_repo_id in connected_pipes:
+ download_kwargs = {
+ "cache_dir": cache_dir,
+ "resume_download": resume_download,
+ "force_download": force_download,
+ "proxies": proxies,
+ "local_files_only": local_files_only,
+ "token": token,
+ "variant": variant,
+ "use_safetensors": use_safetensors,
+ }
+ DiffusionPipeline.download(connected_pipe_repo_id, **download_kwargs)
+
+ return cached_folder
+
+ except FileNotFoundError:
+ # Means we tried to load pipeline with `local_files_only=True` but the files have not been found in local cache.
+ # This can happen in two cases:
+ # 1. If the user passed `local_files_only=True` => we raise the error directly
+ # 2. If we forced `local_files_only=True` when `model_info` failed => we raise the initial error
+ if model_info_call_error is None:
+ # 1. user passed `local_files_only=True`
+ raise
+ else:
+ # 2. we forced `local_files_only=True` when `model_info` failed
+ raise EnvironmentError(
+ f"Cannot load model {pretrained_model_name}: model is not cached locally and an error occurred"
+ " while trying to fetch metadata from the Hub. Please check out the root cause in the stacktrace"
+ " above."
+ ) from model_info_call_error
+
+ @classmethod
+ def _get_signature_keys(cls, obj):
+ parameters = inspect.signature(obj.__init__).parameters
+ required_parameters = {k: v for k, v in parameters.items() if v.default == inspect._empty}
+ optional_parameters = set({k for k, v in parameters.items() if v.default != inspect._empty})
+ expected_modules = set(required_parameters.keys()) - {"self"}
+
+ optional_names = list(optional_parameters)
+ for name in optional_names:
+ if name in cls._optional_components:
+ expected_modules.add(name)
+ optional_parameters.remove(name)
+
+ return expected_modules, optional_parameters
+
+ @property
+ def components(self) -> Dict[str, Any]:
+ r"""
+ The `self.components` property can be useful to run different pipelines with the same weights and
+ configurations without reallocating additional memory.
+
+ Returns (`dict`):
+ A dictionary containing all the modules needed to initialize the pipeline.
+
+ Examples:
+
+ ```py
+ >>> from diffusers import (
+ ... StableDiffusionPipeline,
+ ... StableDiffusionImg2ImgPipeline,
+ ... StableDiffusionInpaintPipeline,
+ ... )
+
+ >>> text2img = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
+ >>> img2img = StableDiffusionImg2ImgPipeline(**text2img.components)
+ >>> inpaint = StableDiffusionInpaintPipeline(**text2img.components)
+ ```
+ """
+ expected_modules, optional_parameters = self._get_signature_keys(self)
+ components = {
+ k: getattr(self, k) for k in self.config.keys() if not k.startswith("_") and k not in optional_parameters
+ }
+
+ if set(components.keys()) != expected_modules:
+ raise ValueError(
+ f"{self} has been incorrectly initialized or {self.__class__} is incorrectly implemented. Expected"
+ f" {expected_modules} to be defined, but {components.keys()} are defined."
+ )
+
+ return components
+
+ @staticmethod
+ def numpy_to_pil(images):
+ """
+ Convert a NumPy image or a batch of images to a PIL image.
+ """
+ return numpy_to_pil(images)
+
+ def progress_bar(self, iterable=None, total=None):
+ if not hasattr(self, "_progress_bar_config"):
+ self._progress_bar_config = {}
+ elif not isinstance(self._progress_bar_config, dict):
+ raise ValueError(
+ f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}."
+ )
+
+ if iterable is not None:
+ return tqdm(iterable, **self._progress_bar_config)
+ elif total is not None:
+ return tqdm(total=total, **self._progress_bar_config)
+ else:
+ raise ValueError("Either `total` or `iterable` has to be defined.")
+
+ def set_progress_bar_config(self, **kwargs):
+ self._progress_bar_config = kwargs
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ r"""
+ Enable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/). When this
+ option is enabled, you should observe lower GPU memory usage and a potential speed up during inference. Speed
+ up during training is not guaranteed.
+
+
+
+ ⚠️ When memory efficient attention and sliced attention are both enabled, memory efficient attention takes
+ precedent.
+
+
+
+ Parameters:
+ attention_op (`Callable`, *optional*):
+ Override the default `None` operator for use as `op` argument to the
+ [`memory_efficient_attention()`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention)
+ function of xFormers.
+
+ Examples:
+
+ ```py
+ >>> import torch
+ >>> from diffusers import DiffusionPipeline
+ >>> from xformers.ops import MemoryEfficientAttentionFlashAttentionOp
+
+ >>> pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16)
+ >>> pipe = pipe.to("cuda")
+ >>> pipe.enable_xformers_memory_efficient_attention(attention_op=MemoryEfficientAttentionFlashAttentionOp)
+ >>> # Workaround for not accepting attention shape using VAE for Flash Attention
+ >>> pipe.vae.enable_xformers_memory_efficient_attention(attention_op=None)
+ ```
+ """
+ self.set_use_memory_efficient_attention_xformers(True, attention_op)
+
+ def disable_xformers_memory_efficient_attention(self):
+ r"""
+ Disable memory efficient attention from [xFormers](https://facebookresearch.github.io/xformers/).
+ """
+ self.set_use_memory_efficient_attention_xformers(False)
+
+ def set_use_memory_efficient_attention_xformers(
+ self, valid: bool, attention_op: Optional[Callable] = None
+ ) -> None:
+ # Recursively walk through all the children.
+ # Any children which exposes the set_use_memory_efficient_attention_xformers method
+ # gets the message
+ def fn_recursive_set_mem_eff(module: torch.nn.Module):
+ if hasattr(module, "set_use_memory_efficient_attention_xformers"):
+ module.set_use_memory_efficient_attention_xformers(valid, attention_op)
+
+ for child in module.children():
+ fn_recursive_set_mem_eff(child)
+
+ module_names, _ = self._get_signature_keys(self)
+ modules = [getattr(self, n, None) for n in module_names]
+ modules = [m for m in modules if isinstance(m, torch.nn.Module)]
+
+ for module in modules:
+ fn_recursive_set_mem_eff(module)
+
+ def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
+ r"""
+ Enable sliced attention computation. When this option is enabled, the attention module splits the input tensor
+ in slices to compute attention in several steps. For more than one attention head, the computation is performed
+ sequentially over each head. This is useful to save some memory in exchange for a small speed decrease.
+
+
+
+ ⚠️ Don't enable attention slicing if you're already using `scaled_dot_product_attention` (SDPA) from PyTorch
+ 2.0 or xFormers. These attention computations are already very memory efficient so you won't need to enable
+ this function. If you enable attention slicing with SDPA or xFormers, it can lead to serious slow downs!
+
+
+
+ Args:
+ slice_size (`str` or `int`, *optional*, defaults to `"auto"`):
+ When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
+ `"max"`, maximum amount of memory will be saved by running only one slice at a time. If a number is
+ provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+ must be a multiple of `slice_size`.
+
+ Examples:
+
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionPipeline
+
+ >>> pipe = StableDiffusionPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5",
+ ... torch_dtype=torch.float16,
+ ... use_safetensors=True,
+ ... )
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> pipe.enable_attention_slicing()
+ >>> image = pipe(prompt).images[0]
+ ```
+ """
+ self.set_attention_slice(slice_size)
+
+ def disable_attention_slicing(self):
+ r"""
+ Disable sliced attention computation. If `enable_attention_slicing` was previously called, attention is
+ computed in one step.
+ """
+ # set slice_size = `None` to disable `attention slicing`
+ self.enable_attention_slicing(None)
+
+ def set_attention_slice(self, slice_size: Optional[int]):
+ module_names, _ = self._get_signature_keys(self)
+ modules = [getattr(self, n, None) for n in module_names]
+ modules = [m for m in modules if isinstance(m, torch.nn.Module) and hasattr(m, "set_attention_slice")]
+
+ for module in modules:
+ module.set_attention_slice(slice_size)
+
+
+class StableDiffusionMixin:
+ r"""
+ Helper for DiffusionPipeline with vae and unet.(mainly for LDM such as stable diffusion)
+ """
+
+ def enable_vae_slicing(self):
+ r"""
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+ """
+ self.vae.enable_slicing()
+
+ def disable_vae_slicing(self):
+ r"""
+ Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+ computing decoding in one step.
+ """
+ self.vae.disable_slicing()
+
+ def enable_vae_tiling(self):
+ r"""
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+ processing larger images.
+ """
+ self.vae.enable_tiling()
+
+ def disable_vae_tiling(self):
+ r"""
+ Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+ computing decoding in one step.
+ """
+ self.vae.disable_tiling()
+
+ def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+ r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497.
+
+ The suffixes after the scaling factors represent the stages where they are being applied.
+
+ Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values
+ that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+ Args:
+ s1 (`float`):
+ Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+ mitigate "oversmoothing effect" in the enhanced denoising process.
+ s2 (`float`):
+ Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+ mitigate "oversmoothing effect" in the enhanced denoising process.
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+ """
+ if not hasattr(self, "unet"):
+ raise ValueError("The pipeline must have `unet` for using FreeU.")
+ self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2)
+
+ def disable_freeu(self):
+ """Disables the FreeU mechanism if enabled."""
+ self.unet.disable_freeu()
+
+ def fuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+ """
+ Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+ key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ Args:
+ unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+ vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+ """
+ self.fusing_unet = False
+ self.fusing_vae = False
+
+ if unet:
+ self.fusing_unet = True
+ self.unet.fuse_qkv_projections()
+ self.unet.set_attn_processor(FusedAttnProcessor2_0())
+
+ if vae:
+ if not isinstance(self.vae, AutoencoderKL):
+ raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.")
+
+ self.fusing_vae = True
+ self.vae.fuse_qkv_projections()
+ self.vae.set_attn_processor(FusedAttnProcessor2_0())
+
+ def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True):
+ """Disable QKV projection fusion if enabled.
+
+
+
+ This API is 🧪 experimental.
+
+
+
+ Args:
+ unet (`bool`, defaults to `True`): To apply fusion on the UNet.
+ vae (`bool`, defaults to `True`): To apply fusion on the VAE.
+
+ """
+ if unet:
+ if not self.fusing_unet:
+ logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.")
+ else:
+ self.unet.unfuse_qkv_projections()
+ self.fusing_unet = False
+
+ if vae:
+ if not self.fusing_vae:
+ logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.")
+ else:
+ self.vae.unfuse_qkv_projections()
+ self.fusing_vae = False
diff --git a/model/diffusers_c/pipelines/pixart_alpha/__init__.py b/model/diffusers_c/pipelines/pixart_alpha/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..0bfa28fcde50a555197fdc594d3fb92957398397
--- /dev/null
+++ b/model/diffusers_c/pipelines/pixart_alpha/__init__.py
@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_pixart_alpha"] = ["PixArtAlphaPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_pixart_alpha import PixArtAlphaPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/pixart_alpha/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/pixart_alpha/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d2cb5b2804b7c75116ebd1b4e68c530d89aae12c
Binary files /dev/null and b/model/diffusers_c/pipelines/pixart_alpha/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/pixart_alpha/__pycache__/pipeline_pixart_alpha.cpython-39.pyc b/model/diffusers_c/pipelines/pixart_alpha/__pycache__/pipeline_pixart_alpha.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..054347f797d404f9b803dcc08b9fb62a797d8f68
Binary files /dev/null and b/model/diffusers_c/pipelines/pixart_alpha/__pycache__/pipeline_pixart_alpha.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/pixart_alpha/pipeline_pixart_alpha.py b/model/diffusers_c/pipelines/pixart_alpha/pipeline_pixart_alpha.py
new file mode 100644
index 0000000000000000000000000000000000000000..c12bca90aa141d6b2f91463fcdbebc9cc660ad44
--- /dev/null
+++ b/model/diffusers_c/pipelines/pixart_alpha/pipeline_pixart_alpha.py
@@ -0,0 +1,979 @@
+# Copyright 2024 PixArt-Alpha Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import html
+import inspect
+import re
+import urllib.parse as ul
+from typing import Callable, List, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import T5EncoderModel, T5Tokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...models import AutoencoderKL, Transformer2DModel
+from ...schedulers import DPMSolverMultistepScheduler
+from ...utils import (
+ BACKENDS_MAPPING,
+ deprecate,
+ is_bs4_available,
+ is_ftfy_available,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+if is_bs4_available():
+ from bs4 import BeautifulSoup
+
+if is_ftfy_available():
+ import ftfy
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import PixArtAlphaPipeline
+
+ >>> # You can replace the checkpoint id with "PixArt-alpha/PixArt-XL-2-512x512" too.
+ >>> pipe = PixArtAlphaPipeline.from_pretrained("PixArt-alpha/PixArt-XL-2-1024-MS", torch_dtype=torch.float16)
+ >>> # Enable memory optimizations.
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = "A small cactus with a happy face in the Sahara desert."
+ >>> image = pipe(prompt).images[0]
+ ```
+"""
+
+ASPECT_RATIO_1024_BIN = {
+ "0.25": [512.0, 2048.0],
+ "0.28": [512.0, 1856.0],
+ "0.32": [576.0, 1792.0],
+ "0.33": [576.0, 1728.0],
+ "0.35": [576.0, 1664.0],
+ "0.4": [640.0, 1600.0],
+ "0.42": [640.0, 1536.0],
+ "0.48": [704.0, 1472.0],
+ "0.5": [704.0, 1408.0],
+ "0.52": [704.0, 1344.0],
+ "0.57": [768.0, 1344.0],
+ "0.6": [768.0, 1280.0],
+ "0.68": [832.0, 1216.0],
+ "0.72": [832.0, 1152.0],
+ "0.78": [896.0, 1152.0],
+ "0.82": [896.0, 1088.0],
+ "0.88": [960.0, 1088.0],
+ "0.94": [960.0, 1024.0],
+ "1.0": [1024.0, 1024.0],
+ "1.07": [1024.0, 960.0],
+ "1.13": [1088.0, 960.0],
+ "1.21": [1088.0, 896.0],
+ "1.29": [1152.0, 896.0],
+ "1.38": [1152.0, 832.0],
+ "1.46": [1216.0, 832.0],
+ "1.67": [1280.0, 768.0],
+ "1.75": [1344.0, 768.0],
+ "2.0": [1408.0, 704.0],
+ "2.09": [1472.0, 704.0],
+ "2.4": [1536.0, 640.0],
+ "2.5": [1600.0, 640.0],
+ "3.0": [1728.0, 576.0],
+ "4.0": [2048.0, 512.0],
+}
+
+ASPECT_RATIO_512_BIN = {
+ "0.25": [256.0, 1024.0],
+ "0.28": [256.0, 928.0],
+ "0.32": [288.0, 896.0],
+ "0.33": [288.0, 864.0],
+ "0.35": [288.0, 832.0],
+ "0.4": [320.0, 800.0],
+ "0.42": [320.0, 768.0],
+ "0.48": [352.0, 736.0],
+ "0.5": [352.0, 704.0],
+ "0.52": [352.0, 672.0],
+ "0.57": [384.0, 672.0],
+ "0.6": [384.0, 640.0],
+ "0.68": [416.0, 608.0],
+ "0.72": [416.0, 576.0],
+ "0.78": [448.0, 576.0],
+ "0.82": [448.0, 544.0],
+ "0.88": [480.0, 544.0],
+ "0.94": [480.0, 512.0],
+ "1.0": [512.0, 512.0],
+ "1.07": [512.0, 480.0],
+ "1.13": [544.0, 480.0],
+ "1.21": [544.0, 448.0],
+ "1.29": [576.0, 448.0],
+ "1.38": [576.0, 416.0],
+ "1.46": [608.0, 416.0],
+ "1.67": [640.0, 384.0],
+ "1.75": [672.0, 384.0],
+ "2.0": [704.0, 352.0],
+ "2.09": [736.0, 352.0],
+ "2.4": [768.0, 320.0],
+ "2.5": [800.0, 320.0],
+ "3.0": [864.0, 288.0],
+ "4.0": [1024.0, 256.0],
+}
+
+ASPECT_RATIO_256_BIN = {
+ "0.25": [128.0, 512.0],
+ "0.28": [128.0, 464.0],
+ "0.32": [144.0, 448.0],
+ "0.33": [144.0, 432.0],
+ "0.35": [144.0, 416.0],
+ "0.4": [160.0, 400.0],
+ "0.42": [160.0, 384.0],
+ "0.48": [176.0, 368.0],
+ "0.5": [176.0, 352.0],
+ "0.52": [176.0, 336.0],
+ "0.57": [192.0, 336.0],
+ "0.6": [192.0, 320.0],
+ "0.68": [208.0, 304.0],
+ "0.72": [208.0, 288.0],
+ "0.78": [224.0, 288.0],
+ "0.82": [224.0, 272.0],
+ "0.88": [240.0, 272.0],
+ "0.94": [240.0, 256.0],
+ "1.0": [256.0, 256.0],
+ "1.07": [256.0, 240.0],
+ "1.13": [272.0, 240.0],
+ "1.21": [272.0, 224.0],
+ "1.29": [288.0, 224.0],
+ "1.38": [288.0, 208.0],
+ "1.46": [304.0, 208.0],
+ "1.67": [320.0, 192.0],
+ "1.75": [336.0, 192.0],
+ "2.0": [352.0, 176.0],
+ "2.09": [368.0, 176.0],
+ "2.4": [384.0, 160.0],
+ "2.5": [400.0, 160.0],
+ "3.0": [432.0, 144.0],
+ "4.0": [512.0, 128.0],
+}
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class PixArtAlphaPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for text-to-image generation using PixArt-Alpha.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`T5EncoderModel`]):
+ Frozen text-encoder. PixArt-Alpha uses
+ [T5](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5EncoderModel), specifically the
+ [t5-v1_1-xxl](https://huggingface.co/PixArt-alpha/PixArt-alpha/tree/main/t5-v1_1-xxl) variant.
+ tokenizer (`T5Tokenizer`):
+ Tokenizer of class
+ [T5Tokenizer](https://huggingface.co/docs/transformers/model_doc/t5#transformers.T5Tokenizer).
+ transformer ([`Transformer2DModel`]):
+ A text conditioned `Transformer2DModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
+ """
+
+ bad_punct_regex = re.compile(
+ r"["
+ + "#®•©™&@·º½¾¿¡§~"
+ + r"\)"
+ + r"\("
+ + r"\]"
+ + r"\["
+ + r"\}"
+ + r"\{"
+ + r"\|"
+ + "\\"
+ + r"\/"
+ + r"\*"
+ + r"]{1,}"
+ ) # noqa
+
+ _optional_components = ["tokenizer", "text_encoder"]
+ model_cpu_offload_seq = "text_encoder->transformer->vae"
+
+ def __init__(
+ self,
+ tokenizer: T5Tokenizer,
+ text_encoder: T5EncoderModel,
+ vae: AutoencoderKL,
+ transformer: Transformer2DModel,
+ scheduler: DPMSolverMultistepScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
+ )
+
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/utils.py
+ def mask_text_embeddings(self, emb, mask):
+ if emb.shape[0] == 1:
+ keep_index = mask.sum().item()
+ return emb[:, :, :keep_index, :], keep_index
+ else:
+ masked_feature = emb * mask[:, None, :, None]
+ return masked_feature, emb.shape[2]
+
+ # Adapted from diffusers.pipelines.deepfloyd_if.pipeline_if.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: str = "",
+ num_images_per_prompt: int = 1,
+ device: Optional[torch.device] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ prompt_attention_mask: Optional[torch.FloatTensor] = None,
+ negative_prompt_attention_mask: Optional[torch.FloatTensor] = None,
+ clean_caption: bool = False,
+ max_sequence_length: int = 120,
+ **kwargs,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds`
+ instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). For
+ PixArt-Alpha, this should be "".
+ do_classifier_free_guidance (`bool`, *optional*, defaults to `True`):
+ whether to use classifier free guidance or not
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ number of images that should be generated per prompt
+ device: (`torch.device`, *optional*):
+ torch device to place the resulting embeddings on
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. For PixArt-Alpha, it's should be the embeddings of the ""
+ string.
+ clean_caption (`bool`, defaults to `False`):
+ If `True`, the function will preprocess and clean the provided caption before encoding.
+ max_sequence_length (`int`, defaults to 120): Maximum sequence length to use for the prompt.
+ """
+
+ if "mask_feature" in kwargs:
+ deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version."
+ deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False)
+
+ if device is None:
+ device = self._execution_device
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # See Section 3.1. of the paper.
+ max_length = max_sequence_length
+
+ if prompt_embeds is None:
+ prompt = self._text_preprocessing(prompt, clean_caption=clean_caption)
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {max_length} tokens: {removed_text}"
+ )
+
+ prompt_attention_mask = text_inputs.attention_mask
+ prompt_attention_mask = prompt_attention_mask.to(device)
+
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=prompt_attention_mask)
+ prompt_embeds = prompt_embeds[0]
+
+ if self.text_encoder is not None:
+ dtype = self.text_encoder.dtype
+ elif self.transformer is not None:
+ dtype = self.transformer.dtype
+ else:
+ dtype = None
+
+ prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+ prompt_attention_mask = prompt_attention_mask.view(bs_embed, -1)
+ prompt_attention_mask = prompt_attention_mask.repeat(num_images_per_prompt, 1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens = [negative_prompt] * batch_size
+ uncond_tokens = self._text_preprocessing(uncond_tokens, clean_caption=clean_caption)
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_attention_mask=True,
+ add_special_tokens=True,
+ return_tensors="pt",
+ )
+ negative_prompt_attention_mask = uncond_input.attention_mask
+ negative_prompt_attention_mask = negative_prompt_attention_mask.to(device)
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device), attention_mask=negative_prompt_attention_mask
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ negative_prompt_attention_mask = negative_prompt_attention_mask.view(bs_embed, -1)
+ negative_prompt_attention_mask = negative_prompt_attention_mask.repeat(num_images_per_prompt, 1)
+ else:
+ negative_prompt_embeds = None
+ negative_prompt_attention_mask = None
+
+ return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ negative_prompt,
+ callback_steps,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ prompt_attention_mask=None,
+ negative_prompt_attention_mask=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and prompt_attention_mask is None:
+ raise ValueError("Must provide `prompt_attention_mask` when specifying `prompt_embeds`.")
+
+ if negative_prompt_embeds is not None and negative_prompt_attention_mask is None:
+ raise ValueError("Must provide `negative_prompt_attention_mask` when specifying `negative_prompt_embeds`.")
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+ if prompt_attention_mask.shape != negative_prompt_attention_mask.shape:
+ raise ValueError(
+ "`prompt_attention_mask` and `negative_prompt_attention_mask` must have the same shape when passed directly, but"
+ f" got: `prompt_attention_mask` {prompt_attention_mask.shape} != `negative_prompt_attention_mask`"
+ f" {negative_prompt_attention_mask.shape}."
+ )
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._text_preprocessing
+ def _text_preprocessing(self, text, clean_caption=False):
+ if clean_caption and not is_bs4_available():
+ logger.warn(BACKENDS_MAPPING["bs4"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if clean_caption and not is_ftfy_available():
+ logger.warn(BACKENDS_MAPPING["ftfy"][-1].format("Setting `clean_caption=True`"))
+ logger.warn("Setting `clean_caption` to False...")
+ clean_caption = False
+
+ if not isinstance(text, (tuple, list)):
+ text = [text]
+
+ def process(text: str):
+ if clean_caption:
+ text = self._clean_caption(text)
+ text = self._clean_caption(text)
+ else:
+ text = text.lower().strip()
+ return text
+
+ return [process(t) for t in text]
+
+ # Copied from diffusers.pipelines.deepfloyd_if.pipeline_if.IFPipeline._clean_caption
+ def _clean_caption(self, caption):
+ caption = str(caption)
+ caption = ul.unquote_plus(caption)
+ caption = caption.strip().lower()
+ caption = re.sub("", "person", caption)
+ # urls:
+ caption = re.sub(
+ r"\b((?:https?:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ caption = re.sub(
+ r"\b((?:www:(?:\/{1,3}|[a-zA-Z0-9%])|[a-zA-Z0-9.\-]+[.](?:com|co|ru|net|org|edu|gov|it)[\w/-]*\b\/?(?!@)))", # noqa
+ "",
+ caption,
+ ) # regex for urls
+ # html:
+ caption = BeautifulSoup(caption, features="html.parser").text
+
+ # @
+ caption = re.sub(r"@[\w\d]+\b", "", caption)
+
+ # 31C0—31EF CJK Strokes
+ # 31F0—31FF Katakana Phonetic Extensions
+ # 3200—32FF Enclosed CJK Letters and Months
+ # 3300—33FF CJK Compatibility
+ # 3400—4DBF CJK Unified Ideographs Extension A
+ # 4DC0—4DFF Yijing Hexagram Symbols
+ # 4E00—9FFF CJK Unified Ideographs
+ caption = re.sub(r"[\u31c0-\u31ef]+", "", caption)
+ caption = re.sub(r"[\u31f0-\u31ff]+", "", caption)
+ caption = re.sub(r"[\u3200-\u32ff]+", "", caption)
+ caption = re.sub(r"[\u3300-\u33ff]+", "", caption)
+ caption = re.sub(r"[\u3400-\u4dbf]+", "", caption)
+ caption = re.sub(r"[\u4dc0-\u4dff]+", "", caption)
+ caption = re.sub(r"[\u4e00-\u9fff]+", "", caption)
+ #######################################################
+
+ # все виды тире / all types of dash --> "-"
+ caption = re.sub(
+ r"[\u002D\u058A\u05BE\u1400\u1806\u2010-\u2015\u2E17\u2E1A\u2E3A\u2E3B\u2E40\u301C\u3030\u30A0\uFE31\uFE32\uFE58\uFE63\uFF0D]+", # noqa
+ "-",
+ caption,
+ )
+
+ # кавычки к одному стандарту
+ caption = re.sub(r"[`´«»“”¨]", '"', caption)
+ caption = re.sub(r"[‘’]", "'", caption)
+
+ # "
+ caption = re.sub(r""?", "", caption)
+ # &
+ caption = re.sub(r"&", "", caption)
+
+ # ip adresses:
+ caption = re.sub(r"\d{1,3}\.\d{1,3}\.\d{1,3}\.\d{1,3}", " ", caption)
+
+ # article ids:
+ caption = re.sub(r"\d:\d\d\s+$", "", caption)
+
+ # \n
+ caption = re.sub(r"\\n", " ", caption)
+
+ # "#123"
+ caption = re.sub(r"#\d{1,3}\b", "", caption)
+ # "#12345.."
+ caption = re.sub(r"#\d{5,}\b", "", caption)
+ # "123456.."
+ caption = re.sub(r"\b\d{6,}\b", "", caption)
+ # filenames:
+ caption = re.sub(r"[\S]+\.(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)", "", caption)
+
+ #
+ caption = re.sub(r"[\"\']{2,}", r'"', caption) # """AUSVERKAUFT"""
+ caption = re.sub(r"[\.]{2,}", r" ", caption) # """AUSVERKAUFT"""
+
+ caption = re.sub(self.bad_punct_regex, r" ", caption) # ***AUSVERKAUFT***, #AUSVERKAUFT
+ caption = re.sub(r"\s+\.\s+", r" ", caption) # " . "
+
+ # this-is-my-cute-cat / this_is_my_cute_cat
+ regex2 = re.compile(r"(?:\-|\_)")
+ if len(re.findall(regex2, caption)) > 3:
+ caption = re.sub(regex2, " ", caption)
+
+ caption = ftfy.fix_text(caption)
+ caption = html.unescape(html.unescape(caption))
+
+ caption = re.sub(r"\b[a-zA-Z]{1,3}\d{3,15}\b", "", caption) # jc6640
+ caption = re.sub(r"\b[a-zA-Z]+\d+[a-zA-Z]+\b", "", caption) # jc6640vc
+ caption = re.sub(r"\b\d+[a-zA-Z]+\d+\b", "", caption) # 6640vc231
+
+ caption = re.sub(r"(worldwide\s+)?(free\s+)?shipping", "", caption)
+ caption = re.sub(r"(free\s)?download(\sfree)?", "", caption)
+ caption = re.sub(r"\bclick\b\s(?:for|on)\s\w+", "", caption)
+ caption = re.sub(r"\b(?:png|jpg|jpeg|bmp|webp|eps|pdf|apk|mp4)(\simage[s]?)?", "", caption)
+ caption = re.sub(r"\bpage\s+\d+\b", "", caption)
+
+ caption = re.sub(r"\b\d*[a-zA-Z]+\d+[a-zA-Z]+\d+[a-zA-Z\d]*\b", r" ", caption) # j2d1a2a...
+
+ caption = re.sub(r"\b\d+\.?\d*[xх×]\d+\.?\d*\b", "", caption)
+
+ caption = re.sub(r"\b\s+\:\s+", r": ", caption)
+ caption = re.sub(r"(\D[,\./])\b", r"\1 ", caption)
+ caption = re.sub(r"\s+", " ", caption)
+
+ caption.strip()
+
+ caption = re.sub(r"^[\"\']([\w\W]+)[\"\']$", r"\1", caption)
+ caption = re.sub(r"^[\'\_,\-\:;]", r"", caption)
+ caption = re.sub(r"[\'\_,\-\:\-\+]$", r"", caption)
+ caption = re.sub(r"^\.\S+$", "", caption)
+
+ return caption.strip()
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @staticmethod
+ def classify_height_width_bin(height: int, width: int, ratios: dict) -> Tuple[int, int]:
+ """Returns binned height and width."""
+ ar = float(height / width)
+ closest_ratio = min(ratios.keys(), key=lambda ratio: abs(float(ratio) - ar))
+ default_hw = ratios[closest_ratio]
+ return int(default_hw[0]), int(default_hw[1])
+
+ @staticmethod
+ def resize_and_crop_tensor(samples: torch.Tensor, new_width: int, new_height: int) -> torch.Tensor:
+ orig_height, orig_width = samples.shape[2], samples.shape[3]
+
+ # Check if resizing is needed
+ if orig_height != new_height or orig_width != new_width:
+ ratio = max(new_height / orig_height, new_width / orig_width)
+ resized_width = int(orig_width * ratio)
+ resized_height = int(orig_height * ratio)
+
+ # Resize
+ samples = F.interpolate(
+ samples, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+ )
+
+ # Center Crop
+ start_x = (resized_width - new_width) // 2
+ end_x = start_x + new_width
+ start_y = (resized_height - new_height) // 2
+ end_y = start_y + new_height
+ samples = samples[:, :, start_y:end_y, start_x:end_x]
+
+ return samples
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ negative_prompt: str = "",
+ num_inference_steps: int = 20,
+ timesteps: List[int] = None,
+ guidance_scale: float = 4.5,
+ num_images_per_prompt: Optional[int] = 1,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ prompt_attention_mask: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_attention_mask: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ clean_caption: bool = True,
+ use_resolution_binning: bool = True,
+ max_sequence_length: int = 120,
+ **kwargs,
+ ) -> Union[ImagePipelineOutput, Tuple]:
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 4.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size):
+ The width in pixels of the generated image.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ prompt_attention_mask (`torch.FloatTensor`, *optional*): Pre-generated attention mask for text embeddings.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. For PixArt-Alpha this negative prompt should be "". If not
+ provided, negative_prompt_embeds will be generated from `negative_prompt` input argument.
+ negative_prompt_attention_mask (`torch.FloatTensor`, *optional*):
+ Pre-generated attention mask for negative text embeddings.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ clean_caption (`bool`, *optional*, defaults to `True`):
+ Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
+ be installed. If the dependencies are not installed, the embeddings will be created from the raw
+ prompt.
+ use_resolution_binning (`bool` defaults to `True`):
+ If set to `True`, the requested height and width are first mapped to the closest resolutions using
+ `ASPECT_RATIO_1024_BIN`. After the produced latents are decoded into images, they are resized back to
+ the requested resolution. Useful for generating non-square images.
+ max_sequence_length (`int` defaults to 120): Maximum sequence length to use with the `prompt`.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images
+ """
+ if "mask_feature" in kwargs:
+ deprecation_message = "The use of `mask_feature` is deprecated. It is no longer used in any computation and that doesn't affect the end results. It will be removed in a future version."
+ deprecate("mask_feature", "1.0.0", deprecation_message, standard_warn=False)
+ # 1. Check inputs. Raise error if not correct
+ height = height or self.transformer.config.sample_size * self.vae_scale_factor
+ width = width or self.transformer.config.sample_size * self.vae_scale_factor
+ if use_resolution_binning:
+ if self.transformer.config.sample_size == 128:
+ aspect_ratio_bin = ASPECT_RATIO_1024_BIN
+ elif self.transformer.config.sample_size == 64:
+ aspect_ratio_bin = ASPECT_RATIO_512_BIN
+ elif self.transformer.config.sample_size == 32:
+ aspect_ratio_bin = ASPECT_RATIO_256_BIN
+ else:
+ raise ValueError("Invalid sample size")
+ orig_height, orig_width = height, width
+ height, width = self.classify_height_width_bin(height, width, ratios=aspect_ratio_bin)
+
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ negative_prompt,
+ callback_steps,
+ prompt_embeds,
+ negative_prompt_embeds,
+ prompt_attention_mask,
+ negative_prompt_attention_mask,
+ )
+
+ # 2. Default height and width to transformer
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ (
+ prompt_embeds,
+ prompt_attention_mask,
+ negative_prompt_embeds,
+ negative_prompt_attention_mask,
+ ) = self.encode_prompt(
+ prompt,
+ do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ prompt_attention_mask=prompt_attention_mask,
+ negative_prompt_attention_mask=negative_prompt_attention_mask,
+ clean_caption=clean_caption,
+ max_sequence_length=max_sequence_length,
+ )
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ prompt_attention_mask = torch.cat([negative_prompt_attention_mask, prompt_attention_mask], dim=0)
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+ # 5. Prepare latents.
+ latent_channels = self.transformer.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ latent_channels,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 6.1 Prepare micro-conditions.
+ added_cond_kwargs = {"resolution": None, "aspect_ratio": None}
+ if self.transformer.config.sample_size == 128:
+ resolution = torch.tensor([height, width]).repeat(batch_size * num_images_per_prompt, 1)
+ aspect_ratio = torch.tensor([float(height / width)]).repeat(batch_size * num_images_per_prompt, 1)
+ resolution = resolution.to(dtype=prompt_embeds.dtype, device=device)
+ aspect_ratio = aspect_ratio.to(dtype=prompt_embeds.dtype, device=device)
+
+ if do_classifier_free_guidance:
+ resolution = torch.cat([resolution, resolution], dim=0)
+ aspect_ratio = torch.cat([aspect_ratio, aspect_ratio], dim=0)
+
+ added_cond_kwargs = {"resolution": resolution, "aspect_ratio": aspect_ratio}
+
+ # 7. Denoising loop
+ num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ current_timestep = t
+ if not torch.is_tensor(current_timestep):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = latent_model_input.device.type == "mps"
+ if isinstance(current_timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ current_timestep = torch.tensor([current_timestep], dtype=dtype, device=latent_model_input.device)
+ elif len(current_timestep.shape) == 0:
+ current_timestep = current_timestep[None].to(latent_model_input.device)
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ current_timestep = current_timestep.expand(latent_model_input.shape[0])
+
+ # predict noise model_output
+ noise_pred = self.transformer(
+ latent_model_input,
+ encoder_hidden_states=prompt_embeds,
+ encoder_attention_mask=prompt_attention_mask,
+ timestep=current_timestep,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # learned sigma
+ if self.transformer.config.out_channels // 2 == latent_channels:
+ noise_pred = noise_pred.chunk(2, dim=1)[0]
+ else:
+ noise_pred = noise_pred
+
+ # compute previous image: x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ if use_resolution_binning:
+ image = self.resize_and_crop_tensor(image, orig_width, orig_height)
+ else:
+ image = latents
+
+ if not output_type == "latent":
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/semantic_stable_diffusion/__init__.py b/model/diffusers_c/pipelines/semantic_stable_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..70f5b1a547c4b90e28109843ae3be2fca2e98c88
--- /dev/null
+++ b/model/diffusers_c/pipelines/semantic_stable_diffusion/__init__.py
@@ -0,0 +1,49 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_output"] = ["SemanticStableDiffusionPipelineOutput"]
+ _import_structure["pipeline_semantic_stable_diffusion"] = ["SemanticStableDiffusionPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_semantic_stable_diffusion import SemanticStableDiffusionPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..125cfef3bd732c183c152b4534db639d99a79097
Binary files /dev/null and b/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/pipeline_output.cpython-39.pyc b/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/pipeline_output.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b8106473ffb64b4fbe863894a75c9af68520ca7d
Binary files /dev/null and b/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/pipeline_output.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/pipeline_semantic_stable_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/pipeline_semantic_stable_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..91beaf355472eb0cb32bd74af1b0ef94b0bb63a0
Binary files /dev/null and b/model/diffusers_c/pipelines/semantic_stable_diffusion/__pycache__/pipeline_semantic_stable_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/semantic_stable_diffusion/pipeline_output.py b/model/diffusers_c/pipelines/semantic_stable_diffusion/pipeline_output.py
new file mode 100644
index 0000000000000000000000000000000000000000..34991299398115f439537b77e1f1fc8a83e0d431
--- /dev/null
+++ b/model/diffusers_c/pipelines/semantic_stable_diffusion/pipeline_output.py
@@ -0,0 +1,25 @@
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL.Image
+
+from ...utils import BaseOutput
+
+
+@dataclass
+class SemanticStableDiffusionPipelineOutput(BaseOutput):
+ """
+ Output class for Stable Diffusion pipelines.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ nsfw_content_detected (`List[bool]`)
+ List indicating whether the corresponding generated image contains “not-safe-for-work” (nsfw) content or
+ `None` if safety checking could not be performed.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+ nsfw_content_detected: Optional[List[bool]]
diff --git a/model/diffusers_c/pipelines/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py b/model/diffusers_c/pipelines/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..96873423faeb3f66a001535d8d3a618abec8bed0
--- /dev/null
+++ b/model/diffusers_c/pipelines/semantic_stable_diffusion/pipeline_semantic_stable_diffusion.py
@@ -0,0 +1,718 @@
+import inspect
+from itertools import repeat
+from typing import Callable, List, Optional, Union
+
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import deprecate, logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import SemanticStableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class SemanticStableDiffusionPipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with latent editing.
+
+ This model inherits from [`DiffusionPipeline`] and builds on the [`StableDiffusionPipeline`]. Check the superclass
+ documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular
+ device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`Q16SafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: int = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ editing_prompt: Optional[Union[str, List[str]]] = None,
+ editing_prompt_embeddings: Optional[torch.Tensor] = None,
+ reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
+ edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
+ edit_warmup_steps: Optional[Union[int, List[int]]] = 10,
+ edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
+ edit_threshold: Optional[Union[float, List[float]]] = 0.9,
+ edit_momentum_scale: Optional[float] = 0.1,
+ edit_mom_beta: Optional[float] = 0.4,
+ edit_weights: Optional[List[float]] = None,
+ sem_guidance: Optional[List[torch.Tensor]] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ editing_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to use for semantic guidance. Semantic guidance is disabled by setting
+ `editing_prompt = None`. Guidance direction of prompt should be specified via
+ `reverse_editing_direction`.
+ editing_prompt_embeddings (`torch.Tensor`, *optional*):
+ Pre-computed embeddings to use for semantic guidance. Guidance direction of embedding should be
+ specified via `reverse_editing_direction`.
+ reverse_editing_direction (`bool` or `List[bool]`, *optional*, defaults to `False`):
+ Whether the corresponding prompt in `editing_prompt` should be increased or decreased.
+ edit_guidance_scale (`float` or `List[float]`, *optional*, defaults to 5):
+ Guidance scale for semantic guidance. If provided as a list, values should correspond to
+ `editing_prompt`.
+ edit_warmup_steps (`float` or `List[float]`, *optional*, defaults to 10):
+ Number of diffusion steps (for each prompt) for which semantic guidance is not applied. Momentum is
+ calculated for those steps and applied once all warmup periods are over.
+ edit_cooldown_steps (`float` or `List[float]`, *optional*, defaults to `None`):
+ Number of diffusion steps (for each prompt) after which semantic guidance is longer applied.
+ edit_threshold (`float` or `List[float]`, *optional*, defaults to 0.9):
+ Threshold of semantic guidance.
+ edit_momentum_scale (`float`, *optional*, defaults to 0.1):
+ Scale of the momentum to be added to the semantic guidance at each diffusion step. If set to 0.0,
+ momentum is disabled. Momentum is already built up during warmup (for diffusion steps smaller than
+ `sld_warmup_steps`). Momentum is only added to latent guidance once all warmup periods are finished.
+ edit_mom_beta (`float`, *optional*, defaults to 0.4):
+ Defines how semantic guidance momentum builds up. `edit_mom_beta` indicates how much of the previous
+ momentum is kept. Momentum is already built up during warmup (for diffusion steps smaller than
+ `edit_warmup_steps`).
+ edit_weights (`List[float]`, *optional*, defaults to `None`):
+ Indicates how much each individual concept should influence the overall guidance. If no weights are
+ provided all concepts are applied equally.
+ sem_guidance (`List[torch.Tensor]`, *optional*):
+ List of pre-generated guidance vectors to be applied at generation. Length of the list has to
+ correspond to `num_inference_steps`.
+
+ Examples:
+
+ ```py
+ >>> import torch
+ >>> from diffusers import SemanticStableDiffusionPipeline
+
+ >>> pipe = SemanticStableDiffusionPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> out = pipe(
+ ... prompt="a photo of the face of a woman",
+ ... num_images_per_prompt=1,
+ ... guidance_scale=7,
+ ... editing_prompt=[
+ ... "smiling, smile", # Concepts to apply
+ ... "glasses, wearing glasses",
+ ... "curls, wavy hair, curly hair",
+ ... "beard, full beard, mustache",
+ ... ],
+ ... reverse_editing_direction=[
+ ... False,
+ ... False,
+ ... False,
+ ... False,
+ ... ], # Direction of guidance i.e. increase all concepts
+ ... edit_warmup_steps=[10, 10, 10, 10], # Warmup period for each concept
+ ... edit_guidance_scale=[4, 5, 5, 5.4], # Guidance scale for each concept
+ ... edit_threshold=[
+ ... 0.99,
+ ... 0.975,
+ ... 0.925,
+ ... 0.96,
+ ... ], # Threshold for each concept. Threshold equals the percentile of the latent space that will be discarded. I.e. threshold=0.99 uses 1% of the latent dimensions
+ ... edit_momentum_scale=0.3, # Momentum scale that will be added to the latent guidance
+ ... edit_mom_beta=0.6, # Momentum beta
+ ... edit_weights=[1, 1, 1, 1, 1], # Weights of the individual concepts against each other
+ ... )
+ >>> image = out.images[0]
+ ```
+
+ Returns:
+ [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`,
+ [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] is returned, otherwise a
+ `tuple` is returned where the first element is a list with the generated images and the second element
+ is a list of `bool`s indicating whether the corresponding generated image contains "not-safe-for-work"
+ (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(prompt, height, width, callback_steps)
+
+ # 2. Define call parameters
+ batch_size = 1 if isinstance(prompt, str) else len(prompt)
+
+ if editing_prompt:
+ enable_edit_guidance = True
+ if isinstance(editing_prompt, str):
+ editing_prompt = [editing_prompt]
+ enabled_editing_prompts = len(editing_prompt)
+ elif editing_prompt_embeddings is not None:
+ enable_edit_guidance = True
+ enabled_editing_prompts = editing_prompt_embeddings.shape[0]
+ else:
+ enabled_editing_prompts = 0
+ enable_edit_guidance = False
+
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+
+ if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
+ removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+ text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0]
+
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = text_embeddings.shape
+ text_embeddings = text_embeddings.repeat(1, num_images_per_prompt, 1)
+ text_embeddings = text_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if enable_edit_guidance:
+ # get safety text embeddings
+ if editing_prompt_embeddings is None:
+ edit_concepts_input = self.tokenizer(
+ [x for item in editing_prompt for x in repeat(item, batch_size)],
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ return_tensors="pt",
+ )
+
+ edit_concepts_input_ids = edit_concepts_input.input_ids
+
+ if edit_concepts_input_ids.shape[-1] > self.tokenizer.model_max_length:
+ removed_text = self.tokenizer.batch_decode(
+ edit_concepts_input_ids[:, self.tokenizer.model_max_length :]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ edit_concepts_input_ids = edit_concepts_input_ids[:, : self.tokenizer.model_max_length]
+ edit_concepts = self.text_encoder(edit_concepts_input_ids.to(self.device))[0]
+ else:
+ edit_concepts = editing_prompt_embeddings.to(self.device).repeat(batch_size, 1, 1)
+
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ bs_embed_edit, seq_len_edit, _ = edit_concepts.shape
+ edit_concepts = edit_concepts.repeat(1, num_images_per_prompt, 1)
+ edit_concepts = edit_concepts.view(bs_embed_edit * num_images_per_prompt, seq_len_edit, -1)
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+ # get unconditional embeddings for classifier free guidance
+
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = text_input_ids.shape[-1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = uncond_embeddings.shape[1]
+ uncond_embeddings = uncond_embeddings.repeat(1, num_images_per_prompt, 1)
+ uncond_embeddings = uncond_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if enable_edit_guidance:
+ text_embeddings = torch.cat([uncond_embeddings, text_embeddings, edit_concepts])
+ else:
+ text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+ # get the initial random noise unless the user supplied it
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=self.device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ text_embeddings.dtype,
+ self.device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # Initialize edit_momentum to None
+ edit_momentum = None
+
+ self.uncond_estimates = None
+ self.text_estimates = None
+ self.edit_estimates = None
+ self.sem_guidance = None
+
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = (
+ torch.cat([latents] * (2 + enabled_editing_prompts)) if do_classifier_free_guidance else latents
+ )
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_out = noise_pred.chunk(2 + enabled_editing_prompts) # [b,4, 64, 64]
+ noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
+ noise_pred_edit_concepts = noise_pred_out[2:]
+
+ # default text guidance
+ noise_guidance = guidance_scale * (noise_pred_text - noise_pred_uncond)
+ # noise_guidance = (noise_pred_text - noise_pred_edit_concepts[0])
+
+ if self.uncond_estimates is None:
+ self.uncond_estimates = torch.zeros((num_inference_steps + 1, *noise_pred_uncond.shape))
+ self.uncond_estimates[i] = noise_pred_uncond.detach().cpu()
+
+ if self.text_estimates is None:
+ self.text_estimates = torch.zeros((num_inference_steps + 1, *noise_pred_text.shape))
+ self.text_estimates[i] = noise_pred_text.detach().cpu()
+
+ if self.edit_estimates is None and enable_edit_guidance:
+ self.edit_estimates = torch.zeros(
+ (num_inference_steps + 1, len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
+ )
+
+ if self.sem_guidance is None:
+ self.sem_guidance = torch.zeros((num_inference_steps + 1, *noise_pred_text.shape))
+
+ if edit_momentum is None:
+ edit_momentum = torch.zeros_like(noise_guidance)
+
+ if enable_edit_guidance:
+ concept_weights = torch.zeros(
+ (len(noise_pred_edit_concepts), noise_guidance.shape[0]),
+ device=self.device,
+ dtype=noise_guidance.dtype,
+ )
+ noise_guidance_edit = torch.zeros(
+ (len(noise_pred_edit_concepts), *noise_guidance.shape),
+ device=self.device,
+ dtype=noise_guidance.dtype,
+ )
+ # noise_guidance_edit = torch.zeros_like(noise_guidance)
+ warmup_inds = []
+ for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
+ self.edit_estimates[i, c] = noise_pred_edit_concept
+ if isinstance(edit_guidance_scale, list):
+ edit_guidance_scale_c = edit_guidance_scale[c]
+ else:
+ edit_guidance_scale_c = edit_guidance_scale
+
+ if isinstance(edit_threshold, list):
+ edit_threshold_c = edit_threshold[c]
+ else:
+ edit_threshold_c = edit_threshold
+ if isinstance(reverse_editing_direction, list):
+ reverse_editing_direction_c = reverse_editing_direction[c]
+ else:
+ reverse_editing_direction_c = reverse_editing_direction
+ if edit_weights:
+ edit_weight_c = edit_weights[c]
+ else:
+ edit_weight_c = 1.0
+ if isinstance(edit_warmup_steps, list):
+ edit_warmup_steps_c = edit_warmup_steps[c]
+ else:
+ edit_warmup_steps_c = edit_warmup_steps
+
+ if isinstance(edit_cooldown_steps, list):
+ edit_cooldown_steps_c = edit_cooldown_steps[c]
+ elif edit_cooldown_steps is None:
+ edit_cooldown_steps_c = i + 1
+ else:
+ edit_cooldown_steps_c = edit_cooldown_steps
+ if i >= edit_warmup_steps_c:
+ warmup_inds.append(c)
+ if i >= edit_cooldown_steps_c:
+ noise_guidance_edit[c, :, :, :, :] = torch.zeros_like(noise_pred_edit_concept)
+ continue
+
+ noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond
+ # tmp_weights = (noise_pred_text - noise_pred_edit_concept).sum(dim=(1, 2, 3))
+ tmp_weights = (noise_guidance - noise_pred_edit_concept).sum(dim=(1, 2, 3))
+
+ tmp_weights = torch.full_like(tmp_weights, edit_weight_c) # * (1 / enabled_editing_prompts)
+ if reverse_editing_direction_c:
+ noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1
+ concept_weights[c, :] = tmp_weights
+
+ noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c
+
+ # torch.quantile function expects float32
+ if noise_guidance_edit_tmp.dtype == torch.float32:
+ tmp = torch.quantile(
+ torch.abs(noise_guidance_edit_tmp).flatten(start_dim=2),
+ edit_threshold_c,
+ dim=2,
+ keepdim=False,
+ )
+ else:
+ tmp = torch.quantile(
+ torch.abs(noise_guidance_edit_tmp).flatten(start_dim=2).to(torch.float32),
+ edit_threshold_c,
+ dim=2,
+ keepdim=False,
+ ).to(noise_guidance_edit_tmp.dtype)
+
+ noise_guidance_edit_tmp = torch.where(
+ torch.abs(noise_guidance_edit_tmp) >= tmp[:, :, None, None],
+ noise_guidance_edit_tmp,
+ torch.zeros_like(noise_guidance_edit_tmp),
+ )
+ noise_guidance_edit[c, :, :, :, :] = noise_guidance_edit_tmp
+
+ # noise_guidance_edit = noise_guidance_edit + noise_guidance_edit_tmp
+
+ warmup_inds = torch.tensor(warmup_inds).to(self.device)
+ if len(noise_pred_edit_concepts) > warmup_inds.shape[0] > 0:
+ concept_weights = concept_weights.to("cpu") # Offload to cpu
+ noise_guidance_edit = noise_guidance_edit.to("cpu")
+
+ concept_weights_tmp = torch.index_select(concept_weights.to(self.device), 0, warmup_inds)
+ concept_weights_tmp = torch.where(
+ concept_weights_tmp < 0, torch.zeros_like(concept_weights_tmp), concept_weights_tmp
+ )
+ concept_weights_tmp = concept_weights_tmp / concept_weights_tmp.sum(dim=0)
+ # concept_weights_tmp = torch.nan_to_num(concept_weights_tmp)
+
+ noise_guidance_edit_tmp = torch.index_select(
+ noise_guidance_edit.to(self.device), 0, warmup_inds
+ )
+ noise_guidance_edit_tmp = torch.einsum(
+ "cb,cbijk->bijk", concept_weights_tmp, noise_guidance_edit_tmp
+ )
+ noise_guidance_edit_tmp = noise_guidance_edit_tmp
+ noise_guidance = noise_guidance + noise_guidance_edit_tmp
+
+ self.sem_guidance[i] = noise_guidance_edit_tmp.detach().cpu()
+
+ del noise_guidance_edit_tmp
+ del concept_weights_tmp
+ concept_weights = concept_weights.to(self.device)
+ noise_guidance_edit = noise_guidance_edit.to(self.device)
+
+ concept_weights = torch.where(
+ concept_weights < 0, torch.zeros_like(concept_weights), concept_weights
+ )
+
+ concept_weights = torch.nan_to_num(concept_weights)
+
+ noise_guidance_edit = torch.einsum("cb,cbijk->bijk", concept_weights, noise_guidance_edit)
+
+ noise_guidance_edit = noise_guidance_edit + edit_momentum_scale * edit_momentum
+
+ edit_momentum = edit_mom_beta * edit_momentum + (1 - edit_mom_beta) * noise_guidance_edit
+
+ if warmup_inds.shape[0] == len(noise_pred_edit_concepts):
+ noise_guidance = noise_guidance + noise_guidance_edit
+ self.sem_guidance[i] = noise_guidance_edit.detach().cpu()
+
+ if sem_guidance is not None:
+ edit_guidance = sem_guidance[i].to(self.device)
+ noise_guidance = noise_guidance + edit_guidance
+
+ noise_pred = noise_pred_uncond + noise_guidance
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 8. Post-processing
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, self.device, text_embeddings.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return SemanticStableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/shap_e/__init__.py b/model/diffusers_c/pipelines/shap_e/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..4ed563c4a51f6e627c06711b60fe3a0709ff22f7
--- /dev/null
+++ b/model/diffusers_c/pipelines/shap_e/__init__.py
@@ -0,0 +1,71 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["camera"] = ["create_pan_cameras"]
+ _import_structure["pipeline_shap_e"] = ["ShapEPipeline"]
+ _import_structure["pipeline_shap_e_img2img"] = ["ShapEImg2ImgPipeline"]
+ _import_structure["renderer"] = [
+ "BoundingBoxVolume",
+ "ImportanceRaySampler",
+ "MLPNeRFModelOutput",
+ "MLPNeRSTFModel",
+ "ShapEParamsProjModel",
+ "ShapERenderer",
+ "StratifiedRaySampler",
+ "VoidNeRFModel",
+ ]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .camera import create_pan_cameras
+ from .pipeline_shap_e import ShapEPipeline
+ from .pipeline_shap_e_img2img import ShapEImg2ImgPipeline
+ from .renderer import (
+ BoundingBoxVolume,
+ ImportanceRaySampler,
+ MLPNeRFModelOutput,
+ MLPNeRSTFModel,
+ ShapEParamsProjModel,
+ ShapERenderer,
+ StratifiedRaySampler,
+ VoidNeRFModel,
+ )
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/shap_e/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/shap_e/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..290dce4c696d195c95e9eddfb28151a6d169f4fa
Binary files /dev/null and b/model/diffusers_c/pipelines/shap_e/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/shap_e/__pycache__/camera.cpython-39.pyc b/model/diffusers_c/pipelines/shap_e/__pycache__/camera.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0e7f49ac86fc1a530ba06049de4491fac72b415b
Binary files /dev/null and b/model/diffusers_c/pipelines/shap_e/__pycache__/camera.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/shap_e/__pycache__/pipeline_shap_e.cpython-39.pyc b/model/diffusers_c/pipelines/shap_e/__pycache__/pipeline_shap_e.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ed3f468ac6d03e219398469d6b2f99f209a0868f
Binary files /dev/null and b/model/diffusers_c/pipelines/shap_e/__pycache__/pipeline_shap_e.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/shap_e/__pycache__/pipeline_shap_e_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/shap_e/__pycache__/pipeline_shap_e_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..990ec10c4e7b37ab50114feba03e75fb12e9fcf9
Binary files /dev/null and b/model/diffusers_c/pipelines/shap_e/__pycache__/pipeline_shap_e_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/shap_e/__pycache__/renderer.cpython-39.pyc b/model/diffusers_c/pipelines/shap_e/__pycache__/renderer.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4266daf648a9d933bc3375abd17c221ab7942e1b
Binary files /dev/null and b/model/diffusers_c/pipelines/shap_e/__pycache__/renderer.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/shap_e/camera.py b/model/diffusers_c/pipelines/shap_e/camera.py
new file mode 100644
index 0000000000000000000000000000000000000000..d4b94c3000d845dedf1dc3e45fa217b9071c5f38
--- /dev/null
+++ b/model/diffusers_c/pipelines/shap_e/camera.py
@@ -0,0 +1,147 @@
+# Copyright 2024 Open AI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Tuple
+
+import numpy as np
+import torch
+
+
+@dataclass
+class DifferentiableProjectiveCamera:
+ """
+ Implements a batch, differentiable, standard pinhole camera
+ """
+
+ origin: torch.Tensor # [batch_size x 3]
+ x: torch.Tensor # [batch_size x 3]
+ y: torch.Tensor # [batch_size x 3]
+ z: torch.Tensor # [batch_size x 3]
+ width: int
+ height: int
+ x_fov: float
+ y_fov: float
+ shape: Tuple[int]
+
+ def __post_init__(self):
+ assert self.x.shape[0] == self.y.shape[0] == self.z.shape[0] == self.origin.shape[0]
+ assert self.x.shape[1] == self.y.shape[1] == self.z.shape[1] == self.origin.shape[1] == 3
+ assert len(self.x.shape) == len(self.y.shape) == len(self.z.shape) == len(self.origin.shape) == 2
+
+ def resolution(self):
+ return torch.from_numpy(np.array([self.width, self.height], dtype=np.float32))
+
+ def fov(self):
+ return torch.from_numpy(np.array([self.x_fov, self.y_fov], dtype=np.float32))
+
+ def get_image_coords(self) -> torch.Tensor:
+ """
+ :return: coords of shape (width * height, 2)
+ """
+ pixel_indices = torch.arange(self.height * self.width)
+ coords = torch.stack(
+ [
+ pixel_indices % self.width,
+ torch.div(pixel_indices, self.width, rounding_mode="trunc"),
+ ],
+ axis=1,
+ )
+ return coords
+
+ @property
+ def camera_rays(self):
+ batch_size, *inner_shape = self.shape
+ inner_batch_size = int(np.prod(inner_shape))
+
+ coords = self.get_image_coords()
+ coords = torch.broadcast_to(coords.unsqueeze(0), [batch_size * inner_batch_size, *coords.shape])
+ rays = self.get_camera_rays(coords)
+
+ rays = rays.view(batch_size, inner_batch_size * self.height * self.width, 2, 3)
+
+ return rays
+
+ def get_camera_rays(self, coords: torch.Tensor) -> torch.Tensor:
+ batch_size, *shape, n_coords = coords.shape
+ assert n_coords == 2
+ assert batch_size == self.origin.shape[0]
+
+ flat = coords.view(batch_size, -1, 2)
+
+ res = self.resolution()
+ fov = self.fov()
+
+ fracs = (flat.float() / (res - 1)) * 2 - 1
+ fracs = fracs * torch.tan(fov / 2)
+
+ fracs = fracs.view(batch_size, -1, 2)
+ directions = (
+ self.z.view(batch_size, 1, 3)
+ + self.x.view(batch_size, 1, 3) * fracs[:, :, :1]
+ + self.y.view(batch_size, 1, 3) * fracs[:, :, 1:]
+ )
+ directions = directions / directions.norm(dim=-1, keepdim=True)
+ rays = torch.stack(
+ [
+ torch.broadcast_to(self.origin.view(batch_size, 1, 3), [batch_size, directions.shape[1], 3]),
+ directions,
+ ],
+ dim=2,
+ )
+ return rays.view(batch_size, *shape, 2, 3)
+
+ def resize_image(self, width: int, height: int) -> "DifferentiableProjectiveCamera":
+ """
+ Creates a new camera for the resized view assuming the aspect ratio does not change.
+ """
+ assert width * self.height == height * self.width, "The aspect ratio should not change."
+ return DifferentiableProjectiveCamera(
+ origin=self.origin,
+ x=self.x,
+ y=self.y,
+ z=self.z,
+ width=width,
+ height=height,
+ x_fov=self.x_fov,
+ y_fov=self.y_fov,
+ )
+
+
+def create_pan_cameras(size: int) -> DifferentiableProjectiveCamera:
+ origins = []
+ xs = []
+ ys = []
+ zs = []
+ for theta in np.linspace(0, 2 * np.pi, num=20):
+ z = np.array([np.sin(theta), np.cos(theta), -0.5])
+ z /= np.sqrt(np.sum(z**2))
+ origin = -z * 4
+ x = np.array([np.cos(theta), -np.sin(theta), 0.0])
+ y = np.cross(z, x)
+ origins.append(origin)
+ xs.append(x)
+ ys.append(y)
+ zs.append(z)
+ return DifferentiableProjectiveCamera(
+ origin=torch.from_numpy(np.stack(origins, axis=0)).float(),
+ x=torch.from_numpy(np.stack(xs, axis=0)).float(),
+ y=torch.from_numpy(np.stack(ys, axis=0)).float(),
+ z=torch.from_numpy(np.stack(zs, axis=0)).float(),
+ width=size,
+ height=size,
+ x_fov=0.7,
+ y_fov=0.7,
+ shape=(1, len(xs)),
+ )
diff --git a/model/diffusers_c/pipelines/shap_e/pipeline_shap_e.py b/model/diffusers_c/pipelines/shap_e/pipeline_shap_e.py
new file mode 100644
index 0000000000000000000000000000000000000000..1ef10e17cbd749eecd4f74a447cb1fe9d91fdc77
--- /dev/null
+++ b/model/diffusers_c/pipelines/shap_e/pipeline_shap_e.py
@@ -0,0 +1,334 @@
+# Copyright 2024 Open AI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...models import PriorTransformer
+from ...schedulers import HeunDiscreteScheduler
+from ...utils import (
+ BaseOutput,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .renderer import ShapERenderer
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import DiffusionPipeline
+ >>> from diffusers.utils import export_to_gif
+
+ >>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+ >>> repo = "openai/shap-e"
+ >>> pipe = DiffusionPipeline.from_pretrained(repo, torch_dtype=torch.float16)
+ >>> pipe = pipe.to(device)
+
+ >>> guidance_scale = 15.0
+ >>> prompt = "a shark"
+
+ >>> images = pipe(
+ ... prompt,
+ ... guidance_scale=guidance_scale,
+ ... num_inference_steps=64,
+ ... frame_size=256,
+ ... ).images
+
+ >>> gif_path = export_to_gif(images[0], "shark_3d.gif")
+ ```
+"""
+
+
+@dataclass
+class ShapEPipelineOutput(BaseOutput):
+ """
+ Output class for [`ShapEPipeline`] and [`ShapEImg2ImgPipeline`].
+
+ Args:
+ images (`torch.FloatTensor`)
+ A list of images for 3D rendering.
+ """
+
+ images: Union[List[List[PIL.Image.Image]], List[List[np.ndarray]]]
+
+
+class ShapEPipeline(DiffusionPipeline):
+ """
+ Pipeline for generating latent representation of a 3D asset and rendering with the NeRF method.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ prior ([`PriorTransformer`]):
+ The canonical unCLIP prior to approximate the image embedding from the text embedding.
+ text_encoder ([`~transformers.CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ scheduler ([`HeunDiscreteScheduler`]):
+ A scheduler to be used in combination with the `prior` model to generate image embedding.
+ shap_e_renderer ([`ShapERenderer`]):
+ Shap-E renderer projects the generated latents into parameters of a MLP to create 3D objects with the NeRF
+ rendering method.
+ """
+
+ model_cpu_offload_seq = "text_encoder->prior"
+ _exclude_from_cpu_offload = ["shap_e_renderer"]
+
+ def __init__(
+ self,
+ prior: PriorTransformer,
+ text_encoder: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ scheduler: HeunDiscreteScheduler,
+ shap_e_renderer: ShapERenderer,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ prior=prior,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ scheduler=scheduler,
+ shap_e_renderer=shap_e_renderer,
+ )
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ ):
+ len(prompt) if isinstance(prompt, list) else 1
+
+ # YiYi Notes: set pad_token_id to be 0, not sure why I can't set in the config file
+ self.tokenizer.pad_token_id = 0
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ text_encoder_output = self.text_encoder(text_input_ids.to(device))
+ prompt_embeds = text_encoder_output.text_embeds
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ # in Shap-E it normalize the prompt_embeds and then later rescale it
+ prompt_embeds = prompt_embeds / torch.linalg.norm(prompt_embeds, dim=-1, keepdim=True)
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # Rescale the features to have unit variance
+ prompt_embeds = math.sqrt(prompt_embeds.shape[1]) * prompt_embeds
+
+ return prompt_embeds
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: str,
+ num_images_per_prompt: int = 1,
+ num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ guidance_scale: float = 4.0,
+ frame_size: int = 64,
+ output_type: Optional[str] = "pil", # pil, np, latent, mesh
+ return_dict: bool = True,
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ frame_size (`int`, *optional*, default to 64):
+ The width and height of each image frame of the generated 3D output.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`), `"latent"` (`torch.Tensor`), or mesh ([`MeshDecoderOutput`]).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] instead of a plain
+ tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images.
+ """
+
+ if isinstance(prompt, str):
+ batch_size = 1
+ elif isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ device = self._execution_device
+
+ batch_size = batch_size * num_images_per_prompt
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+ prompt_embeds = self._encode_prompt(prompt, device, num_images_per_prompt, do_classifier_free_guidance)
+
+ # prior
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ num_embeddings = self.prior.config.num_embeddings
+ embedding_dim = self.prior.config.embedding_dim
+
+ latents = self.prepare_latents(
+ (batch_size, num_embeddings * embedding_dim),
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+
+ # YiYi notes: for testing only to match ldm, we can directly create a latents with desired shape: batch_size, num_embeddings, embedding_dim
+ latents = latents.reshape(latents.shape[0], num_embeddings, embedding_dim)
+
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ scaled_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ noise_pred = self.prior(
+ scaled_model_input,
+ timestep=t,
+ proj_embedding=prompt_embeds,
+ ).predicted_image_embedding
+
+ # remove the variance
+ noise_pred, _ = noise_pred.split(
+ scaled_model_input.shape[2], dim=2
+ ) # batch_size, num_embeddings, embedding_dim
+
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred - noise_pred_uncond)
+
+ latents = self.scheduler.step(
+ noise_pred,
+ timestep=t,
+ sample=latents,
+ ).prev_sample
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if output_type not in ["np", "pil", "latent", "mesh"]:
+ raise ValueError(
+ f"Only the output types `pil`, `np`, `latent` and `mesh` are supported not output_type={output_type}"
+ )
+
+ if output_type == "latent":
+ return ShapEPipelineOutput(images=latents)
+
+ images = []
+ if output_type == "mesh":
+ for i, latent in enumerate(latents):
+ mesh = self.shap_e_renderer.decode_to_mesh(
+ latent[None, :],
+ device,
+ )
+ images.append(mesh)
+
+ else:
+ # np, pil
+ for i, latent in enumerate(latents):
+ image = self.shap_e_renderer.decode_to_image(
+ latent[None, :],
+ device,
+ size=frame_size,
+ )
+ images.append(image)
+
+ images = torch.stack(images)
+
+ images = images.cpu().numpy()
+
+ if output_type == "pil":
+ images = [self.numpy_to_pil(image) for image in images]
+
+ if not return_dict:
+ return (images,)
+
+ return ShapEPipelineOutput(images=images)
diff --git a/model/diffusers_c/pipelines/shap_e/pipeline_shap_e_img2img.py b/model/diffusers_c/pipelines/shap_e/pipeline_shap_e_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..641ec56a1596f7b134fae32d5207eec0adcbaab4
--- /dev/null
+++ b/model/diffusers_c/pipelines/shap_e/pipeline_shap_e_img2img.py
@@ -0,0 +1,321 @@
+# Copyright 2024 Open AI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPVisionModel
+
+from ...models import PriorTransformer
+from ...schedulers import HeunDiscreteScheduler
+from ...utils import (
+ BaseOutput,
+ logging,
+ replace_example_docstring,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .renderer import ShapERenderer
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from PIL import Image
+ >>> import torch
+ >>> from diffusers import DiffusionPipeline
+ >>> from diffusers.utils import export_to_gif, load_image
+
+ >>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+ >>> repo = "openai/shap-e-img2img"
+ >>> pipe = DiffusionPipeline.from_pretrained(repo, torch_dtype=torch.float16)
+ >>> pipe = pipe.to(device)
+
+ >>> guidance_scale = 3.0
+ >>> image_url = "https://hf.co/datasets/diffusers/docs-images/resolve/main/shap-e/corgi.png"
+ >>> image = load_image(image_url).convert("RGB")
+
+ >>> images = pipe(
+ ... image,
+ ... guidance_scale=guidance_scale,
+ ... num_inference_steps=64,
+ ... frame_size=256,
+ ... ).images
+
+ >>> gif_path = export_to_gif(images[0], "corgi_3d.gif")
+ ```
+"""
+
+
+@dataclass
+class ShapEPipelineOutput(BaseOutput):
+ """
+ Output class for [`ShapEPipeline`] and [`ShapEImg2ImgPipeline`].
+
+ Args:
+ images (`torch.FloatTensor`)
+ A list of images for 3D rendering.
+ """
+
+ images: Union[PIL.Image.Image, np.ndarray]
+
+
+class ShapEImg2ImgPipeline(DiffusionPipeline):
+ """
+ Pipeline for generating latent representation of a 3D asset and rendering with the NeRF method from an image.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ image_encoder ([`~transformers.CLIPVisionModel`]):
+ Frozen image-encoder.
+ image_processor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to process images.
+ scheduler ([`HeunDiscreteScheduler`]):
+ A scheduler to be used in combination with the `prior` model to generate image embedding.
+ shap_e_renderer ([`ShapERenderer`]):
+ Shap-E renderer projects the generated latents into parameters of a MLP to create 3D objects with the NeRF
+ rendering method.
+ """
+
+ model_cpu_offload_seq = "image_encoder->prior"
+ _exclude_from_cpu_offload = ["shap_e_renderer"]
+
+ def __init__(
+ self,
+ prior: PriorTransformer,
+ image_encoder: CLIPVisionModel,
+ image_processor: CLIPImageProcessor,
+ scheduler: HeunDiscreteScheduler,
+ shap_e_renderer: ShapERenderer,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ prior=prior,
+ image_encoder=image_encoder,
+ image_processor=image_processor,
+ scheduler=scheduler,
+ shap_e_renderer=shap_e_renderer,
+ )
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def _encode_image(
+ self,
+ image,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ ):
+ if isinstance(image, List) and isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, axis=0) if image[0].ndim == 4 else torch.stack(image, axis=0)
+
+ if not isinstance(image, torch.Tensor):
+ image = self.image_processor(image, return_tensors="pt").pixel_values[0].unsqueeze(0)
+
+ image = image.to(dtype=self.image_encoder.dtype, device=device)
+
+ image_embeds = self.image_encoder(image)["last_hidden_state"]
+ image_embeds = image_embeds[:, 1:, :].contiguous() # batch_size, dim, 256
+
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ negative_image_embeds = torch.zeros_like(image_embeds)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
+ return image_embeds
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image: Union[PIL.Image.Image, List[PIL.Image.Image]],
+ num_images_per_prompt: int = 1,
+ num_inference_steps: int = 25,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ guidance_scale: float = 4.0,
+ frame_size: int = 64,
+ output_type: Optional[str] = "pil", # pil, np, latent, mesh
+ return_dict: bool = True,
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image` or tensor representing an image batch to be used as the starting point. Can also accept image
+ latents as image, but if passing latents directly it is not encoded again.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ guidance_scale (`float`, *optional*, defaults to 4.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ frame_size (`int`, *optional*, default to 64):
+ The width and height of each image frame of the generated 3D output.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`), `"latent"` (`torch.Tensor`), or mesh ([`MeshDecoderOutput`]).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] instead of a plain
+ tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.shap_e.pipeline_shap_e.ShapEPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images.
+ """
+
+ if isinstance(image, PIL.Image.Image):
+ batch_size = 1
+ elif isinstance(image, torch.Tensor):
+ batch_size = image.shape[0]
+ elif isinstance(image, list) and isinstance(image[0], (torch.Tensor, PIL.Image.Image)):
+ batch_size = len(image)
+ else:
+ raise ValueError(
+ f"`image` has to be of type `PIL.Image.Image`, `torch.Tensor`, `List[PIL.Image.Image]` or `List[torch.Tensor]` but is {type(image)}"
+ )
+
+ device = self._execution_device
+
+ batch_size = batch_size * num_images_per_prompt
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+ image_embeds = self._encode_image(image, device, num_images_per_prompt, do_classifier_free_guidance)
+
+ # prior
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ num_embeddings = self.prior.config.num_embeddings
+ embedding_dim = self.prior.config.embedding_dim
+
+ latents = self.prepare_latents(
+ (batch_size, num_embeddings * embedding_dim),
+ image_embeds.dtype,
+ device,
+ generator,
+ latents,
+ self.scheduler,
+ )
+
+ # YiYi notes: for testing only to match ldm, we can directly create a latents with desired shape: batch_size, num_embeddings, embedding_dim
+ latents = latents.reshape(latents.shape[0], num_embeddings, embedding_dim)
+
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ scaled_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ noise_pred = self.prior(
+ scaled_model_input,
+ timestep=t,
+ proj_embedding=image_embeds,
+ ).predicted_image_embedding
+
+ # remove the variance
+ noise_pred, _ = noise_pred.split(
+ scaled_model_input.shape[2], dim=2
+ ) # batch_size, num_embeddings, embedding_dim
+
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred - noise_pred_uncond)
+
+ latents = self.scheduler.step(
+ noise_pred,
+ timestep=t,
+ sample=latents,
+ ).prev_sample
+
+ if output_type not in ["np", "pil", "latent", "mesh"]:
+ raise ValueError(
+ f"Only the output types `pil`, `np`, `latent` and `mesh` are supported not output_type={output_type}"
+ )
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if output_type == "latent":
+ return ShapEPipelineOutput(images=latents)
+
+ images = []
+ if output_type == "mesh":
+ for i, latent in enumerate(latents):
+ mesh = self.shap_e_renderer.decode_to_mesh(
+ latent[None, :],
+ device,
+ )
+ images.append(mesh)
+
+ else:
+ # np, pil
+ for i, latent in enumerate(latents):
+ image = self.shap_e_renderer.decode_to_image(
+ latent[None, :],
+ device,
+ size=frame_size,
+ )
+ images.append(image)
+
+ images = torch.stack(images)
+
+ images = images.cpu().numpy()
+
+ if output_type == "pil":
+ images = [self.numpy_to_pil(image) for image in images]
+
+ if not return_dict:
+ return (images,)
+
+ return ShapEPipelineOutput(images=images)
diff --git a/model/diffusers_c/pipelines/shap_e/renderer.py b/model/diffusers_c/pipelines/shap_e/renderer.py
new file mode 100644
index 0000000000000000000000000000000000000000..047c6f7dd863c706e05fd4963265c6eaffeea978
--- /dev/null
+++ b/model/diffusers_c/pipelines/shap_e/renderer.py
@@ -0,0 +1,1050 @@
+# Copyright 2024 Open AI and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models import ModelMixin
+from ...utils import BaseOutput
+from .camera import create_pan_cameras
+
+
+def sample_pmf(pmf: torch.Tensor, n_samples: int) -> torch.Tensor:
+ r"""
+ Sample from the given discrete probability distribution with replacement.
+
+ The i-th bin is assumed to have mass pmf[i].
+
+ Args:
+ pmf: [batch_size, *shape, n_samples, 1] where (pmf.sum(dim=-2) == 1).all()
+ n_samples: number of samples
+
+ Return:
+ indices sampled with replacement
+ """
+
+ *shape, support_size, last_dim = pmf.shape
+ assert last_dim == 1
+
+ cdf = torch.cumsum(pmf.view(-1, support_size), dim=1)
+ inds = torch.searchsorted(cdf, torch.rand(cdf.shape[0], n_samples, device=cdf.device))
+
+ return inds.view(*shape, n_samples, 1).clamp(0, support_size - 1)
+
+
+def posenc_nerf(x: torch.Tensor, min_deg: int = 0, max_deg: int = 15) -> torch.Tensor:
+ """
+ Concatenate x and its positional encodings, following NeRF.
+
+ Reference: https://arxiv.org/pdf/2210.04628.pdf
+ """
+ if min_deg == max_deg:
+ return x
+
+ scales = 2.0 ** torch.arange(min_deg, max_deg, dtype=x.dtype, device=x.device)
+ *shape, dim = x.shape
+ xb = (x.reshape(-1, 1, dim) * scales.view(1, -1, 1)).reshape(*shape, -1)
+ assert xb.shape[-1] == dim * (max_deg - min_deg)
+ emb = torch.cat([xb, xb + math.pi / 2.0], axis=-1).sin()
+ return torch.cat([x, emb], dim=-1)
+
+
+def encode_position(position):
+ return posenc_nerf(position, min_deg=0, max_deg=15)
+
+
+def encode_direction(position, direction=None):
+ if direction is None:
+ return torch.zeros_like(posenc_nerf(position, min_deg=0, max_deg=8))
+ else:
+ return posenc_nerf(direction, min_deg=0, max_deg=8)
+
+
+def _sanitize_name(x: str) -> str:
+ return x.replace(".", "__")
+
+
+def integrate_samples(volume_range, ts, density, channels):
+ r"""
+ Function integrating the model output.
+
+ Args:
+ volume_range: Specifies the integral range [t0, t1]
+ ts: timesteps
+ density: torch.Tensor [batch_size, *shape, n_samples, 1]
+ channels: torch.Tensor [batch_size, *shape, n_samples, n_channels]
+ returns:
+ channels: integrated rgb output weights: torch.Tensor [batch_size, *shape, n_samples, 1] (density
+ *transmittance)[i] weight for each rgb output at [..., i, :]. transmittance: transmittance of this volume
+ )
+ """
+
+ # 1. Calculate the weights
+ _, _, dt = volume_range.partition(ts)
+ ddensity = density * dt
+
+ mass = torch.cumsum(ddensity, dim=-2)
+ transmittance = torch.exp(-mass[..., -1, :])
+
+ alphas = 1.0 - torch.exp(-ddensity)
+ Ts = torch.exp(torch.cat([torch.zeros_like(mass[..., :1, :]), -mass[..., :-1, :]], dim=-2))
+ # This is the probability of light hitting and reflecting off of
+ # something at depth [..., i, :].
+ weights = alphas * Ts
+
+ # 2. Integrate channels
+ channels = torch.sum(channels * weights, dim=-2)
+
+ return channels, weights, transmittance
+
+
+def volume_query_points(volume, grid_size):
+ indices = torch.arange(grid_size**3, device=volume.bbox_min.device)
+ zs = indices % grid_size
+ ys = torch.div(indices, grid_size, rounding_mode="trunc") % grid_size
+ xs = torch.div(indices, grid_size**2, rounding_mode="trunc") % grid_size
+ combined = torch.stack([xs, ys, zs], dim=1)
+ return (combined.float() / (grid_size - 1)) * (volume.bbox_max - volume.bbox_min) + volume.bbox_min
+
+
+def _convert_srgb_to_linear(u: torch.Tensor):
+ return torch.where(u <= 0.04045, u / 12.92, ((u + 0.055) / 1.055) ** 2.4)
+
+
+def _create_flat_edge_indices(
+ flat_cube_indices: torch.Tensor,
+ grid_size: Tuple[int, int, int],
+):
+ num_xs = (grid_size[0] - 1) * grid_size[1] * grid_size[2]
+ y_offset = num_xs
+ num_ys = grid_size[0] * (grid_size[1] - 1) * grid_size[2]
+ z_offset = num_xs + num_ys
+ return torch.stack(
+ [
+ # Edges spanning x-axis.
+ flat_cube_indices[:, 0] * grid_size[1] * grid_size[2]
+ + flat_cube_indices[:, 1] * grid_size[2]
+ + flat_cube_indices[:, 2],
+ flat_cube_indices[:, 0] * grid_size[1] * grid_size[2]
+ + (flat_cube_indices[:, 1] + 1) * grid_size[2]
+ + flat_cube_indices[:, 2],
+ flat_cube_indices[:, 0] * grid_size[1] * grid_size[2]
+ + flat_cube_indices[:, 1] * grid_size[2]
+ + flat_cube_indices[:, 2]
+ + 1,
+ flat_cube_indices[:, 0] * grid_size[1] * grid_size[2]
+ + (flat_cube_indices[:, 1] + 1) * grid_size[2]
+ + flat_cube_indices[:, 2]
+ + 1,
+ # Edges spanning y-axis.
+ (
+ y_offset
+ + flat_cube_indices[:, 0] * (grid_size[1] - 1) * grid_size[2]
+ + flat_cube_indices[:, 1] * grid_size[2]
+ + flat_cube_indices[:, 2]
+ ),
+ (
+ y_offset
+ + (flat_cube_indices[:, 0] + 1) * (grid_size[1] - 1) * grid_size[2]
+ + flat_cube_indices[:, 1] * grid_size[2]
+ + flat_cube_indices[:, 2]
+ ),
+ (
+ y_offset
+ + flat_cube_indices[:, 0] * (grid_size[1] - 1) * grid_size[2]
+ + flat_cube_indices[:, 1] * grid_size[2]
+ + flat_cube_indices[:, 2]
+ + 1
+ ),
+ (
+ y_offset
+ + (flat_cube_indices[:, 0] + 1) * (grid_size[1] - 1) * grid_size[2]
+ + flat_cube_indices[:, 1] * grid_size[2]
+ + flat_cube_indices[:, 2]
+ + 1
+ ),
+ # Edges spanning z-axis.
+ (
+ z_offset
+ + flat_cube_indices[:, 0] * grid_size[1] * (grid_size[2] - 1)
+ + flat_cube_indices[:, 1] * (grid_size[2] - 1)
+ + flat_cube_indices[:, 2]
+ ),
+ (
+ z_offset
+ + (flat_cube_indices[:, 0] + 1) * grid_size[1] * (grid_size[2] - 1)
+ + flat_cube_indices[:, 1] * (grid_size[2] - 1)
+ + flat_cube_indices[:, 2]
+ ),
+ (
+ z_offset
+ + flat_cube_indices[:, 0] * grid_size[1] * (grid_size[2] - 1)
+ + (flat_cube_indices[:, 1] + 1) * (grid_size[2] - 1)
+ + flat_cube_indices[:, 2]
+ ),
+ (
+ z_offset
+ + (flat_cube_indices[:, 0] + 1) * grid_size[1] * (grid_size[2] - 1)
+ + (flat_cube_indices[:, 1] + 1) * (grid_size[2] - 1)
+ + flat_cube_indices[:, 2]
+ ),
+ ],
+ dim=-1,
+ )
+
+
+class VoidNeRFModel(nn.Module):
+ """
+ Implements the default empty space model where all queries are rendered as background.
+ """
+
+ def __init__(self, background, channel_scale=255.0):
+ super().__init__()
+ background = nn.Parameter(torch.from_numpy(np.array(background)).to(dtype=torch.float32) / channel_scale)
+
+ self.register_buffer("background", background)
+
+ def forward(self, position):
+ background = self.background[None].to(position.device)
+
+ shape = position.shape[:-1]
+ ones = [1] * (len(shape) - 1)
+ n_channels = background.shape[-1]
+ background = torch.broadcast_to(background.view(background.shape[0], *ones, n_channels), [*shape, n_channels])
+
+ return background
+
+
+@dataclass
+class VolumeRange:
+ t0: torch.Tensor
+ t1: torch.Tensor
+ intersected: torch.Tensor
+
+ def __post_init__(self):
+ assert self.t0.shape == self.t1.shape == self.intersected.shape
+
+ def partition(self, ts):
+ """
+ Partitions t0 and t1 into n_samples intervals.
+
+ Args:
+ ts: [batch_size, *shape, n_samples, 1]
+
+ Return:
+
+ lower: [batch_size, *shape, n_samples, 1] upper: [batch_size, *shape, n_samples, 1] delta: [batch_size,
+ *shape, n_samples, 1]
+
+ where
+ ts \\in [lower, upper] deltas = upper - lower
+ """
+
+ mids = (ts[..., 1:, :] + ts[..., :-1, :]) * 0.5
+ lower = torch.cat([self.t0[..., None, :], mids], dim=-2)
+ upper = torch.cat([mids, self.t1[..., None, :]], dim=-2)
+ delta = upper - lower
+ assert lower.shape == upper.shape == delta.shape == ts.shape
+ return lower, upper, delta
+
+
+class BoundingBoxVolume(nn.Module):
+ """
+ Axis-aligned bounding box defined by the two opposite corners.
+ """
+
+ def __init__(
+ self,
+ *,
+ bbox_min,
+ bbox_max,
+ min_dist: float = 0.0,
+ min_t_range: float = 1e-3,
+ ):
+ """
+ Args:
+ bbox_min: the left/bottommost corner of the bounding box
+ bbox_max: the other corner of the bounding box
+ min_dist: all rays should start at least this distance away from the origin.
+ """
+ super().__init__()
+
+ self.min_dist = min_dist
+ self.min_t_range = min_t_range
+
+ self.bbox_min = torch.tensor(bbox_min)
+ self.bbox_max = torch.tensor(bbox_max)
+ self.bbox = torch.stack([self.bbox_min, self.bbox_max])
+ assert self.bbox.shape == (2, 3)
+ assert min_dist >= 0.0
+ assert min_t_range > 0.0
+
+ def intersect(
+ self,
+ origin: torch.Tensor,
+ direction: torch.Tensor,
+ t0_lower: Optional[torch.Tensor] = None,
+ epsilon=1e-6,
+ ):
+ """
+ Args:
+ origin: [batch_size, *shape, 3]
+ direction: [batch_size, *shape, 3]
+ t0_lower: Optional [batch_size, *shape, 1] lower bound of t0 when intersecting this volume.
+ params: Optional meta parameters in case Volume is parametric
+ epsilon: to stabilize calculations
+
+ Return:
+ A tuple of (t0, t1, intersected) where each has a shape [batch_size, *shape, 1]. If a ray intersects with
+ the volume, `o + td` is in the volume for all t in [t0, t1]. If the volume is bounded, t1 is guaranteed to
+ be on the boundary of the volume.
+ """
+
+ batch_size, *shape, _ = origin.shape
+ ones = [1] * len(shape)
+ bbox = self.bbox.view(1, *ones, 2, 3).to(origin.device)
+
+ def _safe_divide(a, b, epsilon=1e-6):
+ return a / torch.where(b < 0, b - epsilon, b + epsilon)
+
+ ts = _safe_divide(bbox - origin[..., None, :], direction[..., None, :], epsilon=epsilon)
+
+ # Cases to think about:
+ #
+ # 1. t1 <= t0: the ray does not pass through the AABB.
+ # 2. t0 < t1 <= 0: the ray intersects but the BB is behind the origin.
+ # 3. t0 <= 0 <= t1: the ray starts from inside the BB
+ # 4. 0 <= t0 < t1: the ray is not inside and intersects with the BB twice.
+ #
+ # 1 and 4 are clearly handled from t0 < t1 below.
+ # Making t0 at least min_dist (>= 0) takes care of 2 and 3.
+ t0 = ts.min(dim=-2).values.max(dim=-1, keepdim=True).values.clamp(self.min_dist)
+ t1 = ts.max(dim=-2).values.min(dim=-1, keepdim=True).values
+ assert t0.shape == t1.shape == (batch_size, *shape, 1)
+ if t0_lower is not None:
+ assert t0.shape == t0_lower.shape
+ t0 = torch.maximum(t0, t0_lower)
+
+ intersected = t0 + self.min_t_range < t1
+ t0 = torch.where(intersected, t0, torch.zeros_like(t0))
+ t1 = torch.where(intersected, t1, torch.ones_like(t1))
+
+ return VolumeRange(t0=t0, t1=t1, intersected=intersected)
+
+
+class StratifiedRaySampler(nn.Module):
+ """
+ Instead of fixed intervals, a sample is drawn uniformly at random from each interval.
+ """
+
+ def __init__(self, depth_mode: str = "linear"):
+ """
+ :param depth_mode: linear samples ts linearly in depth. harmonic ensures
+ closer points are sampled more densely.
+ """
+ self.depth_mode = depth_mode
+ assert self.depth_mode in ("linear", "geometric", "harmonic")
+
+ def sample(
+ self,
+ t0: torch.Tensor,
+ t1: torch.Tensor,
+ n_samples: int,
+ epsilon: float = 1e-3,
+ ) -> torch.Tensor:
+ """
+ Args:
+ t0: start time has shape [batch_size, *shape, 1]
+ t1: finish time has shape [batch_size, *shape, 1]
+ n_samples: number of ts to sample
+ Return:
+ sampled ts of shape [batch_size, *shape, n_samples, 1]
+ """
+ ones = [1] * (len(t0.shape) - 1)
+ ts = torch.linspace(0, 1, n_samples).view(*ones, n_samples).to(t0.dtype).to(t0.device)
+
+ if self.depth_mode == "linear":
+ ts = t0 * (1.0 - ts) + t1 * ts
+ elif self.depth_mode == "geometric":
+ ts = (t0.clamp(epsilon).log() * (1.0 - ts) + t1.clamp(epsilon).log() * ts).exp()
+ elif self.depth_mode == "harmonic":
+ # The original NeRF recommends this interpolation scheme for
+ # spherical scenes, but there could be some weird edge cases when
+ # the observer crosses from the inner to outer volume.
+ ts = 1.0 / (1.0 / t0.clamp(epsilon) * (1.0 - ts) + 1.0 / t1.clamp(epsilon) * ts)
+
+ mids = 0.5 * (ts[..., 1:] + ts[..., :-1])
+ upper = torch.cat([mids, t1], dim=-1)
+ lower = torch.cat([t0, mids], dim=-1)
+ # yiyi notes: add a random seed here for testing, don't forget to remove
+ torch.manual_seed(0)
+ t_rand = torch.rand_like(ts)
+
+ ts = lower + (upper - lower) * t_rand
+ return ts.unsqueeze(-1)
+
+
+class ImportanceRaySampler(nn.Module):
+ """
+ Given the initial estimate of densities, this samples more from regions/bins expected to have objects.
+ """
+
+ def __init__(
+ self,
+ volume_range: VolumeRange,
+ ts: torch.Tensor,
+ weights: torch.Tensor,
+ blur_pool: bool = False,
+ alpha: float = 1e-5,
+ ):
+ """
+ Args:
+ volume_range: the range in which a ray intersects the given volume.
+ ts: earlier samples from the coarse rendering step
+ weights: discretized version of density * transmittance
+ blur_pool: if true, use 2-tap max + 2-tap blur filter from mip-NeRF.
+ alpha: small value to add to weights.
+ """
+ self.volume_range = volume_range
+ self.ts = ts.clone().detach()
+ self.weights = weights.clone().detach()
+ self.blur_pool = blur_pool
+ self.alpha = alpha
+
+ @torch.no_grad()
+ def sample(self, t0: torch.Tensor, t1: torch.Tensor, n_samples: int) -> torch.Tensor:
+ """
+ Args:
+ t0: start time has shape [batch_size, *shape, 1]
+ t1: finish time has shape [batch_size, *shape, 1]
+ n_samples: number of ts to sample
+ Return:
+ sampled ts of shape [batch_size, *shape, n_samples, 1]
+ """
+ lower, upper, _ = self.volume_range.partition(self.ts)
+
+ batch_size, *shape, n_coarse_samples, _ = self.ts.shape
+
+ weights = self.weights
+ if self.blur_pool:
+ padded = torch.cat([weights[..., :1, :], weights, weights[..., -1:, :]], dim=-2)
+ maxes = torch.maximum(padded[..., :-1, :], padded[..., 1:, :])
+ weights = 0.5 * (maxes[..., :-1, :] + maxes[..., 1:, :])
+ weights = weights + self.alpha
+ pmf = weights / weights.sum(dim=-2, keepdim=True)
+ inds = sample_pmf(pmf, n_samples)
+ assert inds.shape == (batch_size, *shape, n_samples, 1)
+ assert (inds >= 0).all() and (inds < n_coarse_samples).all()
+
+ t_rand = torch.rand(inds.shape, device=inds.device)
+ lower_ = torch.gather(lower, -2, inds)
+ upper_ = torch.gather(upper, -2, inds)
+
+ ts = lower_ + (upper_ - lower_) * t_rand
+ ts = torch.sort(ts, dim=-2).values
+ return ts
+
+
+@dataclass
+class MeshDecoderOutput(BaseOutput):
+ """
+ A 3D triangle mesh with optional data at the vertices and faces.
+
+ Args:
+ verts (`torch.Tensor` of shape `(N, 3)`):
+ array of vertext coordinates
+ faces (`torch.Tensor` of shape `(N, 3)`):
+ array of triangles, pointing to indices in verts.
+ vertext_channels (Dict):
+ vertext coordinates for each color channel
+ """
+
+ verts: torch.Tensor
+ faces: torch.Tensor
+ vertex_channels: Dict[str, torch.Tensor]
+
+
+class MeshDecoder(nn.Module):
+ """
+ Construct meshes from Signed distance functions (SDFs) using marching cubes method
+ """
+
+ def __init__(self):
+ super().__init__()
+ cases = torch.zeros(256, 5, 3, dtype=torch.long)
+ masks = torch.zeros(256, 5, dtype=torch.bool)
+
+ self.register_buffer("cases", cases)
+ self.register_buffer("masks", masks)
+
+ def forward(self, field: torch.Tensor, min_point: torch.Tensor, size: torch.Tensor):
+ """
+ For a signed distance field, produce a mesh using marching cubes.
+
+ :param field: a 3D tensor of field values, where negative values correspond
+ to the outside of the shape. The dimensions correspond to the x, y, and z directions, respectively.
+ :param min_point: a tensor of shape [3] containing the point corresponding
+ to (0, 0, 0) in the field.
+ :param size: a tensor of shape [3] containing the per-axis distance from the
+ (0, 0, 0) field corner and the (-1, -1, -1) field corner.
+ """
+ assert len(field.shape) == 3, "input must be a 3D scalar field"
+ dev = field.device
+
+ cases = self.cases.to(dev)
+ masks = self.masks.to(dev)
+
+ min_point = min_point.to(dev)
+ size = size.to(dev)
+
+ grid_size = field.shape
+ grid_size_tensor = torch.tensor(grid_size).to(size)
+
+ # Create bitmasks between 0 and 255 (inclusive) indicating the state
+ # of the eight corners of each cube.
+ bitmasks = (field > 0).to(torch.uint8)
+ bitmasks = bitmasks[:-1, :, :] | (bitmasks[1:, :, :] << 1)
+ bitmasks = bitmasks[:, :-1, :] | (bitmasks[:, 1:, :] << 2)
+ bitmasks = bitmasks[:, :, :-1] | (bitmasks[:, :, 1:] << 4)
+
+ # Compute corner coordinates across the entire grid.
+ corner_coords = torch.empty(*grid_size, 3, device=dev, dtype=field.dtype)
+ corner_coords[range(grid_size[0]), :, :, 0] = torch.arange(grid_size[0], device=dev, dtype=field.dtype)[
+ :, None, None
+ ]
+ corner_coords[:, range(grid_size[1]), :, 1] = torch.arange(grid_size[1], device=dev, dtype=field.dtype)[
+ :, None
+ ]
+ corner_coords[:, :, range(grid_size[2]), 2] = torch.arange(grid_size[2], device=dev, dtype=field.dtype)
+
+ # Compute all vertices across all edges in the grid, even though we will
+ # throw some out later. We have (X-1)*Y*Z + X*(Y-1)*Z + X*Y*(Z-1) vertices.
+ # These are all midpoints, and don't account for interpolation (which is
+ # done later based on the used edge midpoints).
+ edge_midpoints = torch.cat(
+ [
+ ((corner_coords[:-1] + corner_coords[1:]) / 2).reshape(-1, 3),
+ ((corner_coords[:, :-1] + corner_coords[:, 1:]) / 2).reshape(-1, 3),
+ ((corner_coords[:, :, :-1] + corner_coords[:, :, 1:]) / 2).reshape(-1, 3),
+ ],
+ dim=0,
+ )
+
+ # Create a flat array of [X, Y, Z] indices for each cube.
+ cube_indices = torch.zeros(
+ grid_size[0] - 1, grid_size[1] - 1, grid_size[2] - 1, 3, device=dev, dtype=torch.long
+ )
+ cube_indices[range(grid_size[0] - 1), :, :, 0] = torch.arange(grid_size[0] - 1, device=dev)[:, None, None]
+ cube_indices[:, range(grid_size[1] - 1), :, 1] = torch.arange(grid_size[1] - 1, device=dev)[:, None]
+ cube_indices[:, :, range(grid_size[2] - 1), 2] = torch.arange(grid_size[2] - 1, device=dev)
+ flat_cube_indices = cube_indices.reshape(-1, 3)
+
+ # Create a flat array mapping each cube to 12 global edge indices.
+ edge_indices = _create_flat_edge_indices(flat_cube_indices, grid_size)
+
+ # Apply the LUT to figure out the triangles.
+ flat_bitmasks = bitmasks.reshape(-1).long() # must cast to long for indexing to believe this not a mask
+ local_tris = cases[flat_bitmasks]
+ local_masks = masks[flat_bitmasks]
+ # Compute the global edge indices for the triangles.
+ global_tris = torch.gather(edge_indices, 1, local_tris.reshape(local_tris.shape[0], -1)).reshape(
+ local_tris.shape
+ )
+ # Select the used triangles for each cube.
+ selected_tris = global_tris.reshape(-1, 3)[local_masks.reshape(-1)]
+
+ # Now we have a bunch of indices into the full list of possible vertices,
+ # but we want to reduce this list to only the used vertices.
+ used_vertex_indices = torch.unique(selected_tris.view(-1))
+ used_edge_midpoints = edge_midpoints[used_vertex_indices]
+ old_index_to_new_index = torch.zeros(len(edge_midpoints), device=dev, dtype=torch.long)
+ old_index_to_new_index[used_vertex_indices] = torch.arange(
+ len(used_vertex_indices), device=dev, dtype=torch.long
+ )
+
+ # Rewrite the triangles to use the new indices
+ faces = torch.gather(old_index_to_new_index, 0, selected_tris.view(-1)).reshape(selected_tris.shape)
+
+ # Compute the actual interpolated coordinates corresponding to edge midpoints.
+ v1 = torch.floor(used_edge_midpoints).to(torch.long)
+ v2 = torch.ceil(used_edge_midpoints).to(torch.long)
+ s1 = field[v1[:, 0], v1[:, 1], v1[:, 2]]
+ s2 = field[v2[:, 0], v2[:, 1], v2[:, 2]]
+ p1 = (v1.float() / (grid_size_tensor - 1)) * size + min_point
+ p2 = (v2.float() / (grid_size_tensor - 1)) * size + min_point
+ # The signs of s1 and s2 should be different. We want to find
+ # t such that t*s2 + (1-t)*s1 = 0.
+ t = (s1 / (s1 - s2))[:, None]
+ verts = t * p2 + (1 - t) * p1
+
+ return MeshDecoderOutput(verts=verts, faces=faces, vertex_channels=None)
+
+
+@dataclass
+class MLPNeRFModelOutput(BaseOutput):
+ density: torch.Tensor
+ signed_distance: torch.Tensor
+ channels: torch.Tensor
+ ts: torch.Tensor
+
+
+class MLPNeRSTFModel(ModelMixin, ConfigMixin):
+ @register_to_config
+ def __init__(
+ self,
+ d_hidden: int = 256,
+ n_output: int = 12,
+ n_hidden_layers: int = 6,
+ act_fn: str = "swish",
+ insert_direction_at: int = 4,
+ ):
+ super().__init__()
+
+ # Instantiate the MLP
+
+ # Find out the dimension of encoded position and direction
+ dummy = torch.eye(1, 3)
+ d_posenc_pos = encode_position(position=dummy).shape[-1]
+ d_posenc_dir = encode_direction(position=dummy).shape[-1]
+
+ mlp_widths = [d_hidden] * n_hidden_layers
+ input_widths = [d_posenc_pos] + mlp_widths
+ output_widths = mlp_widths + [n_output]
+
+ if insert_direction_at is not None:
+ input_widths[insert_direction_at] += d_posenc_dir
+
+ self.mlp = nn.ModuleList([nn.Linear(d_in, d_out) for d_in, d_out in zip(input_widths, output_widths)])
+
+ if act_fn == "swish":
+ # self.activation = swish
+ # yiyi testing:
+ self.activation = lambda x: F.silu(x)
+ else:
+ raise ValueError(f"Unsupported activation function {act_fn}")
+
+ self.sdf_activation = torch.tanh
+ self.density_activation = torch.nn.functional.relu
+ self.channel_activation = torch.sigmoid
+
+ def map_indices_to_keys(self, output):
+ h_map = {
+ "sdf": (0, 1),
+ "density_coarse": (1, 2),
+ "density_fine": (2, 3),
+ "stf": (3, 6),
+ "nerf_coarse": (6, 9),
+ "nerf_fine": (9, 12),
+ }
+
+ mapped_output = {k: output[..., start:end] for k, (start, end) in h_map.items()}
+
+ return mapped_output
+
+ def forward(self, *, position, direction, ts, nerf_level="coarse", rendering_mode="nerf"):
+ h = encode_position(position)
+
+ h_preact = h
+ h_directionless = None
+ for i, layer in enumerate(self.mlp):
+ if i == self.config.insert_direction_at: # 4 in the config
+ h_directionless = h_preact
+ h_direction = encode_direction(position, direction=direction)
+ h = torch.cat([h, h_direction], dim=-1)
+
+ h = layer(h)
+
+ h_preact = h
+
+ if i < len(self.mlp) - 1:
+ h = self.activation(h)
+
+ h_final = h
+ if h_directionless is None:
+ h_directionless = h_preact
+
+ activation = self.map_indices_to_keys(h_final)
+
+ if nerf_level == "coarse":
+ h_density = activation["density_coarse"]
+ else:
+ h_density = activation["density_fine"]
+
+ if rendering_mode == "nerf":
+ if nerf_level == "coarse":
+ h_channels = activation["nerf_coarse"]
+ else:
+ h_channels = activation["nerf_fine"]
+
+ elif rendering_mode == "stf":
+ h_channels = activation["stf"]
+
+ density = self.density_activation(h_density)
+ signed_distance = self.sdf_activation(activation["sdf"])
+ channels = self.channel_activation(h_channels)
+
+ # yiyi notes: I think signed_distance is not used
+ return MLPNeRFModelOutput(density=density, signed_distance=signed_distance, channels=channels, ts=ts)
+
+
+class ChannelsProj(nn.Module):
+ def __init__(
+ self,
+ *,
+ vectors: int,
+ channels: int,
+ d_latent: int,
+ ):
+ super().__init__()
+ self.proj = nn.Linear(d_latent, vectors * channels)
+ self.norm = nn.LayerNorm(channels)
+ self.d_latent = d_latent
+ self.vectors = vectors
+ self.channels = channels
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ x_bvd = x
+ w_vcd = self.proj.weight.view(self.vectors, self.channels, self.d_latent)
+ b_vc = self.proj.bias.view(1, self.vectors, self.channels)
+ h = torch.einsum("bvd,vcd->bvc", x_bvd, w_vcd)
+ h = self.norm(h)
+
+ h = h + b_vc
+ return h
+
+
+class ShapEParamsProjModel(ModelMixin, ConfigMixin):
+ """
+ project the latent representation of a 3D asset to obtain weights of a multi-layer perceptron (MLP).
+
+ For more details, see the original paper:
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ *,
+ param_names: Tuple[str] = (
+ "nerstf.mlp.0.weight",
+ "nerstf.mlp.1.weight",
+ "nerstf.mlp.2.weight",
+ "nerstf.mlp.3.weight",
+ ),
+ param_shapes: Tuple[Tuple[int]] = (
+ (256, 93),
+ (256, 256),
+ (256, 256),
+ (256, 256),
+ ),
+ d_latent: int = 1024,
+ ):
+ super().__init__()
+
+ # check inputs
+ if len(param_names) != len(param_shapes):
+ raise ValueError("Must provide same number of `param_names` as `param_shapes`")
+ self.projections = nn.ModuleDict({})
+ for k, (vectors, channels) in zip(param_names, param_shapes):
+ self.projections[_sanitize_name(k)] = ChannelsProj(
+ vectors=vectors,
+ channels=channels,
+ d_latent=d_latent,
+ )
+
+ def forward(self, x: torch.Tensor):
+ out = {}
+ start = 0
+ for k, shape in zip(self.config.param_names, self.config.param_shapes):
+ vectors, _ = shape
+ end = start + vectors
+ x_bvd = x[:, start:end]
+ out[k] = self.projections[_sanitize_name(k)](x_bvd).reshape(len(x), *shape)
+ start = end
+ return out
+
+
+class ShapERenderer(ModelMixin, ConfigMixin):
+ @register_to_config
+ def __init__(
+ self,
+ *,
+ param_names: Tuple[str] = (
+ "nerstf.mlp.0.weight",
+ "nerstf.mlp.1.weight",
+ "nerstf.mlp.2.weight",
+ "nerstf.mlp.3.weight",
+ ),
+ param_shapes: Tuple[Tuple[int]] = (
+ (256, 93),
+ (256, 256),
+ (256, 256),
+ (256, 256),
+ ),
+ d_latent: int = 1024,
+ d_hidden: int = 256,
+ n_output: int = 12,
+ n_hidden_layers: int = 6,
+ act_fn: str = "swish",
+ insert_direction_at: int = 4,
+ background: Tuple[float] = (
+ 255.0,
+ 255.0,
+ 255.0,
+ ),
+ ):
+ super().__init__()
+
+ self.params_proj = ShapEParamsProjModel(
+ param_names=param_names,
+ param_shapes=param_shapes,
+ d_latent=d_latent,
+ )
+ self.mlp = MLPNeRSTFModel(d_hidden, n_output, n_hidden_layers, act_fn, insert_direction_at)
+ self.void = VoidNeRFModel(background=background, channel_scale=255.0)
+ self.volume = BoundingBoxVolume(bbox_max=[1.0, 1.0, 1.0], bbox_min=[-1.0, -1.0, -1.0])
+ self.mesh_decoder = MeshDecoder()
+
+ @torch.no_grad()
+ def render_rays(self, rays, sampler, n_samples, prev_model_out=None, render_with_direction=False):
+ """
+ Perform volumetric rendering over a partition of possible t's in the union of rendering volumes (written below
+ with some abuse of notations)
+
+ C(r) := sum(
+ transmittance(t[i]) * integrate(
+ lambda t: density(t) * channels(t) * transmittance(t), [t[i], t[i + 1]],
+ ) for i in range(len(parts))
+ ) + transmittance(t[-1]) * void_model(t[-1]).channels
+
+ where
+
+ 1) transmittance(s) := exp(-integrate(density, [t[0], s])) calculates the probability of light passing through
+ the volume specified by [t[0], s]. (transmittance of 1 means light can pass freely) 2) density and channels are
+ obtained by evaluating the appropriate part.model at time t. 3) [t[i], t[i + 1]] is defined as the range of t
+ where the ray intersects (parts[i].volume \\ union(part.volume for part in parts[:i])) at the surface of the
+ shell (if bounded). If the ray does not intersect, the integral over this segment is evaluated as 0 and
+ transmittance(t[i + 1]) := transmittance(t[i]). 4) The last term is integration to infinity (e.g. [t[-1],
+ math.inf]) that is evaluated by the void_model (i.e. we consider this space to be empty).
+
+ args:
+ rays: [batch_size x ... x 2 x 3] origin and direction. sampler: disjoint volume integrals. n_samples:
+ number of ts to sample. prev_model_outputs: model outputs from the previous rendering step, including
+
+ :return: A tuple of
+ - `channels`
+ - A importance samplers for additional fine-grained rendering
+ - raw model output
+ """
+ origin, direction = rays[..., 0, :], rays[..., 1, :]
+
+ # Integrate over [t[i], t[i + 1]]
+
+ # 1 Intersect the rays with the current volume and sample ts to integrate along.
+ vrange = self.volume.intersect(origin, direction, t0_lower=None)
+ ts = sampler.sample(vrange.t0, vrange.t1, n_samples)
+ ts = ts.to(rays.dtype)
+
+ if prev_model_out is not None:
+ # Append the previous ts now before fprop because previous
+ # rendering used a different model and we can't reuse the output.
+ ts = torch.sort(torch.cat([ts, prev_model_out.ts], dim=-2), dim=-2).values
+
+ batch_size, *_shape, _t0_dim = vrange.t0.shape
+ _, *ts_shape, _ts_dim = ts.shape
+
+ # 2. Get the points along the ray and query the model
+ directions = torch.broadcast_to(direction.unsqueeze(-2), [batch_size, *ts_shape, 3])
+ positions = origin.unsqueeze(-2) + ts * directions
+
+ directions = directions.to(self.mlp.dtype)
+ positions = positions.to(self.mlp.dtype)
+
+ optional_directions = directions if render_with_direction else None
+
+ model_out = self.mlp(
+ position=positions,
+ direction=optional_directions,
+ ts=ts,
+ nerf_level="coarse" if prev_model_out is None else "fine",
+ )
+
+ # 3. Integrate the model results
+ channels, weights, transmittance = integrate_samples(
+ vrange, model_out.ts, model_out.density, model_out.channels
+ )
+
+ # 4. Clean up results that do not intersect with the volume.
+ transmittance = torch.where(vrange.intersected, transmittance, torch.ones_like(transmittance))
+ channels = torch.where(vrange.intersected, channels, torch.zeros_like(channels))
+ # 5. integration to infinity (e.g. [t[-1], math.inf]) that is evaluated by the void_model (i.e. we consider this space to be empty).
+ channels = channels + transmittance * self.void(origin)
+
+ weighted_sampler = ImportanceRaySampler(vrange, ts=model_out.ts, weights=weights)
+
+ return channels, weighted_sampler, model_out
+
+ @torch.no_grad()
+ def decode_to_image(
+ self,
+ latents,
+ device,
+ size: int = 64,
+ ray_batch_size: int = 4096,
+ n_coarse_samples=64,
+ n_fine_samples=128,
+ ):
+ # project the parameters from the generated latents
+ projected_params = self.params_proj(latents)
+
+ # update the mlp layers of the renderer
+ for name, param in self.mlp.state_dict().items():
+ if f"nerstf.{name}" in projected_params.keys():
+ param.copy_(projected_params[f"nerstf.{name}"].squeeze(0))
+
+ # create cameras object
+ camera = create_pan_cameras(size)
+ rays = camera.camera_rays
+ rays = rays.to(device)
+ n_batches = rays.shape[1] // ray_batch_size
+
+ coarse_sampler = StratifiedRaySampler()
+
+ images = []
+
+ for idx in range(n_batches):
+ rays_batch = rays[:, idx * ray_batch_size : (idx + 1) * ray_batch_size]
+
+ # render rays with coarse, stratified samples.
+ _, fine_sampler, coarse_model_out = self.render_rays(rays_batch, coarse_sampler, n_coarse_samples)
+ # Then, render with additional importance-weighted ray samples.
+ channels, _, _ = self.render_rays(
+ rays_batch, fine_sampler, n_fine_samples, prev_model_out=coarse_model_out
+ )
+
+ images.append(channels)
+
+ images = torch.cat(images, dim=1)
+ images = images.view(*camera.shape, camera.height, camera.width, -1).squeeze(0)
+
+ return images
+
+ @torch.no_grad()
+ def decode_to_mesh(
+ self,
+ latents,
+ device,
+ grid_size: int = 128,
+ query_batch_size: int = 4096,
+ texture_channels: Tuple = ("R", "G", "B"),
+ ):
+ # 1. project the parameters from the generated latents
+ projected_params = self.params_proj(latents)
+
+ # 2. update the mlp layers of the renderer
+ for name, param in self.mlp.state_dict().items():
+ if f"nerstf.{name}" in projected_params.keys():
+ param.copy_(projected_params[f"nerstf.{name}"].squeeze(0))
+
+ # 3. decoding with STF rendering
+ # 3.1 query the SDF values at vertices along a regular 128**3 grid
+
+ query_points = volume_query_points(self.volume, grid_size)
+ query_positions = query_points[None].repeat(1, 1, 1).to(device=device, dtype=self.mlp.dtype)
+
+ fields = []
+
+ for idx in range(0, query_positions.shape[1], query_batch_size):
+ query_batch = query_positions[:, idx : idx + query_batch_size]
+
+ model_out = self.mlp(
+ position=query_batch, direction=None, ts=None, nerf_level="fine", rendering_mode="stf"
+ )
+ fields.append(model_out.signed_distance)
+
+ # predicted SDF values
+ fields = torch.cat(fields, dim=1)
+ fields = fields.float()
+
+ assert (
+ len(fields.shape) == 3 and fields.shape[-1] == 1
+ ), f"expected [meta_batch x inner_batch] SDF results, but got {fields.shape}"
+
+ fields = fields.reshape(1, *([grid_size] * 3))
+
+ # create grid 128 x 128 x 128
+ # - force a negative border around the SDFs to close off all the models.
+ full_grid = torch.zeros(
+ 1,
+ grid_size + 2,
+ grid_size + 2,
+ grid_size + 2,
+ device=fields.device,
+ dtype=fields.dtype,
+ )
+ full_grid.fill_(-1.0)
+ full_grid[:, 1:-1, 1:-1, 1:-1] = fields
+ fields = full_grid
+
+ # apply a differentiable implementation of Marching Cubes to construct meshs
+ raw_meshes = []
+ mesh_mask = []
+
+ for field in fields:
+ raw_mesh = self.mesh_decoder(field, self.volume.bbox_min, self.volume.bbox_max - self.volume.bbox_min)
+ mesh_mask.append(True)
+ raw_meshes.append(raw_mesh)
+
+ mesh_mask = torch.tensor(mesh_mask, device=fields.device)
+ max_vertices = max(len(m.verts) for m in raw_meshes)
+
+ # 3.2. query the texture color head at each vertex of the resulting mesh.
+ texture_query_positions = torch.stack(
+ [m.verts[torch.arange(0, max_vertices) % len(m.verts)] for m in raw_meshes],
+ dim=0,
+ )
+ texture_query_positions = texture_query_positions.to(device=device, dtype=self.mlp.dtype)
+
+ textures = []
+
+ for idx in range(0, texture_query_positions.shape[1], query_batch_size):
+ query_batch = texture_query_positions[:, idx : idx + query_batch_size]
+
+ texture_model_out = self.mlp(
+ position=query_batch, direction=None, ts=None, nerf_level="fine", rendering_mode="stf"
+ )
+ textures.append(texture_model_out.channels)
+
+ # predict texture color
+ textures = torch.cat(textures, dim=1)
+
+ textures = _convert_srgb_to_linear(textures)
+ textures = textures.float()
+
+ # 3.3 augument the mesh with texture data
+ assert len(textures.shape) == 3 and textures.shape[-1] == len(
+ texture_channels
+ ), f"expected [meta_batch x inner_batch x texture_channels] field results, but got {textures.shape}"
+
+ for m, texture in zip(raw_meshes, textures):
+ texture = texture[: len(m.verts)]
+ m.vertex_channels = dict(zip(texture_channels, texture.unbind(-1)))
+
+ return raw_meshes[0]
diff --git a/model/diffusers_c/pipelines/stable_cascade/__init__.py b/model/diffusers_c/pipelines/stable_cascade/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5270cb94af01fd94ed6c8e76c243a86cad8ec348
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_cascade/__init__.py
@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_cascade"] = ["StableCascadeDecoderPipeline"]
+ _import_structure["pipeline_stable_cascade_combined"] = ["StableCascadeCombinedPipeline"]
+ _import_structure["pipeline_stable_cascade_prior"] = ["StableCascadePriorPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import * # noqa F403
+ else:
+ from .pipeline_stable_cascade import StableCascadeDecoderPipeline
+ from .pipeline_stable_cascade_combined import StableCascadeCombinedPipeline
+ from .pipeline_stable_cascade_prior import StableCascadePriorPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_cascade/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_cascade/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c36829e2db1b10530b066f8e0431cf35f2d6460b
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_cascade/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade.cpython-39.pyc b/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..60d6bab2daaf47fe72199bf5e26ac9f6d163d973
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade_combined.cpython-39.pyc b/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade_combined.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ab0861993830ea3f8e59e284f4dfaa9dc2a561b2
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade_combined.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade_prior.cpython-39.pyc b/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade_prior.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d86cc88881fd04a663824f2537bb044979b8f701
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_cascade/__pycache__/pipeline_stable_cascade_prior.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade.py b/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade.py
new file mode 100644
index 0000000000000000000000000000000000000000..8a8d5b65e31ac5e2f28e3f02f62dab0259e0efd5
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade.py
@@ -0,0 +1,467 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ...models import StableCascadeUNet
+from ...schedulers import DDPMWuerstchenScheduler
+from ...utils import is_torch_version, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from ..wuerstchen.modeling_paella_vq_model import PaellaVQModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableCascadePriorPipeline, StableCascadeDecoderPipeline
+
+ >>> prior_pipe = StableCascadePriorPipeline.from_pretrained(
+ ... "stabilityai/stable-cascade-prior", torch_dtype=torch.bfloat16
+ ... ).to("cuda")
+ >>> gen_pipe = StableCascadeDecoderPipeline.from_pretrain(
+ ... "stabilityai/stable-cascade", torch_dtype=torch.float16
+ ... ).to("cuda")
+
+ >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet"
+ >>> prior_output = pipe(prompt)
+ >>> images = gen_pipe(prior_output.image_embeddings, prompt=prompt)
+ ```
+"""
+
+
+class StableCascadeDecoderPipeline(DiffusionPipeline):
+ """
+ Pipeline for generating images from the Stable Cascade model.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ tokenizer (`CLIPTokenizer`):
+ The CLIP tokenizer.
+ text_encoder (`CLIPTextModel`):
+ The CLIP text encoder.
+ decoder ([`StableCascadeUNet`]):
+ The Stable Cascade decoder unet.
+ vqgan ([`PaellaVQModel`]):
+ The VQGAN model.
+ scheduler ([`DDPMWuerstchenScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ latent_dim_scale (float, `optional`, defaults to 10.67):
+ Multiplier to determine the VQ latent space size from the image embeddings. If the image embeddings are
+ height=24 and width=24, the VQ latent shape needs to be height=int(24*10.67)=256 and
+ width=int(24*10.67)=256 in order to match the training conditions.
+ """
+
+ unet_name = "decoder"
+ text_encoder_name = "text_encoder"
+ model_cpu_offload_seq = "text_encoder->decoder->vqgan"
+ _callback_tensor_inputs = [
+ "latents",
+ "prompt_embeds_pooled",
+ "negative_prompt_embeds",
+ "image_embeddings",
+ ]
+
+ def __init__(
+ self,
+ decoder: StableCascadeUNet,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModel,
+ scheduler: DDPMWuerstchenScheduler,
+ vqgan: PaellaVQModel,
+ latent_dim_scale: float = 10.67,
+ ) -> None:
+ super().__init__()
+ self.register_modules(
+ decoder=decoder,
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ scheduler=scheduler,
+ vqgan=vqgan,
+ )
+ self.register_to_config(latent_dim_scale=latent_dim_scale)
+
+ def prepare_latents(self, image_embeddings, num_images_per_prompt, dtype, device, generator, latents, scheduler):
+ batch_size, channels, height, width = image_embeddings.shape
+ latents_shape = (
+ batch_size * num_images_per_prompt,
+ 4,
+ int(height * self.config.latent_dim_scale),
+ int(width * self.config.latent_dim_scale),
+ )
+
+ if latents is None:
+ latents = randn_tensor(latents_shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def encode_prompt(
+ self,
+ device,
+ batch_size,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ prompt=None,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ prompt_embeds_pooled: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds_pooled: Optional[torch.FloatTensor] = None,
+ ):
+ if prompt_embeds is None:
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ attention_mask = text_inputs.attention_mask
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+ attention_mask = attention_mask[:, : self.tokenizer.model_max_length]
+
+ text_encoder_output = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask.to(device), output_hidden_states=True
+ )
+ prompt_embeds = text_encoder_output.hidden_states[-1]
+ if prompt_embeds_pooled is None:
+ prompt_embeds_pooled = text_encoder_output.text_embeds.unsqueeze(1)
+
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+ prompt_embeds_pooled = prompt_embeds_pooled.to(dtype=self.text_encoder.dtype, device=device)
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ prompt_embeds_pooled = prompt_embeds_pooled.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if negative_prompt_embeds is None and do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=uncond_input.attention_mask.to(device),
+ output_hidden_states=True,
+ )
+
+ negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.hidden_states[-1]
+ negative_prompt_embeds_pooled = negative_prompt_embeds_text_encoder_output.text_embeds.unsqueeze(1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ seq_len = negative_prompt_embeds_pooled.shape[1]
+ negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.to(
+ dtype=self.text_encoder.dtype, device=device
+ )
+ negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ # done duplicates
+
+ return prompt_embeds, prompt_embeds_pooled, negative_prompt_embeds, negative_prompt_embeds_pooled
+
+ def check_inputs(
+ self,
+ prompt,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image_embeddings: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ prompt: Union[str, List[str]] = None,
+ num_inference_steps: int = 10,
+ guidance_scale: float = 0.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ image_embedding (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ Image Embeddings either extracted from an image or generated by a Prior Model.
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ num_inference_steps (`int`, *optional*, defaults to 12):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 0.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `decoder_guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting
+ `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely
+ linked to the text `prompt`, usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `decoder_guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True,
+ otherwise a `tuple`. When returning a tuple, the first element is a list with the generated image
+ embeddings.
+ """
+
+ # 0. Define commonly used variables
+ device = self._execution_device
+ dtype = self.decoder.dtype
+ self._guidance_scale = guidance_scale
+ if is_torch_version("<", "2.2.0") and dtype == torch.bfloat16:
+ raise ValueError("`StableCascadeDecoderPipeline` requires torch>=2.2.0 when using `torch.bfloat16` dtype.")
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ )
+ if isinstance(image_embeddings, list):
+ image_embeddings = torch.cat(image_embeddings, dim=0)
+ batch_size = image_embeddings.shape[0]
+
+ # 2. Encode caption
+ if prompt_embeds is None and negative_prompt_embeds is None:
+ prompt_embeds, _, negative_prompt_embeds, _ = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ batch_size=batch_size,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+ prompt_embeds_pooled = (
+ torch.cat([prompt_embeds, negative_prompt_embeds]) if self.do_classifier_free_guidance else prompt_embeds
+ )
+ effnet = (
+ torch.cat([image_embeddings, torch.zeros_like(image_embeddings)])
+ if self.do_classifier_free_guidance
+ else image_embeddings
+ )
+
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latents
+ latents = self.prepare_latents(
+ image_embeddings, num_images_per_prompt, dtype, device, generator, latents, self.scheduler
+ )
+
+ # 6. Run denoising loop
+ self._num_timesteps = len(timesteps[:-1])
+ for i, t in enumerate(self.progress_bar(timesteps[:-1])):
+ timestep_ratio = t.expand(latents.size(0)).to(dtype)
+
+ # 7. Denoise latents
+ predicted_latents = self.decoder(
+ sample=torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents,
+ timestep_ratio=torch.cat([timestep_ratio] * 2) if self.do_classifier_free_guidance else timestep_ratio,
+ clip_text_pooled=prompt_embeds_pooled,
+ effnet=effnet,
+ return_dict=False,
+ )[0]
+
+ # 8. Check for classifier free guidance and apply it
+ if self.do_classifier_free_guidance:
+ predicted_latents_text, predicted_latents_uncond = predicted_latents.chunk(2)
+ predicted_latents = torch.lerp(predicted_latents_uncond, predicted_latents_text, self.guidance_scale)
+
+ # 9. Renoise latents to next timestep
+ latents = self.scheduler.step(
+ model_output=predicted_latents,
+ timestep=timestep_ratio,
+ sample=latents,
+ generator=generator,
+ ).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ if output_type not in ["pt", "np", "pil", "latent"]:
+ raise ValueError(
+ f"Only the output types `pt`, `np`, `pil` and `latent` are supported not output_type={output_type}"
+ )
+
+ if not output_type == "latent":
+ # 10. Scale and decode the image latents with vq-vae
+ latents = self.vqgan.config.scale_factor * latents
+ images = self.vqgan.decode(latents).sample.clamp(0, 1)
+ if output_type == "np":
+ images = images.permute(0, 2, 3, 1).cpu().float().numpy() # float() as bfloat16-> numpy doesnt work
+ elif output_type == "pil":
+ images = images.permute(0, 2, 3, 1).cpu().float().numpy() # float() as bfloat16-> numpy doesnt work
+ images = self.numpy_to_pil(images)
+ else:
+ images = latents
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return images
+ return ImagePipelineOutput(images)
diff --git a/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade_combined.py b/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade_combined.py
new file mode 100644
index 0000000000000000000000000000000000000000..cd3592b49ac0adcc9a506f56a69722a6dbbd0ba2
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade_combined.py
@@ -0,0 +1,294 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Callable, Dict, List, Optional, Union
+
+import PIL
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...models import StableCascadeUNet
+from ...schedulers import DDPMWuerstchenScheduler
+from ...utils import replace_example_docstring
+from ..pipeline_utils import DiffusionPipeline
+from ..wuerstchen.modeling_paella_vq_model import PaellaVQModel
+from .pipeline_stable_cascade import StableCascadeDecoderPipeline
+from .pipeline_stable_cascade_prior import StableCascadePriorPipeline
+
+
+TEXT2IMAGE_EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusions import StableCascadeCombinedPipeline
+
+ >>> pipe = StableCascadeCombinedPipeline.from_pretrained("stabilityai/stable-cascade-combined", torch_dtype=torch.bfloat16).to(
+ ... "cuda"
+ ... )
+ >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet"
+ >>> images = pipe(prompt=prompt)
+ ```
+"""
+
+
+class StableCascadeCombinedPipeline(DiffusionPipeline):
+ """
+ Combined Pipeline for text-to-image generation using Stable Cascade.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ tokenizer (`CLIPTokenizer`):
+ The decoder tokenizer to be used for text inputs.
+ text_encoder (`CLIPTextModel`):
+ The decoder text encoder to be used for text inputs.
+ decoder (`StableCascadeUNet`):
+ The decoder model to be used for decoder image generation pipeline.
+ scheduler (`DDPMWuerstchenScheduler`):
+ The scheduler to be used for decoder image generation pipeline.
+ vqgan (`PaellaVQModel`):
+ The VQGAN model to be used for decoder image generation pipeline.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `image_encoder`.
+ image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ prior_prior (`StableCascadeUNet`):
+ The prior model to be used for prior pipeline.
+ prior_scheduler (`DDPMWuerstchenScheduler`):
+ The scheduler to be used for prior pipeline.
+ """
+
+ _load_connected_pipes = True
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModel,
+ decoder: StableCascadeUNet,
+ scheduler: DDPMWuerstchenScheduler,
+ vqgan: PaellaVQModel,
+ prior_prior: StableCascadeUNet,
+ prior_text_encoder: CLIPTextModel,
+ prior_tokenizer: CLIPTokenizer,
+ prior_scheduler: DDPMWuerstchenScheduler,
+ prior_feature_extractor: Optional[CLIPImageProcessor] = None,
+ prior_image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ decoder=decoder,
+ scheduler=scheduler,
+ vqgan=vqgan,
+ prior_text_encoder=prior_text_encoder,
+ prior_tokenizer=prior_tokenizer,
+ prior_prior=prior_prior,
+ prior_scheduler=prior_scheduler,
+ prior_feature_extractor=prior_feature_extractor,
+ prior_image_encoder=prior_image_encoder,
+ )
+ self.prior_pipe = StableCascadePriorPipeline(
+ prior=prior_prior,
+ text_encoder=prior_text_encoder,
+ tokenizer=prior_tokenizer,
+ scheduler=prior_scheduler,
+ image_encoder=prior_image_encoder,
+ feature_extractor=prior_feature_extractor,
+ )
+ self.decoder_pipe = StableCascadeDecoderPipeline(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ decoder=decoder,
+ scheduler=scheduler,
+ vqgan=vqgan,
+ )
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
+
+ def enable_model_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+ """
+ self.prior_pipe.enable_model_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_model_cpu_offload(gpu_id=gpu_id)
+
+ def enable_sequential_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗
+ Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a
+ GPU only when their specific submodule's `forward` method is called. Offloading happens on a submodule basis.
+ Memory savings are higher than using `enable_model_cpu_offload`, but performance is lower.
+ """
+ self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+
+ def progress_bar(self, iterable=None, total=None):
+ self.prior_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.progress_bar(iterable=iterable, total=total)
+
+ def set_progress_bar_config(self, **kwargs):
+ self.prior_pipe.set_progress_bar_config(**kwargs)
+ self.decoder_pipe.set_progress_bar_config(**kwargs)
+
+ @torch.no_grad()
+ @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ images: Union[torch.Tensor, PIL.Image.Image, List[torch.Tensor], List[PIL.Image.Image]] = None,
+ height: int = 512,
+ width: int = 512,
+ prior_num_inference_steps: int = 60,
+ prior_timesteps: Optional[List[float]] = None,
+ prior_guidance_scale: float = 4.0,
+ num_inference_steps: int = 12,
+ decoder_timesteps: Optional[List[float]] = None,
+ decoder_guidance_scale: float = 0.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ prior_callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ prior_callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation for the prior and decoder.
+ images (`torch.Tensor`, `PIL.Image.Image`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, *optional*):
+ The images to guide the image generation for the prior.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.*
+ prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `prior_guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting
+ `prior_guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked
+ to the text `prompt`, usually at the expense of lower image quality.
+ prior_num_inference_steps (`Union[int, Dict[float, int]]`, *optional*, defaults to 60):
+ The number of prior denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. For more specific timestep spacing, you can pass customized
+ `prior_timesteps`
+ num_inference_steps (`int`, *optional*, defaults to 12):
+ The number of decoder denoising steps. More denoising steps usually lead to a higher quality image at
+ the expense of slower inference. For more specific timestep spacing, you can pass customized
+ `timesteps`
+ decoder_guidance_scale (`float`, *optional*, defaults to 0.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ prior_callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `prior_callback_on_step_end(self: DiffusionPipeline, step: int, timestep:
+ int, callback_kwargs: Dict)`.
+ prior_callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the
+ list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in
+ the `._callback_tensor_inputs` attribute of your pipeine class.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True,
+ otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images.
+ """
+
+ prior_outputs = self.prior_pipe(
+ prompt=prompt if prompt_embeds is None else None,
+ images=images,
+ height=height,
+ width=width,
+ num_inference_steps=prior_num_inference_steps,
+ guidance_scale=prior_guidance_scale,
+ negative_prompt=negative_prompt if negative_prompt_embeds is None else None,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ num_images_per_prompt=num_images_per_prompt,
+ generator=generator,
+ latents=latents,
+ output_type="pt",
+ return_dict=True,
+ callback_on_step_end=prior_callback_on_step_end,
+ callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs,
+ )
+ image_embeddings = prior_outputs.image_embeddings
+ prompt_embeds = prior_outputs.get("prompt_embeds", None)
+ negative_prompt_embeds = prior_outputs.get("negative_prompt_embeds", None)
+
+ outputs = self.decoder_pipe(
+ image_embeddings=image_embeddings,
+ prompt=prompt if prompt_embeds is None else None,
+ num_inference_steps=num_inference_steps,
+ guidance_scale=decoder_guidance_scale,
+ negative_prompt=negative_prompt if negative_prompt_embeds is None else None,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ generator=generator,
+ output_type=output_type,
+ return_dict=return_dict,
+ callback_on_step_end=callback_on_step_end,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ )
+
+ return outputs
diff --git a/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade_prior.py b/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade_prior.py
new file mode 100644
index 0000000000000000000000000000000000000000..8df0d2a991b5f1cb17897d2dbc92f47b0f100d32
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_cascade/pipeline_stable_cascade_prior.py
@@ -0,0 +1,614 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from math import ceil
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...models import StableCascadeUNet
+from ...schedulers import DDPMWuerstchenScheduler
+from ...utils import BaseOutput, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+DEFAULT_STAGE_C_TIMESTEPS = list(np.linspace(1.0, 2 / 3, 20)) + list(np.linspace(2 / 3, 0.0, 11))[1:]
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableCascadePriorPipeline
+
+ >>> prior_pipe = StableCascadePriorPipeline.from_pretrained(
+ ... "stabilityai/stable-cascade-prior", torch_dtype=torch.bfloat16
+ ... ).to("cuda")
+
+ >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet"
+ >>> prior_output = pipe(prompt)
+ ```
+"""
+
+
+@dataclass
+class StableCascadePriorPipelineOutput(BaseOutput):
+ """
+ Output class for WuerstchenPriorPipeline.
+
+ Args:
+ image_embeddings (`torch.FloatTensor` or `np.ndarray`)
+ Prior image embeddings for text prompt
+ prompt_embeds (`torch.FloatTensor`):
+ Text embeddings for the prompt.
+ negative_prompt_embeds (`torch.FloatTensor`):
+ Text embeddings for the negative prompt.
+ """
+
+ image_embeddings: Union[torch.FloatTensor, np.ndarray]
+ prompt_embeds: Union[torch.FloatTensor, np.ndarray]
+ negative_prompt_embeds: Union[torch.FloatTensor, np.ndarray]
+
+
+class StableCascadePriorPipeline(DiffusionPipeline):
+ """
+ Pipeline for generating image prior for Stable Cascade.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ prior ([`StableCascadeUNet`]):
+ The Stable Cascade prior to approximate the image embedding from the text and/or image embedding.
+ text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)).
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `image_encoder`.
+ image_encoder ([`CLIPVisionModelWithProjection`]):
+ Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ scheduler ([`DDPMWuerstchenScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ resolution_multiple ('float', *optional*, defaults to 42.67):
+ Default resolution for multiple images generated.
+ """
+
+ unet_name = "prior"
+ text_encoder_name = "text_encoder"
+ model_cpu_offload_seq = "image_encoder->text_encoder->prior"
+ _optional_components = ["image_encoder", "feature_extractor"]
+ _callback_tensor_inputs = ["latents", "text_encoder_hidden_states", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModelWithProjection,
+ prior: StableCascadeUNet,
+ scheduler: DDPMWuerstchenScheduler,
+ resolution_multiple: float = 42.67,
+ feature_extractor: Optional[CLIPImageProcessor] = None,
+ image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ ) -> None:
+ super().__init__()
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ image_encoder=image_encoder,
+ feature_extractor=feature_extractor,
+ prior=prior,
+ scheduler=scheduler,
+ )
+ self.register_to_config(resolution_multiple=resolution_multiple)
+
+ def prepare_latents(
+ self, batch_size, height, width, num_images_per_prompt, dtype, device, generator, latents, scheduler
+ ):
+ latent_shape = (
+ num_images_per_prompt * batch_size,
+ self.prior.config.in_channels,
+ ceil(height / self.config.resolution_multiple),
+ ceil(width / self.config.resolution_multiple),
+ )
+
+ if latents is None:
+ latents = randn_tensor(latent_shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != latent_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latent_shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def encode_prompt(
+ self,
+ device,
+ batch_size,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ prompt=None,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ prompt_embeds_pooled: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds_pooled: Optional[torch.FloatTensor] = None,
+ ):
+ if prompt_embeds is None:
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ attention_mask = text_inputs.attention_mask
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+ attention_mask = attention_mask[:, : self.tokenizer.model_max_length]
+
+ text_encoder_output = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask.to(device), output_hidden_states=True
+ )
+ prompt_embeds = text_encoder_output.hidden_states[-1]
+ if prompt_embeds_pooled is None:
+ prompt_embeds_pooled = text_encoder_output.text_embeds.unsqueeze(1)
+
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+ prompt_embeds_pooled = prompt_embeds_pooled.to(dtype=self.text_encoder.dtype, device=device)
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ prompt_embeds_pooled = prompt_embeds_pooled.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if negative_prompt_embeds is None and do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=uncond_input.attention_mask.to(device),
+ output_hidden_states=True,
+ )
+
+ negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.hidden_states[-1]
+ negative_prompt_embeds_pooled = negative_prompt_embeds_text_encoder_output.text_embeds.unsqueeze(1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ seq_len = negative_prompt_embeds_pooled.shape[1]
+ negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.to(
+ dtype=self.text_encoder.dtype, device=device
+ )
+ negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds_pooled = negative_prompt_embeds_pooled.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ # done duplicates
+
+ return prompt_embeds, prompt_embeds_pooled, negative_prompt_embeds, negative_prompt_embeds_pooled
+
+ def encode_image(self, images, device, dtype, batch_size, num_images_per_prompt):
+ image_embeds = []
+ for image in images:
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+ image = image.to(device=device, dtype=dtype)
+ image_embed = self.image_encoder(image).image_embeds.unsqueeze(1)
+ image_embeds.append(image_embed)
+ image_embeds = torch.cat(image_embeds, dim=1)
+
+ image_embeds = image_embeds.repeat(batch_size * num_images_per_prompt, 1, 1)
+ negative_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, negative_image_embeds
+
+ def check_inputs(
+ self,
+ prompt,
+ images=None,
+ image_embeds=None,
+ negative_prompt=None,
+ prompt_embeds=None,
+ prompt_embeds_pooled=None,
+ negative_prompt_embeds=None,
+ negative_prompt_embeds_pooled=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if prompt_embeds_pooled is not None and negative_prompt_embeds_pooled is not None:
+ if prompt_embeds_pooled.shape != negative_prompt_embeds_pooled.shape:
+ raise ValueError(
+ "`prompt_embeds_pooled` and `negative_prompt_embeds_pooled` must have the same shape when passed"
+ f"directly, but got: `prompt_embeds_pooled` {prompt_embeds_pooled.shape} !="
+ f"`negative_prompt_embeds_pooled` {negative_prompt_embeds_pooled.shape}."
+ )
+
+ if image_embeds is not None and images is not None:
+ raise ValueError(
+ f"Cannot forward both `images`: {images} and `image_embeds`: {image_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+
+ if images:
+ for i, image in enumerate(images):
+ if not isinstance(image, torch.Tensor) and not isinstance(image, PIL.Image.Image):
+ raise TypeError(
+ f"'images' must contain images of type 'torch.Tensor' or 'PIL.Image.Image, but got"
+ f"{type(image)} for image number {i}."
+ )
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ def get_t_condioning(self, t, alphas_cumprod):
+ s = torch.tensor([0.003])
+ clamp_range = [0, 1]
+ min_var = torch.cos(s / (1 + s) * torch.pi * 0.5) ** 2
+ var = alphas_cumprod[t]
+ var = var.clamp(*clamp_range)
+ s, min_var = s.to(var.device), min_var.to(var.device)
+ ratio = (((var * min_var) ** 0.5).acos() / (torch.pi * 0.5)) * (1 + s) - s
+ return ratio
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ images: Union[torch.Tensor, PIL.Image.Image, List[torch.Tensor], List[PIL.Image.Image]] = None,
+ height: int = 1024,
+ width: int = 1024,
+ num_inference_steps: int = 20,
+ timesteps: List[float] = None,
+ guidance_scale: float = 4.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ prompt_embeds_pooled: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds_pooled: Optional[torch.FloatTensor] = None,
+ image_embeds: Optional[torch.FloatTensor] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pt",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ height (`int`, *optional*, defaults to 1024):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 1024):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 60):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 8.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `decoder_guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting
+ `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely
+ linked to the text `prompt`, usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `decoder_guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ prompt_embeds_pooled (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ negative_prompt_embeds_pooled (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds_pooled will be generated from `negative_prompt` input
+ argument.
+ image_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated image embeddings. Can be used to easily tweak image inputs, *e.g.* prompt weighting.
+ If not provided, image embeddings will be generated from `image` input argument if existing.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`StableCascadePriorPipelineOutput`] or `tuple` [`StableCascadePriorPipelineOutput`] if
+ `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
+ generated image embeddings.
+ """
+
+ # 0. Define commonly used variables
+ device = self._execution_device
+ dtype = next(self.prior.parameters()).dtype
+ self._guidance_scale = guidance_scale
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ images=images,
+ image_embeds=image_embeds,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ prompt_embeds_pooled=prompt_embeds_pooled,
+ negative_prompt_embeds=negative_prompt_embeds,
+ negative_prompt_embeds_pooled=negative_prompt_embeds_pooled,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ )
+
+ # 2. Encode caption + images
+ (
+ prompt_embeds,
+ prompt_embeds_pooled,
+ negative_prompt_embeds,
+ negative_prompt_embeds_pooled,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ batch_size=batch_size,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ prompt_embeds_pooled=prompt_embeds_pooled,
+ negative_prompt_embeds=negative_prompt_embeds,
+ negative_prompt_embeds_pooled=negative_prompt_embeds_pooled,
+ )
+
+ if images is not None:
+ image_embeds_pooled, uncond_image_embeds_pooled = self.encode_image(
+ images=images,
+ device=device,
+ dtype=dtype,
+ batch_size=batch_size,
+ num_images_per_prompt=num_images_per_prompt,
+ )
+ elif image_embeds is not None:
+ image_embeds_pooled = image_embeds.repeat(batch_size * num_images_per_prompt, 1, 1)
+ uncond_image_embeds_pooled = torch.zeros_like(image_embeds_pooled)
+ else:
+ image_embeds_pooled = torch.zeros(
+ batch_size * num_images_per_prompt,
+ 1,
+ self.prior.config.clip_image_in_channels,
+ device=device,
+ dtype=dtype,
+ )
+ uncond_image_embeds_pooled = torch.zeros(
+ batch_size * num_images_per_prompt,
+ 1,
+ self.prior.config.clip_image_in_channels,
+ device=device,
+ dtype=dtype,
+ )
+
+ if self.do_classifier_free_guidance:
+ image_embeds = torch.cat([image_embeds_pooled, uncond_image_embeds_pooled], dim=0)
+ else:
+ image_embeds = image_embeds_pooled
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ text_encoder_hidden_states = (
+ torch.cat([prompt_embeds, negative_prompt_embeds]) if negative_prompt_embeds is not None else prompt_embeds
+ )
+ text_encoder_pooled = (
+ torch.cat([prompt_embeds_pooled, negative_prompt_embeds_pooled])
+ if negative_prompt_embeds is not None
+ else prompt_embeds_pooled
+ )
+
+ # 4. Prepare and set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latents
+ latents = self.prepare_latents(
+ batch_size, height, width, num_images_per_prompt, dtype, device, generator, latents, self.scheduler
+ )
+
+ if isinstance(self.scheduler, DDPMWuerstchenScheduler):
+ timesteps = timesteps[:-1]
+ else:
+ if self.scheduler.config.clip_sample:
+ self.scheduler.config.clip_sample = False # disample sample clipping
+ logger.warning(" set `clip_sample` to be False")
+ # 6. Run denoising loop
+ if hasattr(self.scheduler, "betas"):
+ alphas = 1.0 - self.scheduler.betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ else:
+ alphas_cumprod = []
+
+ self._num_timesteps = len(timesteps)
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ if not isinstance(self.scheduler, DDPMWuerstchenScheduler):
+ if len(alphas_cumprod) > 0:
+ timestep_ratio = self.get_t_condioning(t.long().cpu(), alphas_cumprod)
+ timestep_ratio = timestep_ratio.expand(latents.size(0)).to(dtype).to(device)
+ else:
+ timestep_ratio = t.float().div(self.scheduler.timesteps[-1]).expand(latents.size(0)).to(dtype)
+ else:
+ timestep_ratio = t.expand(latents.size(0)).to(dtype)
+ # 7. Denoise image embeddings
+ predicted_image_embedding = self.prior(
+ sample=torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents,
+ timestep_ratio=torch.cat([timestep_ratio] * 2) if self.do_classifier_free_guidance else timestep_ratio,
+ clip_text_pooled=text_encoder_pooled,
+ clip_text=text_encoder_hidden_states,
+ clip_img=image_embeds,
+ return_dict=False,
+ )[0]
+
+ # 8. Check for classifier free guidance and apply it
+ if self.do_classifier_free_guidance:
+ predicted_image_embedding_text, predicted_image_embedding_uncond = predicted_image_embedding.chunk(2)
+ predicted_image_embedding = torch.lerp(
+ predicted_image_embedding_uncond, predicted_image_embedding_text, self.guidance_scale
+ )
+
+ # 9. Renoise latents to next timestep
+ if not isinstance(self.scheduler, DDPMWuerstchenScheduler):
+ timestep_ratio = t
+ latents = self.scheduler.step(
+ model_output=predicted_image_embedding, timestep=timestep_ratio, sample=latents, generator=generator
+ ).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if output_type == "np":
+ latents = latents.cpu().float().numpy() # float() as bfloat16-> numpy doesnt work
+ prompt_embeds = prompt_embeds.cpu().float().numpy() # float() as bfloat16-> numpy doesnt work
+ negative_prompt_embeds = (
+ negative_prompt_embeds.cpu().float().numpy() if negative_prompt_embeds is not None else None
+ ) # float() as bfloat16-> numpy doesnt work
+
+ if not return_dict:
+ return (latents, prompt_embeds, negative_prompt_embeds)
+
+ return StableCascadePriorPipelineOutput(latents, prompt_embeds, negative_prompt_embeds)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/README.md b/model/diffusers_c/pipelines/stable_diffusion/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..5b6424308f020ef901aab854b901abaf59b23e37
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/README.md
@@ -0,0 +1,176 @@
+# Stable Diffusion
+
+## Overview
+
+Stable Diffusion was proposed in [Stable Diffusion Announcement](https://stability.ai/blog/stable-diffusion-announcement) by Patrick Esser and Robin Rombach and the Stability AI team.
+
+The summary of the model is the following:
+
+*Stable Diffusion is a text-to-image model that will empower billions of people to create stunning art within seconds. It is a breakthrough in speed and quality meaning that it can run on consumer GPUs. You can see some of the amazing output that has been created by this model without pre or post-processing on this page. The model itself builds upon the work of the team at CompVis and Runway in their widely used latent diffusion model combined with insights from the conditional diffusion models by our lead generative AI developer Katherine Crowson, Dall-E 2 by Open AI, Imagen by Google Brain and many others. We are delighted that AI media generation is a cooperative field and hope it can continue this way to bring the gift of creativity to all.*
+
+## Tips:
+
+- Stable Diffusion has the same architecture as [Latent Diffusion](https://arxiv.org/abs/2112.10752) but uses a frozen CLIP Text Encoder instead of training the text encoder jointly with the diffusion model.
+- An in-detail explanation of the Stable Diffusion model can be found under [Stable Diffusion with 🧨 Diffusers](https://huggingface.co/blog/stable_diffusion).
+- If you don't want to rely on the Hugging Face Hub and having to pass a authentication token, you can
+download the weights with `git lfs install; git clone https://huggingface.co/runwayml/stable-diffusion-v1-5` and instead pass the local path to the cloned folder to `from_pretrained` as shown below.
+- Stable Diffusion can work with a variety of different samplers as is shown below.
+
+## Available Pipelines:
+
+| Pipeline | Tasks | Colab
+|---|---|:---:|
+| [pipeline_stable_diffusion.py](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py) | *Text-to-Image Generation* | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/training_example.ipynb)
+| [pipeline_stable_diffusion_img2img](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py) | *Image-to-Image Text-Guided Generation* | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/image_2_image_using_diffusers.ipynb)
+| [pipeline_stable_diffusion_inpaint](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py) | *Text-Guided Image Inpainting* | [](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/in_painting_with_stable_diffusion_using_diffusers.ipynb)
+
+## Examples:
+
+### Using Stable Diffusion without being logged into the Hub.
+
+If you want to download the model weights using a single Python line, you need to be logged in via `huggingface-cli login`.
+
+```python
+from diffusers import DiffusionPipeline
+
+pipeline = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
+```
+
+This however can make it difficult to build applications on top of `diffusers` as you will always have to pass the token around. A potential way to solve this issue is by downloading the weights to a local path `"./stable-diffusion-v1-5"`:
+
+```
+git lfs install
+git clone https://huggingface.co/runwayml/stable-diffusion-v1-5
+```
+
+and simply passing the local path to `from_pretrained`:
+
+```python
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained("./stable-diffusion-v1-5")
+```
+
+### Text-to-Image with default PLMS scheduler
+
+```python
+# make sure you're logged in with `huggingface-cli login`
+from diffusers import StableDiffusionPipeline
+
+pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5")
+pipe = pipe.to("cuda")
+
+prompt = "a photo of an astronaut riding a horse on mars"
+image = pipe(prompt).images[0]
+
+image.save("astronaut_rides_horse.png")
+```
+
+### Text-to-Image with DDIM scheduler
+
+```python
+# make sure you're logged in with `huggingface-cli login`
+from diffusers import StableDiffusionPipeline, DDIMScheduler
+
+scheduler = DDIMScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler")
+
+pipe = StableDiffusionPipeline.from_pretrained(
+ "runwayml/stable-diffusion-v1-5",
+ scheduler=scheduler,
+).to("cuda")
+
+prompt = "a photo of an astronaut riding a horse on mars"
+image = pipe(prompt).images[0]
+
+image.save("astronaut_rides_horse.png")
+```
+
+### Text-to-Image with K-LMS scheduler
+
+```python
+# make sure you're logged in with `huggingface-cli login`
+from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler
+
+lms = LMSDiscreteScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler")
+
+pipe = StableDiffusionPipeline.from_pretrained(
+ "runwayml/stable-diffusion-v1-5",
+ scheduler=lms,
+).to("cuda")
+
+prompt = "a photo of an astronaut riding a horse on mars"
+image = pipe(prompt).images[0]
+
+image.save("astronaut_rides_horse.png")
+```
+
+### CycleDiffusion using Stable Diffusion and DDIM scheduler
+
+```python
+import requests
+import torch
+from PIL import Image
+from io import BytesIO
+
+from diffusers import CycleDiffusionPipeline, DDIMScheduler
+
+
+# load the scheduler. CycleDiffusion only supports stochastic schedulers.
+
+# load the pipeline
+# make sure you're logged in with `huggingface-cli login`
+model_id_or_path = "CompVis/stable-diffusion-v1-4"
+scheduler = DDIMScheduler.from_pretrained(model_id_or_path, subfolder="scheduler")
+pipe = CycleDiffusionPipeline.from_pretrained(model_id_or_path, scheduler=scheduler).to("cuda")
+
+# let's download an initial image
+url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/An%20astronaut%20riding%20a%20horse.png"
+response = requests.get(url)
+init_image = Image.open(BytesIO(response.content)).convert("RGB")
+init_image = init_image.resize((512, 512))
+init_image.save("horse.png")
+
+# let's specify a prompt
+source_prompt = "An astronaut riding a horse"
+prompt = "An astronaut riding an elephant"
+
+# call the pipeline
+image = pipe(
+ prompt=prompt,
+ source_prompt=source_prompt,
+ image=init_image,
+ num_inference_steps=100,
+ eta=0.1,
+ strength=0.8,
+ guidance_scale=2,
+ source_guidance_scale=1,
+).images[0]
+
+image.save("horse_to_elephant.png")
+
+# let's try another example
+# See more samples at the original repo: https://github.com/ChenWu98/cycle-diffusion
+url = "https://raw.githubusercontent.com/ChenWu98/cycle-diffusion/main/data/dalle2/A%20black%20colored%20car.png"
+response = requests.get(url)
+init_image = Image.open(BytesIO(response.content)).convert("RGB")
+init_image = init_image.resize((512, 512))
+init_image.save("black.png")
+
+source_prompt = "A black colored car"
+prompt = "A blue colored car"
+
+# call the pipeline
+torch.manual_seed(0)
+image = pipe(
+ prompt=prompt,
+ source_prompt=source_prompt,
+ image=init_image,
+ num_inference_steps=100,
+ eta=0.1,
+ strength=0.85,
+ guidance_scale=3,
+ source_guidance_scale=1,
+).images[0]
+
+image.save("black_to_blue.png")
+```
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__init__.py b/model/diffusers_c/pipelines/stable_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..0eda32d333b9115b2e1f4b9ccbbda7157b915142
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/__init__.py
@@ -0,0 +1,203 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_flax_available,
+ is_k_diffusion_available,
+ is_k_diffusion_version,
+ is_onnx_available,
+ is_torch_available,
+ is_transformers_available,
+ is_transformers_version,
+)
+
+
+_dummy_objects = {}
+_additional_imports = {}
+_import_structure = {"pipeline_output": ["StableDiffusionPipelineOutput"]}
+
+if is_transformers_available() and is_flax_available():
+ _import_structure["pipeline_output"].extend(["FlaxStableDiffusionPipelineOutput"])
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["clip_image_project_model"] = ["CLIPImageProjection"]
+ _import_structure["pipeline_cycle_diffusion"] = ["CycleDiffusionPipeline"]
+ _import_structure["pipeline_stable_diffusion"] = ["StableDiffusionPipeline"]
+ _import_structure["pipeline_stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"]
+ _import_structure["pipeline_stable_diffusion_gligen"] = ["StableDiffusionGLIGENPipeline"]
+ _import_structure["pipeline_stable_diffusion_gligen_text_image"] = ["StableDiffusionGLIGENTextImagePipeline"]
+ _import_structure["pipeline_stable_diffusion_img2img"] = ["StableDiffusionImg2ImgPipeline"]
+ _import_structure["pipeline_stable_diffusion_inpaint"] = ["StableDiffusionInpaintPipeline"]
+ _import_structure["pipeline_stable_diffusion_inpaint_legacy"] = ["StableDiffusionInpaintPipelineLegacy"]
+ _import_structure["pipeline_stable_diffusion_instruct_pix2pix"] = ["StableDiffusionInstructPix2PixPipeline"]
+ _import_structure["pipeline_stable_diffusion_latent_upscale"] = ["StableDiffusionLatentUpscalePipeline"]
+ _import_structure["pipeline_stable_diffusion_model_editing"] = ["StableDiffusionModelEditingPipeline"]
+ _import_structure["pipeline_stable_diffusion_paradigms"] = ["StableDiffusionParadigmsPipeline"]
+ _import_structure["pipeline_stable_diffusion_upscale"] = ["StableDiffusionUpscalePipeline"]
+ _import_structure["pipeline_stable_unclip"] = ["StableUnCLIPPipeline"]
+ _import_structure["pipeline_stable_unclip_img2img"] = ["StableUnCLIPImg2ImgPipeline"]
+ _import_structure["safety_checker"] = ["StableDiffusionSafetyChecker"]
+ _import_structure["stable_unclip_image_normalizer"] = ["StableUnCLIPImageNormalizer"]
+try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ StableDiffusionImageVariationPipeline,
+ )
+
+ _dummy_objects.update({"StableDiffusionImageVariationPipeline": StableDiffusionImageVariationPipeline})
+else:
+ _import_structure["pipeline_stable_diffusion_image_variation"] = ["StableDiffusionImageVariationPipeline"]
+try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.26.0")):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ StableDiffusionDepth2ImgPipeline,
+ )
+
+ _dummy_objects.update(
+ {
+ "StableDiffusionDepth2ImgPipeline": StableDiffusionDepth2ImgPipeline,
+ }
+ )
+else:
+ _import_structure["pipeline_stable_diffusion_depth2img"] = ["StableDiffusionDepth2ImgPipeline"]
+
+try:
+ if not (is_transformers_available() and is_onnx_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_onnx_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_onnx_objects))
+else:
+ _import_structure["pipeline_onnx_stable_diffusion"] = [
+ "OnnxStableDiffusionPipeline",
+ "StableDiffusionOnnxPipeline",
+ ]
+ _import_structure["pipeline_onnx_stable_diffusion_img2img"] = ["OnnxStableDiffusionImg2ImgPipeline"]
+ _import_structure["pipeline_onnx_stable_diffusion_inpaint"] = ["OnnxStableDiffusionInpaintPipeline"]
+ _import_structure["pipeline_onnx_stable_diffusion_inpaint_legacy"] = ["OnnxStableDiffusionInpaintPipelineLegacy"]
+ _import_structure["pipeline_onnx_stable_diffusion_upscale"] = ["OnnxStableDiffusionUpscalePipeline"]
+
+if is_transformers_available() and is_flax_available():
+ from ...schedulers.scheduling_pndm_flax import PNDMSchedulerState
+
+ _additional_imports.update({"PNDMSchedulerState": PNDMSchedulerState})
+ _import_structure["pipeline_flax_stable_diffusion"] = ["FlaxStableDiffusionPipeline"]
+ _import_structure["pipeline_flax_stable_diffusion_img2img"] = ["FlaxStableDiffusionImg2ImgPipeline"]
+ _import_structure["pipeline_flax_stable_diffusion_inpaint"] = ["FlaxStableDiffusionInpaintPipeline"]
+ _import_structure["safety_checker_flax"] = ["FlaxStableDiffusionSafetyChecker"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+
+ else:
+ from .clip_image_project_model import CLIPImageProjection
+ from .pipeline_stable_diffusion import (
+ StableDiffusionPipeline,
+ StableDiffusionPipelineOutput,
+ StableDiffusionSafetyChecker,
+ )
+ from .pipeline_stable_diffusion_img2img import StableDiffusionImg2ImgPipeline
+ from .pipeline_stable_diffusion_inpaint import StableDiffusionInpaintPipeline
+ from .pipeline_stable_diffusion_instruct_pix2pix import (
+ StableDiffusionInstructPix2PixPipeline,
+ )
+ from .pipeline_stable_diffusion_latent_upscale import (
+ StableDiffusionLatentUpscalePipeline,
+ )
+ from .pipeline_stable_diffusion_upscale import StableDiffusionUpscalePipeline
+ from .pipeline_stable_unclip import StableUnCLIPPipeline
+ from .pipeline_stable_unclip_img2img import StableUnCLIPImg2ImgPipeline
+ from .safety_checker import StableDiffusionSafetyChecker
+ from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
+
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ StableDiffusionImageVariationPipeline,
+ )
+ else:
+ from .pipeline_stable_diffusion_image_variation import (
+ StableDiffusionImageVariationPipeline,
+ )
+
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.26.0")):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import StableDiffusionDepth2ImgPipeline
+ else:
+ from .pipeline_stable_diffusion_depth2img import (
+ StableDiffusionDepth2ImgPipeline,
+ )
+
+ try:
+ if not (is_transformers_available() and is_onnx_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_onnx_objects import *
+ else:
+ from .pipeline_onnx_stable_diffusion import (
+ OnnxStableDiffusionPipeline,
+ StableDiffusionOnnxPipeline,
+ )
+ from .pipeline_onnx_stable_diffusion_img2img import (
+ OnnxStableDiffusionImg2ImgPipeline,
+ )
+ from .pipeline_onnx_stable_diffusion_inpaint import (
+ OnnxStableDiffusionInpaintPipeline,
+ )
+ from .pipeline_onnx_stable_diffusion_upscale import (
+ OnnxStableDiffusionUpscalePipeline,
+ )
+
+ try:
+ if not (is_transformers_available() and is_flax_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_flax_objects import *
+ else:
+ from .pipeline_flax_stable_diffusion import FlaxStableDiffusionPipeline
+ from .pipeline_flax_stable_diffusion_img2img import (
+ FlaxStableDiffusionImg2ImgPipeline,
+ )
+ from .pipeline_flax_stable_diffusion_inpaint import (
+ FlaxStableDiffusionInpaintPipeline,
+ )
+ from .pipeline_output import FlaxStableDiffusionPipelineOutput
+ from .safety_checker_flax import FlaxStableDiffusionSafetyChecker
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
+ for name, value in _additional_imports.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..946a67e8e59ebdaf3ac9cd2b1bb3b91383aec135
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/clip_image_project_model.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/clip_image_project_model.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7e1acd389b9144bb1c85cefe112acc7775552723
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/clip_image_project_model.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_output.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_output.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3862321ec6002207a0750c65864abc45579083d9
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_output.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..23644f87fa0f687bf72f2553dcdf7847ef8ceec9
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_depth2img.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_depth2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3d2c81c5e0f77a8dbebfbe88ca90ddffb1d6008f
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_depth2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_image_variation.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_image_variation.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1834f191a72978f5e1aae1df359414b9ffbfa164
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_image_variation.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..02a8377d191802a6f5affdbe614de60d1c06da6c
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_inpaint.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_inpaint.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..19f72093b7ddf2db506da9a954f4822dd7ac33d9
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_inpaint.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_instruct_pix2pix.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_instruct_pix2pix.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..186030f0b6824eda3cdf7b20167be9f10f611def
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_instruct_pix2pix.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_latent_upscale.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_latent_upscale.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..91c512a9398d780a919de9bc589e2f2c0b2d876c
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_latent_upscale.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_upscale.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_upscale.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5ed20e069a6c54eb8d86110e5c7e5f59a1013bad
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_diffusion_upscale.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..99f8aba0cdc6d98a1933a2ddf829429b63f46211
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b4eac3a72961d808475da6d477dfd4ee8865fd51
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/pipeline_stable_unclip_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/safety_checker.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/safety_checker.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..43b26f4bfe02a8c6d242bc2f1c2e2c0f176f38c7
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/safety_checker.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/__pycache__/stable_unclip_image_normalizer.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/stable_unclip_image_normalizer.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..890bd5bdb325807ab052e815de1a1a99ec24e90f
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion/__pycache__/stable_unclip_image_normalizer.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion/clip_image_project_model.py b/model/diffusers_c/pipelines/stable_diffusion/clip_image_project_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..71f9d9714e6b8b7772888d959ae1ae24fae37f77
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/clip_image_project_model.py
@@ -0,0 +1,29 @@
+# Copyright 2024 The GLIGEN Authors and HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models.modeling_utils import ModelMixin
+
+
+class CLIPImageProjection(ModelMixin, ConfigMixin):
+ @register_to_config
+ def __init__(self, hidden_size: int = 768):
+ super().__init__()
+ self.hidden_size = hidden_size
+ self.project = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
+
+ def forward(self, x):
+ return self.project(x)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/convert_from_ckpt.py b/model/diffusers_c/pipelines/stable_diffusion/convert_from_ckpt.py
new file mode 100644
index 0000000000000000000000000000000000000000..438b270f8cb620aaac05c9a8be316c12bf6c2e22
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/convert_from_ckpt.py
@@ -0,0 +1,1859 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Conversion script for the Stable Diffusion checkpoints."""
+
+import re
+from contextlib import nullcontext
+from io import BytesIO
+from typing import Dict, Optional, Union
+
+import requests
+import torch
+import yaml
+from transformers import (
+ AutoFeatureExtractor,
+ BertTokenizerFast,
+ CLIPImageProcessor,
+ CLIPTextConfig,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionConfig,
+ CLIPVisionModelWithProjection,
+)
+
+from ...models import (
+ AutoencoderKL,
+ ControlNetModel,
+ PriorTransformer,
+ UNet2DConditionModel,
+)
+from ...schedulers import (
+ DDIMScheduler,
+ DDPMScheduler,
+ DPMSolverMultistepScheduler,
+ EulerAncestralDiscreteScheduler,
+ EulerDiscreteScheduler,
+ HeunDiscreteScheduler,
+ LMSDiscreteScheduler,
+ PNDMScheduler,
+ UnCLIPScheduler,
+)
+from ...utils import is_accelerate_available, logging
+from ..latent_diffusion.pipeline_latent_diffusion import LDMBertConfig, LDMBertModel
+from ..paint_by_example import PaintByExampleImageEncoder
+from ..pipeline_utils import DiffusionPipeline
+from .safety_checker import StableDiffusionSafetyChecker
+from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
+
+
+if is_accelerate_available():
+ from accelerate import init_empty_weights
+ from accelerate.utils import set_module_tensor_to_device
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def shave_segments(path, n_shave_prefix_segments=1):
+ """
+ Removes segments. Positive values shave the first segments, negative shave the last segments.
+ """
+ if n_shave_prefix_segments >= 0:
+ return ".".join(path.split(".")[n_shave_prefix_segments:])
+ else:
+ return ".".join(path.split(".")[:n_shave_prefix_segments])
+
+
+def renew_resnet_paths(old_list, n_shave_prefix_segments=0):
+ """
+ Updates paths inside resnets to the new naming scheme (local renaming)
+ """
+ mapping = []
+ for old_item in old_list:
+ new_item = old_item.replace("in_layers.0", "norm1")
+ new_item = new_item.replace("in_layers.2", "conv1")
+
+ new_item = new_item.replace("out_layers.0", "norm2")
+ new_item = new_item.replace("out_layers.3", "conv2")
+
+ new_item = new_item.replace("emb_layers.1", "time_emb_proj")
+ new_item = new_item.replace("skip_connection", "conv_shortcut")
+
+ new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+ mapping.append({"old": old_item, "new": new_item})
+
+ return mapping
+
+
+def renew_vae_resnet_paths(old_list, n_shave_prefix_segments=0):
+ """
+ Updates paths inside resnets to the new naming scheme (local renaming)
+ """
+ mapping = []
+ for old_item in old_list:
+ new_item = old_item
+
+ new_item = new_item.replace("nin_shortcut", "conv_shortcut")
+ new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+ mapping.append({"old": old_item, "new": new_item})
+
+ return mapping
+
+
+def renew_attention_paths(old_list, n_shave_prefix_segments=0):
+ """
+ Updates paths inside attentions to the new naming scheme (local renaming)
+ """
+ mapping = []
+ for old_item in old_list:
+ new_item = old_item
+
+ # new_item = new_item.replace('norm.weight', 'group_norm.weight')
+ # new_item = new_item.replace('norm.bias', 'group_norm.bias')
+
+ # new_item = new_item.replace('proj_out.weight', 'proj_attn.weight')
+ # new_item = new_item.replace('proj_out.bias', 'proj_attn.bias')
+
+ # new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+ mapping.append({"old": old_item, "new": new_item})
+
+ return mapping
+
+
+def renew_vae_attention_paths(old_list, n_shave_prefix_segments=0):
+ """
+ Updates paths inside attentions to the new naming scheme (local renaming)
+ """
+ mapping = []
+ for old_item in old_list:
+ new_item = old_item
+
+ new_item = new_item.replace("norm.weight", "group_norm.weight")
+ new_item = new_item.replace("norm.bias", "group_norm.bias")
+
+ new_item = new_item.replace("q.weight", "to_q.weight")
+ new_item = new_item.replace("q.bias", "to_q.bias")
+
+ new_item = new_item.replace("k.weight", "to_k.weight")
+ new_item = new_item.replace("k.bias", "to_k.bias")
+
+ new_item = new_item.replace("v.weight", "to_v.weight")
+ new_item = new_item.replace("v.bias", "to_v.bias")
+
+ new_item = new_item.replace("proj_out.weight", "to_out.0.weight")
+ new_item = new_item.replace("proj_out.bias", "to_out.0.bias")
+
+ new_item = shave_segments(new_item, n_shave_prefix_segments=n_shave_prefix_segments)
+
+ mapping.append({"old": old_item, "new": new_item})
+
+ return mapping
+
+
+def assign_to_checkpoint(
+ paths, checkpoint, old_checkpoint, attention_paths_to_split=None, additional_replacements=None, config=None
+):
+ """
+ This does the final conversion step: take locally converted weights and apply a global renaming to them. It splits
+ attention layers, and takes into account additional replacements that may arise.
+
+ Assigns the weights to the new checkpoint.
+ """
+ assert isinstance(paths, list), "Paths should be a list of dicts containing 'old' and 'new' keys."
+
+ # Splits the attention layers into three variables.
+ if attention_paths_to_split is not None:
+ for path, path_map in attention_paths_to_split.items():
+ old_tensor = old_checkpoint[path]
+ channels = old_tensor.shape[0] // 3
+
+ target_shape = (-1, channels) if len(old_tensor.shape) == 3 else (-1)
+
+ num_heads = old_tensor.shape[0] // config["num_head_channels"] // 3
+
+ old_tensor = old_tensor.reshape((num_heads, 3 * channels // num_heads) + old_tensor.shape[1:])
+ query, key, value = old_tensor.split(channels // num_heads, dim=1)
+
+ checkpoint[path_map["query"]] = query.reshape(target_shape)
+ checkpoint[path_map["key"]] = key.reshape(target_shape)
+ checkpoint[path_map["value"]] = value.reshape(target_shape)
+
+ for path in paths:
+ new_path = path["new"]
+
+ # These have already been assigned
+ if attention_paths_to_split is not None and new_path in attention_paths_to_split:
+ continue
+
+ # Global renaming happens here
+ new_path = new_path.replace("middle_block.0", "mid_block.resnets.0")
+ new_path = new_path.replace("middle_block.1", "mid_block.attentions.0")
+ new_path = new_path.replace("middle_block.2", "mid_block.resnets.1")
+
+ if additional_replacements is not None:
+ for replacement in additional_replacements:
+ new_path = new_path.replace(replacement["old"], replacement["new"])
+
+ # proj_attn.weight has to be converted from conv 1D to linear
+ is_attn_weight = "proj_attn.weight" in new_path or ("attentions" in new_path and "to_" in new_path)
+ shape = old_checkpoint[path["old"]].shape
+ if is_attn_weight and len(shape) == 3:
+ checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0]
+ elif is_attn_weight and len(shape) == 4:
+ checkpoint[new_path] = old_checkpoint[path["old"]][:, :, 0, 0]
+ else:
+ checkpoint[new_path] = old_checkpoint[path["old"]]
+
+
+def conv_attn_to_linear(checkpoint):
+ keys = list(checkpoint.keys())
+ attn_keys = ["query.weight", "key.weight", "value.weight"]
+ for key in keys:
+ if ".".join(key.split(".")[-2:]) in attn_keys:
+ if checkpoint[key].ndim > 2:
+ checkpoint[key] = checkpoint[key][:, :, 0, 0]
+ elif "proj_attn.weight" in key:
+ if checkpoint[key].ndim > 2:
+ checkpoint[key] = checkpoint[key][:, :, 0]
+
+
+def create_unet_diffusers_config(original_config, image_size: int, controlnet=False):
+ """
+ Creates a config for the diffusers based on the config of the LDM model.
+ """
+ if controlnet:
+ unet_params = original_config["model"]["params"]["control_stage_config"]["params"]
+ else:
+ if (
+ "unet_config" in original_config["model"]["params"]
+ and original_config["model"]["params"]["unet_config"] is not None
+ ):
+ unet_params = original_config["model"]["params"]["unet_config"]["params"]
+ else:
+ unet_params = original_config["model"]["params"]["network_config"]["params"]
+
+ vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"]
+
+ block_out_channels = [unet_params["model_channels"] * mult for mult in unet_params["channel_mult"]]
+
+ down_block_types = []
+ resolution = 1
+ for i in range(len(block_out_channels)):
+ block_type = "CrossAttnDownBlock2D" if resolution in unet_params["attention_resolutions"] else "DownBlock2D"
+ down_block_types.append(block_type)
+ if i != len(block_out_channels) - 1:
+ resolution *= 2
+
+ up_block_types = []
+ for i in range(len(block_out_channels)):
+ block_type = "CrossAttnUpBlock2D" if resolution in unet_params["attention_resolutions"] else "UpBlock2D"
+ up_block_types.append(block_type)
+ resolution //= 2
+
+ if unet_params["transformer_depth"] is not None:
+ transformer_layers_per_block = (
+ unet_params["transformer_depth"]
+ if isinstance(unet_params["transformer_depth"], int)
+ else list(unet_params["transformer_depth"])
+ )
+ else:
+ transformer_layers_per_block = 1
+
+ vae_scale_factor = 2 ** (len(vae_params["ch_mult"]) - 1)
+
+ head_dim = unet_params["num_heads"] if "num_heads" in unet_params else None
+ use_linear_projection = (
+ unet_params["use_linear_in_transformer"] if "use_linear_in_transformer" in unet_params else False
+ )
+ if use_linear_projection:
+ # stable diffusion 2-base-512 and 2-768
+ if head_dim is None:
+ head_dim_mult = unet_params["model_channels"] // unet_params["num_head_channels"]
+ head_dim = [head_dim_mult * c for c in list(unet_params["channel_mult"])]
+
+ class_embed_type = None
+ addition_embed_type = None
+ addition_time_embed_dim = None
+ projection_class_embeddings_input_dim = None
+ context_dim = None
+
+ if unet_params["context_dim"] is not None:
+ context_dim = (
+ unet_params["context_dim"]
+ if isinstance(unet_params["context_dim"], int)
+ else unet_params["context_dim"][0]
+ )
+
+ if "num_classes" in unet_params:
+ if unet_params["num_classes"] == "sequential":
+ if context_dim in [2048, 1280]:
+ # SDXL
+ addition_embed_type = "text_time"
+ addition_time_embed_dim = 256
+ else:
+ class_embed_type = "projection"
+ assert "adm_in_channels" in unet_params
+ projection_class_embeddings_input_dim = unet_params["adm_in_channels"]
+
+ config = {
+ "sample_size": image_size // vae_scale_factor,
+ "in_channels": unet_params["in_channels"],
+ "down_block_types": tuple(down_block_types),
+ "block_out_channels": tuple(block_out_channels),
+ "layers_per_block": unet_params["num_res_blocks"],
+ "cross_attention_dim": context_dim,
+ "attention_head_dim": head_dim,
+ "use_linear_projection": use_linear_projection,
+ "class_embed_type": class_embed_type,
+ "addition_embed_type": addition_embed_type,
+ "addition_time_embed_dim": addition_time_embed_dim,
+ "projection_class_embeddings_input_dim": projection_class_embeddings_input_dim,
+ "transformer_layers_per_block": transformer_layers_per_block,
+ }
+
+ if "disable_self_attentions" in unet_params:
+ config["only_cross_attention"] = unet_params["disable_self_attentions"]
+
+ if "num_classes" in unet_params and isinstance(unet_params["num_classes"], int):
+ config["num_class_embeds"] = unet_params["num_classes"]
+
+ if controlnet:
+ config["conditioning_channels"] = unet_params["hint_channels"]
+ else:
+ config["out_channels"] = unet_params["out_channels"]
+ config["up_block_types"] = tuple(up_block_types)
+
+ return config
+
+
+def create_vae_diffusers_config(original_config, image_size: int):
+ """
+ Creates a config for the diffusers based on the config of the LDM model.
+ """
+ vae_params = original_config["model"]["params"]["first_stage_config"]["params"]["ddconfig"]
+ _ = original_config["model"]["params"]["first_stage_config"]["params"]["embed_dim"]
+
+ block_out_channels = [vae_params["ch"] * mult for mult in vae_params["ch_mult"]]
+ down_block_types = ["DownEncoderBlock2D"] * len(block_out_channels)
+ up_block_types = ["UpDecoderBlock2D"] * len(block_out_channels)
+
+ config = {
+ "sample_size": image_size,
+ "in_channels": vae_params["in_channels"],
+ "out_channels": vae_params["out_ch"],
+ "down_block_types": tuple(down_block_types),
+ "up_block_types": tuple(up_block_types),
+ "block_out_channels": tuple(block_out_channels),
+ "latent_channels": vae_params["z_channels"],
+ "layers_per_block": vae_params["num_res_blocks"],
+ }
+ return config
+
+
+def create_diffusers_schedular(original_config):
+ schedular = DDIMScheduler(
+ num_train_timesteps=original_config["model"]["params"]["timesteps"],
+ beta_start=original_config["model"]["params"]["linear_start"],
+ beta_end=original_config["model"]["params"]["linear_end"],
+ beta_schedule="scaled_linear",
+ )
+ return schedular
+
+
+def create_ldm_bert_config(original_config):
+ bert_params = original_config["model"]["params"]["cond_stage_config"]["params"]
+ config = LDMBertConfig(
+ d_model=bert_params.n_embed,
+ encoder_layers=bert_params.n_layer,
+ encoder_ffn_dim=bert_params.n_embed * 4,
+ )
+ return config
+
+
+def convert_ldm_unet_checkpoint(
+ checkpoint, config, path=None, extract_ema=False, controlnet=False, skip_extract_state_dict=False
+):
+ """
+ Takes a state dict and a config, and returns a converted checkpoint.
+ """
+
+ if skip_extract_state_dict:
+ unet_state_dict = checkpoint
+ else:
+ # extract state_dict for UNet
+ unet_state_dict = {}
+ keys = list(checkpoint.keys())
+
+ if controlnet:
+ unet_key = "control_model."
+ else:
+ unet_key = "model.diffusion_model."
+
+ # at least a 100 parameters have to start with `model_ema` in order for the checkpoint to be EMA
+ if sum(k.startswith("model_ema") for k in keys) > 100 and extract_ema:
+ logger.warning(f"Checkpoint {path} has both EMA and non-EMA weights.")
+ logger.warning(
+ "In this conversion only the EMA weights are extracted. If you want to instead extract the non-EMA"
+ " weights (useful to continue fine-tuning), please make sure to remove the `--extract_ema` flag."
+ )
+ for key in keys:
+ if key.startswith("model.diffusion_model"):
+ flat_ema_key = "model_ema." + "".join(key.split(".")[1:])
+ unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(flat_ema_key)
+ else:
+ if sum(k.startswith("model_ema") for k in keys) > 100:
+ logger.warning(
+ "In this conversion only the non-EMA weights are extracted. If you want to instead extract the EMA"
+ " weights (usually better for inference), please make sure to add the `--extract_ema` flag."
+ )
+
+ for key in keys:
+ if key.startswith(unet_key):
+ unet_state_dict[key.replace(unet_key, "")] = checkpoint.pop(key)
+
+ new_checkpoint = {}
+
+ new_checkpoint["time_embedding.linear_1.weight"] = unet_state_dict["time_embed.0.weight"]
+ new_checkpoint["time_embedding.linear_1.bias"] = unet_state_dict["time_embed.0.bias"]
+ new_checkpoint["time_embedding.linear_2.weight"] = unet_state_dict["time_embed.2.weight"]
+ new_checkpoint["time_embedding.linear_2.bias"] = unet_state_dict["time_embed.2.bias"]
+
+ if config["class_embed_type"] is None:
+ # No parameters to port
+ ...
+ elif config["class_embed_type"] == "timestep" or config["class_embed_type"] == "projection":
+ new_checkpoint["class_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"]
+ new_checkpoint["class_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"]
+ new_checkpoint["class_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"]
+ new_checkpoint["class_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]
+ else:
+ raise NotImplementedError(f"Not implemented `class_embed_type`: {config['class_embed_type']}")
+
+ if config["addition_embed_type"] == "text_time":
+ new_checkpoint["add_embedding.linear_1.weight"] = unet_state_dict["label_emb.0.0.weight"]
+ new_checkpoint["add_embedding.linear_1.bias"] = unet_state_dict["label_emb.0.0.bias"]
+ new_checkpoint["add_embedding.linear_2.weight"] = unet_state_dict["label_emb.0.2.weight"]
+ new_checkpoint["add_embedding.linear_2.bias"] = unet_state_dict["label_emb.0.2.bias"]
+
+ # Relevant to StableDiffusionUpscalePipeline
+ if "num_class_embeds" in config:
+ if (config["num_class_embeds"] is not None) and ("label_emb.weight" in unet_state_dict):
+ new_checkpoint["class_embedding.weight"] = unet_state_dict["label_emb.weight"]
+
+ new_checkpoint["conv_in.weight"] = unet_state_dict["input_blocks.0.0.weight"]
+ new_checkpoint["conv_in.bias"] = unet_state_dict["input_blocks.0.0.bias"]
+
+ if not controlnet:
+ new_checkpoint["conv_norm_out.weight"] = unet_state_dict["out.0.weight"]
+ new_checkpoint["conv_norm_out.bias"] = unet_state_dict["out.0.bias"]
+ new_checkpoint["conv_out.weight"] = unet_state_dict["out.2.weight"]
+ new_checkpoint["conv_out.bias"] = unet_state_dict["out.2.bias"]
+
+ # Retrieves the keys for the input blocks only
+ num_input_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "input_blocks" in layer})
+ input_blocks = {
+ layer_id: [key for key in unet_state_dict if f"input_blocks.{layer_id}" in key]
+ for layer_id in range(num_input_blocks)
+ }
+
+ # Retrieves the keys for the middle blocks only
+ num_middle_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "middle_block" in layer})
+ middle_blocks = {
+ layer_id: [key for key in unet_state_dict if f"middle_block.{layer_id}" in key]
+ for layer_id in range(num_middle_blocks)
+ }
+
+ # Retrieves the keys for the output blocks only
+ num_output_blocks = len({".".join(layer.split(".")[:2]) for layer in unet_state_dict if "output_blocks" in layer})
+ output_blocks = {
+ layer_id: [key for key in unet_state_dict if f"output_blocks.{layer_id}" in key]
+ for layer_id in range(num_output_blocks)
+ }
+
+ for i in range(1, num_input_blocks):
+ block_id = (i - 1) // (config["layers_per_block"] + 1)
+ layer_in_block_id = (i - 1) % (config["layers_per_block"] + 1)
+
+ resnets = [
+ key for key in input_blocks[i] if f"input_blocks.{i}.0" in key and f"input_blocks.{i}.0.op" not in key
+ ]
+ attentions = [key for key in input_blocks[i] if f"input_blocks.{i}.1" in key]
+
+ if f"input_blocks.{i}.0.op.weight" in unet_state_dict:
+ new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.weight"] = unet_state_dict.pop(
+ f"input_blocks.{i}.0.op.weight"
+ )
+ new_checkpoint[f"down_blocks.{block_id}.downsamplers.0.conv.bias"] = unet_state_dict.pop(
+ f"input_blocks.{i}.0.op.bias"
+ )
+
+ paths = renew_resnet_paths(resnets)
+ meta_path = {"old": f"input_blocks.{i}.0", "new": f"down_blocks.{block_id}.resnets.{layer_in_block_id}"}
+ assign_to_checkpoint(
+ paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+ )
+
+ if len(attentions):
+ paths = renew_attention_paths(attentions)
+
+ meta_path = {"old": f"input_blocks.{i}.1", "new": f"down_blocks.{block_id}.attentions.{layer_in_block_id}"}
+ assign_to_checkpoint(
+ paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+ )
+
+ resnet_0 = middle_blocks[0]
+ attentions = middle_blocks[1]
+ resnet_1 = middle_blocks[2]
+
+ resnet_0_paths = renew_resnet_paths(resnet_0)
+ assign_to_checkpoint(resnet_0_paths, new_checkpoint, unet_state_dict, config=config)
+
+ resnet_1_paths = renew_resnet_paths(resnet_1)
+ assign_to_checkpoint(resnet_1_paths, new_checkpoint, unet_state_dict, config=config)
+
+ attentions_paths = renew_attention_paths(attentions)
+ meta_path = {"old": "middle_block.1", "new": "mid_block.attentions.0"}
+ assign_to_checkpoint(
+ attentions_paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+ )
+
+ for i in range(num_output_blocks):
+ block_id = i // (config["layers_per_block"] + 1)
+ layer_in_block_id = i % (config["layers_per_block"] + 1)
+ output_block_layers = [shave_segments(name, 2) for name in output_blocks[i]]
+ output_block_list = {}
+
+ for layer in output_block_layers:
+ layer_id, layer_name = layer.split(".")[0], shave_segments(layer, 1)
+ if layer_id in output_block_list:
+ output_block_list[layer_id].append(layer_name)
+ else:
+ output_block_list[layer_id] = [layer_name]
+
+ if len(output_block_list) > 1:
+ resnets = [key for key in output_blocks[i] if f"output_blocks.{i}.0" in key]
+ attentions = [key for key in output_blocks[i] if f"output_blocks.{i}.1" in key]
+
+ resnet_0_paths = renew_resnet_paths(resnets)
+ paths = renew_resnet_paths(resnets)
+
+ meta_path = {"old": f"output_blocks.{i}.0", "new": f"up_blocks.{block_id}.resnets.{layer_in_block_id}"}
+ assign_to_checkpoint(
+ paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+ )
+
+ output_block_list = {k: sorted(v) for k, v in output_block_list.items()}
+ if ["conv.bias", "conv.weight"] in output_block_list.values():
+ index = list(output_block_list.values()).index(["conv.bias", "conv.weight"])
+ new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.weight"] = unet_state_dict[
+ f"output_blocks.{i}.{index}.conv.weight"
+ ]
+ new_checkpoint[f"up_blocks.{block_id}.upsamplers.0.conv.bias"] = unet_state_dict[
+ f"output_blocks.{i}.{index}.conv.bias"
+ ]
+
+ # Clear attentions as they have been attributed above.
+ if len(attentions) == 2:
+ attentions = []
+
+ if len(attentions):
+ paths = renew_attention_paths(attentions)
+ meta_path = {
+ "old": f"output_blocks.{i}.1",
+ "new": f"up_blocks.{block_id}.attentions.{layer_in_block_id}",
+ }
+ assign_to_checkpoint(
+ paths, new_checkpoint, unet_state_dict, additional_replacements=[meta_path], config=config
+ )
+ else:
+ resnet_0_paths = renew_resnet_paths(output_block_layers, n_shave_prefix_segments=1)
+ for path in resnet_0_paths:
+ old_path = ".".join(["output_blocks", str(i), path["old"]])
+ new_path = ".".join(["up_blocks", str(block_id), "resnets", str(layer_in_block_id), path["new"]])
+
+ new_checkpoint[new_path] = unet_state_dict[old_path]
+
+ if controlnet:
+ # conditioning embedding
+
+ orig_index = 0
+
+ new_checkpoint["controlnet_cond_embedding.conv_in.weight"] = unet_state_dict.pop(
+ f"input_hint_block.{orig_index}.weight"
+ )
+ new_checkpoint["controlnet_cond_embedding.conv_in.bias"] = unet_state_dict.pop(
+ f"input_hint_block.{orig_index}.bias"
+ )
+
+ orig_index += 2
+
+ diffusers_index = 0
+
+ while diffusers_index < 6:
+ new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_index}.weight"] = unet_state_dict.pop(
+ f"input_hint_block.{orig_index}.weight"
+ )
+ new_checkpoint[f"controlnet_cond_embedding.blocks.{diffusers_index}.bias"] = unet_state_dict.pop(
+ f"input_hint_block.{orig_index}.bias"
+ )
+ diffusers_index += 1
+ orig_index += 2
+
+ new_checkpoint["controlnet_cond_embedding.conv_out.weight"] = unet_state_dict.pop(
+ f"input_hint_block.{orig_index}.weight"
+ )
+ new_checkpoint["controlnet_cond_embedding.conv_out.bias"] = unet_state_dict.pop(
+ f"input_hint_block.{orig_index}.bias"
+ )
+
+ # down blocks
+ for i in range(num_input_blocks):
+ new_checkpoint[f"controlnet_down_blocks.{i}.weight"] = unet_state_dict.pop(f"zero_convs.{i}.0.weight")
+ new_checkpoint[f"controlnet_down_blocks.{i}.bias"] = unet_state_dict.pop(f"zero_convs.{i}.0.bias")
+
+ # mid block
+ new_checkpoint["controlnet_mid_block.weight"] = unet_state_dict.pop("middle_block_out.0.weight")
+ new_checkpoint["controlnet_mid_block.bias"] = unet_state_dict.pop("middle_block_out.0.bias")
+
+ return new_checkpoint
+
+
+def convert_ldm_vae_checkpoint(checkpoint, config):
+ # extract state dict for VAE
+ vae_state_dict = {}
+ keys = list(checkpoint.keys())
+ vae_key = "first_stage_model." if any(k.startswith("first_stage_model.") for k in keys) else ""
+ for key in keys:
+ if key.startswith(vae_key):
+ vae_state_dict[key.replace(vae_key, "")] = checkpoint.get(key)
+
+ new_checkpoint = {}
+
+ new_checkpoint["encoder.conv_in.weight"] = vae_state_dict["encoder.conv_in.weight"]
+ new_checkpoint["encoder.conv_in.bias"] = vae_state_dict["encoder.conv_in.bias"]
+ new_checkpoint["encoder.conv_out.weight"] = vae_state_dict["encoder.conv_out.weight"]
+ new_checkpoint["encoder.conv_out.bias"] = vae_state_dict["encoder.conv_out.bias"]
+ new_checkpoint["encoder.conv_norm_out.weight"] = vae_state_dict["encoder.norm_out.weight"]
+ new_checkpoint["encoder.conv_norm_out.bias"] = vae_state_dict["encoder.norm_out.bias"]
+
+ new_checkpoint["decoder.conv_in.weight"] = vae_state_dict["decoder.conv_in.weight"]
+ new_checkpoint["decoder.conv_in.bias"] = vae_state_dict["decoder.conv_in.bias"]
+ new_checkpoint["decoder.conv_out.weight"] = vae_state_dict["decoder.conv_out.weight"]
+ new_checkpoint["decoder.conv_out.bias"] = vae_state_dict["decoder.conv_out.bias"]
+ new_checkpoint["decoder.conv_norm_out.weight"] = vae_state_dict["decoder.norm_out.weight"]
+ new_checkpoint["decoder.conv_norm_out.bias"] = vae_state_dict["decoder.norm_out.bias"]
+
+ new_checkpoint["quant_conv.weight"] = vae_state_dict["quant_conv.weight"]
+ new_checkpoint["quant_conv.bias"] = vae_state_dict["quant_conv.bias"]
+ new_checkpoint["post_quant_conv.weight"] = vae_state_dict["post_quant_conv.weight"]
+ new_checkpoint["post_quant_conv.bias"] = vae_state_dict["post_quant_conv.bias"]
+
+ # Retrieves the keys for the encoder down blocks only
+ num_down_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "encoder.down" in layer})
+ down_blocks = {
+ layer_id: [key for key in vae_state_dict if f"down.{layer_id}" in key] for layer_id in range(num_down_blocks)
+ }
+
+ # Retrieves the keys for the decoder up blocks only
+ num_up_blocks = len({".".join(layer.split(".")[:3]) for layer in vae_state_dict if "decoder.up" in layer})
+ up_blocks = {
+ layer_id: [key for key in vae_state_dict if f"up.{layer_id}" in key] for layer_id in range(num_up_blocks)
+ }
+
+ for i in range(num_down_blocks):
+ resnets = [key for key in down_blocks[i] if f"down.{i}" in key and f"down.{i}.downsample" not in key]
+
+ if f"encoder.down.{i}.downsample.conv.weight" in vae_state_dict:
+ new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.weight"] = vae_state_dict.pop(
+ f"encoder.down.{i}.downsample.conv.weight"
+ )
+ new_checkpoint[f"encoder.down_blocks.{i}.downsamplers.0.conv.bias"] = vae_state_dict.pop(
+ f"encoder.down.{i}.downsample.conv.bias"
+ )
+
+ paths = renew_vae_resnet_paths(resnets)
+ meta_path = {"old": f"down.{i}.block", "new": f"down_blocks.{i}.resnets"}
+ assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+ mid_resnets = [key for key in vae_state_dict if "encoder.mid.block" in key]
+ num_mid_res_blocks = 2
+ for i in range(1, num_mid_res_blocks + 1):
+ resnets = [key for key in mid_resnets if f"encoder.mid.block_{i}" in key]
+
+ paths = renew_vae_resnet_paths(resnets)
+ meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
+ assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+ mid_attentions = [key for key in vae_state_dict if "encoder.mid.attn" in key]
+ paths = renew_vae_attention_paths(mid_attentions)
+ meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+ assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+ conv_attn_to_linear(new_checkpoint)
+
+ for i in range(num_up_blocks):
+ block_id = num_up_blocks - 1 - i
+ resnets = [
+ key for key in up_blocks[block_id] if f"up.{block_id}" in key and f"up.{block_id}.upsample" not in key
+ ]
+
+ if f"decoder.up.{block_id}.upsample.conv.weight" in vae_state_dict:
+ new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.weight"] = vae_state_dict[
+ f"decoder.up.{block_id}.upsample.conv.weight"
+ ]
+ new_checkpoint[f"decoder.up_blocks.{i}.upsamplers.0.conv.bias"] = vae_state_dict[
+ f"decoder.up.{block_id}.upsample.conv.bias"
+ ]
+
+ paths = renew_vae_resnet_paths(resnets)
+ meta_path = {"old": f"up.{block_id}.block", "new": f"up_blocks.{i}.resnets"}
+ assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+ mid_resnets = [key for key in vae_state_dict if "decoder.mid.block" in key]
+ num_mid_res_blocks = 2
+ for i in range(1, num_mid_res_blocks + 1):
+ resnets = [key for key in mid_resnets if f"decoder.mid.block_{i}" in key]
+
+ paths = renew_vae_resnet_paths(resnets)
+ meta_path = {"old": f"mid.block_{i}", "new": f"mid_block.resnets.{i - 1}"}
+ assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+
+ mid_attentions = [key for key in vae_state_dict if "decoder.mid.attn" in key]
+ paths = renew_vae_attention_paths(mid_attentions)
+ meta_path = {"old": "mid.attn_1", "new": "mid_block.attentions.0"}
+ assign_to_checkpoint(paths, new_checkpoint, vae_state_dict, additional_replacements=[meta_path], config=config)
+ conv_attn_to_linear(new_checkpoint)
+ return new_checkpoint
+
+
+def convert_ldm_bert_checkpoint(checkpoint, config):
+ def _copy_attn_layer(hf_attn_layer, pt_attn_layer):
+ hf_attn_layer.q_proj.weight.data = pt_attn_layer.to_q.weight
+ hf_attn_layer.k_proj.weight.data = pt_attn_layer.to_k.weight
+ hf_attn_layer.v_proj.weight.data = pt_attn_layer.to_v.weight
+
+ hf_attn_layer.out_proj.weight = pt_attn_layer.to_out.weight
+ hf_attn_layer.out_proj.bias = pt_attn_layer.to_out.bias
+
+ def _copy_linear(hf_linear, pt_linear):
+ hf_linear.weight = pt_linear.weight
+ hf_linear.bias = pt_linear.bias
+
+ def _copy_layer(hf_layer, pt_layer):
+ # copy layer norms
+ _copy_linear(hf_layer.self_attn_layer_norm, pt_layer[0][0])
+ _copy_linear(hf_layer.final_layer_norm, pt_layer[1][0])
+
+ # copy attn
+ _copy_attn_layer(hf_layer.self_attn, pt_layer[0][1])
+
+ # copy MLP
+ pt_mlp = pt_layer[1][1]
+ _copy_linear(hf_layer.fc1, pt_mlp.net[0][0])
+ _copy_linear(hf_layer.fc2, pt_mlp.net[2])
+
+ def _copy_layers(hf_layers, pt_layers):
+ for i, hf_layer in enumerate(hf_layers):
+ if i != 0:
+ i += i
+ pt_layer = pt_layers[i : i + 2]
+ _copy_layer(hf_layer, pt_layer)
+
+ hf_model = LDMBertModel(config).eval()
+
+ # copy embeds
+ hf_model.model.embed_tokens.weight = checkpoint.transformer.token_emb.weight
+ hf_model.model.embed_positions.weight.data = checkpoint.transformer.pos_emb.emb.weight
+
+ # copy layer norm
+ _copy_linear(hf_model.model.layer_norm, checkpoint.transformer.norm)
+
+ # copy hidden layers
+ _copy_layers(hf_model.model.layers, checkpoint.transformer.attn_layers.layers)
+
+ _copy_linear(hf_model.to_logits, checkpoint.transformer.to_logits)
+
+ return hf_model
+
+
+def convert_ldm_clip_checkpoint(checkpoint, local_files_only=False, text_encoder=None):
+ if text_encoder is None:
+ config_name = "openai/clip-vit-large-patch14"
+ try:
+ config = CLIPTextConfig.from_pretrained(config_name, local_files_only=local_files_only)
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the configuration in the following path: 'openai/clip-vit-large-patch14'."
+ )
+
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+ with ctx():
+ text_model = CLIPTextModel(config)
+ else:
+ text_model = text_encoder
+
+ keys = list(checkpoint.keys())
+
+ text_model_dict = {}
+
+ remove_prefixes = ["cond_stage_model.transformer", "conditioner.embedders.0.transformer"]
+
+ for key in keys:
+ for prefix in remove_prefixes:
+ if key.startswith(prefix):
+ text_model_dict[key[len(prefix + ".") :]] = checkpoint[key]
+
+ if is_accelerate_available():
+ for param_name, param in text_model_dict.items():
+ set_module_tensor_to_device(text_model, param_name, "cpu", value=param)
+ else:
+ if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)):
+ text_model_dict.pop("text_model.embeddings.position_ids", None)
+
+ text_model.load_state_dict(text_model_dict)
+
+ return text_model
+
+
+textenc_conversion_lst = [
+ ("positional_embedding", "text_model.embeddings.position_embedding.weight"),
+ ("token_embedding.weight", "text_model.embeddings.token_embedding.weight"),
+ ("ln_final.weight", "text_model.final_layer_norm.weight"),
+ ("ln_final.bias", "text_model.final_layer_norm.bias"),
+ ("text_projection", "text_projection.weight"),
+]
+textenc_conversion_map = {x[0]: x[1] for x in textenc_conversion_lst}
+
+textenc_transformer_conversion_lst = [
+ # (stable-diffusion, HF Diffusers)
+ ("resblocks.", "text_model.encoder.layers."),
+ ("ln_1", "layer_norm1"),
+ ("ln_2", "layer_norm2"),
+ (".c_fc.", ".fc1."),
+ (".c_proj.", ".fc2."),
+ (".attn", ".self_attn"),
+ ("ln_final.", "transformer.text_model.final_layer_norm."),
+ ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
+ ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
+]
+protected = {re.escape(x[0]): x[1] for x in textenc_transformer_conversion_lst}
+textenc_pattern = re.compile("|".join(protected.keys()))
+
+
+def convert_paint_by_example_checkpoint(checkpoint, local_files_only=False):
+ config = CLIPVisionConfig.from_pretrained("openai/clip-vit-large-patch14", local_files_only=local_files_only)
+ model = PaintByExampleImageEncoder(config)
+
+ keys = list(checkpoint.keys())
+
+ text_model_dict = {}
+
+ for key in keys:
+ if key.startswith("cond_stage_model.transformer"):
+ text_model_dict[key[len("cond_stage_model.transformer.") :]] = checkpoint[key]
+
+ # load clip vision
+ model.model.load_state_dict(text_model_dict)
+
+ # load mapper
+ keys_mapper = {
+ k[len("cond_stage_model.mapper.res") :]: v
+ for k, v in checkpoint.items()
+ if k.startswith("cond_stage_model.mapper")
+ }
+
+ MAPPING = {
+ "attn.c_qkv": ["attn1.to_q", "attn1.to_k", "attn1.to_v"],
+ "attn.c_proj": ["attn1.to_out.0"],
+ "ln_1": ["norm1"],
+ "ln_2": ["norm3"],
+ "mlp.c_fc": ["ff.net.0.proj"],
+ "mlp.c_proj": ["ff.net.2"],
+ }
+
+ mapped_weights = {}
+ for key, value in keys_mapper.items():
+ prefix = key[: len("blocks.i")]
+ suffix = key.split(prefix)[-1].split(".")[-1]
+ name = key.split(prefix)[-1].split(suffix)[0][1:-1]
+ mapped_names = MAPPING[name]
+
+ num_splits = len(mapped_names)
+ for i, mapped_name in enumerate(mapped_names):
+ new_name = ".".join([prefix, mapped_name, suffix])
+ shape = value.shape[0] // num_splits
+ mapped_weights[new_name] = value[i * shape : (i + 1) * shape]
+
+ model.mapper.load_state_dict(mapped_weights)
+
+ # load final layer norm
+ model.final_layer_norm.load_state_dict(
+ {
+ "bias": checkpoint["cond_stage_model.final_ln.bias"],
+ "weight": checkpoint["cond_stage_model.final_ln.weight"],
+ }
+ )
+
+ # load final proj
+ model.proj_out.load_state_dict(
+ {
+ "bias": checkpoint["proj_out.bias"],
+ "weight": checkpoint["proj_out.weight"],
+ }
+ )
+
+ # load uncond vector
+ model.uncond_vector.data = torch.nn.Parameter(checkpoint["learnable_vector"])
+ return model
+
+
+def convert_open_clip_checkpoint(
+ checkpoint,
+ config_name,
+ prefix="cond_stage_model.model.",
+ has_projection=False,
+ local_files_only=False,
+ **config_kwargs,
+):
+ # text_model = CLIPTextModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="text_encoder")
+ # text_model = CLIPTextModelWithProjection.from_pretrained(
+ # "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", projection_dim=1280
+ # )
+ try:
+ config = CLIPTextConfig.from_pretrained(config_name, **config_kwargs, local_files_only=local_files_only)
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the configuration in the following path: '{config_name}'."
+ )
+
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+ with ctx():
+ text_model = CLIPTextModelWithProjection(config) if has_projection else CLIPTextModel(config)
+
+ keys = list(checkpoint.keys())
+
+ keys_to_ignore = []
+ if config_name == "stabilityai/stable-diffusion-2" and config.num_hidden_layers == 23:
+ # make sure to remove all keys > 22
+ keys_to_ignore += [k for k in keys if k.startswith("cond_stage_model.model.transformer.resblocks.23")]
+ keys_to_ignore += ["cond_stage_model.model.text_projection"]
+
+ text_model_dict = {}
+
+ if prefix + "text_projection" in checkpoint:
+ d_model = int(checkpoint[prefix + "text_projection"].shape[0])
+ else:
+ d_model = 1024
+
+ text_model_dict["text_model.embeddings.position_ids"] = text_model.text_model.embeddings.get_buffer("position_ids")
+
+ for key in keys:
+ if key in keys_to_ignore:
+ continue
+ if key[len(prefix) :] in textenc_conversion_map:
+ if key.endswith("text_projection"):
+ value = checkpoint[key].T.contiguous()
+ else:
+ value = checkpoint[key]
+
+ text_model_dict[textenc_conversion_map[key[len(prefix) :]]] = value
+
+ if key.startswith(prefix + "transformer."):
+ new_key = key[len(prefix + "transformer.") :]
+ if new_key.endswith(".in_proj_weight"):
+ new_key = new_key[: -len(".in_proj_weight")]
+ new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key)
+ text_model_dict[new_key + ".q_proj.weight"] = checkpoint[key][:d_model, :]
+ text_model_dict[new_key + ".k_proj.weight"] = checkpoint[key][d_model : d_model * 2, :]
+ text_model_dict[new_key + ".v_proj.weight"] = checkpoint[key][d_model * 2 :, :]
+ elif new_key.endswith(".in_proj_bias"):
+ new_key = new_key[: -len(".in_proj_bias")]
+ new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key)
+ text_model_dict[new_key + ".q_proj.bias"] = checkpoint[key][:d_model]
+ text_model_dict[new_key + ".k_proj.bias"] = checkpoint[key][d_model : d_model * 2]
+ text_model_dict[new_key + ".v_proj.bias"] = checkpoint[key][d_model * 2 :]
+ else:
+ new_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], new_key)
+
+ text_model_dict[new_key] = checkpoint[key]
+
+ if is_accelerate_available():
+ for param_name, param in text_model_dict.items():
+ set_module_tensor_to_device(text_model, param_name, "cpu", value=param)
+ else:
+ if not (hasattr(text_model, "embeddings") and hasattr(text_model.embeddings.position_ids)):
+ text_model_dict.pop("text_model.embeddings.position_ids", None)
+
+ text_model.load_state_dict(text_model_dict)
+
+ return text_model
+
+
+def stable_unclip_image_encoder(original_config, local_files_only=False):
+ """
+ Returns the image processor and clip image encoder for the img2img unclip pipeline.
+
+ We currently know of two types of stable unclip models which separately use the clip and the openclip image
+ encoders.
+ """
+
+ image_embedder_config = original_config["model"]["params"]["embedder_config"]
+
+ sd_clip_image_embedder_class = image_embedder_config["target"]
+ sd_clip_image_embedder_class = sd_clip_image_embedder_class.split(".")[-1]
+
+ if sd_clip_image_embedder_class == "ClipImageEmbedder":
+ clip_model_name = image_embedder_config.params.model
+
+ if clip_model_name == "ViT-L/14":
+ feature_extractor = CLIPImageProcessor()
+ image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+ "openai/clip-vit-large-patch14", local_files_only=local_files_only
+ )
+ else:
+ raise NotImplementedError(f"Unknown CLIP checkpoint name in stable diffusion checkpoint {clip_model_name}")
+
+ elif sd_clip_image_embedder_class == "FrozenOpenCLIPImageEmbedder":
+ feature_extractor = CLIPImageProcessor()
+ image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+ "laion/CLIP-ViT-H-14-laion2B-s32B-b79K", local_files_only=local_files_only
+ )
+ else:
+ raise NotImplementedError(
+ f"Unknown CLIP image embedder class in stable diffusion checkpoint {sd_clip_image_embedder_class}"
+ )
+
+ return feature_extractor, image_encoder
+
+
+def stable_unclip_image_noising_components(
+ original_config, clip_stats_path: Optional[str] = None, device: Optional[str] = None
+):
+ """
+ Returns the noising components for the img2img and txt2img unclip pipelines.
+
+ Converts the stability noise augmentor into
+ 1. a `StableUnCLIPImageNormalizer` for holding the CLIP stats
+ 2. a `DDPMScheduler` for holding the noise schedule
+
+ If the noise augmentor config specifies a clip stats path, the `clip_stats_path` must be provided.
+ """
+ noise_aug_config = original_config["model"]["params"]["noise_aug_config"]
+ noise_aug_class = noise_aug_config["target"]
+ noise_aug_class = noise_aug_class.split(".")[-1]
+
+ if noise_aug_class == "CLIPEmbeddingNoiseAugmentation":
+ noise_aug_config = noise_aug_config.params
+ embedding_dim = noise_aug_config.timestep_dim
+ max_noise_level = noise_aug_config.noise_schedule_config.timesteps
+ beta_schedule = noise_aug_config.noise_schedule_config.beta_schedule
+
+ image_normalizer = StableUnCLIPImageNormalizer(embedding_dim=embedding_dim)
+ image_noising_scheduler = DDPMScheduler(num_train_timesteps=max_noise_level, beta_schedule=beta_schedule)
+
+ if "clip_stats_path" in noise_aug_config:
+ if clip_stats_path is None:
+ raise ValueError("This stable unclip config requires a `clip_stats_path`")
+
+ clip_mean, clip_std = torch.load(clip_stats_path, map_location=device)
+ clip_mean = clip_mean[None, :]
+ clip_std = clip_std[None, :]
+
+ clip_stats_state_dict = {
+ "mean": clip_mean,
+ "std": clip_std,
+ }
+
+ image_normalizer.load_state_dict(clip_stats_state_dict)
+ else:
+ raise NotImplementedError(f"Unknown noise augmentor class: {noise_aug_class}")
+
+ return image_normalizer, image_noising_scheduler
+
+
+def convert_controlnet_checkpoint(
+ checkpoint,
+ original_config,
+ checkpoint_path,
+ image_size,
+ upcast_attention,
+ extract_ema,
+ use_linear_projection=None,
+ cross_attention_dim=None,
+):
+ ctrlnet_config = create_unet_diffusers_config(original_config, image_size=image_size, controlnet=True)
+ ctrlnet_config["upcast_attention"] = upcast_attention
+
+ ctrlnet_config.pop("sample_size")
+
+ if use_linear_projection is not None:
+ ctrlnet_config["use_linear_projection"] = use_linear_projection
+
+ if cross_attention_dim is not None:
+ ctrlnet_config["cross_attention_dim"] = cross_attention_dim
+
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+ with ctx():
+ controlnet = ControlNetModel(**ctrlnet_config)
+
+ # Some controlnet ckpt files are distributed independently from the rest of the
+ # model components i.e. https://huggingface.co/thibaud/controlnet-sd21/
+ if "time_embed.0.weight" in checkpoint:
+ skip_extract_state_dict = True
+ else:
+ skip_extract_state_dict = False
+
+ converted_ctrl_checkpoint = convert_ldm_unet_checkpoint(
+ checkpoint,
+ ctrlnet_config,
+ path=checkpoint_path,
+ extract_ema=extract_ema,
+ controlnet=True,
+ skip_extract_state_dict=skip_extract_state_dict,
+ )
+
+ if is_accelerate_available():
+ for param_name, param in converted_ctrl_checkpoint.items():
+ set_module_tensor_to_device(controlnet, param_name, "cpu", value=param)
+ else:
+ controlnet.load_state_dict(converted_ctrl_checkpoint)
+
+ return controlnet
+
+
+def download_from_original_stable_diffusion_ckpt(
+ checkpoint_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+ original_config_file: str = None,
+ image_size: Optional[int] = None,
+ prediction_type: str = None,
+ model_type: str = None,
+ extract_ema: bool = False,
+ scheduler_type: str = "pndm",
+ num_in_channels: Optional[int] = None,
+ upcast_attention: Optional[bool] = None,
+ device: str = None,
+ from_safetensors: bool = False,
+ stable_unclip: Optional[str] = None,
+ stable_unclip_prior: Optional[str] = None,
+ clip_stats_path: Optional[str] = None,
+ controlnet: Optional[bool] = None,
+ adapter: Optional[bool] = None,
+ load_safety_checker: bool = True,
+ pipeline_class: DiffusionPipeline = None,
+ local_files_only=False,
+ vae_path=None,
+ vae=None,
+ text_encoder=None,
+ text_encoder_2=None,
+ tokenizer=None,
+ tokenizer_2=None,
+ config_files=None,
+) -> DiffusionPipeline:
+ """
+ Load a Stable Diffusion pipeline object from a CompVis-style `.ckpt`/`.safetensors` file and (ideally) a `.yaml`
+ config file.
+
+ Although many of the arguments can be automatically inferred, some of these rely on brittle checks against the
+ global step count, which will likely fail for models that have undergone further fine-tuning. Therefore, it is
+ recommended that you override the default values and/or supply an `original_config_file` wherever possible.
+
+ Args:
+ checkpoint_path_or_dict (`str` or `dict`): Path to `.ckpt` file, or the state dict.
+ original_config_file (`str`):
+ Path to `.yaml` config file corresponding to the original architecture. If `None`, will be automatically
+ inferred by looking for a key that only exists in SD2.0 models.
+ image_size (`int`, *optional*, defaults to 512):
+ The image size that the model was trained on. Use 512 for Stable Diffusion v1.X and Stable Diffusion v2
+ Base. Use 768 for Stable Diffusion v2.
+ prediction_type (`str`, *optional*):
+ The prediction type that the model was trained on. Use `'epsilon'` for Stable Diffusion v1.X and Stable
+ Diffusion v2 Base. Use `'v_prediction'` for Stable Diffusion v2.
+ num_in_channels (`int`, *optional*, defaults to None):
+ The number of input channels. If `None`, it will be automatically inferred.
+ scheduler_type (`str`, *optional*, defaults to 'pndm'):
+ Type of scheduler to use. Should be one of `["pndm", "lms", "heun", "euler", "euler-ancestral", "dpm",
+ "ddim"]`.
+ model_type (`str`, *optional*, defaults to `None`):
+ The pipeline type. `None` to automatically infer, or one of `["FrozenOpenCLIPEmbedder",
+ "FrozenCLIPEmbedder", "PaintByExample"]`.
+ is_img2img (`bool`, *optional*, defaults to `False`):
+ Whether the model should be loaded as an img2img pipeline.
+ extract_ema (`bool`, *optional*, defaults to `False`): Only relevant for
+ checkpoints that have both EMA and non-EMA weights. Whether to extract the EMA weights or not. Defaults to
+ `False`. Pass `True` to extract the EMA weights. EMA weights usually yield higher quality images for
+ inference. Non-EMA weights are usually better to continue fine-tuning.
+ upcast_attention (`bool`, *optional*, defaults to `None`):
+ Whether the attention computation should always be upcasted. This is necessary when running stable
+ diffusion 2.1.
+ device (`str`, *optional*, defaults to `None`):
+ The device to use. Pass `None` to determine automatically.
+ from_safetensors (`str`, *optional*, defaults to `False`):
+ If `checkpoint_path` is in `safetensors` format, load checkpoint with safetensors instead of PyTorch.
+ load_safety_checker (`bool`, *optional*, defaults to `True`):
+ Whether to load the safety checker or not. Defaults to `True`.
+ pipeline_class (`str`, *optional*, defaults to `None`):
+ The pipeline class to use. Pass `None` to determine automatically.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ Whether or not to only look at local files (i.e., do not try to download the model).
+ vae (`AutoencoderKL`, *optional*, defaults to `None`):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. If
+ this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed.
+ text_encoder (`CLIPTextModel`, *optional*, defaults to `None`):
+ An instance of [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel)
+ to use, specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)
+ variant. If this parameter is `None`, the function will load a new instance of [CLIP] by itself, if needed.
+ tokenizer (`CLIPTokenizer`, *optional*, defaults to `None`):
+ An instance of
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer)
+ to use. If this parameter is `None`, the function will load a new instance of [CLIPTokenizer] by itself, if
+ needed.
+ config_files (`Dict[str, str]`, *optional*, defaults to `None`):
+ A dictionary mapping from config file names to their contents. If this parameter is `None`, the function
+ will load the config files by itself, if needed. Valid keys are:
+ - `v1`: Config file for Stable Diffusion v1
+ - `v2`: Config file for Stable Diffusion v2
+ - `xl`: Config file for Stable Diffusion XL
+ - `xl_refiner`: Config file for Stable Diffusion XL Refiner
+ return: A StableDiffusionPipeline object representing the passed-in `.ckpt`/`.safetensors` file.
+ """
+
+ # import pipelines here to avoid circular import error when using from_single_file method
+ from diffusers import (
+ LDMTextToImagePipeline,
+ PaintByExamplePipeline,
+ StableDiffusionControlNetPipeline,
+ StableDiffusionInpaintPipeline,
+ StableDiffusionPipeline,
+ StableDiffusionUpscalePipeline,
+ StableDiffusionXLControlNetInpaintPipeline,
+ StableDiffusionXLImg2ImgPipeline,
+ StableDiffusionXLInpaintPipeline,
+ StableDiffusionXLPipeline,
+ StableUnCLIPImg2ImgPipeline,
+ StableUnCLIPPipeline,
+ )
+
+ if prediction_type == "v-prediction":
+ prediction_type = "v_prediction"
+
+ if isinstance(checkpoint_path_or_dict, str):
+ if from_safetensors:
+ from safetensors.torch import load_file as safe_load
+
+ checkpoint = safe_load(checkpoint_path_or_dict, device="cpu")
+ else:
+ if device is None:
+ device = "cuda" if torch.cuda.is_available() else "cpu"
+ checkpoint = torch.load(checkpoint_path_or_dict, map_location=device)
+ else:
+ checkpoint = torch.load(checkpoint_path_or_dict, map_location=device)
+ elif isinstance(checkpoint_path_or_dict, dict):
+ checkpoint = checkpoint_path_or_dict
+
+ # Sometimes models don't have the global_step item
+ if "global_step" in checkpoint:
+ global_step = checkpoint["global_step"]
+ else:
+ logger.debug("global_step key not found in model")
+ global_step = None
+
+ # NOTE: this while loop isn't great but this controlnet checkpoint has one additional
+ # "state_dict" key https://huggingface.co/thibaud/controlnet-canny-sd21
+ while "state_dict" in checkpoint:
+ checkpoint = checkpoint["state_dict"]
+
+ if original_config_file is None:
+ key_name_v2_1 = "model.diffusion_model.input_blocks.2.1.transformer_blocks.0.attn2.to_k.weight"
+ key_name_sd_xl_base = "conditioner.embedders.1.model.transformer.resblocks.9.mlp.c_proj.bias"
+ key_name_sd_xl_refiner = "conditioner.embedders.0.model.transformer.resblocks.9.mlp.c_proj.bias"
+ is_upscale = pipeline_class == StableDiffusionUpscalePipeline
+
+ config_url = None
+
+ # model_type = "v1"
+ if config_files is not None and "v1" in config_files:
+ original_config_file = config_files["v1"]
+ else:
+ config_url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/configs/stable-diffusion/v1-inference.yaml"
+
+ if key_name_v2_1 in checkpoint and checkpoint[key_name_v2_1].shape[-1] == 1024:
+ # model_type = "v2"
+ if config_files is not None and "v2" in config_files:
+ original_config_file = config_files["v2"]
+ else:
+ config_url = "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/v2-inference-v.yaml"
+ if global_step == 110000:
+ # v2.1 needs to upcast attention
+ upcast_attention = True
+ elif key_name_sd_xl_base in checkpoint:
+ # only base xl has two text embedders
+ if config_files is not None and "xl" in config_files:
+ original_config_file = config_files["xl"]
+ else:
+ config_url = "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_base.yaml"
+ elif key_name_sd_xl_refiner in checkpoint:
+ # only refiner xl has embedder and one text embedders
+ if config_files is not None and "xl_refiner" in config_files:
+ original_config_file = config_files["xl_refiner"]
+ else:
+ config_url = "https://raw.githubusercontent.com/Stability-AI/generative-models/main/configs/inference/sd_xl_refiner.yaml"
+
+ if is_upscale:
+ config_url = "https://raw.githubusercontent.com/Stability-AI/stablediffusion/main/configs/stable-diffusion/x4-upscaling.yaml"
+
+ if config_url is not None:
+ original_config_file = BytesIO(requests.get(config_url).content)
+ else:
+ with open(original_config_file, "r") as f:
+ original_config_file = f.read()
+ else:
+ with open(original_config_file, "r") as f:
+ original_config_file = f.read()
+
+ original_config = yaml.safe_load(original_config_file)
+
+ # Convert the text model.
+ if (
+ model_type is None
+ and "cond_stage_config" in original_config["model"]["params"]
+ and original_config["model"]["params"]["cond_stage_config"] is not None
+ ):
+ model_type = original_config["model"]["params"]["cond_stage_config"]["target"].split(".")[-1]
+ logger.debug(f"no `model_type` given, `model_type` inferred as: {model_type}")
+ elif model_type is None and original_config["model"]["params"]["network_config"] is not None:
+ if original_config["model"]["params"]["network_config"]["params"]["context_dim"] == 2048:
+ model_type = "SDXL"
+ else:
+ model_type = "SDXL-Refiner"
+ if image_size is None:
+ image_size = 1024
+
+ if pipeline_class is None:
+ # Check if we have a SDXL or SD model and initialize default pipeline
+ if model_type not in ["SDXL", "SDXL-Refiner"]:
+ pipeline_class = StableDiffusionPipeline if not controlnet else StableDiffusionControlNetPipeline
+ else:
+ pipeline_class = StableDiffusionXLPipeline if model_type == "SDXL" else StableDiffusionXLImg2ImgPipeline
+
+ if num_in_channels is None and pipeline_class in [
+ StableDiffusionInpaintPipeline,
+ StableDiffusionXLInpaintPipeline,
+ StableDiffusionXLControlNetInpaintPipeline,
+ ]:
+ num_in_channels = 9
+ if num_in_channels is None and pipeline_class == StableDiffusionUpscalePipeline:
+ num_in_channels = 7
+ elif num_in_channels is None:
+ num_in_channels = 4
+
+ if "unet_config" in original_config["model"]["params"]:
+ original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels
+
+ if (
+ "parameterization" in original_config["model"]["params"]
+ and original_config["model"]["params"]["parameterization"] == "v"
+ ):
+ if prediction_type is None:
+ # NOTE: For stable diffusion 2 base it is recommended to pass `prediction_type=="epsilon"`
+ # as it relies on a brittle global step parameter here
+ prediction_type = "epsilon" if global_step == 875000 else "v_prediction"
+ if image_size is None:
+ # NOTE: For stable diffusion 2 base one has to pass `image_size==512`
+ # as it relies on a brittle global step parameter here
+ image_size = 512 if global_step == 875000 else 768
+ else:
+ if prediction_type is None:
+ prediction_type = "epsilon"
+ if image_size is None:
+ image_size = 512
+
+ if controlnet is None and "control_stage_config" in original_config["model"]["params"]:
+ path = checkpoint_path_or_dict if isinstance(checkpoint_path_or_dict, str) else ""
+ controlnet = convert_controlnet_checkpoint(
+ checkpoint, original_config, path, image_size, upcast_attention, extract_ema
+ )
+
+ if "timesteps" in original_config["model"]["params"]:
+ num_train_timesteps = original_config["model"]["params"]["timesteps"]
+ else:
+ num_train_timesteps = 1000
+
+ if model_type in ["SDXL", "SDXL-Refiner"]:
+ scheduler_dict = {
+ "beta_schedule": "scaled_linear",
+ "beta_start": 0.00085,
+ "beta_end": 0.012,
+ "interpolation_type": "linear",
+ "num_train_timesteps": num_train_timesteps,
+ "prediction_type": "epsilon",
+ "sample_max_value": 1.0,
+ "set_alpha_to_one": False,
+ "skip_prk_steps": True,
+ "steps_offset": 1,
+ "timestep_spacing": "leading",
+ }
+ scheduler = EulerDiscreteScheduler.from_config(scheduler_dict)
+ scheduler_type = "euler"
+ else:
+ if "linear_start" in original_config["model"]["params"]:
+ beta_start = original_config["model"]["params"]["linear_start"]
+ else:
+ beta_start = 0.02
+
+ if "linear_end" in original_config["model"]["params"]:
+ beta_end = original_config["model"]["params"]["linear_end"]
+ else:
+ beta_end = 0.085
+ scheduler = DDIMScheduler(
+ beta_end=beta_end,
+ beta_schedule="scaled_linear",
+ beta_start=beta_start,
+ num_train_timesteps=num_train_timesteps,
+ steps_offset=1,
+ clip_sample=False,
+ set_alpha_to_one=False,
+ prediction_type=prediction_type,
+ )
+ # make sure scheduler works correctly with DDIM
+ scheduler.register_to_config(clip_sample=False)
+
+ if scheduler_type == "pndm":
+ config = dict(scheduler.config)
+ config["skip_prk_steps"] = True
+ scheduler = PNDMScheduler.from_config(config)
+ elif scheduler_type == "lms":
+ scheduler = LMSDiscreteScheduler.from_config(scheduler.config)
+ elif scheduler_type == "heun":
+ scheduler = HeunDiscreteScheduler.from_config(scheduler.config)
+ elif scheduler_type == "euler":
+ scheduler = EulerDiscreteScheduler.from_config(scheduler.config)
+ elif scheduler_type == "euler-ancestral":
+ scheduler = EulerAncestralDiscreteScheduler.from_config(scheduler.config)
+ elif scheduler_type == "dpm":
+ scheduler = DPMSolverMultistepScheduler.from_config(scheduler.config)
+ elif scheduler_type == "ddim":
+ scheduler = scheduler
+ else:
+ raise ValueError(f"Scheduler of type {scheduler_type} doesn't exist!")
+
+ if pipeline_class == StableDiffusionUpscalePipeline:
+ image_size = original_config["model"]["params"]["unet_config"]["params"]["image_size"]
+
+ # Convert the UNet2DConditionModel model.
+ unet_config = create_unet_diffusers_config(original_config, image_size=image_size)
+ unet_config["upcast_attention"] = upcast_attention
+
+ path = checkpoint_path_or_dict if isinstance(checkpoint_path_or_dict, str) else ""
+ converted_unet_checkpoint = convert_ldm_unet_checkpoint(
+ checkpoint, unet_config, path=path, extract_ema=extract_ema
+ )
+
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+ with ctx():
+ unet = UNet2DConditionModel(**unet_config)
+
+ if is_accelerate_available():
+ if model_type not in ["SDXL", "SDXL-Refiner"]: # SBM Delay this.
+ for param_name, param in converted_unet_checkpoint.items():
+ set_module_tensor_to_device(unet, param_name, "cpu", value=param)
+ else:
+ unet.load_state_dict(converted_unet_checkpoint)
+
+ # Convert the VAE model.
+ if vae_path is None and vae is None:
+ vae_config = create_vae_diffusers_config(original_config, image_size=image_size)
+ converted_vae_checkpoint = convert_ldm_vae_checkpoint(checkpoint, vae_config)
+
+ if (
+ "model" in original_config
+ and "params" in original_config["model"]
+ and "scale_factor" in original_config["model"]["params"]
+ ):
+ vae_scaling_factor = original_config["model"]["params"]["scale_factor"]
+ else:
+ vae_scaling_factor = 0.18215 # default SD scaling factor
+
+ vae_config["scaling_factor"] = vae_scaling_factor
+
+ ctx = init_empty_weights if is_accelerate_available() else nullcontext
+ with ctx():
+ vae = AutoencoderKL(**vae_config)
+
+ if is_accelerate_available():
+ for param_name, param in converted_vae_checkpoint.items():
+ set_module_tensor_to_device(vae, param_name, "cpu", value=param)
+ else:
+ vae.load_state_dict(converted_vae_checkpoint)
+ elif vae is None:
+ vae = AutoencoderKL.from_pretrained(vae_path, local_files_only=local_files_only)
+
+ if model_type == "FrozenOpenCLIPEmbedder":
+ config_name = "stabilityai/stable-diffusion-2"
+ config_kwargs = {"subfolder": "text_encoder"}
+
+ if text_encoder is None:
+ text_model = convert_open_clip_checkpoint(
+ checkpoint, config_name, local_files_only=local_files_only, **config_kwargs
+ )
+ else:
+ text_model = text_encoder
+
+ try:
+ tokenizer = CLIPTokenizer.from_pretrained(
+ "stabilityai/stable-diffusion-2", subfolder="tokenizer", local_files_only=local_files_only
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'stabilityai/stable-diffusion-2'."
+ )
+
+ if stable_unclip is None:
+ if controlnet:
+ pipe = pipeline_class(
+ vae=vae,
+ text_encoder=text_model,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ controlnet=controlnet,
+ safety_checker=None,
+ feature_extractor=None,
+ )
+ if hasattr(pipe, "requires_safety_checker"):
+ pipe.requires_safety_checker = False
+
+ elif pipeline_class == StableDiffusionUpscalePipeline:
+ scheduler = DDIMScheduler.from_pretrained(
+ "stabilityai/stable-diffusion-x4-upscaler", subfolder="scheduler"
+ )
+ low_res_scheduler = DDPMScheduler.from_pretrained(
+ "stabilityai/stable-diffusion-x4-upscaler", subfolder="low_res_scheduler"
+ )
+
+ pipe = pipeline_class(
+ vae=vae,
+ text_encoder=text_model,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ low_res_scheduler=low_res_scheduler,
+ safety_checker=None,
+ feature_extractor=None,
+ )
+
+ else:
+ pipe = pipeline_class(
+ vae=vae,
+ text_encoder=text_model,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=None,
+ feature_extractor=None,
+ )
+ if hasattr(pipe, "requires_safety_checker"):
+ pipe.requires_safety_checker = False
+
+ else:
+ image_normalizer, image_noising_scheduler = stable_unclip_image_noising_components(
+ original_config, clip_stats_path=clip_stats_path, device=device
+ )
+
+ if stable_unclip == "img2img":
+ feature_extractor, image_encoder = stable_unclip_image_encoder(original_config)
+
+ pipe = StableUnCLIPImg2ImgPipeline(
+ # image encoding components
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ # image noising components
+ image_normalizer=image_normalizer,
+ image_noising_scheduler=image_noising_scheduler,
+ # regular denoising components
+ tokenizer=tokenizer,
+ text_encoder=text_model,
+ unet=unet,
+ scheduler=scheduler,
+ # vae
+ vae=vae,
+ )
+ elif stable_unclip == "txt2img":
+ if stable_unclip_prior is None or stable_unclip_prior == "karlo":
+ karlo_model = "kakaobrain/karlo-v1-alpha"
+ prior = PriorTransformer.from_pretrained(
+ karlo_model, subfolder="prior", local_files_only=local_files_only
+ )
+
+ try:
+ prior_tokenizer = CLIPTokenizer.from_pretrained(
+ "openai/clip-vit-large-patch14", local_files_only=local_files_only
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'."
+ )
+ prior_text_model = CLIPTextModelWithProjection.from_pretrained(
+ "openai/clip-vit-large-patch14", local_files_only=local_files_only
+ )
+
+ prior_scheduler = UnCLIPScheduler.from_pretrained(
+ karlo_model, subfolder="prior_scheduler", local_files_only=local_files_only
+ )
+ prior_scheduler = DDPMScheduler.from_config(prior_scheduler.config)
+ else:
+ raise NotImplementedError(f"unknown prior for stable unclip model: {stable_unclip_prior}")
+
+ pipe = StableUnCLIPPipeline(
+ # prior components
+ prior_tokenizer=prior_tokenizer,
+ prior_text_encoder=prior_text_model,
+ prior=prior,
+ prior_scheduler=prior_scheduler,
+ # image noising components
+ image_normalizer=image_normalizer,
+ image_noising_scheduler=image_noising_scheduler,
+ # regular denoising components
+ tokenizer=tokenizer,
+ text_encoder=text_model,
+ unet=unet,
+ scheduler=scheduler,
+ # vae
+ vae=vae,
+ )
+ else:
+ raise NotImplementedError(f"unknown `stable_unclip` type: {stable_unclip}")
+ elif model_type == "PaintByExample":
+ vision_model = convert_paint_by_example_checkpoint(checkpoint)
+ try:
+ tokenizer = CLIPTokenizer.from_pretrained(
+ "openai/clip-vit-large-patch14", local_files_only=local_files_only
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'."
+ )
+ try:
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
+ "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the feature_extractor in the following path: 'CompVis/stable-diffusion-safety-checker'."
+ )
+ pipe = PaintByExamplePipeline(
+ vae=vae,
+ image_encoder=vision_model,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=None,
+ feature_extractor=feature_extractor,
+ )
+ elif model_type == "FrozenCLIPEmbedder":
+ text_model = convert_ldm_clip_checkpoint(
+ checkpoint, local_files_only=local_files_only, text_encoder=text_encoder
+ )
+ try:
+ tokenizer = (
+ CLIPTokenizer.from_pretrained("openai/clip-vit-large-patch14", local_files_only=local_files_only)
+ if tokenizer is None
+ else tokenizer
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'."
+ )
+
+ if load_safety_checker:
+ safety_checker = StableDiffusionSafetyChecker.from_pretrained(
+ "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only
+ )
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
+ "CompVis/stable-diffusion-safety-checker", local_files_only=local_files_only
+ )
+ else:
+ safety_checker = None
+ feature_extractor = None
+
+ if controlnet:
+ pipe = pipeline_class(
+ vae=vae,
+ text_encoder=text_model,
+ tokenizer=tokenizer,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ else:
+ pipe = pipeline_class(
+ vae=vae,
+ text_encoder=text_model,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ elif model_type in ["SDXL", "SDXL-Refiner"]:
+ is_refiner = model_type == "SDXL-Refiner"
+
+ if (is_refiner is False) and (tokenizer is None):
+ try:
+ tokenizer = CLIPTokenizer.from_pretrained(
+ "openai/clip-vit-large-patch14", local_files_only=local_files_only
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'openai/clip-vit-large-patch14'."
+ )
+
+ if (is_refiner is False) and (text_encoder is None):
+ text_encoder = convert_ldm_clip_checkpoint(checkpoint, local_files_only=local_files_only)
+
+ if tokenizer_2 is None:
+ try:
+ tokenizer_2 = CLIPTokenizer.from_pretrained(
+ "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k", pad_token="!", local_files_only=local_files_only
+ )
+ except Exception:
+ raise ValueError(
+ f"With local_files_only set to {local_files_only}, you must first locally save the tokenizer in the following path: 'laion/CLIP-ViT-bigG-14-laion2B-39B-b160k' with `pad_token` set to '!'."
+ )
+
+ if text_encoder_2 is None:
+ config_name = "laion/CLIP-ViT-bigG-14-laion2B-39B-b160k"
+ config_kwargs = {"projection_dim": 1280}
+ prefix = "conditioner.embedders.0.model." if is_refiner else "conditioner.embedders.1.model."
+
+ text_encoder_2 = convert_open_clip_checkpoint(
+ checkpoint,
+ config_name,
+ prefix=prefix,
+ has_projection=True,
+ local_files_only=local_files_only,
+ **config_kwargs,
+ )
+
+ if is_accelerate_available(): # SBM Now move model to cpu.
+ for param_name, param in converted_unet_checkpoint.items():
+ set_module_tensor_to_device(unet, param_name, "cpu", value=param)
+
+ if controlnet:
+ pipe = pipeline_class(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ text_encoder_2=text_encoder_2,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ force_zeros_for_empty_prompt=True,
+ )
+ elif adapter:
+ pipe = pipeline_class(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ text_encoder_2=text_encoder_2,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ adapter=adapter,
+ scheduler=scheduler,
+ force_zeros_for_empty_prompt=True,
+ )
+
+ else:
+ pipeline_kwargs = {
+ "vae": vae,
+ "text_encoder": text_encoder,
+ "tokenizer": tokenizer,
+ "text_encoder_2": text_encoder_2,
+ "tokenizer_2": tokenizer_2,
+ "unet": unet,
+ "scheduler": scheduler,
+ }
+
+ if (pipeline_class == StableDiffusionXLImg2ImgPipeline) or (
+ pipeline_class == StableDiffusionXLInpaintPipeline
+ ):
+ pipeline_kwargs.update({"requires_aesthetics_score": is_refiner})
+
+ if is_refiner:
+ pipeline_kwargs.update({"force_zeros_for_empty_prompt": False})
+
+ pipe = pipeline_class(**pipeline_kwargs)
+ else:
+ text_config = create_ldm_bert_config(original_config)
+ text_model = convert_ldm_bert_checkpoint(checkpoint, text_config)
+ tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased", local_files_only=local_files_only)
+ pipe = LDMTextToImagePipeline(vqvae=vae, bert=text_model, tokenizer=tokenizer, unet=unet, scheduler=scheduler)
+
+ return pipe
+
+
+def download_controlnet_from_original_ckpt(
+ checkpoint_path: str,
+ original_config_file: str,
+ image_size: int = 512,
+ extract_ema: bool = False,
+ num_in_channels: Optional[int] = None,
+ upcast_attention: Optional[bool] = None,
+ device: str = None,
+ from_safetensors: bool = False,
+ use_linear_projection: Optional[bool] = None,
+ cross_attention_dim: Optional[bool] = None,
+) -> DiffusionPipeline:
+ if from_safetensors:
+ from safetensors import safe_open
+
+ checkpoint = {}
+ with safe_open(checkpoint_path, framework="pt", device="cpu") as f:
+ for key in f.keys():
+ checkpoint[key] = f.get_tensor(key)
+ else:
+ if device is None:
+ device = "cuda" if torch.cuda.is_available() else "cpu"
+ checkpoint = torch.load(checkpoint_path, map_location=device)
+ else:
+ checkpoint = torch.load(checkpoint_path, map_location=device)
+
+ # NOTE: this while loop isn't great but this controlnet checkpoint has one additional
+ # "state_dict" key https://huggingface.co/thibaud/controlnet-canny-sd21
+ while "state_dict" in checkpoint:
+ checkpoint = checkpoint["state_dict"]
+
+ original_config = yaml.safe_load(original_config_file)
+
+ if num_in_channels is not None:
+ original_config["model"]["params"]["unet_config"]["params"]["in_channels"] = num_in_channels
+
+ if "control_stage_config" not in original_config["model"]["params"]:
+ raise ValueError("`control_stage_config` not present in original config")
+
+ controlnet = convert_controlnet_checkpoint(
+ checkpoint,
+ original_config,
+ checkpoint_path,
+ image_size,
+ upcast_attention,
+ extract_ema,
+ use_linear_projection=use_linear_projection,
+ cross_attention_dim=cross_attention_dim,
+ )
+
+ return controlnet
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..55ff51c627b508989cb007707a92d8d9d77fbe19
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion.py
@@ -0,0 +1,473 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from functools import partial
+from typing import Dict, List, Optional, Union
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.core.frozen_dict import FrozenDict
+from flax.jax_utils import unreplicate
+from flax.training.common_utils import shard
+from packaging import version
+from PIL import Image
+from transformers import CLIPImageProcessor, CLIPTokenizer, FlaxCLIPTextModel
+
+from ...models import FlaxAutoencoderKL, FlaxUNet2DConditionModel
+from ...schedulers import (
+ FlaxDDIMScheduler,
+ FlaxDPMSolverMultistepScheduler,
+ FlaxLMSDiscreteScheduler,
+ FlaxPNDMScheduler,
+)
+from ...utils import deprecate, logging, replace_example_docstring
+from ..pipeline_flax_utils import FlaxDiffusionPipeline
+from .pipeline_output import FlaxStableDiffusionPipelineOutput
+from .safety_checker_flax import FlaxStableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+# Set to True to use python for loop instead of jax.fori_loop for easier debugging
+DEBUG = False
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import jax
+ >>> import numpy as np
+ >>> from flax.jax_utils import replicate
+ >>> from flax.training.common_utils import shard
+
+ >>> from diffusers import FlaxStableDiffusionPipeline
+
+ >>> pipeline, params = FlaxStableDiffusionPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", revision="bf16", dtype=jax.numpy.bfloat16
+ ... )
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+
+ >>> prng_seed = jax.random.PRNGKey(0)
+ >>> num_inference_steps = 50
+
+ >>> num_samples = jax.device_count()
+ >>> prompt = num_samples * [prompt]
+ >>> prompt_ids = pipeline.prepare_inputs(prompt)
+ # shard inputs and rng
+
+ >>> params = replicate(params)
+ >>> prng_seed = jax.random.split(prng_seed, jax.device_count())
+ >>> prompt_ids = shard(prompt_ids)
+
+ >>> images = pipeline(prompt_ids, params, prng_seed, num_inference_steps, jit=True).images
+ >>> images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:])))
+ ```
+"""
+
+
+class FlaxStableDiffusionPipeline(FlaxDiffusionPipeline):
+ r"""
+ Flax-based pipeline for text-to-image generation using Stable Diffusion.
+
+ This model inherits from [`FlaxDiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`FlaxAutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.FlaxCLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`FlaxUNet2DConditionModel`]):
+ A `FlaxUNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`FlaxDDIMScheduler`], [`FlaxLMSDiscreteScheduler`], [`FlaxPNDMScheduler`], or
+ [`FlaxDPMSolverMultistepScheduler`].
+ safety_checker ([`FlaxStableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ def __init__(
+ self,
+ vae: FlaxAutoencoderKL,
+ text_encoder: FlaxCLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: FlaxUNet2DConditionModel,
+ scheduler: Union[
+ FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler
+ ],
+ safety_checker: FlaxStableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ super().__init__()
+ self.dtype = dtype
+
+ if safety_checker is None:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ def prepare_inputs(self, prompt: Union[str, List[str]]):
+ if not isinstance(prompt, (str, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ text_input = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ return text_input.input_ids
+
+ def _get_has_nsfw_concepts(self, features, params):
+ has_nsfw_concepts = self.safety_checker(features, params)
+ return has_nsfw_concepts
+
+ def _run_safety_checker(self, images, safety_model_params, jit=False):
+ # safety_model_params should already be replicated when jit is True
+ pil_images = [Image.fromarray(image) for image in images]
+ features = self.feature_extractor(pil_images, return_tensors="np").pixel_values
+
+ if jit:
+ features = shard(features)
+ has_nsfw_concepts = _p_get_has_nsfw_concepts(self, features, safety_model_params)
+ has_nsfw_concepts = unshard(has_nsfw_concepts)
+ safety_model_params = unreplicate(safety_model_params)
+ else:
+ has_nsfw_concepts = self._get_has_nsfw_concepts(features, safety_model_params)
+
+ images_was_copied = False
+ for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
+ if has_nsfw_concept:
+ if not images_was_copied:
+ images_was_copied = True
+ images = images.copy()
+
+ images[idx] = np.zeros(images[idx].shape, dtype=np.uint8) # black image
+
+ if any(has_nsfw_concepts):
+ warnings.warn(
+ "Potential NSFW content was detected in one or more images. A black image will be returned"
+ " instead. Try again with a different prompt and/or seed."
+ )
+
+ return images, has_nsfw_concepts
+
+ def _generate(
+ self,
+ prompt_ids: jnp.array,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ num_inference_steps: int,
+ height: int,
+ width: int,
+ guidance_scale: float,
+ latents: Optional[jnp.ndarray] = None,
+ neg_prompt_ids: Optional[jnp.ndarray] = None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ # get prompt text embeddings
+ prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0]
+
+ # TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0`
+ # implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0`
+ batch_size = prompt_ids.shape[0]
+
+ max_length = prompt_ids.shape[-1]
+
+ if neg_prompt_ids is None:
+ uncond_input = self.tokenizer(
+ [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np"
+ ).input_ids
+ else:
+ uncond_input = neg_prompt_ids
+ negative_prompt_embeds = self.text_encoder(uncond_input, params=params["text_encoder"])[0]
+ context = jnp.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ # Ensure model output will be `float32` before going into the scheduler
+ guidance_scale = jnp.array([guidance_scale], dtype=jnp.float32)
+
+ latents_shape = (
+ batch_size,
+ self.unet.config.in_channels,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if latents is None:
+ latents = jax.random.normal(prng_seed, shape=latents_shape, dtype=jnp.float32)
+ else:
+ if latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+
+ def loop_body(step, args):
+ latents, scheduler_state = args
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ latents_input = jnp.concatenate([latents] * 2)
+
+ t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step]
+ timestep = jnp.broadcast_to(t, latents_input.shape[0])
+
+ latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet.apply(
+ {"params": params["unet"]},
+ jnp.array(latents_input),
+ jnp.array(timestep, dtype=jnp.int32),
+ encoder_hidden_states=context,
+ ).sample
+ # perform guidance
+ noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple()
+ return latents, scheduler_state
+
+ scheduler_state = self.scheduler.set_timesteps(
+ params["scheduler"], num_inference_steps=num_inference_steps, shape=latents.shape
+ )
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * params["scheduler"].init_noise_sigma
+
+ if DEBUG:
+ # run with python for loop
+ for i in range(num_inference_steps):
+ latents, scheduler_state = loop_body(i, (latents, scheduler_state))
+ else:
+ latents, _ = jax.lax.fori_loop(0, num_inference_steps, loop_body, (latents, scheduler_state))
+
+ # scale and decode the image latents with vae
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample
+
+ image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1)
+ return image
+
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt_ids: jnp.array,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ num_inference_steps: int = 50,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ guidance_scale: Union[float, jnp.ndarray] = 7.5,
+ latents: jnp.ndarray = None,
+ neg_prompt_ids: jnp.ndarray = None,
+ return_dict: bool = True,
+ jit: bool = False,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ latents (`jnp.ndarray`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ array is generated by sampling using the supplied random `generator`.
+ jit (`bool`, defaults to `False`):
+ Whether to run `pmap` versions of the generation and safety scoring functions.
+
+
+
+ This argument exists because `__call__` is not yet end-to-end pmap-able. It will be removed in a
+ future release.
+
+
+
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] instead of
+ a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated images
+ and the second element is a list of `bool`s indicating whether the corresponding generated image
+ contains "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ if isinstance(guidance_scale, float):
+ # Convert to a tensor so each device gets a copy. Follow the prompt_ids for
+ # shape information, as they may be sharded (when `jit` is `True`), or not.
+ guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0])
+ if len(prompt_ids.shape) > 2:
+ # Assume sharded
+ guidance_scale = guidance_scale[:, None]
+
+ if jit:
+ images = _p_generate(
+ self,
+ prompt_ids,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ )
+ else:
+ images = self._generate(
+ prompt_ids,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ )
+
+ if self.safety_checker is not None:
+ safety_params = params["safety_checker"]
+ images_uint8_casted = (images * 255).round().astype("uint8")
+ num_devices, batch_size = images.shape[:2]
+
+ images_uint8_casted = np.asarray(images_uint8_casted).reshape(num_devices * batch_size, height, width, 3)
+ images_uint8_casted, has_nsfw_concept = self._run_safety_checker(images_uint8_casted, safety_params, jit)
+ images = np.asarray(images).copy()
+
+ # block images
+ if any(has_nsfw_concept):
+ for i, is_nsfw in enumerate(has_nsfw_concept):
+ if is_nsfw:
+ images[i, 0] = np.asarray(images_uint8_casted[i])
+
+ images = images.reshape(num_devices, batch_size, height, width, 3)
+ else:
+ images = np.asarray(images)
+ has_nsfw_concept = False
+
+ if not return_dict:
+ return (images, has_nsfw_concept)
+
+ return FlaxStableDiffusionPipelineOutput(images=images, nsfw_content_detected=has_nsfw_concept)
+
+
+# Static argnums are pipe, num_inference_steps, height, width. A change would trigger recompilation.
+# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`).
+@partial(
+ jax.pmap,
+ in_axes=(None, 0, 0, 0, None, None, None, 0, 0, 0),
+ static_broadcasted_argnums=(0, 4, 5, 6),
+)
+def _p_generate(
+ pipe,
+ prompt_ids,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+):
+ return pipe._generate(
+ prompt_ids,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ )
+
+
+@partial(jax.pmap, static_broadcasted_argnums=(0,))
+def _p_get_has_nsfw_concepts(pipe, features, params):
+ return pipe._get_has_nsfw_concepts(features, params)
+
+
+def unshard(x: jnp.ndarray):
+ # einops.rearrange(x, 'd b ... -> (d b) ...')
+ num_devices, batch_size = x.shape[:2]
+ rest = x.shape[2:]
+ return x.reshape(num_devices * batch_size, *rest)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_img2img.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..1d73d35f68d65458b94ff3b2ad334f2ee2c75ece
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_img2img.py
@@ -0,0 +1,532 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from functools import partial
+from typing import Dict, List, Optional, Union
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.core.frozen_dict import FrozenDict
+from flax.jax_utils import unreplicate
+from flax.training.common_utils import shard
+from PIL import Image
+from transformers import CLIPImageProcessor, CLIPTokenizer, FlaxCLIPTextModel
+
+from ...models import FlaxAutoencoderKL, FlaxUNet2DConditionModel
+from ...schedulers import (
+ FlaxDDIMScheduler,
+ FlaxDPMSolverMultistepScheduler,
+ FlaxLMSDiscreteScheduler,
+ FlaxPNDMScheduler,
+)
+from ...utils import PIL_INTERPOLATION, logging, replace_example_docstring
+from ..pipeline_flax_utils import FlaxDiffusionPipeline
+from .pipeline_output import FlaxStableDiffusionPipelineOutput
+from .safety_checker_flax import FlaxStableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+# Set to True to use python for loop instead of jax.fori_loop for easier debugging
+DEBUG = False
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import jax
+ >>> import numpy as np
+ >>> import jax.numpy as jnp
+ >>> from flax.jax_utils import replicate
+ >>> from flax.training.common_utils import shard
+ >>> import requests
+ >>> from io import BytesIO
+ >>> from PIL import Image
+ >>> from diffusers import FlaxStableDiffusionImg2ImgPipeline
+
+
+ >>> def create_key(seed=0):
+ ... return jax.random.PRNGKey(seed)
+
+
+ >>> rng = create_key(0)
+
+ >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+ >>> response = requests.get(url)
+ >>> init_img = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> init_img = init_img.resize((768, 512))
+
+ >>> prompts = "A fantasy landscape, trending on artstation"
+
+ >>> pipeline, params = FlaxStableDiffusionImg2ImgPipeline.from_pretrained(
+ ... "CompVis/stable-diffusion-v1-4",
+ ... revision="flax",
+ ... dtype=jnp.bfloat16,
+ ... )
+
+ >>> num_samples = jax.device_count()
+ >>> rng = jax.random.split(rng, jax.device_count())
+ >>> prompt_ids, processed_image = pipeline.prepare_inputs(
+ ... prompt=[prompts] * num_samples, image=[init_img] * num_samples
+ ... )
+ >>> p_params = replicate(params)
+ >>> prompt_ids = shard(prompt_ids)
+ >>> processed_image = shard(processed_image)
+
+ >>> output = pipeline(
+ ... prompt_ids=prompt_ids,
+ ... image=processed_image,
+ ... params=p_params,
+ ... prng_seed=rng,
+ ... strength=0.75,
+ ... num_inference_steps=50,
+ ... jit=True,
+ ... height=512,
+ ... width=768,
+ ... ).images
+
+ >>> output_images = pipeline.numpy_to_pil(np.asarray(output.reshape((num_samples,) + output.shape[-3:])))
+ ```
+"""
+
+
+class FlaxStableDiffusionImg2ImgPipeline(FlaxDiffusionPipeline):
+ r"""
+ Flax-based pipeline for text-guided image-to-image generation using Stable Diffusion.
+
+ This model inherits from [`FlaxDiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`FlaxAutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.FlaxCLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`FlaxUNet2DConditionModel`]):
+ A `FlaxUNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`FlaxDDIMScheduler`], [`FlaxLMSDiscreteScheduler`], [`FlaxPNDMScheduler`], or
+ [`FlaxDPMSolverMultistepScheduler`].
+ safety_checker ([`FlaxStableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ def __init__(
+ self,
+ vae: FlaxAutoencoderKL,
+ text_encoder: FlaxCLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: FlaxUNet2DConditionModel,
+ scheduler: Union[
+ FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler
+ ],
+ safety_checker: FlaxStableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ super().__init__()
+ self.dtype = dtype
+
+ if safety_checker is None:
+ logger.warn(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ def prepare_inputs(self, prompt: Union[str, List[str]], image: Union[Image.Image, List[Image.Image]]):
+ if not isinstance(prompt, (str, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if not isinstance(image, (Image.Image, list)):
+ raise ValueError(f"image has to be of type `PIL.Image.Image` or list but is {type(image)}")
+
+ if isinstance(image, Image.Image):
+ image = [image]
+
+ processed_images = jnp.concatenate([preprocess(img, jnp.float32) for img in image])
+
+ text_input = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ return text_input.input_ids, processed_images
+
+ def _get_has_nsfw_concepts(self, features, params):
+ has_nsfw_concepts = self.safety_checker(features, params)
+ return has_nsfw_concepts
+
+ def _run_safety_checker(self, images, safety_model_params, jit=False):
+ # safety_model_params should already be replicated when jit is True
+ pil_images = [Image.fromarray(image) for image in images]
+ features = self.feature_extractor(pil_images, return_tensors="np").pixel_values
+
+ if jit:
+ features = shard(features)
+ has_nsfw_concepts = _p_get_has_nsfw_concepts(self, features, safety_model_params)
+ has_nsfw_concepts = unshard(has_nsfw_concepts)
+ safety_model_params = unreplicate(safety_model_params)
+ else:
+ has_nsfw_concepts = self._get_has_nsfw_concepts(features, safety_model_params)
+
+ images_was_copied = False
+ for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
+ if has_nsfw_concept:
+ if not images_was_copied:
+ images_was_copied = True
+ images = images.copy()
+
+ images[idx] = np.zeros(images[idx].shape, dtype=np.uint8) # black image
+
+ if any(has_nsfw_concepts):
+ warnings.warn(
+ "Potential NSFW content was detected in one or more images. A black image will be returned"
+ " instead. Try again with a different prompt and/or seed."
+ )
+
+ return images, has_nsfw_concepts
+
+ def get_timestep_start(self, num_inference_steps, strength):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+
+ return t_start
+
+ def _generate(
+ self,
+ prompt_ids: jnp.ndarray,
+ image: jnp.ndarray,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ start_timestep: int,
+ num_inference_steps: int,
+ height: int,
+ width: int,
+ guidance_scale: float,
+ noise: Optional[jnp.ndarray] = None,
+ neg_prompt_ids: Optional[jnp.ndarray] = None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ # get prompt text embeddings
+ prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0]
+
+ # TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0`
+ # implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0`
+ batch_size = prompt_ids.shape[0]
+
+ max_length = prompt_ids.shape[-1]
+
+ if neg_prompt_ids is None:
+ uncond_input = self.tokenizer(
+ [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np"
+ ).input_ids
+ else:
+ uncond_input = neg_prompt_ids
+ negative_prompt_embeds = self.text_encoder(uncond_input, params=params["text_encoder"])[0]
+ context = jnp.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ latents_shape = (
+ batch_size,
+ self.unet.config.in_channels,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if noise is None:
+ noise = jax.random.normal(prng_seed, shape=latents_shape, dtype=jnp.float32)
+ else:
+ if noise.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {noise.shape}, expected {latents_shape}")
+
+ # Create init_latents
+ init_latent_dist = self.vae.apply({"params": params["vae"]}, image, method=self.vae.encode).latent_dist
+ init_latents = init_latent_dist.sample(key=prng_seed).transpose((0, 3, 1, 2))
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ def loop_body(step, args):
+ latents, scheduler_state = args
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ latents_input = jnp.concatenate([latents] * 2)
+
+ t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step]
+ timestep = jnp.broadcast_to(t, latents_input.shape[0])
+
+ latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet.apply(
+ {"params": params["unet"]},
+ jnp.array(latents_input),
+ jnp.array(timestep, dtype=jnp.int32),
+ encoder_hidden_states=context,
+ ).sample
+ # perform guidance
+ noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple()
+ return latents, scheduler_state
+
+ scheduler_state = self.scheduler.set_timesteps(
+ params["scheduler"], num_inference_steps=num_inference_steps, shape=latents_shape
+ )
+
+ latent_timestep = scheduler_state.timesteps[start_timestep : start_timestep + 1].repeat(batch_size)
+
+ latents = self.scheduler.add_noise(params["scheduler"], init_latents, noise, latent_timestep)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * params["scheduler"].init_noise_sigma
+
+ if DEBUG:
+ # run with python for loop
+ for i in range(start_timestep, num_inference_steps):
+ latents, scheduler_state = loop_body(i, (latents, scheduler_state))
+ else:
+ latents, _ = jax.lax.fori_loop(start_timestep, num_inference_steps, loop_body, (latents, scheduler_state))
+
+ # scale and decode the image latents with vae
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample
+
+ image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1)
+ return image
+
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt_ids: jnp.ndarray,
+ image: jnp.ndarray,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ strength: float = 0.8,
+ num_inference_steps: int = 50,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ guidance_scale: Union[float, jnp.ndarray] = 7.5,
+ noise: jnp.ndarray = None,
+ neg_prompt_ids: jnp.ndarray = None,
+ return_dict: bool = True,
+ jit: bool = False,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt_ids (`jnp.ndarray`):
+ The prompt or prompts to guide image generation.
+ image (`jnp.ndarray`):
+ Array representing an image batch to be used as the starting point.
+ params (`Dict` or `FrozenDict`):
+ Dictionary containing the model parameters/weights.
+ prng_seed (`jax.Array` or `jax.Array`):
+ Array containing random number generator key.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ noise (`jnp.ndarray`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. The array is generated by
+ sampling using the supplied random `generator`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] instead of
+ a plain tuple.
+ jit (`bool`, defaults to `False`):
+ Whether to run `pmap` versions of the generation and safety scoring functions.
+
+
+
+ This argument exists because `__call__` is not yet end-to-end pmap-able. It will be removed in a
+ future release.
+
+
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated images
+ and the second element is a list of `bool`s indicating whether the corresponding generated image
+ contains "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ if isinstance(guidance_scale, float):
+ # Convert to a tensor so each device gets a copy. Follow the prompt_ids for
+ # shape information, as they may be sharded (when `jit` is `True`), or not.
+ guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0])
+ if len(prompt_ids.shape) > 2:
+ # Assume sharded
+ guidance_scale = guidance_scale[:, None]
+
+ start_timestep = self.get_timestep_start(num_inference_steps, strength)
+
+ if jit:
+ images = _p_generate(
+ self,
+ prompt_ids,
+ image,
+ params,
+ prng_seed,
+ start_timestep,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ noise,
+ neg_prompt_ids,
+ )
+ else:
+ images = self._generate(
+ prompt_ids,
+ image,
+ params,
+ prng_seed,
+ start_timestep,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ noise,
+ neg_prompt_ids,
+ )
+
+ if self.safety_checker is not None:
+ safety_params = params["safety_checker"]
+ images_uint8_casted = (images * 255).round().astype("uint8")
+ num_devices, batch_size = images.shape[:2]
+
+ images_uint8_casted = np.asarray(images_uint8_casted).reshape(num_devices * batch_size, height, width, 3)
+ images_uint8_casted, has_nsfw_concept = self._run_safety_checker(images_uint8_casted, safety_params, jit)
+ images = np.asarray(images)
+
+ # block images
+ if any(has_nsfw_concept):
+ for i, is_nsfw in enumerate(has_nsfw_concept):
+ if is_nsfw:
+ images[i] = np.asarray(images_uint8_casted[i])
+
+ images = images.reshape(num_devices, batch_size, height, width, 3)
+ else:
+ images = np.asarray(images)
+ has_nsfw_concept = False
+
+ if not return_dict:
+ return (images, has_nsfw_concept)
+
+ return FlaxStableDiffusionPipelineOutput(images=images, nsfw_content_detected=has_nsfw_concept)
+
+
+# Static argnums are pipe, start_timestep, num_inference_steps, height, width. A change would trigger recompilation.
+# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`).
+@partial(
+ jax.pmap,
+ in_axes=(None, 0, 0, 0, 0, None, None, None, None, 0, 0, 0),
+ static_broadcasted_argnums=(0, 5, 6, 7, 8),
+)
+def _p_generate(
+ pipe,
+ prompt_ids,
+ image,
+ params,
+ prng_seed,
+ start_timestep,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ noise,
+ neg_prompt_ids,
+):
+ return pipe._generate(
+ prompt_ids,
+ image,
+ params,
+ prng_seed,
+ start_timestep,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ noise,
+ neg_prompt_ids,
+ )
+
+
+@partial(jax.pmap, static_broadcasted_argnums=(0,))
+def _p_get_has_nsfw_concepts(pipe, features, params):
+ return pipe._get_has_nsfw_concepts(features, params)
+
+
+def unshard(x: jnp.ndarray):
+ # einops.rearrange(x, 'd b ... -> (d b) ...')
+ num_devices, batch_size = x.shape[:2]
+ rest = x.shape[2:]
+ return x.reshape(num_devices * batch_size, *rest)
+
+
+def preprocess(image, dtype):
+ w, h = image.size
+ w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32
+ image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
+ image = jnp.array(image).astype(dtype) / 255.0
+ image = image[None].transpose(0, 3, 1, 2)
+ return 2.0 * image - 1.0
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_inpaint.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_inpaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..f6bb0ac299b3bc382fd0a79b1d0a78e907ebd790
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_flax_stable_diffusion_inpaint.py
@@ -0,0 +1,589 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from functools import partial
+from typing import Dict, List, Optional, Union
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+from flax.core.frozen_dict import FrozenDict
+from flax.jax_utils import unreplicate
+from flax.training.common_utils import shard
+from packaging import version
+from PIL import Image
+from transformers import CLIPImageProcessor, CLIPTokenizer, FlaxCLIPTextModel
+
+from ...models import FlaxAutoencoderKL, FlaxUNet2DConditionModel
+from ...schedulers import (
+ FlaxDDIMScheduler,
+ FlaxDPMSolverMultistepScheduler,
+ FlaxLMSDiscreteScheduler,
+ FlaxPNDMScheduler,
+)
+from ...utils import PIL_INTERPOLATION, deprecate, logging, replace_example_docstring
+from ..pipeline_flax_utils import FlaxDiffusionPipeline
+from .pipeline_output import FlaxStableDiffusionPipelineOutput
+from .safety_checker_flax import FlaxStableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+# Set to True to use python for loop instead of jax.fori_loop for easier debugging
+DEBUG = False
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import jax
+ >>> import numpy as np
+ >>> from flax.jax_utils import replicate
+ >>> from flax.training.common_utils import shard
+ >>> import PIL
+ >>> import requests
+ >>> from io import BytesIO
+ >>> from diffusers import FlaxStableDiffusionInpaintPipeline
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+ >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+ >>> init_image = download_image(img_url).resize((512, 512))
+ >>> mask_image = download_image(mask_url).resize((512, 512))
+
+ >>> pipeline, params = FlaxStableDiffusionInpaintPipeline.from_pretrained(
+ ... "xvjiarui/stable-diffusion-2-inpainting"
+ ... )
+
+ >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+ >>> prng_seed = jax.random.PRNGKey(0)
+ >>> num_inference_steps = 50
+
+ >>> num_samples = jax.device_count()
+ >>> prompt = num_samples * [prompt]
+ >>> init_image = num_samples * [init_image]
+ >>> mask_image = num_samples * [mask_image]
+ >>> prompt_ids, processed_masked_images, processed_masks = pipeline.prepare_inputs(
+ ... prompt, init_image, mask_image
+ ... )
+ # shard inputs and rng
+
+ >>> params = replicate(params)
+ >>> prng_seed = jax.random.split(prng_seed, jax.device_count())
+ >>> prompt_ids = shard(prompt_ids)
+ >>> processed_masked_images = shard(processed_masked_images)
+ >>> processed_masks = shard(processed_masks)
+
+ >>> images = pipeline(
+ ... prompt_ids, processed_masks, processed_masked_images, params, prng_seed, num_inference_steps, jit=True
+ ... ).images
+ >>> images = pipeline.numpy_to_pil(np.asarray(images.reshape((num_samples,) + images.shape[-3:])))
+ ```
+"""
+
+
+class FlaxStableDiffusionInpaintPipeline(FlaxDiffusionPipeline):
+ r"""
+ Flax-based pipeline for text-guided image inpainting using Stable Diffusion.
+
+
+
+ 🧪 This is an experimental feature!
+
+
+
+ This model inherits from [`FlaxDiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`FlaxAutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.FlaxCLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`FlaxUNet2DConditionModel`]):
+ A `FlaxUNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`FlaxDDIMScheduler`], [`FlaxLMSDiscreteScheduler`], [`FlaxPNDMScheduler`], or
+ [`FlaxDPMSolverMultistepScheduler`].
+ safety_checker ([`FlaxStableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ def __init__(
+ self,
+ vae: FlaxAutoencoderKL,
+ text_encoder: FlaxCLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: FlaxUNet2DConditionModel,
+ scheduler: Union[
+ FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler
+ ],
+ safety_checker: FlaxStableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ super().__init__()
+ self.dtype = dtype
+
+ if safety_checker is None:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ def prepare_inputs(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[Image.Image, List[Image.Image]],
+ mask: Union[Image.Image, List[Image.Image]],
+ ):
+ if not isinstance(prompt, (str, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if not isinstance(image, (Image.Image, list)):
+ raise ValueError(f"image has to be of type `PIL.Image.Image` or list but is {type(image)}")
+
+ if isinstance(image, Image.Image):
+ image = [image]
+
+ if not isinstance(mask, (Image.Image, list)):
+ raise ValueError(f"image has to be of type `PIL.Image.Image` or list but is {type(image)}")
+
+ if isinstance(mask, Image.Image):
+ mask = [mask]
+
+ processed_images = jnp.concatenate([preprocess_image(img, jnp.float32) for img in image])
+ processed_masks = jnp.concatenate([preprocess_mask(m, jnp.float32) for m in mask])
+ # processed_masks[processed_masks < 0.5] = 0
+ processed_masks = processed_masks.at[processed_masks < 0.5].set(0)
+ # processed_masks[processed_masks >= 0.5] = 1
+ processed_masks = processed_masks.at[processed_masks >= 0.5].set(1)
+
+ processed_masked_images = processed_images * (processed_masks < 0.5)
+
+ text_input = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ return text_input.input_ids, processed_masked_images, processed_masks
+
+ def _get_has_nsfw_concepts(self, features, params):
+ has_nsfw_concepts = self.safety_checker(features, params)
+ return has_nsfw_concepts
+
+ def _run_safety_checker(self, images, safety_model_params, jit=False):
+ # safety_model_params should already be replicated when jit is True
+ pil_images = [Image.fromarray(image) for image in images]
+ features = self.feature_extractor(pil_images, return_tensors="np").pixel_values
+
+ if jit:
+ features = shard(features)
+ has_nsfw_concepts = _p_get_has_nsfw_concepts(self, features, safety_model_params)
+ has_nsfw_concepts = unshard(has_nsfw_concepts)
+ safety_model_params = unreplicate(safety_model_params)
+ else:
+ has_nsfw_concepts = self._get_has_nsfw_concepts(features, safety_model_params)
+
+ images_was_copied = False
+ for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
+ if has_nsfw_concept:
+ if not images_was_copied:
+ images_was_copied = True
+ images = images.copy()
+
+ images[idx] = np.zeros(images[idx].shape, dtype=np.uint8) # black image
+
+ if any(has_nsfw_concepts):
+ warnings.warn(
+ "Potential NSFW content was detected in one or more images. A black image will be returned"
+ " instead. Try again with a different prompt and/or seed."
+ )
+
+ return images, has_nsfw_concepts
+
+ def _generate(
+ self,
+ prompt_ids: jnp.ndarray,
+ mask: jnp.ndarray,
+ masked_image: jnp.ndarray,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ num_inference_steps: int,
+ height: int,
+ width: int,
+ guidance_scale: float,
+ latents: Optional[jnp.ndarray] = None,
+ neg_prompt_ids: Optional[jnp.ndarray] = None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ # get prompt text embeddings
+ prompt_embeds = self.text_encoder(prompt_ids, params=params["text_encoder"])[0]
+
+ # TODO: currently it is assumed `do_classifier_free_guidance = guidance_scale > 1.0`
+ # implement this conditional `do_classifier_free_guidance = guidance_scale > 1.0`
+ batch_size = prompt_ids.shape[0]
+
+ max_length = prompt_ids.shape[-1]
+
+ if neg_prompt_ids is None:
+ uncond_input = self.tokenizer(
+ [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="np"
+ ).input_ids
+ else:
+ uncond_input = neg_prompt_ids
+ negative_prompt_embeds = self.text_encoder(uncond_input, params=params["text_encoder"])[0]
+ context = jnp.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ latents_shape = (
+ batch_size,
+ self.vae.config.latent_channels,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if latents is None:
+ latents = jax.random.normal(prng_seed, shape=latents_shape, dtype=self.dtype)
+ else:
+ if latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+
+ prng_seed, mask_prng_seed = jax.random.split(prng_seed)
+
+ masked_image_latent_dist = self.vae.apply(
+ {"params": params["vae"]}, masked_image, method=self.vae.encode
+ ).latent_dist
+ masked_image_latents = masked_image_latent_dist.sample(key=mask_prng_seed).transpose((0, 3, 1, 2))
+ masked_image_latents = self.vae.config.scaling_factor * masked_image_latents
+ del mask_prng_seed
+
+ mask = jax.image.resize(mask, (*mask.shape[:-2], *masked_image_latents.shape[-2:]), method="nearest")
+
+ # 8. Check that sizes of mask, masked image and latents match
+ num_channels_latents = self.vae.config.latent_channels
+ num_channels_mask = mask.shape[1]
+ num_channels_masked_image = masked_image_latents.shape[1]
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+
+ def loop_body(step, args):
+ latents, mask, masked_image_latents, scheduler_state = args
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ latents_input = jnp.concatenate([latents] * 2)
+ mask_input = jnp.concatenate([mask] * 2)
+ masked_image_latents_input = jnp.concatenate([masked_image_latents] * 2)
+
+ t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step]
+ timestep = jnp.broadcast_to(t, latents_input.shape[0])
+
+ latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t)
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latents_input = jnp.concatenate([latents_input, mask_input, masked_image_latents_input], axis=1)
+
+ # predict the noise residual
+ noise_pred = self.unet.apply(
+ {"params": params["unet"]},
+ jnp.array(latents_input),
+ jnp.array(timestep, dtype=jnp.int32),
+ encoder_hidden_states=context,
+ ).sample
+ # perform guidance
+ noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple()
+ return latents, mask, masked_image_latents, scheduler_state
+
+ scheduler_state = self.scheduler.set_timesteps(
+ params["scheduler"], num_inference_steps=num_inference_steps, shape=latents.shape
+ )
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * params["scheduler"].init_noise_sigma
+
+ if DEBUG:
+ # run with python for loop
+ for i in range(num_inference_steps):
+ latents, mask, masked_image_latents, scheduler_state = loop_body(
+ i, (latents, mask, masked_image_latents, scheduler_state)
+ )
+ else:
+ latents, _, _, _ = jax.lax.fori_loop(
+ 0, num_inference_steps, loop_body, (latents, mask, masked_image_latents, scheduler_state)
+ )
+
+ # scale and decode the image latents with vae
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample
+
+ image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1)
+ return image
+
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt_ids: jnp.ndarray,
+ mask: jnp.ndarray,
+ masked_image: jnp.ndarray,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ num_inference_steps: int = 50,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ guidance_scale: Union[float, jnp.ndarray] = 7.5,
+ latents: jnp.ndarray = None,
+ neg_prompt_ids: jnp.ndarray = None,
+ return_dict: bool = True,
+ jit: bool = False,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ latents (`jnp.ndarray`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ array is generated by sampling using the supplied random `generator`.
+ jit (`bool`, defaults to `False`):
+ Whether to run `pmap` versions of the generation and safety scoring functions.
+
+
+
+ This argument exists because `__call__` is not yet end-to-end pmap-able. It will be removed in a
+ future release.
+
+
+
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] instead of
+ a plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.FlaxStableDiffusionPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated images
+ and the second element is a list of `bool`s indicating whether the corresponding generated image
+ contains "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ masked_image = jax.image.resize(masked_image, (*masked_image.shape[:-2], height, width), method="bicubic")
+ mask = jax.image.resize(mask, (*mask.shape[:-2], height, width), method="nearest")
+
+ if isinstance(guidance_scale, float):
+ # Convert to a tensor so each device gets a copy. Follow the prompt_ids for
+ # shape information, as they may be sharded (when `jit` is `True`), or not.
+ guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0])
+ if len(prompt_ids.shape) > 2:
+ # Assume sharded
+ guidance_scale = guidance_scale[:, None]
+
+ if jit:
+ images = _p_generate(
+ self,
+ prompt_ids,
+ mask,
+ masked_image,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ )
+ else:
+ images = self._generate(
+ prompt_ids,
+ mask,
+ masked_image,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ )
+
+ if self.safety_checker is not None:
+ safety_params = params["safety_checker"]
+ images_uint8_casted = (images * 255).round().astype("uint8")
+ num_devices, batch_size = images.shape[:2]
+
+ images_uint8_casted = np.asarray(images_uint8_casted).reshape(num_devices * batch_size, height, width, 3)
+ images_uint8_casted, has_nsfw_concept = self._run_safety_checker(images_uint8_casted, safety_params, jit)
+ images = np.asarray(images)
+
+ # block images
+ if any(has_nsfw_concept):
+ for i, is_nsfw in enumerate(has_nsfw_concept):
+ if is_nsfw:
+ images[i] = np.asarray(images_uint8_casted[i])
+
+ images = images.reshape(num_devices, batch_size, height, width, 3)
+ else:
+ images = np.asarray(images)
+ has_nsfw_concept = False
+
+ if not return_dict:
+ return (images, has_nsfw_concept)
+
+ return FlaxStableDiffusionPipelineOutput(images=images, nsfw_content_detected=has_nsfw_concept)
+
+
+# Static argnums are pipe, num_inference_steps, height, width. A change would trigger recompilation.
+# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`).
+@partial(
+ jax.pmap,
+ in_axes=(None, 0, 0, 0, 0, 0, None, None, None, 0, 0, 0),
+ static_broadcasted_argnums=(0, 6, 7, 8),
+)
+def _p_generate(
+ pipe,
+ prompt_ids,
+ mask,
+ masked_image,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+):
+ return pipe._generate(
+ prompt_ids,
+ mask,
+ masked_image,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ )
+
+
+@partial(jax.pmap, static_broadcasted_argnums=(0,))
+def _p_get_has_nsfw_concepts(pipe, features, params):
+ return pipe._get_has_nsfw_concepts(features, params)
+
+
+def unshard(x: jnp.ndarray):
+ # einops.rearrange(x, 'd b ... -> (d b) ...')
+ num_devices, batch_size = x.shape[:2]
+ rest = x.shape[2:]
+ return x.reshape(num_devices * batch_size, *rest)
+
+
+def preprocess_image(image, dtype):
+ w, h = image.size
+ w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32
+ image = image.resize((w, h), resample=PIL_INTERPOLATION["lanczos"])
+ image = jnp.array(image).astype(dtype) / 255.0
+ image = image[None].transpose(0, 3, 1, 2)
+ return 2.0 * image - 1.0
+
+
+def preprocess_mask(mask, dtype):
+ w, h = mask.size
+ w, h = (x - x % 32 for x in (w, h)) # resize to integer multiple of 32
+ mask = mask.resize((w, h))
+ mask = jnp.array(mask.convert("L")).astype(dtype) / 255.0
+ mask = jnp.expand_dims(mask, axis=(0, 1))
+
+ return mask
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..311347dccaa8218227598b015d5c4a8d97eecef8
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py
@@ -0,0 +1,487 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import CLIPImageProcessor, CLIPTokenizer
+
+from ...configuration_utils import FrozenDict
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import deprecate, logging
+from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel
+from ..pipeline_utils import DiffusionPipeline
+from . import StableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__)
+
+
+class OnnxStableDiffusionPipeline(DiffusionPipeline):
+ vae_encoder: OnnxRuntimeModel
+ vae_decoder: OnnxRuntimeModel
+ text_encoder: OnnxRuntimeModel
+ tokenizer: CLIPTokenizer
+ unet: OnnxRuntimeModel
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+ safety_checker: OnnxRuntimeModel
+ feature_extractor: CLIPImageProcessor
+
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _is_onnx = True
+
+ def __init__(
+ self,
+ vae_encoder: OnnxRuntimeModel,
+ vae_decoder: OnnxRuntimeModel,
+ text_encoder: OnnxRuntimeModel,
+ tokenizer: CLIPTokenizer,
+ unet: OnnxRuntimeModel,
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+ safety_checker: OnnxRuntimeModel,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae_encoder=vae_encoder,
+ vae_decoder=vae_decoder,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ def _encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ num_images_per_prompt: Optional[int],
+ do_classifier_free_guidance: bool,
+ negative_prompt: Optional[str],
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids
+
+ if not np.array_equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0]
+
+ prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt] * batch_size
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0]
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ def check_inputs(
+ self,
+ prompt: Union[str, List[str]],
+ height: Optional[int],
+ width: Optional[int],
+ callback_steps: int,
+ negative_prompt: Optional[str] = None,
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = 512,
+ width: Optional[int] = 512,
+ num_inference_steps: Optional[int] = 50,
+ guidance_scale: Optional[float] = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: Optional[float] = 0.0,
+ generator: Optional[np.random.RandomState] = None,
+ latents: Optional[np.ndarray] = None,
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ image (`PIL.Image.Image` or List[`PIL.Image.Image`] or `torch.FloatTensor`):
+ `Image`, or tensor representing an image batch which will be upscaled. *
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. Ignored when not using guidance (i.e., ignored if `guidance_scale`
+ is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`np.random.RandomState`, *optional*):
+ One or a list of [numpy generator(s)](TODO) to make generation deterministic.
+ latents (`np.ndarray`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+
+ # check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ # define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if generator is None:
+ generator = np.random
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ prompt_embeds = self._encode_prompt(
+ prompt,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ # get the initial random noise unless the user supplied it
+ latents_dtype = prompt_embeds.dtype
+ latents_shape = (batch_size * num_images_per_prompt, 4, height // 8, width // 8)
+ if latents is None:
+ latents = generator.randn(*latents_shape).astype(latents_dtype)
+ elif latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+
+ # set timesteps
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ latents = latents * np.float64(self.scheduler.init_noise_sigma)
+
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ timestep_dtype = next(
+ (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)"
+ )
+ timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]
+
+ for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t)
+ latent_model_input = latent_model_input.cpu().numpy()
+
+ # predict the noise residual
+ timestep = np.array([t], dtype=timestep_dtype)
+ noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)
+ noise_pred = noise_pred[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ scheduler_output = self.scheduler.step(
+ torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs
+ )
+ latents = scheduler_output.prev_sample.numpy()
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ latents = 1 / 0.18215 * latents
+ # image = self.vae_decoder(latent_sample=latents)[0]
+ # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
+ image = np.concatenate(
+ [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])]
+ )
+
+ image = np.clip(image / 2 + 0.5, 0, 1)
+ image = image.transpose((0, 2, 3, 1))
+
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(
+ self.numpy_to_pil(image), return_tensors="np"
+ ).pixel_values.astype(image.dtype)
+
+ images, has_nsfw_concept = [], []
+ for i in range(image.shape[0]):
+ image_i, has_nsfw_concept_i = self.safety_checker(
+ clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
+ )
+ images.append(image_i)
+ has_nsfw_concept.append(has_nsfw_concept_i[0])
+ image = np.concatenate(images)
+ else:
+ has_nsfw_concept = None
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+
+class StableDiffusionOnnxPipeline(OnnxStableDiffusionPipeline):
+ def __init__(
+ self,
+ vae_encoder: OnnxRuntimeModel,
+ vae_decoder: OnnxRuntimeModel,
+ text_encoder: OnnxRuntimeModel,
+ tokenizer: CLIPTokenizer,
+ unet: OnnxRuntimeModel,
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+ safety_checker: OnnxRuntimeModel,
+ feature_extractor: CLIPImageProcessor,
+ ):
+ deprecation_message = "Please use `OnnxStableDiffusionPipeline` instead of `StableDiffusionOnnxPipeline`."
+ deprecate("StableDiffusionOnnxPipeline", "1.0.0", deprecation_message)
+ super().__init__(
+ vae_encoder=vae_encoder,
+ vae_decoder=vae_decoder,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..c39409886bd9d27faf71d6e343b5e45fa8b6a51c
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_img2img.py
@@ -0,0 +1,549 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTokenizer
+
+from ...configuration_utils import FrozenDict
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import PIL_INTERPOLATION, deprecate, logging
+from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel
+from ..pipeline_utils import DiffusionPipeline
+from . import StableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess with 8->64
+def preprocess(image):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+class OnnxStableDiffusionImg2ImgPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for text-guided image to image generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ """
+
+ vae_encoder: OnnxRuntimeModel
+ vae_decoder: OnnxRuntimeModel
+ text_encoder: OnnxRuntimeModel
+ tokenizer: CLIPTokenizer
+ unet: OnnxRuntimeModel
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+ safety_checker: OnnxRuntimeModel
+ feature_extractor: CLIPImageProcessor
+
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _is_onnx = True
+
+ def __init__(
+ self,
+ vae_encoder: OnnxRuntimeModel,
+ vae_decoder: OnnxRuntimeModel,
+ text_encoder: OnnxRuntimeModel,
+ tokenizer: CLIPTokenizer,
+ unet: OnnxRuntimeModel,
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+ safety_checker: OnnxRuntimeModel,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae_encoder=vae_encoder,
+ vae_decoder=vae_decoder,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ num_images_per_prompt: Optional[int],
+ do_classifier_free_guidance: bool,
+ negative_prompt: Optional[str],
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids
+
+ if not np.array_equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0]
+
+ prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt] * batch_size
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0]
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ def check_inputs(
+ self,
+ prompt: Union[str, List[str]],
+ callback_steps: int,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[np.ndarray, PIL.Image.Image] = None,
+ strength: float = 0.8,
+ num_inference_steps: Optional[int] = 50,
+ guidance_scale: Optional[float] = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: Optional[float] = 0.0,
+ generator: Optional[np.random.RandomState] = None,
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`np.ndarray` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ strength (`float`, *optional*, defaults to 0.8):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter will be modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`np.random.RandomState`, *optional*):
+ A np.random.RandomState to make generation deterministic.
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+
+ # check inputs. Raise error if not correct
+ self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+ # define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if generator is None:
+ generator = np.random
+
+ # set timesteps
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ image = preprocess(image).cpu().numpy()
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ prompt_embeds = self._encode_prompt(
+ prompt,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ latents_dtype = prompt_embeds.dtype
+ image = image.astype(latents_dtype)
+ # encode the init image into latents and scale the latents
+ init_latents = self.vae_encoder(sample=image)[0]
+ init_latents = 0.18215 * init_latents
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ if len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ deprecation_message = (
+ f"You have passed {len(prompt)} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = len(prompt) // init_latents.shape[0]
+ init_latents = np.concatenate([init_latents] * additional_image_per_prompt * num_images_per_prompt, axis=0)
+ elif len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {len(prompt)} text prompts."
+ )
+ else:
+ init_latents = np.concatenate([init_latents] * num_images_per_prompt, axis=0)
+
+ # get the original timestep using init_timestep
+ offset = self.scheduler.config.get("steps_offset", 0)
+ init_timestep = int(num_inference_steps * strength) + offset
+ init_timestep = min(init_timestep, num_inference_steps)
+
+ timesteps = self.scheduler.timesteps.numpy()[-init_timestep]
+ timesteps = np.array([timesteps] * batch_size * num_images_per_prompt)
+
+ # add noise to latents using the timesteps
+ noise = generator.randn(*init_latents.shape).astype(latents_dtype)
+ init_latents = self.scheduler.add_noise(
+ torch.from_numpy(init_latents), torch.from_numpy(noise), torch.from_numpy(timesteps)
+ )
+ init_latents = init_latents.numpy()
+
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ latents = init_latents
+
+ t_start = max(num_inference_steps - init_timestep + offset, 0)
+ timesteps = self.scheduler.timesteps[t_start:].numpy()
+
+ timestep_dtype = next(
+ (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)"
+ )
+ timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]
+
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t)
+ latent_model_input = latent_model_input.cpu().numpy()
+
+ # predict the noise residual
+ timestep = np.array([t], dtype=timestep_dtype)
+ noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)[
+ 0
+ ]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ scheduler_output = self.scheduler.step(
+ torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs
+ )
+ latents = scheduler_output.prev_sample.numpy()
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ latents = 1 / 0.18215 * latents
+ # image = self.vae_decoder(latent_sample=latents)[0]
+ # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
+ image = np.concatenate(
+ [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])]
+ )
+
+ image = np.clip(image / 2 + 0.5, 0, 1)
+ image = image.transpose((0, 2, 3, 1))
+
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(
+ self.numpy_to_pil(image), return_tensors="np"
+ ).pixel_values.astype(image.dtype)
+ # safety_checker does not support batched inputs yet
+ images, has_nsfw_concept = [], []
+ for i in range(image.shape[0]):
+ image_i, has_nsfw_concept_i = self.safety_checker(
+ clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
+ )
+ images.append(image_i)
+ has_nsfw_concept.append(has_nsfw_concept_i[0])
+ image = np.concatenate(images)
+ else:
+ has_nsfw_concept = None
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..18d8050826cc3b755f7d42dd1903b5c80f2c6672
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_inpaint.py
@@ -0,0 +1,563 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTokenizer
+
+from ...configuration_utils import FrozenDict
+from ...schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler
+from ...utils import PIL_INTERPOLATION, deprecate, logging
+from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel
+from ..pipeline_utils import DiffusionPipeline
+from . import StableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+NUM_UNET_INPUT_CHANNELS = 9
+NUM_LATENT_CHANNELS = 4
+
+
+def prepare_mask_and_masked_image(image, mask, latents_shape):
+ image = np.array(image.convert("RGB").resize((latents_shape[1] * 8, latents_shape[0] * 8)))
+ image = image[None].transpose(0, 3, 1, 2)
+ image = image.astype(np.float32) / 127.5 - 1.0
+
+ image_mask = np.array(mask.convert("L").resize((latents_shape[1] * 8, latents_shape[0] * 8)))
+ masked_image = image * (image_mask < 127.5)
+
+ mask = mask.resize((latents_shape[1], latents_shape[0]), PIL_INTERPOLATION["nearest"])
+ mask = np.array(mask.convert("L"))
+ mask = mask.astype(np.float32) / 255.0
+ mask = mask[None, None]
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ return mask, masked_image
+
+
+class OnnxStableDiffusionInpaintPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for text-guided image inpainting using Stable Diffusion. *This is an experimental feature*.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ """
+
+ vae_encoder: OnnxRuntimeModel
+ vae_decoder: OnnxRuntimeModel
+ text_encoder: OnnxRuntimeModel
+ tokenizer: CLIPTokenizer
+ unet: OnnxRuntimeModel
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler]
+ safety_checker: OnnxRuntimeModel
+ feature_extractor: CLIPImageProcessor
+
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _is_onnx = True
+
+ def __init__(
+ self,
+ vae_encoder: OnnxRuntimeModel,
+ vae_decoder: OnnxRuntimeModel,
+ text_encoder: OnnxRuntimeModel,
+ tokenizer: CLIPTokenizer,
+ unet: OnnxRuntimeModel,
+ scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],
+ safety_checker: OnnxRuntimeModel,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+ logger.info("`OnnxStableDiffusionInpaintPipeline` is experimental and will very likely change in the future.")
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae_encoder=vae_encoder,
+ vae_decoder=vae_decoder,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ num_images_per_prompt: Optional[int],
+ do_classifier_free_guidance: bool,
+ negative_prompt: Optional[str],
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids
+
+ if not np.array_equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0]
+
+ prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt] * batch_size
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0]
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_onnx_stable_diffusion.OnnxStableDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt: Union[str, List[str]],
+ height: Optional[int],
+ width: Optional[int],
+ callback_steps: int,
+ negative_prompt: Optional[str] = None,
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: PIL.Image.Image,
+ mask_image: PIL.Image.Image,
+ height: Optional[int] = 512,
+ width: Optional[int] = 512,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[np.random.RandomState] = None,
+ latents: Optional[np.ndarray] = None,
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
+ be masked out with `mask_image` and repainted according to `prompt`.
+ mask_image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+ repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
+ to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
+ instead of 3, so the expected shape would be `(B, H, W, 1)`.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`np.random.RandomState`, *optional*):
+ A np.random.RandomState to make generation deterministic.
+ latents (`np.ndarray`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+
+ # check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ # define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if generator is None:
+ generator = np.random
+
+ # set timesteps
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ prompt_embeds = self._encode_prompt(
+ prompt,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ num_channels_latents = NUM_LATENT_CHANNELS
+ latents_shape = (batch_size * num_images_per_prompt, num_channels_latents, height // 8, width // 8)
+ latents_dtype = prompt_embeds.dtype
+ if latents is None:
+ latents = generator.randn(*latents_shape).astype(latents_dtype)
+ else:
+ if latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+
+ # prepare mask and masked_image
+ mask, masked_image = prepare_mask_and_masked_image(image, mask_image, latents_shape[-2:])
+ mask = mask.astype(latents.dtype)
+ masked_image = masked_image.astype(latents.dtype)
+
+ masked_image_latents = self.vae_encoder(sample=masked_image)[0]
+ masked_image_latents = 0.18215 * masked_image_latents
+
+ # duplicate mask and masked_image_latents for each generation per prompt
+ mask = mask.repeat(batch_size * num_images_per_prompt, 0)
+ masked_image_latents = masked_image_latents.repeat(batch_size * num_images_per_prompt, 0)
+
+ mask = np.concatenate([mask] * 2) if do_classifier_free_guidance else mask
+ masked_image_latents = (
+ np.concatenate([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+ )
+
+ num_channels_mask = mask.shape[1]
+ num_channels_masked_image = masked_image_latents.shape[1]
+
+ unet_input_channels = NUM_UNET_INPUT_CHANNELS
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != unet_input_channels:
+ raise ValueError(
+ "Incorrect configuration settings! The config of `pipeline.unet` expects"
+ f" {unet_input_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+
+ # set timesteps
+ self.scheduler.set_timesteps(num_inference_steps)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * np.float64(self.scheduler.init_noise_sigma)
+
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ timestep_dtype = next(
+ (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)"
+ )
+ timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]
+
+ for i, t in enumerate(self.progress_bar(self.scheduler.timesteps)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents
+ # concat latents, mask, masked_image_latnets in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(torch.from_numpy(latent_model_input), t)
+ latent_model_input = latent_model_input.cpu().numpy()
+ latent_model_input = np.concatenate([latent_model_input, mask, masked_image_latents], axis=1)
+
+ # predict the noise residual
+ timestep = np.array([t], dtype=timestep_dtype)
+ noise_pred = self.unet(sample=latent_model_input, timestep=timestep, encoder_hidden_states=prompt_embeds)[
+ 0
+ ]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ scheduler_output = self.scheduler.step(
+ torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs
+ )
+ latents = scheduler_output.prev_sample.numpy()
+
+ # call the callback, if provided
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ latents = 1 / 0.18215 * latents
+ # image = self.vae_decoder(latent_sample=latents)[0]
+ # it seems likes there is a strange result for using half-precision vae decoder if batchsize>1
+ image = np.concatenate(
+ [self.vae_decoder(latent_sample=latents[i : i + 1])[0] for i in range(latents.shape[0])]
+ )
+
+ image = np.clip(image / 2 + 0.5, 0, 1)
+ image = image.transpose((0, 2, 3, 1))
+
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(
+ self.numpy_to_pil(image), return_tensors="np"
+ ).pixel_values.astype(image.dtype)
+ # safety_checker does not support batched inputs yet
+ images, has_nsfw_concept = [], []
+ for i in range(image.shape[0]):
+ image_i, has_nsfw_concept_i = self.safety_checker(
+ clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
+ )
+ images.append(image_i)
+ has_nsfw_concept.append(has_nsfw_concept_i[0])
+ image = np.concatenate(images)
+ else:
+ has_nsfw_concept = None
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py
new file mode 100644
index 0000000000000000000000000000000000000000..58d83de0d396b7cc4e5ba6c1099370205e060c58
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion_upscale.py
@@ -0,0 +1,586 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTokenizer
+
+from ...configuration_utils import FrozenDict
+from ...schedulers import DDPMScheduler, KarrasDiffusionSchedulers
+from ...utils import deprecate, logging
+from ..onnx_utils import ORT_TO_NP_TYPE, OnnxRuntimeModel
+from ..pipeline_utils import DiffusionPipeline
+from . import StableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__)
+
+
+def preprocess(image):
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 32
+
+ image = [np.array(i.resize((w, h)))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+
+ return image
+
+
+class OnnxStableDiffusionUpscalePipeline(DiffusionPipeline):
+ vae: OnnxRuntimeModel
+ text_encoder: OnnxRuntimeModel
+ tokenizer: CLIPTokenizer
+ unet: OnnxRuntimeModel
+ low_res_scheduler: DDPMScheduler
+ scheduler: KarrasDiffusionSchedulers
+ safety_checker: OnnxRuntimeModel
+ feature_extractor: CLIPImageProcessor
+
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _is_onnx = True
+
+ def __init__(
+ self,
+ vae: OnnxRuntimeModel,
+ text_encoder: OnnxRuntimeModel,
+ tokenizer: Any,
+ unet: OnnxRuntimeModel,
+ low_res_scheduler: DDPMScheduler,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: Optional[OnnxRuntimeModel] = None,
+ feature_extractor: Optional[CLIPImageProcessor] = None,
+ max_noise_level: int = 350,
+ num_latent_channels=4,
+ num_unet_input_channels=7,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ low_res_scheduler=low_res_scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.register_to_config(
+ max_noise_level=max_noise_level,
+ num_latent_channels=num_latent_channels,
+ num_unet_input_channels=num_unet_input_channels,
+ )
+
+ def check_inputs(
+ self,
+ prompt: Union[str, List[str]],
+ image,
+ noise_level,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, np.ndarray)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `np.ndarray`, `PIL.Image.Image` or `list` but is {type(image)}"
+ )
+
+ # verify batch size of prompt and image are same if image is a list or tensor or numpy array
+ if isinstance(image, list) or isinstance(image, np.ndarray):
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if isinstance(image, list):
+ image_batch_size = len(image)
+ else:
+ image_batch_size = image.shape[0]
+ if batch_size != image_batch_size:
+ raise ValueError(
+ f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}."
+ " Please make sure that passed `prompt` matches the batch size of `image`."
+ )
+
+ # check noise level
+ if noise_level > self.config.max_noise_level:
+ raise ValueError(f"`noise_level` has to be <= {self.config.max_noise_level} but is {noise_level}")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height, width)
+ if latents is None:
+ latents = generator.randn(*shape).astype(dtype)
+ elif latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+
+ return latents
+
+ def decode_latents(self, latents):
+ latents = 1 / 0.08333 * latents
+ image = self.vae(latent_sample=latents)[0]
+ image = np.clip(image / 2 + 0.5, 0, 1)
+ image = image.transpose((0, 2, 3, 1))
+ return image
+
+ def _encode_prompt(
+ self,
+ prompt: Union[str, List[str]],
+ num_images_per_prompt: Optional[int],
+ do_classifier_free_guidance: bool,
+ negative_prompt: Optional[str],
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="np").input_ids
+
+ if not np.array_equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = self.text_encoder(input_ids=text_input_ids.astype(np.int32))[0]
+
+ prompt_embeds = np.repeat(prompt_embeds, num_images_per_prompt, axis=0)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt] * batch_size
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ negative_prompt_embeds = self.text_encoder(input_ids=uncond_input.input_ids.astype(np.int32))[0]
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = np.repeat(negative_prompt_embeds, num_images_per_prompt, axis=0)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = np.concatenate([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: Union[np.ndarray, PIL.Image.Image, List[PIL.Image.Image]],
+ num_inference_steps: int = 75,
+ guidance_scale: float = 9.0,
+ noise_level: int = 20,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[np.random.RandomState, List[np.random.RandomState]]] = None,
+ latents: Optional[np.ndarray] = None,
+ prompt_embeds: Optional[np.ndarray] = None,
+ negative_prompt_embeds: Optional[np.ndarray] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, np.ndarray], None]] = None,
+ callback_steps: Optional[int] = 1,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ image (`np.ndarray` or `PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, that will be used as the starting point for the
+ process.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter will be modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ noise_level (`float`, defaults to 0.2):
+ Deteremines the amount of noise to add to the initial image before performing upscaling.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`np.random.RandomState`, *optional*):
+ A np.random.RandomState to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`np.ndarray`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: np.ndarray)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ image,
+ noise_level,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if generator is None:
+ generator = np.random
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ prompt_embeds = self._encode_prompt(
+ prompt,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ latents_dtype = prompt_embeds.dtype
+ image = preprocess(image).cpu().numpy()
+ height, width = image.shape[2:]
+
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ self.num_latent_channels,
+ height,
+ width,
+ latents_dtype,
+ generator,
+ )
+ image = image.astype(latents_dtype)
+
+ self.scheduler.set_timesteps(num_inference_steps)
+ timesteps = self.scheduler.timesteps
+
+ # Scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * np.float64(self.scheduler.init_noise_sigma)
+
+ # 5. Add noise to image
+ noise_level = np.array([noise_level]).astype(np.int64)
+ noise = generator.randn(*image.shape).astype(latents_dtype)
+
+ image = self.low_res_scheduler.add_noise(
+ torch.from_numpy(image), torch.from_numpy(noise), torch.from_numpy(noise_level)
+ )
+ image = image.numpy()
+
+ batch_multiplier = 2 if do_classifier_free_guidance else 1
+ image = np.concatenate([image] * batch_multiplier * num_images_per_prompt)
+ noise_level = np.concatenate([noise_level] * image.shape[0])
+
+ # 7. Check that sizes of image and latents match
+ num_channels_image = image.shape[1]
+ if self.num_latent_channels + num_channels_image != self.num_unet_input_channels:
+ raise ValueError(
+ "Incorrect configuration settings! The config of `pipeline.unet` expects"
+ f" {self.num_unet_input_channels} but received `num_channels_latents`: {self.num_latent_channels} +"
+ f" `num_channels_image`: {num_channels_image} "
+ f" = {self.num_latent_channels + num_channels_image}. Please verify the config of"
+ " `pipeline.unet` or your `image` input."
+ )
+
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ timestep_dtype = next(
+ (input.type for input in self.unet.model.get_inputs() if input.name == "timestep"), "tensor(float)"
+ )
+ timestep_dtype = ORT_TO_NP_TYPE[timestep_dtype]
+
+ # 9. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents
+
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+ latent_model_input = np.concatenate([latent_model_input, image], axis=1)
+
+ # timestep to tensor
+ timestep = np.array([t], dtype=timestep_dtype)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=timestep,
+ encoder_hidden_states=prompt_embeds,
+ class_labels=noise_level,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = np.split(noise_pred, 2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(
+ torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs
+ ).prev_sample
+ latents = latents.numpy()
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 10. Post-processing
+ image = self.decode_latents(latents)
+
+ if self.safety_checker is not None:
+ safety_checker_input = self.feature_extractor(
+ self.numpy_to_pil(image), return_tensors="np"
+ ).pixel_values.astype(image.dtype)
+
+ images, has_nsfw_concept = [], []
+ for i in range(image.shape[0]):
+ image_i, has_nsfw_concept_i = self.safety_checker(
+ clip_input=safety_checker_input[i : i + 1], images=image[i : i + 1]
+ )
+ images.append(image_i)
+ has_nsfw_concept.append(has_nsfw_concept_i[0])
+ image = np.concatenate(images)
+ else:
+ has_nsfw_concept = None
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_output.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_output.py
new file mode 100644
index 0000000000000000000000000000000000000000..5fb9b1a1412d96b69144a4c2e960dcc8b75a615c
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_output.py
@@ -0,0 +1,45 @@
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL.Image
+
+from ...utils import BaseOutput, is_flax_available
+
+
+@dataclass
+class StableDiffusionPipelineOutput(BaseOutput):
+ """
+ Output class for Stable Diffusion pipelines.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ nsfw_content_detected (`List[bool]`)
+ List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or
+ `None` if safety checking could not be performed.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+ nsfw_content_detected: Optional[List[bool]]
+
+
+if is_flax_available():
+ import flax
+
+ @flax.struct.dataclass
+ class FlaxStableDiffusionPipelineOutput(BaseOutput):
+ """
+ Output class for Flax-based Stable Diffusion pipelines.
+
+ Args:
+ images (`np.ndarray`):
+ Denoised images of array shape of `(batch_size, height, width, num_channels)`.
+ nsfw_content_detected (`List[bool]`):
+ List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content
+ or `None` if safety checking could not be performed.
+ """
+
+ images: np.ndarray
+ nsfw_content_detected: List[bool]
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..9e4e6c186ffa85fbc5ed4b32f920b7180f02d350
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion.py
@@ -0,0 +1,1032 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...configuration_utils import FrozenDict
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import StableDiffusionPipelineOutput
+from .safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionPipeline
+
+ >>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> image = pipe(prompt).images[0]
+ ```
+"""
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def guidance_rescale(self):
+ return self._guidance_rescale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @property
+ def interrupt(self):
+ return self._interrupt
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ guidance_rescale (`float`, *optional*, defaults to 0.0):
+ Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
+ using zero terminal SNR.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+ # to deal with lora scaling and other possible forward hooks
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._guidance_rescale = guidance_rescale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+ self._interrupt = False
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 6.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if (ip_adapter_image is not None or ip_adapter_image_embeds is not None)
+ else None
+ )
+
+ # 6.2 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ if self.interrupt:
+ continue
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..c410acbedad44ef721fe860e43cf7a0ed01b5e4e
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_depth2img.py
@@ -0,0 +1,860 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import contextlib
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPTextModel, CLIPTokenizer, DPTFeatureExtractor, DPTForDepthEstimation
+
+from ...configuration_utils import FrozenDict
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import PIL_INTERPOLATION, USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
+def preprocess(image):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+class StableDiffusionDepth2ImgPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin):
+ r"""
+ Pipeline for text-guided depth-based image-to-image generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "depth_mask"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ depth_estimator: DPTForDepthEstimation,
+ feature_extractor: DPTFeatureExtractor,
+ ):
+ super().__init__()
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ depth_estimator=depth_estimator,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.prepare_latents
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+
+ return latents
+
+ def prepare_depth_map(self, image, depth_map, batch_size, do_classifier_free_guidance, dtype, device):
+ if isinstance(image, PIL.Image.Image):
+ image = [image]
+ else:
+ image = list(image)
+
+ if isinstance(image[0], PIL.Image.Image):
+ width, height = image[0].size
+ elif isinstance(image[0], np.ndarray):
+ width, height = image[0].shape[:-1]
+ else:
+ height, width = image[0].shape[-2:]
+
+ if depth_map is None:
+ pixel_values = self.feature_extractor(images=image, return_tensors="pt").pixel_values
+ pixel_values = pixel_values.to(device=device)
+ # The DPT-Hybrid model uses batch-norm layers which are not compatible with fp16.
+ # So we use `torch.autocast` here for half precision inference.
+ context_manger = torch.autocast("cuda", dtype=dtype) if device.type == "cuda" else contextlib.nullcontext()
+ with context_manger:
+ depth_map = self.depth_estimator(pixel_values).predicted_depth
+ else:
+ depth_map = depth_map.to(device=device, dtype=dtype)
+
+ depth_map = torch.nn.functional.interpolate(
+ depth_map.unsqueeze(1),
+ size=(height // self.vae_scale_factor, width // self.vae_scale_factor),
+ mode="bicubic",
+ align_corners=False,
+ )
+
+ depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
+ depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
+ depth_map = 2.0 * (depth_map - depth_min) / (depth_max - depth_min) - 1.0
+ depth_map = depth_map.to(dtype)
+
+ # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+ if depth_map.shape[0] < batch_size:
+ repeat_by = batch_size // depth_map.shape[0]
+ depth_map = depth_map.repeat(repeat_by, 1, 1, 1)
+
+ depth_map = torch.cat([depth_map] * 2) if do_classifier_free_guidance else depth_map
+ return depth_map
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ depth_map: Optional[torch.FloatTensor] = None,
+ strength: float = 0.8,
+ num_inference_steps: Optional[int] = 50,
+ guidance_scale: Optional[float] = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: Optional[float] = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image` or tensor representing an image batch to be used as the starting point. Can accept image
+ latents as `image` only if `depth_map` is not `None`.
+ depth_map (`torch.FloatTensor`, *optional*):
+ Depth prediction to be used as additional conditioning for the image generation process. If not
+ defined, it automatically predicts the depth with `self.depth_estimator`.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+ Examples:
+
+ ```py
+ >>> import torch
+ >>> import requests
+ >>> from PIL import Image
+
+ >>> from diffusers import StableDiffusionDepth2ImgPipeline
+
+ >>> pipe = StableDiffusionDepth2ImgPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-2-depth",
+ ... torch_dtype=torch.float16,
+ ... )
+ >>> pipe.to("cuda")
+
+
+ >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+ >>> init_image = Image.open(requests.get(url, stream=True).raw)
+ >>> prompt = "two tigers"
+ >>> n_propmt = "bad, deformed, ugly, bad anotomy"
+ >>> image = pipe(prompt=prompt, image=init_image, negative_prompt=n_propmt, strength=0.7).images[0]
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare depth mask
+ depth_mask = self.prepare_depth_map(
+ image,
+ depth_map,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ prompt_embeds.dtype,
+ device,
+ )
+
+ # 5. Preprocess image
+ image = self.image_processor.preprocess(image)
+
+ # 6. Set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+ # 7. Prepare latent variables
+ latents = self.prepare_latents(
+ image, latent_timestep, batch_size, num_images_per_prompt, prompt_embeds.dtype, device, generator
+ )
+
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 9. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+ latent_model_input = torch.cat([latent_model_input, depth_mask], dim=1)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+ depth_mask = callback_outputs.pop("depth_mask", depth_mask)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = latents
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py
new file mode 100644
index 0000000000000000000000000000000000000000..1333cb825750ace60cc0398bbca26691b9337dca
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_image_variation.py
@@ -0,0 +1,420 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, List, Optional, Union
+
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+
+from ...configuration_utils import FrozenDict
+from ...image_processor import VaeImageProcessor
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import deprecate, logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from . import StableDiffusionPipelineOutput
+from .safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class StableDiffusionImageVariationPipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+ Pipeline to generate image variations from an input image using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ image_encoder ([`~transformers.CLIPVisionModelWithProjection`]):
+ Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ # TODO: feature_extractor is required to encode images (if they are in PIL format),
+ # we should give a descriptive message if the pipeline doesn't have one.
+ _optional_components = ["safety_checker"]
+ model_cpu_offload_seq = "image_encoder->unet->vae"
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ image_encoder: CLIPVisionModelWithProjection,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warn(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ image_encoder=image_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ def _encode_image(self, image, device, num_images_per_prompt, do_classifier_free_guidance):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(images=image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ image_embeddings = self.image_encoder(image).image_embeds
+ image_embeddings = image_embeddings.unsqueeze(1)
+
+ # duplicate image embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = image_embeddings.shape
+ image_embeddings = image_embeddings.repeat(1, num_images_per_prompt, 1)
+ image_embeddings = image_embeddings.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = torch.zeros_like(image_embeddings)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ image_embeddings = torch.cat([negative_prompt_embeds, image_embeddings])
+
+ return image_embeddings
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(self, image, height, width, callback_steps):
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+ f" {type(image)}"
+ )
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+ Image or images to guide image generation. If you provide a tensor, it needs to be compatible with
+ [`CLIPImageProcessor`](https://huggingface.co/lambdalabs/sd-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json).
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+
+ Examples:
+
+ ```py
+ from diffusers import StableDiffusionImageVariationPipeline
+ from PIL import Image
+ from io import BytesIO
+ import requests
+
+ pipe = StableDiffusionImageVariationPipeline.from_pretrained(
+ "lambdalabs/sd-image-variations-diffusers", revision="v2.0"
+ )
+ pipe = pipe.to("cuda")
+
+ url = "https://lh3.googleusercontent.com/y-iFOHfLTwkuQSUegpwDdgKmOjRSTvPxat63dQLB25xkTs4lhIbRUFeNBWZzYf370g=s1200"
+
+ response = requests.get(url)
+ image = Image.open(BytesIO(response.content)).convert("RGB")
+
+ out = pipe(image, num_images_per_prompt=3, guidance_scale=15)
+ out["images"][0].save("result.jpg")
+ ```
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(image, height, width, callback_steps)
+
+ # 2. Define call parameters
+ if isinstance(image, PIL.Image.Image):
+ batch_size = 1
+ elif isinstance(image, list):
+ batch_size = len(image)
+ else:
+ batch_size = image.shape[0]
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input image
+ image_embeddings = self._encode_image(image, device, num_images_per_prompt, do_classifier_free_guidance)
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ image_embeddings.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=image_embeddings).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ self.maybe_free_model_hooks()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, image_embeddings.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..b43e0eb2abcdc4a402ba9f9469dfb18d29f38ee5
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py
@@ -0,0 +1,1113 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...configuration_utils import FrozenDict
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ PIL_INTERPOLATION,
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from . import StableDiffusionPipelineOutput
+from .safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import requests
+ >>> import torch
+ >>> from PIL import Image
+ >>> from io import BytesIO
+
+ >>> from diffusers import StableDiffusionImg2ImgPipeline
+
+ >>> device = "cuda"
+ >>> model_id_or_path = "runwayml/stable-diffusion-v1-5"
+ >>> pipe = StableDiffusionImg2ImgPipeline.from_pretrained(model_id_or_path, torch_dtype=torch.float16)
+ >>> pipe = pipe.to(device)
+
+ >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+
+ >>> response = requests.get(url)
+ >>> init_image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> init_image = init_image.resize((768, 512))
+
+ >>> prompt = "A fantasy landscape, trending on artstation"
+
+ >>> images = pipe(prompt=prompt, image=init_image, strength=0.75, guidance_scale=7.5).images
+ >>> images[0].save("fantasy_landscape.png")
+ ```
+"""
+
+
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+def preprocess(image):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionImg2ImgPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-guided image-to-image generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+
+ return latents
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @property
+ def interrupt(self):
+ return self._interrupt
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ strength: float = 0.8,
+ num_inference_steps: Optional[int] = 50,
+ timesteps: List[int] = None,
+ guidance_scale: Optional[float] = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: Optional[float] = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: int = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
+ numpy array and pytorch tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
+ or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
+ list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
+ latents as `image`, but if passing latents directly it is not encoded again.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+ self._interrupt = False
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Preprocess image
+ image = self.image_processor.preprocess(image)
+
+ # 5. set timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+ # 6. Prepare latent variables
+ latents = self.prepare_latents(
+ image,
+ latent_timestep,
+ batch_size,
+ num_images_per_prompt,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 7.2 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ if self.interrupt:
+ continue
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..221d5c2cfd3ff5420c3fbbeb16734688bd67e3ad
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_inpaint.py
@@ -0,0 +1,1430 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...configuration_utils import FrozenDict
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AsymmetricAutoencoderKL, AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from . import StableDiffusionPipelineOutput
+from .safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool = False):
+ """
+ Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
+ converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
+ ``image`` and ``1`` for the ``mask``.
+
+ The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+ binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+ Args:
+ image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+ ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+ mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+ ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+
+
+ Raises:
+ ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+ should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+ TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+ (ot the other way around).
+
+ Returns:
+ tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
+ dimensions: ``batch x channels x height x width``.
+ """
+ deprecation_message = "The prepare_mask_and_masked_image method is deprecated and will be removed in a future version. Please use VaeImageProcessor.preprocess instead"
+ deprecate(
+ "prepare_mask_and_masked_image",
+ "0.30.0",
+ deprecation_message,
+ )
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ if mask is None:
+ raise ValueError("`mask_image` input cannot be undefined.")
+
+ if isinstance(image, torch.Tensor):
+ if not isinstance(mask, torch.Tensor):
+ raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
+
+ # Batch single image
+ if image.ndim == 3:
+ assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
+ image = image.unsqueeze(0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Single batched mask, no channel dim or single mask not batched but channel dim
+ if mask.shape[0] == 1:
+ mask = mask.unsqueeze(0)
+
+ # Batched masks no channel dim
+ else:
+ mask = mask.unsqueeze(1)
+
+ assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+ assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+ assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+
+ # Check image is in [-1, 1]
+ if image.min() < -1 or image.max() > 1:
+ raise ValueError("Image should be in [-1, 1] range")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("Mask should be in [0, 1] range")
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ # Image as float32
+ image = image.to(dtype=torch.float32)
+ elif isinstance(mask, torch.Tensor):
+ raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+ else:
+ # preprocess image
+ if isinstance(image, (PIL.Image.Image, np.ndarray)):
+ image = [image]
+ if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+ # resize all images w.r.t passed height an width
+ image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
+ image = [np.array(i.convert("RGB"))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+ image = np.concatenate([i[None, :] for i in image], axis=0)
+
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ # preprocess mask
+ if isinstance(mask, (PIL.Image.Image, np.ndarray)):
+ mask = [mask]
+
+ if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
+ mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
+ mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+ mask = mask.astype(np.float32) / 255.0
+ elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+ mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+ mask = torch.from_numpy(mask)
+
+ masked_image = image * (mask < 0.5)
+
+ # n.b. ensure backwards compatibility as old function does not return image
+ if return_image:
+ return mask, masked_image, image
+
+ return mask, masked_image
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionInpaintPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-guided image inpainting using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+
+ Args:
+ vae ([`AutoencoderKL`, `AsymmetricAutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds", "mask", "masked_image_latents"]
+
+ def __init__(
+ self,
+ vae: Union[AutoencoderKL, AsymmetricAutoencoderKL],
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "skip_prk_steps") and scheduler.config.skip_prk_steps is False:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration"
+ " `skip_prk_steps`. `skip_prk_steps` should be set to True in the configuration file. Please make"
+ " sure to update the config accordingly as not setting `skip_prk_steps` in the config might lead to"
+ " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face"
+ " Hub, it would be very nice if you could open a Pull request for the"
+ " `scheduler/scheduler_config.json` file"
+ )
+ deprecate("skip_prk_steps not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["skip_prk_steps"] = True
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ # Check shapes, assume num_channels_latents == 4, num_channels_mask == 1, num_channels_masked == 4
+ if unet.config.in_channels != 9:
+ logger.info(f"You have loaded a UNet with {unet.config.in_channels} input channels which.")
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.mask_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ mask_image,
+ height,
+ width,
+ strength,
+ callback_steps,
+ output_type,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ padding_mask_crop=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+ if padding_mask_crop is not None:
+ if not isinstance(image, PIL.Image.Image):
+ raise ValueError(
+ f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+ )
+ if not isinstance(mask_image, PIL.Image.Image):
+ raise ValueError(
+ f"The mask image should be a PIL image when inpainting mask crop, but is of type"
+ f" {type(mask_image)}."
+ )
+ if output_type != "pil":
+ raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def prepare_latents(
+ self,
+ batch_size,
+ num_channels_latents,
+ height,
+ width,
+ dtype,
+ device,
+ generator,
+ latents=None,
+ image=None,
+ timestep=None,
+ is_strength_max=True,
+ return_noise=False,
+ return_image_latents=False,
+ ):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if (image is None or timestep is None) and not is_strength_max:
+ raise ValueError(
+ "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
+ "However, either the image or the noise timestep has not been provided."
+ )
+
+ if return_image_latents or (latents is None and not is_strength_max):
+ image = image.to(device=device, dtype=dtype)
+
+ if image.shape[1] == 4:
+ image_latents = image
+ else:
+ image_latents = self._encode_vae_image(image=image, generator=generator)
+ image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
+
+ if latents is None:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ # if strength is 1. then initialise the latents to noise, else initial to image + noise
+ latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
+ # if pure noise then scale the initial latents by the Scheduler's init sigma
+ latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+ else:
+ noise = latents.to(device)
+ latents = noise * self.scheduler.init_noise_sigma
+
+ outputs = (latents,)
+
+ if return_noise:
+ outputs += (noise,)
+
+ if return_image_latents:
+ outputs += (image_latents,)
+
+ return outputs
+
+ def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+ if isinstance(generator, list):
+ image_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(image.shape[0])
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ image_latents = self.vae.config.scaling_factor * image_latents
+
+ return image_latents
+
+ def prepare_mask_latents(
+ self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+ ):
+ # resize the mask to latents shape as we concatenate the mask to the latents
+ # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+ # and half precision
+ mask = torch.nn.functional.interpolate(
+ mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+ )
+ mask = mask.to(device=device, dtype=dtype)
+
+ masked_image = masked_image.to(device=device, dtype=dtype)
+
+ if masked_image.shape[1] == 4:
+ masked_image_latents = masked_image
+ else:
+ masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+
+ # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+ if mask.shape[0] < batch_size:
+ if not batch_size % mask.shape[0] == 0:
+ raise ValueError(
+ "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+ f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+ " of masks that you pass is divisible by the total requested batch size."
+ )
+ mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+ if masked_image_latents.shape[0] < batch_size:
+ if not batch_size % masked_image_latents.shape[0] == 0:
+ raise ValueError(
+ "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+ f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+ " Make sure the number of images that you pass is divisible by the total requested batch size."
+ )
+ masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)
+
+ mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+ masked_image_latents = (
+ torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+ )
+
+ # aligning device to prevent device errors when concating it with the latent model input
+ masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+ return mask, masked_image_latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @property
+ def interrupt(self):
+ return self._interrupt
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ mask_image: PipelineImageInput = None,
+ masked_image_latents: torch.FloatTensor = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ padding_mask_crop: Optional[int] = None,
+ strength: float = 1.0,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: int = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, numpy array or tensor representing an image batch to be inpainted (which parts of the image to
+ be masked out with `mask_image` and repainted according to `prompt`). For both numpy array and pytorch
+ tensor, the expected value range is between `[0, 1]` If it's a tensor or a list or tensors, the
+ expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a list of arrays, the
+ expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image latents as `image`, but
+ if passing latents directly it is not encoded again.
+ mask_image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image`, numpy array or tensor representing an image batch to mask `image`. White pixels in the mask
+ are repainted while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a
+ single channel (luminance) before use. If it's a numpy array or pytorch tensor, it should contain one
+ color channel (L) instead of 3, so the expected shape for pytorch tensor would be `(B, 1, H, W)`, `(B,
+ H, W)`, `(1, H, W)`, `(H, W)`. And for numpy array would be for `(B, H, W, 1)`, `(B, H, W)`, `(H, W,
+ 1)`, or `(H, W)`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ padding_mask_crop (`int`, *optional*, defaults to `None`):
+ The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to image and mask_image. If
+ `padding_mask_crop` is not `None`, it will first find a rectangular region with the same aspect ration of the image and
+ contains all masked area, and then expand that area based on `padding_mask_crop`. The image and mask_image will then be cropped based on
+ the expanded area before resizing to the original image size for inpainting. This is useful when the masked area is small while the image is large
+ and contain information inreleant for inpainging, such as background.
+ strength (`float`, *optional*, defaults to 1.0):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+ Examples:
+
+ ```py
+ >>> import PIL
+ >>> import requests
+ >>> import torch
+ >>> from io import BytesIO
+
+ >>> from diffusers import StableDiffusionInpaintPipeline
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+ >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+ >>> init_image = download_image(img_url).resize((512, 512))
+ >>> mask_image = download_image(mask_url).resize((512, 512))
+
+ >>> pipe = StableDiffusionInpaintPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+ >>> image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ image,
+ mask_image,
+ height,
+ width,
+ strength,
+ callback_steps,
+ output_type,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ padding_mask_crop,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+ self._interrupt = False
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. set timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ timesteps, num_inference_steps = self.get_timesteps(
+ num_inference_steps=num_inference_steps, strength=strength, device=device
+ )
+ # check that number of inference steps is not < 1 - as this doesn't make sense
+ if num_inference_steps < 1:
+ raise ValueError(
+ f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+ f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+ )
+ # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+ is_strength_max = strength == 1.0
+
+ # 5. Preprocess mask and image
+
+ if padding_mask_crop is not None:
+ crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop)
+ resize_mode = "fill"
+ else:
+ crops_coords = None
+ resize_mode = "default"
+
+ original_image = image
+ init_image = self.image_processor.preprocess(
+ image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
+ )
+ init_image = init_image.to(dtype=torch.float32)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ num_channels_unet = self.unet.config.in_channels
+ return_image_latents = num_channels_unet == 4
+
+ latents_outputs = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ image=init_image,
+ timestep=latent_timestep,
+ is_strength_max=is_strength_max,
+ return_noise=True,
+ return_image_latents=return_image_latents,
+ )
+
+ if return_image_latents:
+ latents, noise, image_latents = latents_outputs
+ else:
+ latents, noise = latents_outputs
+
+ # 7. Prepare mask latent variables
+ mask_condition = self.mask_processor.preprocess(
+ mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
+ )
+
+ if masked_image_latents is None:
+ masked_image = init_image * (mask_condition < 0.5)
+ else:
+ masked_image = masked_image_latents
+
+ mask, masked_image_latents = self.prepare_mask_latents(
+ mask_condition,
+ masked_image,
+ batch_size * num_images_per_prompt,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ self.do_classifier_free_guidance,
+ )
+
+ # 8. Check that sizes of mask, masked image and latents match
+ if num_channels_unet == 9:
+ # default case for runwayml/stable-diffusion-inpainting
+ num_channels_mask = mask.shape[1]
+ num_channels_masked_image = masked_image_latents.shape[1]
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+ elif num_channels_unet != 4:
+ raise ValueError(
+ f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
+ )
+
+ # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 9.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 9.2 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 10. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ if self.interrupt:
+ continue
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ if num_channels_unet == 9:
+ latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+ if num_channels_unet == 4:
+ init_latents_proper = image_latents
+ if self.do_classifier_free_guidance:
+ init_mask, _ = mask.chunk(2)
+ else:
+ init_mask = mask
+
+ if i < len(timesteps) - 1:
+ noise_timestep = timesteps[i + 1]
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_proper, noise, torch.tensor([noise_timestep])
+ )
+
+ latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+ mask = callback_outputs.pop("mask", mask)
+ masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ condition_kwargs = {}
+ if isinstance(self.vae, AsymmetricAutoencoderKL):
+ init_image = init_image.to(device=device, dtype=masked_image_latents.dtype)
+ init_image_condition = init_image.clone()
+ init_image = self._encode_vae_image(init_image, generator=generator)
+ mask_condition = mask_condition.to(device=device, dtype=masked_image_latents.dtype)
+ condition_kwargs = {"image": init_image_condition, "mask": mask_condition}
+ image = self.vae.decode(
+ latents / self.vae.config.scaling_factor, return_dict=False, generator=generator, **condition_kwargs
+ )[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ if padding_mask_crop is not None:
+ image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image]
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py
new file mode 100644
index 0000000000000000000000000000000000000000..89d4278937fe7d05a2317f0b12f1800e4954ba8d
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_instruct_pix2pix.py
@@ -0,0 +1,827 @@
+# Copyright 2024 The InstructPix2Pix Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import PIL_INTERPOLATION, deprecate, logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from . import StableDiffusionPipelineOutput
+from .safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
+def preprocess(image):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class StableDiffusionInstructPix2PixPipeline(
+ DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin
+):
+ r"""
+ Pipeline for pixel-level image editing by following text instructions (based on Stable Diffusion).
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "image_latents"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ num_inference_steps: int = 100,
+ guidance_scale: float = 7.5,
+ image_guidance_scale: float = 1.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor` `np.ndarray`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image` or tensor representing an image batch to be repainted according to `prompt`. Can also accept
+ image latents as `image`, but if passing latents directly it is not encoded again.
+ num_inference_steps (`int`, *optional*, defaults to 100):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ image_guidance_scale (`float`, *optional*, defaults to 1.5):
+ Push the generated image towards the inital `image`. Image guidance scale is enabled by setting
+ `image_guidance_scale > 1`. Higher image guidance scale encourages generated images that are closely
+ linked to the source `image`, usually at the expense of lower image quality. This pipeline requires a
+ value of at least `1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ ```py
+ >>> import PIL
+ >>> import requests
+ >>> import torch
+ >>> from io import BytesIO
+
+ >>> from diffusers import StableDiffusionInstructPix2PixPipeline
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = "https://huggingface.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png"
+
+ >>> image = download_image(img_url).resize((512, 512))
+
+ >>> pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained(
+ ... "timbrooks/instruct-pix2pix", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "make the mountains snowy"
+ >>> image = pipe(prompt=prompt, image=image).images[0]
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 0. Check inputs
+ self.check_inputs(
+ prompt,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+ self._guidance_scale = guidance_scale
+ self._image_guidance_scale = image_guidance_scale
+
+ device = self._execution_device
+
+ if ip_adapter_image is not None:
+ output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+ image_embeds, negative_image_embeds = self.encode_image(
+ ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+ )
+ if self.do_classifier_free_guidance:
+ image_embeds = torch.cat([image_embeds, negative_image_embeds, negative_image_embeds])
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ # 1. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # check if scheduler is in sigmas space
+ scheduler_is_in_sigma_space = hasattr(self.scheduler, "sigmas")
+
+ # 2. Encode input prompt
+ prompt_embeds = self._encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ # 3. Preprocess image
+ image = self.image_processor.preprocess(image)
+
+ # 4. set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare Image latents
+ image_latents = self.prepare_image_latents(
+ image,
+ batch_size,
+ num_images_per_prompt,
+ prompt_embeds.dtype,
+ device,
+ self.do_classifier_free_guidance,
+ )
+
+ height, width = image_latents.shape[-2:]
+ height = height * self.vae_scale_factor
+ width = width * self.vae_scale_factor
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 7. Check that shapes of latents and image match the UNet channels
+ num_channels_image = image_latents.shape[1]
+ if num_channels_latents + num_channels_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_image`: {num_channels_image} "
+ f" = {num_channels_latents+num_channels_image}. Please verify the config of"
+ " `pipeline.unet` or your `image` input."
+ )
+
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 8.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
+ # 9. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Expand the latents if we are doing classifier free guidance.
+ # The latents are expanded 3 times because for pix2pix the guidance\
+ # is applied for both the text and the input image.
+ latent_model_input = torch.cat([latents] * 3) if self.do_classifier_free_guidance else latents
+
+ # concat latents, image_latents in the channel dimension
+ scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+ scaled_latent_model_input = torch.cat([scaled_latent_model_input, image_latents], dim=1)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ scaled_latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # Hack:
+ # For karras style schedulers the model does classifer free guidance using the
+ # predicted_original_sample instead of the noise_pred. So we need to compute the
+ # predicted_original_sample here if we are using a karras style scheduler.
+ if scheduler_is_in_sigma_space:
+ step_index = (self.scheduler.timesteps == t).nonzero()[0].item()
+ sigma = self.scheduler.sigmas[step_index]
+ noise_pred = latent_model_input - sigma * noise_pred
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_text, noise_pred_image, noise_pred_uncond = noise_pred.chunk(3)
+ noise_pred = (
+ noise_pred_uncond
+ + self.guidance_scale * (noise_pred_text - noise_pred_image)
+ + self.image_guidance_scale * (noise_pred_image - noise_pred_uncond)
+ )
+
+ # Hack:
+ # For karras style schedulers the model does classifer free guidance using the
+ # predicted_original_sample instead of the noise_pred. But the scheduler.step function
+ # expects the noise_pred and computes the predicted_original_sample internally. So we
+ # need to overwrite the noise_pred here such that the value of the computed
+ # predicted_original_sample is correct.
+ if scheduler_is_in_sigma_space:
+ noise_pred = (noise_pred - latents) / (-sigma)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+ image_latents = callback_outputs.pop("image_latents", image_latents)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ """
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ else:
+ prompt_embeds_dtype = self.unet.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ # pix2pix has two negative embeddings, and unlike in other pipelines latents are ordered [prompt_embeds, negative_prompt_embeds, negative_prompt_embeds]
+ prompt_embeds = torch.cat([prompt_embeds, negative_prompt_embeds, negative_prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def check_inputs(
+ self,
+ prompt,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def prepare_image_latents(
+ self, image, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None
+ ):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ image_latents = image
+ else:
+ image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax")
+
+ if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+ # expand image_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // image_latents.shape[0]
+ image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ image_latents = torch.cat([image_latents], dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_image_latents = torch.zeros_like(image_latents)
+ image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0)
+
+ return image_latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def image_guidance_scale(self):
+ return self._image_guidance_scale
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self.guidance_scale > 1.0 and self.image_guidance_scale >= 1.0
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
new file mode 100644
index 0000000000000000000000000000000000000000..918dffe5199db356c543903db024339696c10cf0
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_latent_upscale.py
@@ -0,0 +1,495 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import warnings
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...schedulers import EulerDiscreteScheduler
+from ...utils import deprecate, logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.preprocess
+def preprocess(image):
+ warnings.warn(
+ "The preprocess method is deprecated and will be removed in a future version. Please"
+ " use VaeImageProcessor.preprocess instead",
+ FutureWarning,
+ )
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64
+
+ image = [np.array(i.resize((w, h)))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+class StableDiffusionLatentUpscalePipeline(DiffusionPipeline, StableDiffusionMixin, FromSingleFileMixin):
+ r"""
+ Pipeline for upscaling Stable Diffusion output image resolution by a factor of 2.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A [`EulerDiscreteScheduler`] to be used in combination with `unet` to denoise the encoded image latents.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: EulerDiscreteScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, resample="bicubic")
+
+ def _encode_prompt(self, prompt, device, do_classifier_free_guidance, negative_prompt):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `list(int)`):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ """
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_length=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ text_encoder_out = self.text_encoder(
+ text_input_ids.to(device),
+ output_hidden_states=True,
+ )
+ text_embeddings = text_encoder_out.hidden_states[-1]
+ text_pooler_out = text_encoder_out.pooler_output
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = text_input_ids.shape[-1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_length=True,
+ return_tensors="pt",
+ )
+
+ uncond_encoder_out = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+
+ uncond_embeddings = uncond_encoder_out.hidden_states[-1]
+ uncond_pooler_out = uncond_encoder_out.pooler_output
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+ text_pooler_out = torch.cat([uncond_pooler_out, text_pooler_out])
+
+ return text_embeddings, text_pooler_out
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def check_inputs(self, prompt, image, callback_steps):
+ if not isinstance(prompt, str) and not isinstance(prompt, list):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or `list` but is {type(image)}"
+ )
+
+ # verify batch size of prompt and image are same if image is a list or tensor
+ if isinstance(image, list) or isinstance(image, torch.Tensor):
+ if isinstance(prompt, str):
+ batch_size = 1
+ else:
+ batch_size = len(prompt)
+ if isinstance(image, list):
+ image_batch_size = len(image)
+ else:
+ image_batch_size = image.shape[0] if image.ndim == 4 else 1
+ if batch_size != image_batch_size:
+ raise ValueError(
+ f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}."
+ " Please make sure that passed `prompt` matches the batch size of `image`."
+ )
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height, width)
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ image: PipelineImageInput = None,
+ num_inference_steps: int = 75,
+ guidance_scale: float = 9.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image upscaling.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image` or tensor representing an image batch to be upscaled. If it's a tensor, it can be either a
+ latent output from a Stable Diffusion model or an image tensor in the range `[-1, 1]`. It is considered
+ a `latent` if `image.shape[1]` is `4`; otherwise, it is considered to be an image representation and
+ encoded using this pipeline's `vae` encoder.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Examples:
+ ```py
+ >>> from diffusers import StableDiffusionLatentUpscalePipeline, StableDiffusionPipeline
+ >>> import torch
+
+
+ >>> pipeline = StableDiffusionPipeline.from_pretrained(
+ ... "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16
+ ... )
+ >>> pipeline.to("cuda")
+
+ >>> model_id = "stabilityai/sd-x2-latent-upscaler"
+ >>> upscaler = StableDiffusionLatentUpscalePipeline.from_pretrained(model_id, torch_dtype=torch.float16)
+ >>> upscaler.to("cuda")
+
+ >>> prompt = "a photo of an astronaut high resolution, unreal engine, ultra realistic"
+ >>> generator = torch.manual_seed(33)
+
+ >>> low_res_latents = pipeline(prompt, generator=generator, output_type="latent").images
+
+ >>> with torch.no_grad():
+ ... image = pipeline.decode_latents(low_res_latents)
+ >>> image = pipeline.numpy_to_pil(image)[0]
+
+ >>> image.save("../images/a1.png")
+
+ >>> upscaled_image = upscaler(
+ ... prompt=prompt,
+ ... image=low_res_latents,
+ ... num_inference_steps=20,
+ ... guidance_scale=0,
+ ... generator=generator,
+ ... ).images[0]
+
+ >>> upscaled_image.save("../images/a2.png")
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images.
+ """
+
+ # 1. Check inputs
+ self.check_inputs(prompt, image, callback_steps)
+
+ # 2. Define call parameters
+ batch_size = 1 if isinstance(prompt, str) else len(prompt)
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ if guidance_scale == 0:
+ prompt = [""] * batch_size
+
+ # 3. Encode input prompt
+ text_embeddings, text_pooler_out = self._encode_prompt(
+ prompt, device, do_classifier_free_guidance, negative_prompt
+ )
+
+ # 4. Preprocess image
+ image = self.image_processor.preprocess(image)
+ image = image.to(dtype=text_embeddings.dtype, device=device)
+ if image.shape[1] == 3:
+ # encode image if not in latent-space yet
+ image = self.vae.encode(image).latent_dist.sample() * self.vae.config.scaling_factor
+
+ # 5. set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ batch_multiplier = 2 if do_classifier_free_guidance else 1
+ image = image[None, :] if image.ndim == 3 else image
+ image = torch.cat([image] * batch_multiplier)
+
+ # 5. Add noise to image (set to be 0):
+ # (see below notes from the author):
+ # "the This step theoretically can make the model work better on out-of-distribution inputs, but mostly just seems to make it match the input less, so it's turned off by default."
+ noise_level = torch.tensor([0.0], dtype=torch.float32, device=device)
+ noise_level = torch.cat([noise_level] * image.shape[0])
+ inv_noise_level = (noise_level**2 + 1) ** (-0.5)
+
+ image_cond = F.interpolate(image, scale_factor=2, mode="nearest") * inv_noise_level[:, None, None, None]
+ image_cond = image_cond.to(text_embeddings.dtype)
+
+ noise_level_embed = torch.cat(
+ [
+ torch.ones(text_pooler_out.shape[0], 64, dtype=text_pooler_out.dtype, device=device),
+ torch.zeros(text_pooler_out.shape[0], 64, dtype=text_pooler_out.dtype, device=device),
+ ],
+ dim=1,
+ )
+
+ timestep_condition = torch.cat([noise_level_embed, text_pooler_out], dim=1)
+
+ # 6. Prepare latent variables
+ height, width = image.shape[2:]
+ num_channels_latents = self.vae.config.latent_channels
+ latents = self.prepare_latents(
+ batch_size,
+ num_channels_latents,
+ height * 2, # 2x upscale
+ width * 2,
+ text_embeddings.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 7. Check that sizes of image and latents match
+ num_channels_image = image.shape[1]
+ if num_channels_latents + num_channels_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_image`: {num_channels_image} "
+ f" = {num_channels_latents+num_channels_image}. Please verify the config of"
+ " `pipeline.unet` or your `image` input."
+ )
+
+ # 9. Denoising loop
+ num_warmup_steps = 0
+
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ sigma = self.scheduler.sigmas[i]
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ scaled_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ scaled_model_input = torch.cat([scaled_model_input, image_cond], dim=1)
+ # preconditioning parameter based on Karras et al. (2022) (table 1)
+ timestep = torch.log(sigma) * 0.25
+
+ noise_pred = self.unet(
+ scaled_model_input,
+ timestep,
+ encoder_hidden_states=text_embeddings,
+ timestep_cond=timestep_condition,
+ ).sample
+
+ # in original repo, the output contains a variance channel that's not used
+ noise_pred = noise_pred[:, :-1]
+
+ # apply preconditioning, based on table 1 in Karras et al. (2022)
+ inv_sigma = 1 / (sigma**2 + 1)
+ noise_pred = inv_sigma * latent_model_input + self.scheduler.scale_model_input(sigma, t) * noise_pred
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = latents
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py
new file mode 100644
index 0000000000000000000000000000000000000000..2d04cf41d9b53d8d70f09dbbef524efd2d796a52
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_diffusion_upscale.py
@@ -0,0 +1,808 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import warnings
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import DDPMScheduler, KarrasDiffusionSchedulers
+from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from . import StableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def preprocess(image):
+ warnings.warn(
+ "The preprocess method is deprecated and will be removed in a future version. Please"
+ " use VaeImageProcessor.preprocess instead",
+ FutureWarning,
+ )
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 64 for x in (w, h)) # resize to integer multiple of 64
+
+ image = [np.array(i.resize((w, h)))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+class StableDiffusionUpscalePipeline(
+ DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
+ r"""
+ Pipeline for text-guided image super-resolution using Stable Diffusion 2.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ low_res_scheduler ([`SchedulerMixin`]):
+ A scheduler used to add initial noise to the low resolution conditioning image. It must be an instance of
+ [`DDPMScheduler`].
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["watermarker", "safety_checker", "feature_extractor"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ low_res_scheduler: DDPMScheduler,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: Optional[Any] = None,
+ feature_extractor: Optional[CLIPImageProcessor] = None,
+ watermarker: Optional[Any] = None,
+ max_noise_level: int = 350,
+ ):
+ super().__init__()
+
+ if hasattr(
+ vae, "config"
+ ): # check if vae has a config attribute `scaling_factor` and if it is set to 0.08333, else set it to 0.08333 and deprecate
+ is_vae_scaling_factor_set_to_0_08333 = (
+ hasattr(vae.config, "scaling_factor") and vae.config.scaling_factor == 0.08333
+ )
+ if not is_vae_scaling_factor_set_to_0_08333:
+ deprecation_message = (
+ "The configuration file of the vae does not contain `scaling_factor` or it is set to"
+ f" {vae.config.scaling_factor}, which seems highly unlikely. If your checkpoint is a fine-tuned"
+ " version of `stabilityai/stable-diffusion-x4-upscaler` you should change 'scaling_factor' to"
+ " 0.08333 Please make sure to update the config accordingly, as not doing so might lead to"
+ " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging"
+ " Face Hub, it would be very nice if you could open a Pull Request for the `vae/config.json` file"
+ )
+ deprecate("wrong scaling_factor", "1.0.0", deprecation_message, standard_warn=False)
+ vae.register_to_config(scaling_factor=0.08333)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ low_res_scheduler=low_res_scheduler,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ watermarker=watermarker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, resample="bicubic")
+ self.register_to_config(max_noise_level=max_noise_level)
+
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is not None:
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, nsfw_detected, watermark_detected = self.safety_checker(
+ images=image,
+ clip_input=safety_checker_input.pixel_values.to(dtype=dtype),
+ )
+ else:
+ nsfw_detected = None
+ watermark_detected = None
+
+ if hasattr(self, "unet_offload_hook") and self.unet_offload_hook is not None:
+ self.unet_offload_hook.offload()
+
+ return image, nsfw_detected, watermark_detected
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ noise_level,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, np.ndarray)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `np.ndarray`, `PIL.Image.Image` or `list` but is {type(image)}"
+ )
+
+ # verify batch size of prompt and image are same if image is a list or tensor or numpy array
+ if isinstance(image, list) or isinstance(image, torch.Tensor) or isinstance(image, np.ndarray):
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if isinstance(image, list):
+ image_batch_size = len(image)
+ else:
+ image_batch_size = image.shape[0]
+ if batch_size != image_batch_size:
+ raise ValueError(
+ f"`prompt` has batch size {batch_size} and `image` has batch size {image_batch_size}."
+ " Please make sure that passed `prompt` matches the batch size of `image`."
+ )
+
+ # check noise level
+ if noise_level > self.config.max_noise_level:
+ raise ValueError(f"`noise_level` has to be <= {self.config.max_noise_level} but is {noise_level}")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height, width)
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: PipelineImageInput = None,
+ num_inference_steps: int = 75,
+ guidance_scale: float = 9.0,
+ noise_level: int = 20,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: int = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
+ `Image` or tensor representing an image batch to be upscaled.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+ ```py
+ >>> import requests
+ >>> from PIL import Image
+ >>> from io import BytesIO
+ >>> from diffusers import StableDiffusionUpscalePipeline
+ >>> import torch
+
+ >>> # load model and scheduler
+ >>> model_id = "stabilityai/stable-diffusion-x4-upscaler"
+ >>> pipeline = StableDiffusionUpscalePipeline.from_pretrained(
+ ... model_id, revision="fp16", torch_dtype=torch.float16
+ ... )
+ >>> pipeline = pipeline.to("cuda")
+
+ >>> # let's download an image
+ >>> url = "https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/sd2-upscale/low_res_cat.png"
+ >>> response = requests.get(url)
+ >>> low_res_img = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> low_res_img = low_res_img.resize((128, 128))
+ >>> prompt = "a white cat"
+
+ >>> upscaled_image = pipeline(prompt=prompt, image=low_res_img).images[0]
+ >>> upscaled_image.save("upsampled_cat.png")
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ image,
+ noise_level,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Preprocess image
+ image = self.image_processor.preprocess(image)
+ image = image.to(dtype=prompt_embeds.dtype, device=device)
+
+ # 5. set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Add noise to image
+ noise_level = torch.tensor([noise_level], dtype=torch.long, device=device)
+ noise = randn_tensor(image.shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
+ image = self.low_res_scheduler.add_noise(image, noise, noise_level)
+
+ batch_multiplier = 2 if do_classifier_free_guidance else 1
+ image = torch.cat([image] * batch_multiplier * num_images_per_prompt)
+ noise_level = torch.cat([noise_level] * image.shape[0])
+
+ # 6. Prepare latent variables
+ height, width = image.shape[2:]
+ num_channels_latents = self.vae.config.latent_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 7. Check that sizes of image and latents match
+ num_channels_image = image.shape[1]
+ if num_channels_latents + num_channels_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_image`: {num_channels_image} "
+ f" = {num_channels_latents+num_channels_image}. Please verify the config of"
+ " `pipeline.unet` or your `image` input."
+ )
+
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 9. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+ latent_model_input = torch.cat([latent_model_input, image], dim=1)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ class_labels=noise_level,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+
+ # Ensure latents are always the same type as the VAE
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+
+ image, has_nsfw_concept, _ = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # 11. Apply watermark
+ if output_type == "pil" and self.watermarker is not None:
+ image = self.watermarker.apply_watermark(image)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_unclip.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_unclip.py
new file mode 100644
index 0000000000000000000000000000000000000000..c62e0f4ec50f259d9923275c08c82e7dcb8608cb
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_unclip.py
@@ -0,0 +1,932 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+from transformers.models.clip.modeling_clip import CLIPTextModelOutput
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, PriorTransformer, UNet2DConditionModel
+from ...models.embeddings import get_timestep_embedding
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin
+from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableUnCLIPPipeline
+
+ >>> pipe = StableUnCLIPPipeline.from_pretrained(
+ ... "fusing/stable-unclip-2-1-l", torch_dtype=torch.float16
+ ... ) # TODO update model path
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> images = pipe(prompt).images
+ >>> images[0].save("astronaut_horse.png")
+ ```
+"""
+
+
+class StableUnCLIPPipeline(DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin):
+ """
+ Pipeline for text-to-image generation using stable unCLIP.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ prior_tokenizer ([`CLIPTokenizer`]):
+ A [`CLIPTokenizer`].
+ prior_text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen [`CLIPTextModelWithProjection`] text-encoder.
+ prior ([`PriorTransformer`]):
+ The canonincal unCLIP prior to approximate the image embedding from the text embedding.
+ prior_scheduler ([`KarrasDiffusionSchedulers`]):
+ Scheduler used in the prior denoising process.
+ image_normalizer ([`StableUnCLIPImageNormalizer`]):
+ Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image
+ embeddings after the noise has been applied.
+ image_noising_scheduler ([`KarrasDiffusionSchedulers`]):
+ Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined
+ by the `noise_level`.
+ tokenizer ([`CLIPTokenizer`]):
+ A [`CLIPTokenizer`].
+ text_encoder ([`CLIPTextModel`]):
+ Frozen [`CLIPTextModel`] text-encoder.
+ unet ([`UNet2DConditionModel`]):
+ A [`UNet2DConditionModel`] to denoise the encoded image latents.
+ scheduler ([`KarrasDiffusionSchedulers`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ """
+
+ _exclude_from_cpu_offload = ["prior", "image_normalizer"]
+ model_cpu_offload_seq = "text_encoder->prior_text_encoder->unet->vae"
+
+ # prior components
+ prior_tokenizer: CLIPTokenizer
+ prior_text_encoder: CLIPTextModelWithProjection
+ prior: PriorTransformer
+ prior_scheduler: KarrasDiffusionSchedulers
+
+ # image noising components
+ image_normalizer: StableUnCLIPImageNormalizer
+ image_noising_scheduler: KarrasDiffusionSchedulers
+
+ # regular denoising components
+ tokenizer: CLIPTokenizer
+ text_encoder: CLIPTextModel
+ unet: UNet2DConditionModel
+ scheduler: KarrasDiffusionSchedulers
+
+ vae: AutoencoderKL
+
+ def __init__(
+ self,
+ # prior components
+ prior_tokenizer: CLIPTokenizer,
+ prior_text_encoder: CLIPTextModelWithProjection,
+ prior: PriorTransformer,
+ prior_scheduler: KarrasDiffusionSchedulers,
+ # image noising components
+ image_normalizer: StableUnCLIPImageNormalizer,
+ image_noising_scheduler: KarrasDiffusionSchedulers,
+ # regular denoising components
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModelWithProjection,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ # vae
+ vae: AutoencoderKL,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ prior_tokenizer=prior_tokenizer,
+ prior_text_encoder=prior_text_encoder,
+ prior=prior,
+ prior_scheduler=prior_scheduler,
+ image_normalizer=image_normalizer,
+ image_noising_scheduler=image_noising_scheduler,
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ vae=vae,
+ )
+
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline._encode_prompt with _encode_prompt->_encode_prior_prompt, tokenizer->prior_tokenizer, text_encoder->prior_text_encoder
+ def _encode_prior_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None,
+ text_attention_mask: Optional[torch.Tensor] = None,
+ ):
+ if text_model_output is None:
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.prior_tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.prior_tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ text_mask = text_inputs.attention_mask.bool().to(device)
+
+ untruncated_ids = self.prior_tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.prior_tokenizer.batch_decode(
+ untruncated_ids[:, self.prior_tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.prior_tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.prior_tokenizer.model_max_length]
+
+ prior_text_encoder_output = self.prior_text_encoder(text_input_ids.to(device))
+
+ prompt_embeds = prior_text_encoder_output.text_embeds
+ text_enc_hid_states = prior_text_encoder_output.last_hidden_state
+
+ else:
+ batch_size = text_model_output[0].shape[0]
+ prompt_embeds, text_enc_hid_states = text_model_output[0], text_model_output[1]
+ text_mask = text_attention_mask
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_enc_hid_states = text_enc_hid_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens = [""] * batch_size
+
+ uncond_input = self.prior_tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.prior_tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_text_mask = uncond_input.attention_mask.bool().to(device)
+ negative_prompt_embeds_prior_text_encoder_output = self.prior_text_encoder(
+ uncond_input.input_ids.to(device)
+ )
+
+ negative_prompt_embeds = negative_prompt_embeds_prior_text_encoder_output.text_embeds
+ uncond_text_enc_hid_states = negative_prompt_embeds_prior_text_encoder_output.last_hidden_state
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_enc_hid_states.shape[1]
+ uncond_text_enc_hid_states = uncond_text_enc_hid_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_enc_hid_states = uncond_text_enc_hid_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_enc_hid_states = torch.cat([uncond_text_enc_hid_states, text_enc_hid_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_enc_hid_states, text_mask
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs with prepare_extra_step_kwargs->prepare_prior_extra_step_kwargs, scheduler->prior_scheduler
+ def prepare_prior_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the prior_scheduler step, since not all prior_schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other prior_schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.prior_scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the prior_scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.prior_scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ noise_level,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two."
+ )
+
+ if prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+
+ if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ "Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined."
+ )
+
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps:
+ raise ValueError(
+ f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive."
+ )
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def noise_image_embeddings(
+ self,
+ image_embeds: torch.Tensor,
+ noise_level: int,
+ noise: Optional[torch.FloatTensor] = None,
+ generator: Optional[torch.Generator] = None,
+ ):
+ """
+ Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher
+ `noise_level` increases the variance in the final un-noised images.
+
+ The noise is applied in two ways:
+ 1. A noise schedule is applied directly to the embeddings.
+ 2. A vector of sinusoidal time embeddings are appended to the output.
+
+ In both cases, the amount of noise is controlled by the same `noise_level`.
+
+ The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
+ """
+ if noise is None:
+ noise = randn_tensor(
+ image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype
+ )
+
+ noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device)
+
+ self.image_normalizer.to(image_embeds.device)
+ image_embeds = self.image_normalizer.scale(image_embeds)
+
+ image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise)
+
+ image_embeds = self.image_normalizer.unscale(image_embeds)
+
+ noise_level = get_timestep_embedding(
+ timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0
+ )
+
+ # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors,
+ # but we might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ noise_level = noise_level.to(image_embeds.dtype)
+
+ image_embeds = torch.cat((image_embeds, noise_level), 1)
+
+ return image_embeds
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ # regular denoising process args
+ prompt: Optional[Union[str, List[str]]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 20,
+ guidance_scale: float = 10.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[torch.Generator] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ noise_level: int = 0,
+ # prior args
+ prior_num_inference_steps: int = 25,
+ prior_guidance_scale: float = 4.0,
+ prior_latents: Optional[torch.FloatTensor] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 20):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 10.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ noise_level (`int`, *optional*, defaults to `0`):
+ The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in
+ the final un-noised images. See [`StableUnCLIPPipeline.noise_image_embeddings`] for more details.
+ prior_num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps in the prior denoising process. More denoising steps usually lead to a
+ higher quality image at the expense of slower inference.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ prior_latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ embedding generation in the prior denoising process. Can be used to tweak the same generation with
+ different prompts. If not provided, a latents tensor is generated by sampling using the supplied random
+ `generator`.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning
+ a tuple, the first element is a list with the generated images.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt=prompt,
+ height=height,
+ width=width,
+ callback_steps=callback_steps,
+ noise_level=noise_level,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ batch_size = batch_size * num_images_per_prompt
+
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ prior_do_classifier_free_guidance = prior_guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prior_prompt_embeds, prior_text_encoder_hidden_states, prior_text_mask = self._encode_prior_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=prior_do_classifier_free_guidance,
+ )
+
+ # 4. Prepare prior timesteps
+ self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device)
+ prior_timesteps_tensor = self.prior_scheduler.timesteps
+
+ # 5. Prepare prior latent variables
+ embedding_dim = self.prior.config.embedding_dim
+ prior_latents = self.prepare_latents(
+ (batch_size, embedding_dim),
+ prior_prompt_embeds.dtype,
+ device,
+ generator,
+ prior_latents,
+ self.prior_scheduler,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ prior_extra_step_kwargs = self.prepare_prior_extra_step_kwargs(generator, eta)
+
+ # 7. Prior denoising loop
+ for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([prior_latents] * 2) if prior_do_classifier_free_guidance else prior_latents
+ latent_model_input = self.prior_scheduler.scale_model_input(latent_model_input, t)
+
+ predicted_image_embedding = self.prior(
+ latent_model_input,
+ timestep=t,
+ proj_embedding=prior_prompt_embeds,
+ encoder_hidden_states=prior_text_encoder_hidden_states,
+ attention_mask=prior_text_mask,
+ ).predicted_image_embedding
+
+ if prior_do_classifier_free_guidance:
+ predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2)
+ predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * (
+ predicted_image_embedding_text - predicted_image_embedding_uncond
+ )
+
+ prior_latents = self.prior_scheduler.step(
+ predicted_image_embedding,
+ timestep=t,
+ sample=prior_latents,
+ **prior_extra_step_kwargs,
+ return_dict=False,
+ )[0]
+
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, prior_latents)
+
+ prior_latents = self.prior.post_process_latents(prior_latents)
+
+ image_embeds = prior_latents
+
+ # done prior
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 8. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 9. Prepare image embeddings
+ image_embeds = self.noise_image_embeddings(
+ image_embeds=image_embeds,
+ noise_level=noise_level,
+ generator=generator,
+ )
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = torch.zeros_like(image_embeds)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ image_embeds = torch.cat([negative_prompt_embeds, image_embeds])
+
+ # 10. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 11. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ latents = self.prepare_latents(
+ shape=shape,
+ dtype=prompt_embeds.dtype,
+ device=device,
+ generator=generator,
+ latents=latents,
+ scheduler=self.scheduler,
+ )
+
+ # 12. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 13. Denoising loop
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ class_labels=image_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = latents
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..9b85d9e6b1a49598d46d757b77f1a75ce52026f2
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/pipeline_stable_unclip_img2img.py
@@ -0,0 +1,839 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.embeddings import get_timestep_embedding
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput, StableDiffusionMixin
+from .stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import requests
+ >>> import torch
+ >>> from PIL import Image
+ >>> from io import BytesIO
+
+ >>> from diffusers import StableUnCLIPImg2ImgPipeline
+
+ >>> pipe = StableUnCLIPImg2ImgPipeline.from_pretrained(
+ ... "fusing/stable-unclip-2-1-l-img2img", torch_dtype=torch.float16
+ ... ) # TODO update model path
+ >>> pipe = pipe.to("cuda")
+
+ >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+
+ >>> response = requests.get(url)
+ >>> init_image = Image.open(BytesIO(response.content)).convert("RGB")
+ >>> init_image = init_image.resize((768, 512))
+
+ >>> prompt = "A fantasy landscape, trending on artstation"
+
+ >>> images = pipe(prompt, init_image).images
+ >>> images[0].save("fantasy_landscape.png")
+ ```
+"""
+
+
+class StableUnCLIPImg2ImgPipeline(
+ DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin
+):
+ """
+ Pipeline for text-guided image-to-image generation using stable unCLIP.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ feature_extractor ([`CLIPImageProcessor`]):
+ Feature extractor for image pre-processing before being encoded.
+ image_encoder ([`CLIPVisionModelWithProjection`]):
+ CLIP vision model for encoding images.
+ image_normalizer ([`StableUnCLIPImageNormalizer`]):
+ Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image
+ embeddings after the noise has been applied.
+ image_noising_scheduler ([`KarrasDiffusionSchedulers`]):
+ Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined
+ by the `noise_level`.
+ tokenizer (`~transformers.CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`)].
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen [`~transformers.CLIPTextModel`] text-encoder.
+ unet ([`UNet2DConditionModel`]):
+ A [`UNet2DConditionModel`] to denoise the encoded image latents.
+ scheduler ([`KarrasDiffusionSchedulers`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _exclude_from_cpu_offload = ["image_normalizer"]
+
+ # image encoding components
+ feature_extractor: CLIPImageProcessor
+ image_encoder: CLIPVisionModelWithProjection
+
+ # image noising components
+ image_normalizer: StableUnCLIPImageNormalizer
+ image_noising_scheduler: KarrasDiffusionSchedulers
+
+ # regular denoising components
+ tokenizer: CLIPTokenizer
+ text_encoder: CLIPTextModel
+ unet: UNet2DConditionModel
+ scheduler: KarrasDiffusionSchedulers
+
+ vae: AutoencoderKL
+
+ def __init__(
+ self,
+ # image encoding components
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection,
+ # image noising components
+ image_normalizer: StableUnCLIPImageNormalizer,
+ image_noising_scheduler: KarrasDiffusionSchedulers,
+ # regular denoising components
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModel,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ # vae
+ vae: AutoencoderKL,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ image_normalizer=image_normalizer,
+ image_noising_scheduler=image_noising_scheduler,
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ vae=vae,
+ )
+
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ def _encode_image(
+ self,
+ image,
+ device,
+ batch_size,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ noise_level,
+ generator,
+ image_embeds,
+ ):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if isinstance(image, PIL.Image.Image):
+ # the image embedding should repeated so it matches the total batch size of the prompt
+ repeat_by = batch_size
+ else:
+ # assume the image input is already properly batched and just needs to be repeated so
+ # it matches the num_images_per_prompt.
+ #
+ # NOTE(will) this is probably missing a few number of side cases. I.e. batched/non-batched
+ # `image_embeds`. If those happen to be common use cases, let's think harder about
+ # what the expected dimensions of inputs should be and how we handle the encoding.
+ repeat_by = num_images_per_prompt
+
+ if image_embeds is None:
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(images=image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ image_embeds = self.image_encoder(image).image_embeds
+
+ image_embeds = self.noise_image_embeddings(
+ image_embeds=image_embeds,
+ noise_level=noise_level,
+ generator=generator,
+ )
+
+ # duplicate image embeddings for each generation per prompt, using mps friendly method
+ image_embeds = image_embeds.unsqueeze(1)
+ bs_embed, seq_len, _ = image_embeds.shape
+ image_embeds = image_embeds.repeat(1, repeat_by, 1)
+ image_embeds = image_embeds.view(bs_embed * repeat_by, seq_len, -1)
+ image_embeds = image_embeds.squeeze(1)
+
+ if do_classifier_free_guidance:
+ negative_prompt_embeds = torch.zeros_like(image_embeds)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ image_embeds = torch.cat([negative_prompt_embeds, image_embeds])
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ height,
+ width,
+ callback_steps,
+ noise_level,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ image_embeds=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Please make sure to define only one of the two."
+ )
+
+ if prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+
+ if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ "Provide either `negative_prompt` or `negative_prompt_embeds`. Cannot leave both `negative_prompt` and `negative_prompt_embeds` undefined."
+ )
+
+ if prompt is not None and negative_prompt is not None:
+ if type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps:
+ raise ValueError(
+ f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive."
+ )
+
+ if image is not None and image_embeds is not None:
+ raise ValueError(
+ "Provide either `image` or `image_embeds`. Please make sure to define only one of the two."
+ )
+
+ if image is None and image_embeds is None:
+ raise ValueError(
+ "Provide either `image` or `image_embeds`. Cannot leave both `image` and `image_embeds` undefined."
+ )
+
+ if image is not None:
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+ f" {type(image)}"
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_unclip.StableUnCLIPPipeline.noise_image_embeddings
+ def noise_image_embeddings(
+ self,
+ image_embeds: torch.Tensor,
+ noise_level: int,
+ noise: Optional[torch.FloatTensor] = None,
+ generator: Optional[torch.Generator] = None,
+ ):
+ """
+ Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher
+ `noise_level` increases the variance in the final un-noised images.
+
+ The noise is applied in two ways:
+ 1. A noise schedule is applied directly to the embeddings.
+ 2. A vector of sinusoidal time embeddings are appended to the output.
+
+ In both cases, the amount of noise is controlled by the same `noise_level`.
+
+ The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
+ """
+ if noise is None:
+ noise = randn_tensor(
+ image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype
+ )
+
+ noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device)
+
+ self.image_normalizer.to(image_embeds.device)
+ image_embeds = self.image_normalizer.scale(image_embeds)
+
+ image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise)
+
+ image_embeds = self.image_normalizer.unscale(image_embeds)
+
+ noise_level = get_timestep_embedding(
+ timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0
+ )
+
+ # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors,
+ # but we might actually be running in fp16. so we need to cast here.
+ # there might be better ways to encapsulate this.
+ noise_level = noise_level.to(image_embeds.dtype)
+
+ image_embeds = torch.cat((image_embeds, noise_level), 1)
+
+ return image_embeds
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 20,
+ guidance_scale: float = 10,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[torch.Generator] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ noise_level: int = 0,
+ image_embeds: Optional[torch.FloatTensor] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, either `prompt_embeds` will be
+ used or prompt is initialized to `""`.
+ image (`torch.FloatTensor` or `PIL.Image.Image`):
+ `Image` or tensor representing an image batch. The image is encoded to its CLIP embedding which the
+ `unet` is conditioned on. The image is _not_ encoded by the `vae` and then used as the latents in the
+ denoising process like it is in the standard Stable Diffusion text-guided image variation process.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 20):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 10.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ noise_level (`int`, *optional*, defaults to `0`):
+ The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in
+ the final un-noised images. See [`StableUnCLIPPipeline.noise_image_embeddings`] for more details.
+ image_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated CLIP embeddings to condition the `unet` on. These latents are not used in the denoising
+ process. If you want to provide pre-generated latents, pass them to `__call__` as `latents`.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When returning
+ a tuple, the first element is a list with the generated images.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ if prompt is None and prompt_embeds is None:
+ prompt = len(image) * [""] if isinstance(image, list) else ""
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt=prompt,
+ image=image,
+ height=height,
+ width=width,
+ callback_steps=callback_steps,
+ noise_level=noise_level,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ image_embeds=image_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ batch_size = batch_size * num_images_per_prompt
+
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Encoder input image
+ noise_level = torch.tensor([noise_level], device=device)
+ image_embeds = self._encode_image(
+ image=image,
+ device=device,
+ batch_size=batch_size,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ noise_level=noise_level,
+ generator=generator,
+ image_embeds=image_embeds,
+ )
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size=batch_size,
+ num_channels_latents=num_channels_latents,
+ height=height,
+ width=width,
+ dtype=prompt_embeds.dtype,
+ device=device,
+ generator=generator,
+ latents=latents,
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 8. Denoising loop
+ for i, t in enumerate(self.progress_bar(timesteps)):
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ class_labels=image_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 9. Post-processing
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = latents
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion/safety_checker.py b/model/diffusers_c/pipelines/stable_diffusion/safety_checker.py
new file mode 100644
index 0000000000000000000000000000000000000000..6cc4d26f29b44ffc6ac3c0822e344bbc785e22e2
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/safety_checker.py
@@ -0,0 +1,125 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+import torch
+import torch.nn as nn
+from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel
+
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+def cosine_distance(image_embeds, text_embeds):
+ normalized_image_embeds = nn.functional.normalize(image_embeds)
+ normalized_text_embeds = nn.functional.normalize(text_embeds)
+ return torch.mm(normalized_image_embeds, normalized_text_embeds.t())
+
+
+class StableDiffusionSafetyChecker(PreTrainedModel):
+ config_class = CLIPConfig
+
+ _no_split_modules = ["CLIPEncoderLayer"]
+
+ def __init__(self, config: CLIPConfig):
+ super().__init__(config)
+
+ self.vision_model = CLIPVisionModel(config.vision_config)
+ self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False)
+
+ self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False)
+ self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False)
+
+ self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False)
+ self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False)
+
+ @torch.no_grad()
+ def forward(self, clip_input, images):
+ pooled_output = self.vision_model(clip_input)[1] # pooled_output
+ image_embeds = self.visual_projection(pooled_output)
+
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy()
+ cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy()
+
+ result = []
+ batch_size = image_embeds.shape[0]
+ for i in range(batch_size):
+ result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []}
+
+ # increase this value to create a stronger `nfsw` filter
+ # at the cost of increasing the possibility of filtering benign images
+ adjustment = 0.0
+
+ for concept_idx in range(len(special_cos_dist[0])):
+ concept_cos = special_cos_dist[i][concept_idx]
+ concept_threshold = self.special_care_embeds_weights[concept_idx].item()
+ result_img["special_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3)
+ if result_img["special_scores"][concept_idx] > 0:
+ result_img["special_care"].append({concept_idx, result_img["special_scores"][concept_idx]})
+ adjustment = 0.01
+
+ for concept_idx in range(len(cos_dist[0])):
+ concept_cos = cos_dist[i][concept_idx]
+ concept_threshold = self.concept_embeds_weights[concept_idx].item()
+ result_img["concept_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3)
+ if result_img["concept_scores"][concept_idx] > 0:
+ result_img["bad_concepts"].append(concept_idx)
+
+ result.append(result_img)
+
+ has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result]
+
+ for idx, has_nsfw_concept in enumerate(has_nsfw_concepts):
+ if has_nsfw_concept:
+ if torch.is_tensor(images) or torch.is_tensor(images[0]):
+ images[idx] = torch.zeros_like(images[idx]) # black image
+ else:
+ images[idx] = np.zeros(images[idx].shape) # black image
+
+ if any(has_nsfw_concepts):
+ logger.warning(
+ "Potential NSFW content was detected in one or more images. A black image will be returned instead."
+ " Try again with a different prompt and/or seed."
+ )
+
+ return images, has_nsfw_concepts
+
+ @torch.no_grad()
+ def forward_onnx(self, clip_input: torch.FloatTensor, images: torch.FloatTensor):
+ pooled_output = self.vision_model(clip_input)[1] # pooled_output
+ image_embeds = self.visual_projection(pooled_output)
+
+ special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds)
+ cos_dist = cosine_distance(image_embeds, self.concept_embeds)
+
+ # increase this value to create a stronger `nsfw` filter
+ # at the cost of increasing the possibility of filtering benign images
+ adjustment = 0.0
+
+ special_scores = special_cos_dist - self.special_care_embeds_weights + adjustment
+ # special_scores = special_scores.round(decimals=3)
+ special_care = torch.any(special_scores > 0, dim=1)
+ special_adjustment = special_care * 0.01
+ special_adjustment = special_adjustment.unsqueeze(1).expand(-1, cos_dist.shape[1])
+
+ concept_scores = (cos_dist - self.concept_embeds_weights) + special_adjustment
+ # concept_scores = concept_scores.round(decimals=3)
+ has_nsfw_concepts = torch.any(concept_scores > 0, dim=1)
+
+ images[has_nsfw_concepts] = 0.0 # black image
+
+ return images, has_nsfw_concepts
diff --git a/model/diffusers_c/pipelines/stable_diffusion/safety_checker_flax.py b/model/diffusers_c/pipelines/stable_diffusion/safety_checker_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..571a4f2d771015784926bcbf240d1f367b4fc293
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/safety_checker_flax.py
@@ -0,0 +1,112 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Tuple
+
+import jax
+import jax.numpy as jnp
+from flax import linen as nn
+from flax.core.frozen_dict import FrozenDict
+from transformers import CLIPConfig, FlaxPreTrainedModel
+from transformers.models.clip.modeling_flax_clip import FlaxCLIPVisionModule
+
+
+def jax_cosine_distance(emb_1, emb_2, eps=1e-12):
+ norm_emb_1 = jnp.divide(emb_1.T, jnp.clip(jnp.linalg.norm(emb_1, axis=1), a_min=eps)).T
+ norm_emb_2 = jnp.divide(emb_2.T, jnp.clip(jnp.linalg.norm(emb_2, axis=1), a_min=eps)).T
+ return jnp.matmul(norm_emb_1, norm_emb_2.T)
+
+
+class FlaxStableDiffusionSafetyCheckerModule(nn.Module):
+ config: CLIPConfig
+ dtype: jnp.dtype = jnp.float32
+
+ def setup(self):
+ self.vision_model = FlaxCLIPVisionModule(self.config.vision_config)
+ self.visual_projection = nn.Dense(self.config.projection_dim, use_bias=False, dtype=self.dtype)
+
+ self.concept_embeds = self.param("concept_embeds", jax.nn.initializers.ones, (17, self.config.projection_dim))
+ self.special_care_embeds = self.param(
+ "special_care_embeds", jax.nn.initializers.ones, (3, self.config.projection_dim)
+ )
+
+ self.concept_embeds_weights = self.param("concept_embeds_weights", jax.nn.initializers.ones, (17,))
+ self.special_care_embeds_weights = self.param("special_care_embeds_weights", jax.nn.initializers.ones, (3,))
+
+ def __call__(self, clip_input):
+ pooled_output = self.vision_model(clip_input)[1]
+ image_embeds = self.visual_projection(pooled_output)
+
+ special_cos_dist = jax_cosine_distance(image_embeds, self.special_care_embeds)
+ cos_dist = jax_cosine_distance(image_embeds, self.concept_embeds)
+
+ # increase this value to create a stronger `nfsw` filter
+ # at the cost of increasing the possibility of filtering benign image inputs
+ adjustment = 0.0
+
+ special_scores = special_cos_dist - self.special_care_embeds_weights[None, :] + adjustment
+ special_scores = jnp.round(special_scores, 3)
+ is_special_care = jnp.any(special_scores > 0, axis=1, keepdims=True)
+ # Use a lower threshold if an image has any special care concept
+ special_adjustment = is_special_care * 0.01
+
+ concept_scores = cos_dist - self.concept_embeds_weights[None, :] + special_adjustment
+ concept_scores = jnp.round(concept_scores, 3)
+ has_nsfw_concepts = jnp.any(concept_scores > 0, axis=1)
+
+ return has_nsfw_concepts
+
+
+class FlaxStableDiffusionSafetyChecker(FlaxPreTrainedModel):
+ config_class = CLIPConfig
+ main_input_name = "clip_input"
+ module_class = FlaxStableDiffusionSafetyCheckerModule
+
+ def __init__(
+ self,
+ config: CLIPConfig,
+ input_shape: Optional[Tuple] = None,
+ seed: int = 0,
+ dtype: jnp.dtype = jnp.float32,
+ _do_init: bool = True,
+ **kwargs,
+ ):
+ if input_shape is None:
+ input_shape = (1, 224, 224, 3)
+ module = self.module_class(config=config, dtype=dtype, **kwargs)
+ super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
+
+ def init_weights(self, rng: jax.Array, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
+ # init input tensor
+ clip_input = jax.random.normal(rng, input_shape)
+
+ params_rng, dropout_rng = jax.random.split(rng)
+ rngs = {"params": params_rng, "dropout": dropout_rng}
+
+ random_params = self.module.init(rngs, clip_input)["params"]
+
+ return random_params
+
+ def __call__(
+ self,
+ clip_input,
+ params: dict = None,
+ ):
+ clip_input = jnp.transpose(clip_input, (0, 2, 3, 1))
+
+ return self.module.apply(
+ {"params": params or self.params},
+ jnp.array(clip_input, dtype=jnp.float32),
+ rngs={},
+ )
diff --git a/model/diffusers_c/pipelines/stable_diffusion/stable_unclip_image_normalizer.py b/model/diffusers_c/pipelines/stable_diffusion/stable_unclip_image_normalizer.py
new file mode 100644
index 0000000000000000000000000000000000000000..3fc6b3a3f8b0d1012a13482052adfc61a2682aba
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion/stable_unclip_image_normalizer.py
@@ -0,0 +1,57 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Union
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models.modeling_utils import ModelMixin
+
+
+class StableUnCLIPImageNormalizer(ModelMixin, ConfigMixin):
+ """
+ This class is used to hold the mean and standard deviation of the CLIP embedder used in stable unCLIP.
+
+ It is used to normalize the image embeddings before the noise is applied and un-normalize the noised image
+ embeddings.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ embedding_dim: int = 768,
+ ):
+ super().__init__()
+
+ self.mean = nn.Parameter(torch.zeros(1, embedding_dim))
+ self.std = nn.Parameter(torch.ones(1, embedding_dim))
+
+ def to(
+ self,
+ torch_device: Optional[Union[str, torch.device]] = None,
+ torch_dtype: Optional[torch.dtype] = None,
+ ):
+ self.mean = nn.Parameter(self.mean.to(torch_device).to(torch_dtype))
+ self.std = nn.Parameter(self.std.to(torch_device).to(torch_dtype))
+ return self
+
+ def scale(self, embeds):
+ embeds = (embeds - self.mean) * 1.0 / self.std
+ return embeds
+
+ def unscale(self, embeds):
+ embeds = (embeds * self.std) + self.mean
+ return embeds
diff --git a/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..cce556fceb2379be482d383e380a09836d25ce3b
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__init__.py
@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_attend_and_excite"] = ["StableDiffusionAttendAndExcitePipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_stable_diffusion_attend_and_excite import StableDiffusionAttendAndExcitePipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..02982692aa41f03ce730a953b997170c787b6d4d
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__pycache__/pipeline_stable_diffusion_attend_and_excite.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__pycache__/pipeline_stable_diffusion_attend_and_excite.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9b02bd3dac0592fd7875fa536c136800f1ebaeb4
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/__pycache__/pipeline_stable_diffusion_attend_and_excite.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py b/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py
new file mode 100644
index 0000000000000000000000000000000000000000..03c80b46b806bbfb46716a1106fe303c2cd1db7b
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_attend_and_excite/pipeline_stable_diffusion_attend_and_excite.py
@@ -0,0 +1,1088 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import math
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from torch.nn import functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention_processor import Attention
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__)
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionAttendAndExcitePipeline
+
+ >>> pipe = StableDiffusionAttendAndExcitePipeline.from_pretrained(
+ ... "CompVis/stable-diffusion-v1-4", torch_dtype=torch.float16
+ ... ).to("cuda")
+
+
+ >>> prompt = "a cat and a frog"
+
+ >>> # use get_indices function to find out indices of the tokens you want to alter
+ >>> pipe.get_indices(prompt)
+ {0: '<|startoftext|>', 1: 'a', 2: 'cat', 3: 'and', 4: 'a', 5: 'frog', 6: '<|endoftext|>'}
+
+ >>> token_indices = [2, 5]
+ >>> seed = 6141
+ >>> generator = torch.Generator("cuda").manual_seed(seed)
+
+ >>> images = pipe(
+ ... prompt=prompt,
+ ... token_indices=token_indices,
+ ... guidance_scale=7.5,
+ ... generator=generator,
+ ... num_inference_steps=50,
+ ... max_iter_to_alter=25,
+ ... ).images
+
+ >>> image = images[0]
+ >>> image.save(f"../images/{prompt}_{seed}.png")
+ ```
+"""
+
+
+class AttentionStore:
+ @staticmethod
+ def get_empty_store():
+ return {"down": [], "mid": [], "up": []}
+
+ def __call__(self, attn, is_cross: bool, place_in_unet: str):
+ if self.cur_att_layer >= 0 and is_cross:
+ if attn.shape[1] == np.prod(self.attn_res):
+ self.step_store[place_in_unet].append(attn)
+
+ self.cur_att_layer += 1
+ if self.cur_att_layer == self.num_att_layers:
+ self.cur_att_layer = 0
+ self.between_steps()
+
+ def between_steps(self):
+ self.attention_store = self.step_store
+ self.step_store = self.get_empty_store()
+
+ def get_average_attention(self):
+ average_attention = self.attention_store
+ return average_attention
+
+ def aggregate_attention(self, from_where: List[str]) -> torch.Tensor:
+ """Aggregates the attention across the different layers and heads at the specified resolution."""
+ out = []
+ attention_maps = self.get_average_attention()
+ for location in from_where:
+ for item in attention_maps[location]:
+ cross_maps = item.reshape(-1, self.attn_res[0], self.attn_res[1], item.shape[-1])
+ out.append(cross_maps)
+ out = torch.cat(out, dim=0)
+ out = out.sum(0) / out.shape[0]
+ return out
+
+ def reset(self):
+ self.cur_att_layer = 0
+ self.step_store = self.get_empty_store()
+ self.attention_store = {}
+
+ def __init__(self, attn_res):
+ """
+ Initialize an empty AttentionStore :param step_index: used to visualize only a specific step in the diffusion
+ process
+ """
+ self.num_att_layers = -1
+ self.cur_att_layer = 0
+ self.step_store = self.get_empty_store()
+ self.attention_store = {}
+ self.curr_step_index = 0
+ self.attn_res = attn_res
+
+
+class AttendExciteAttnProcessor:
+ def __init__(self, attnstore, place_in_unet):
+ super().__init__()
+ self.attnstore = attnstore
+ self.place_in_unet = place_in_unet
+
+ def __call__(self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None):
+ batch_size, sequence_length, _ = hidden_states.shape
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+ query = attn.to_q(hidden_states)
+
+ is_cross = encoder_hidden_states is not None
+ encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ query = attn.head_to_batch_dim(query)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ attention_probs = attn.get_attention_scores(query, key, attention_mask)
+
+ # only need to store attention maps during the Attend and Excite process
+ if attention_probs.requires_grad:
+ self.attnstore(attention_probs, is_cross, self.place_in_unet)
+
+ hidden_states = torch.bmm(attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ return hidden_states
+
+
+class StableDiffusionAttendAndExcitePipeline(DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion and Attend-and-Excite.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ indices,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ indices_is_list_ints = isinstance(indices, list) and isinstance(indices[0], int)
+ indices_is_list_list_ints = (
+ isinstance(indices, list) and isinstance(indices[0], list) and isinstance(indices[0][0], int)
+ )
+
+ if not indices_is_list_ints and not indices_is_list_list_ints:
+ raise TypeError("`indices` must be a list of ints or a list of a list of ints")
+
+ if indices_is_list_ints:
+ indices_batch_size = 1
+ elif indices_is_list_list_ints:
+ indices_batch_size = len(indices)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if indices_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"indices batch size must be same as prompt batch size. indices batch size: {indices_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @staticmethod
+ def _compute_max_attention_per_index(
+ attention_maps: torch.Tensor,
+ indices: List[int],
+ ) -> List[torch.Tensor]:
+ """Computes the maximum attention value for each of the tokens we wish to alter."""
+ attention_for_text = attention_maps[:, :, 1:-1]
+ attention_for_text *= 100
+ attention_for_text = torch.nn.functional.softmax(attention_for_text, dim=-1)
+
+ # Shift indices since we removed the first token
+ indices = [index - 1 for index in indices]
+
+ # Extract the maximum values
+ max_indices_list = []
+ for i in indices:
+ image = attention_for_text[:, :, i]
+ smoothing = GaussianSmoothing().to(attention_maps.device)
+ input = F.pad(image.unsqueeze(0).unsqueeze(0), (1, 1, 1, 1), mode="reflect")
+ image = smoothing(input).squeeze(0).squeeze(0)
+ max_indices_list.append(image.max())
+ return max_indices_list
+
+ def _aggregate_and_get_max_attention_per_token(
+ self,
+ indices: List[int],
+ ):
+ """Aggregates the attention for each token and computes the max activation value for each token to alter."""
+ attention_maps = self.attention_store.aggregate_attention(
+ from_where=("up", "down", "mid"),
+ )
+ max_attention_per_index = self._compute_max_attention_per_index(
+ attention_maps=attention_maps,
+ indices=indices,
+ )
+ return max_attention_per_index
+
+ @staticmethod
+ def _compute_loss(max_attention_per_index: List[torch.Tensor]) -> torch.Tensor:
+ """Computes the attend-and-excite loss using the maximum attention value for each token."""
+ losses = [max(0, 1.0 - curr_max) for curr_max in max_attention_per_index]
+ loss = max(losses)
+ return loss
+
+ @staticmethod
+ def _update_latent(latents: torch.Tensor, loss: torch.Tensor, step_size: float) -> torch.Tensor:
+ """Update the latent according to the computed loss."""
+ grad_cond = torch.autograd.grad(loss.requires_grad_(True), [latents], retain_graph=True)[0]
+ latents = latents - step_size * grad_cond
+ return latents
+
+ def _perform_iterative_refinement_step(
+ self,
+ latents: torch.Tensor,
+ indices: List[int],
+ loss: torch.Tensor,
+ threshold: float,
+ text_embeddings: torch.Tensor,
+ step_size: float,
+ t: int,
+ max_refinement_steps: int = 20,
+ ):
+ """
+ Performs the iterative latent refinement introduced in the paper. Here, we continuously update the latent code
+ according to our loss objective until the given threshold is reached for all tokens.
+ """
+ iteration = 0
+ target_loss = max(0, 1.0 - threshold)
+ while loss > target_loss:
+ iteration += 1
+
+ latents = latents.clone().detach().requires_grad_(True)
+ self.unet(latents, t, encoder_hidden_states=text_embeddings).sample
+ self.unet.zero_grad()
+
+ # Get max activation value for each subject token
+ max_attention_per_index = self._aggregate_and_get_max_attention_per_token(
+ indices=indices,
+ )
+
+ loss = self._compute_loss(max_attention_per_index)
+
+ if loss != 0:
+ latents = self._update_latent(latents, loss, step_size)
+
+ logger.info(f"\t Try {iteration}. loss: {loss}")
+
+ if iteration >= max_refinement_steps:
+ logger.info(f"\t Exceeded max number of iterations ({max_refinement_steps})! ")
+ break
+
+ # Run one more time but don't compute gradients and update the latents.
+ # We just need to compute the new loss - the grad update will occur below
+ latents = latents.clone().detach().requires_grad_(True)
+ _ = self.unet(latents, t, encoder_hidden_states=text_embeddings).sample
+ self.unet.zero_grad()
+
+ # Get max activation value for each subject token
+ max_attention_per_index = self._aggregate_and_get_max_attention_per_token(
+ indices=indices,
+ )
+ loss = self._compute_loss(max_attention_per_index)
+ logger.info(f"\t Finished with loss of: {loss}")
+ return loss, latents, max_attention_per_index
+
+ def register_attention_control(self):
+ attn_procs = {}
+ cross_att_count = 0
+ for name in self.unet.attn_processors.keys():
+ if name.startswith("mid_block"):
+ place_in_unet = "mid"
+ elif name.startswith("up_blocks"):
+ place_in_unet = "up"
+ elif name.startswith("down_blocks"):
+ place_in_unet = "down"
+ else:
+ continue
+
+ cross_att_count += 1
+ attn_procs[name] = AttendExciteAttnProcessor(attnstore=self.attention_store, place_in_unet=place_in_unet)
+
+ self.unet.set_attn_processor(attn_procs)
+ self.attention_store.num_att_layers = cross_att_count
+
+ def get_indices(self, prompt: str) -> Dict[str, int]:
+ """Utility function to list the indices of the tokens you wish to alte"""
+ ids = self.tokenizer(prompt).input_ids
+ indices = {i: tok for tok, i in zip(self.tokenizer.convert_ids_to_tokens(ids), range(len(ids)))}
+ return indices
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ token_indices: Union[List[int], List[List[int]]],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: int = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ max_iter_to_alter: int = 25,
+ thresholds: dict = {0: 0.05, 10: 0.5, 20: 0.8},
+ scale_factor: int = 20,
+ attn_res: Optional[Tuple[int]] = (16, 16),
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ token_indices (`List[int]`):
+ The token indices to alter with attend-and-excite.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ max_iter_to_alter (`int`, *optional*, defaults to `25`):
+ Number of denoising steps to apply attend-and-excite. The `max_iter_to_alter` denoising steps are when
+ attend-and-excite is applied. For example, if `max_iter_to_alter` is `25` and there are a total of `30`
+ denoising steps, the first `25` denoising steps applies attend-and-excite and the last `5` will not.
+ thresholds (`dict`, *optional*, defaults to `{0: 0.05, 10: 0.5, 20: 0.8}`):
+ Dictionary defining the iterations and desired thresholds to apply iterative latent refinement in.
+ scale_factor (`int`, *optional*, default to 20):
+ Scale factor to control the step size of each attend-and-excite update.
+ attn_res (`tuple`, *optional*, default computed from width and height):
+ The 2D resolution of the semantic attention map.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ token_indices,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ if attn_res is None:
+ attn_res = int(np.ceil(width / 32)), int(np.ceil(height / 32))
+ self.attention_store = AttentionStore(attn_res)
+ self.register_attention_control()
+
+ # default config for step size from original repo
+ scale_range = np.linspace(1.0, 0.5, len(self.scheduler.timesteps))
+ step_size = scale_factor * np.sqrt(scale_range)
+
+ text_embeddings = (
+ prompt_embeds[batch_size * num_images_per_prompt :] if do_classifier_free_guidance else prompt_embeds
+ )
+
+ if isinstance(token_indices[0], int):
+ token_indices = [token_indices]
+
+ indices = []
+
+ for ind in token_indices:
+ indices = indices + [ind] * num_images_per_prompt
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Attend and excite process
+ with torch.enable_grad():
+ latents = latents.clone().detach().requires_grad_(True)
+ updated_latents = []
+ for latent, index, text_embedding in zip(latents, indices, text_embeddings):
+ # Forward pass of denoising with text conditioning
+ latent = latent.unsqueeze(0)
+ text_embedding = text_embedding.unsqueeze(0)
+
+ self.unet(
+ latent,
+ t,
+ encoder_hidden_states=text_embedding,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+ self.unet.zero_grad()
+
+ # Get max activation value for each subject token
+ max_attention_per_index = self._aggregate_and_get_max_attention_per_token(
+ indices=index,
+ )
+
+ loss = self._compute_loss(max_attention_per_index=max_attention_per_index)
+
+ # If this is an iterative refinement step, verify we have reached the desired threshold for all
+ if i in thresholds.keys() and loss > 1.0 - thresholds[i]:
+ loss, latent, max_attention_per_index = self._perform_iterative_refinement_step(
+ latents=latent,
+ indices=index,
+ loss=loss,
+ threshold=thresholds[i],
+ text_embeddings=text_embedding,
+ step_size=step_size[i],
+ t=t,
+ )
+
+ # Perform gradient update
+ if i < max_iter_to_alter:
+ if loss != 0:
+ latent = self._update_latent(
+ latents=latent,
+ loss=loss,
+ step_size=step_size[i],
+ )
+ logger.info(f"Iteration {i} | Loss: {loss:0.4f}")
+
+ updated_latents.append(latent)
+
+ latents = torch.cat(updated_latents, dim=0)
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 8. Post-processing
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+
+class GaussianSmoothing(torch.nn.Module):
+ """
+ Arguments:
+ Apply gaussian smoothing on a 1d, 2d or 3d tensor. Filtering is performed seperately for each channel in the input
+ using a depthwise convolution.
+ channels (int, sequence): Number of channels of the input tensors. Output will
+ have this number of channels as well.
+ kernel_size (int, sequence): Size of the gaussian kernel. sigma (float, sequence): Standard deviation of the
+ gaussian kernel. dim (int, optional): The number of dimensions of the data.
+ Default value is 2 (spatial).
+ """
+
+ # channels=1, kernel_size=kernel_size, sigma=sigma, dim=2
+ def __init__(
+ self,
+ channels: int = 1,
+ kernel_size: int = 3,
+ sigma: float = 0.5,
+ dim: int = 2,
+ ):
+ super().__init__()
+
+ if isinstance(kernel_size, int):
+ kernel_size = [kernel_size] * dim
+ if isinstance(sigma, float):
+ sigma = [sigma] * dim
+
+ # The gaussian kernel is the product of the
+ # gaussian function of each dimension.
+ kernel = 1
+ meshgrids = torch.meshgrid([torch.arange(size, dtype=torch.float32) for size in kernel_size])
+ for size, std, mgrid in zip(kernel_size, sigma, meshgrids):
+ mean = (size - 1) / 2
+ kernel *= 1 / (std * math.sqrt(2 * math.pi)) * torch.exp(-(((mgrid - mean) / (2 * std)) ** 2))
+
+ # Make sure sum of values in gaussian kernel equals 1.
+ kernel = kernel / torch.sum(kernel)
+
+ # Reshape to depthwise convolutional weight
+ kernel = kernel.view(1, 1, *kernel.size())
+ kernel = kernel.repeat(channels, *[1] * (kernel.dim() - 1))
+
+ self.register_buffer("weight", kernel)
+ self.groups = channels
+
+ if dim == 1:
+ self.conv = F.conv1d
+ elif dim == 2:
+ self.conv = F.conv2d
+ elif dim == 3:
+ self.conv = F.conv3d
+ else:
+ raise RuntimeError("Only 1, 2 and 3 dimensions are supported. Received {}.".format(dim))
+
+ def forward(self, input):
+ """
+ Arguments:
+ Apply gaussian filter to input.
+ input (torch.Tensor): Input to apply gaussian filter on.
+ Returns:
+ filtered (torch.Tensor): Filtered output.
+ """
+ return self.conv(input, weight=self.weight.to(input.dtype), groups=self.groups)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_diffedit/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_diffedit/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e2145edb96c6be124abf9e9a21b9a5e8a3f3d641
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_diffedit/__init__.py
@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_diffedit"] = ["StableDiffusionDiffEditPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_stable_diffusion_diffedit import StableDiffusionDiffEditPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_diffedit/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_diffedit/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2e2553b129899256aa93503f66001ffc6e265578
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_diffedit/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_diffedit/__pycache__/pipeline_stable_diffusion_diffedit.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_diffedit/__pycache__/pipeline_stable_diffusion_diffedit.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..eb53e842cff4518abcf188009c6188316682d919
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_diffedit/__pycache__/pipeline_stable_diffusion_diffedit.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py b/model/diffusers_c/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py
new file mode 100644
index 0000000000000000000000000000000000000000..4c90ce0646c4b5f2da563c7e042e13f4cf8045ec
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_diffedit/pipeline_stable_diffusion_diffedit.py
@@ -0,0 +1,1530 @@
+# Copyright 2024 DiffEdit Authors and Pix2Pix Zero Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ...configuration_utils import FrozenDict
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import DDIMInverseScheduler, KarrasDiffusionSchedulers
+from ...utils import (
+ PIL_INTERPOLATION,
+ USE_PEFT_BACKEND,
+ BaseOutput,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class DiffEditInversionPipelineOutput(BaseOutput):
+ """
+ Output class for Stable Diffusion pipelines.
+
+ Args:
+ latents (`torch.FloatTensor`)
+ inverted latents tensor
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `num_timesteps * batch_size` or numpy array of shape `(num_timesteps,
+ batch_size, height, width, num_channels)`. PIL images or numpy array present the denoised images of the
+ diffusion pipeline.
+ """
+
+ latents: torch.FloatTensor
+ images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+EXAMPLE_DOC_STRING = """
+
+ ```py
+ >>> import PIL
+ >>> import requests
+ >>> import torch
+ >>> from io import BytesIO
+
+ >>> from diffusers import StableDiffusionDiffEditPipeline
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
+
+ >>> init_image = download_image(img_url).resize((768, 768))
+
+ >>> pipe = StableDiffusionDiffEditPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+ >>> pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
+ >>> pipeline.enable_model_cpu_offload()
+
+ >>> mask_prompt = "A bowl of fruits"
+ >>> prompt = "A bowl of pears"
+
+ >>> mask_image = pipe.generate_mask(image=init_image, source_prompt=prompt, target_prompt=mask_prompt)
+ >>> image_latents = pipe.invert(image=init_image, prompt=mask_prompt).latents
+ >>> image = pipe(prompt=prompt, mask_image=mask_image, image_latents=image_latents).images[0]
+ ```
+"""
+
+EXAMPLE_INVERT_DOC_STRING = """
+ ```py
+ >>> import PIL
+ >>> import requests
+ >>> import torch
+ >>> from io import BytesIO
+
+ >>> from diffusers import StableDiffusionDiffEditPipeline
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = "https://github.com/Xiang-cd/DiffEdit-stable-diffusion/raw/main/assets/origin.png"
+
+ >>> init_image = download_image(img_url).resize((768, 768))
+
+ >>> pipe = StableDiffusionDiffEditPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-2-1", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> pipeline.scheduler = DDIMScheduler.from_config(pipeline.scheduler.config)
+ >>> pipeline.inverse_scheduler = DDIMInverseScheduler.from_config(pipeline.scheduler.config)
+ >>> pipeline.enable_model_cpu_offload()
+
+ >>> prompt = "A bowl of fruits"
+
+ >>> inverted_latents = pipe.invert(image=init_image, prompt=prompt).latents
+ ```
+"""
+
+
+def auto_corr_loss(hidden_states, generator=None):
+ reg_loss = 0.0
+ for i in range(hidden_states.shape[0]):
+ for j in range(hidden_states.shape[1]):
+ noise = hidden_states[i : i + 1, j : j + 1, :, :]
+ while True:
+ roll_amount = torch.randint(noise.shape[2] // 2, (1,), generator=generator).item()
+ reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=2)).mean() ** 2
+ reg_loss += (noise * torch.roll(noise, shifts=roll_amount, dims=3)).mean() ** 2
+
+ if noise.shape[2] <= 8:
+ break
+ noise = torch.nn.functional.avg_pool2d(noise, kernel_size=2)
+ return reg_loss
+
+
+def kl_divergence(hidden_states):
+ return hidden_states.var() + hidden_states.mean() ** 2 - 1 - torch.log(hidden_states.var() + 1e-7)
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.preprocess
+def preprocess(image):
+ deprecation_message = "The preprocess method is deprecated and will be removed in diffusers 1.0.0. Please use VaeImageProcessor.preprocess(...) instead"
+ deprecate("preprocess", "1.0.0", deprecation_message, standard_warn=False)
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ w, h = image[0].size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+
+ image = [np.array(i.resize((w, h), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = 2.0 * image - 1.0
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ image = torch.cat(image, dim=0)
+ return image
+
+
+def preprocess_mask(mask, batch_size: int = 1):
+ if not isinstance(mask, torch.Tensor):
+ # preprocess mask
+ if isinstance(mask, PIL.Image.Image) or isinstance(mask, np.ndarray):
+ mask = [mask]
+
+ if isinstance(mask, list):
+ if isinstance(mask[0], PIL.Image.Image):
+ mask = [np.array(m.convert("L")).astype(np.float32) / 255.0 for m in mask]
+ if isinstance(mask[0], np.ndarray):
+ mask = np.stack(mask, axis=0) if mask[0].ndim < 3 else np.concatenate(mask, axis=0)
+ mask = torch.from_numpy(mask)
+ elif isinstance(mask[0], torch.Tensor):
+ mask = torch.stack(mask, dim=0) if mask[0].ndim < 3 else torch.cat(mask, dim=0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Single batched mask, no channel dim or single mask not batched but channel dim
+ if mask.shape[0] == 1:
+ mask = mask.unsqueeze(0)
+
+ # Batched masks no channel dim
+ else:
+ mask = mask.unsqueeze(1)
+
+ # Check mask shape
+ if batch_size > 1:
+ if mask.shape[0] == 1:
+ mask = torch.cat([mask] * batch_size)
+ elif mask.shape[0] > 1 and mask.shape[0] != batch_size:
+ raise ValueError(
+ f"`mask_image` with batch size {mask.shape[0]} cannot be broadcasted to batch size {batch_size} "
+ f"inferred by prompt inputs"
+ )
+
+ if mask.shape[1] != 1:
+ raise ValueError(f"`mask_image` must have 1 channel, but has {mask.shape[1]} channels")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("`mask_image` should be in [0, 1] range")
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ return mask
+
+
+class StableDiffusionDiffEditPipeline(
+ DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin
+):
+ r"""
+
+
+ This is an experimental feature!
+
+
+
+ Pipeline for text-guided image inpainting using Stable Diffusion and DiffEdit.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading and saving methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+ inverse_scheduler ([`DDIMInverseScheduler`]):
+ A scheduler to be used in combination with `unet` to fill in the unmasked part of the input latents.
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "inverse_scheduler"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ inverse_scheduler: DDIMInverseScheduler,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "skip_prk_steps") and scheduler.config.skip_prk_steps is False:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration"
+ " `skip_prk_steps`. `skip_prk_steps` should be set to True in the configuration file. Please make"
+ " sure to update the config accordingly as not setting `skip_prk_steps` in the config might lead to"
+ " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face"
+ " Hub, it would be very nice if you could open a Pull request for the"
+ " `scheduler/scheduler_config.json` file"
+ )
+ deprecate("skip_prk_steps not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["skip_prk_steps"] = True
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ inverse_scheduler=inverse_scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def check_inputs(
+ self,
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if (strength is None) or (strength is not None and (strength < 0 or strength > 1)):
+ raise ValueError(
+ f"The value of `strength` should in [0.0, 1.0] but is, but is {strength} of type {type(strength)}."
+ )
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def check_source_inputs(
+ self,
+ source_prompt=None,
+ source_negative_prompt=None,
+ source_prompt_embeds=None,
+ source_negative_prompt_embeds=None,
+ ):
+ if source_prompt is not None and source_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `source_prompt`: {source_prompt} and `source_prompt_embeds`: {source_prompt_embeds}."
+ " Please make sure to only forward one of the two."
+ )
+ elif source_prompt is None and source_prompt_embeds is None:
+ raise ValueError(
+ "Provide either `source_image` or `source_prompt_embeds`. Cannot leave all both of the arguments undefined."
+ )
+ elif source_prompt is not None and (
+ not isinstance(source_prompt, str) and not isinstance(source_prompt, list)
+ ):
+ raise ValueError(f"`source_prompt` has to be of type `str` or `list` but is {type(source_prompt)}")
+
+ if source_negative_prompt is not None and source_negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `source_negative_prompt`: {source_negative_prompt} and `source_negative_prompt_embeds`:"
+ f" {source_negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if source_prompt_embeds is not None and source_negative_prompt_embeds is not None:
+ if source_prompt_embeds.shape != source_negative_prompt_embeds.shape:
+ raise ValueError(
+ "`source_prompt_embeds` and `source_negative_prompt_embeds` must have the same shape when passed"
+ f" directly, but got: `source_prompt_embeds` {source_prompt_embeds.shape} !="
+ f" `source_negative_prompt_embeds` {source_negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ def get_inverse_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+
+ # safety for t_start overflow to prevent empty timsteps slice
+ if t_start == 0:
+ return self.inverse_scheduler.timesteps, num_inference_steps
+ timesteps = self.inverse_scheduler.timesteps[:-t_start]
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def prepare_image_latents(self, image, batch_size, dtype, device, generator=None):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ if image.shape[1] == 4:
+ latents = image
+
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if isinstance(generator, list):
+ latents = [
+ self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
+ ]
+ latents = torch.cat(latents, dim=0)
+ else:
+ latents = self.vae.encode(image).latent_dist.sample(generator)
+
+ latents = self.vae.config.scaling_factor * latents
+
+ if batch_size != latents.shape[0]:
+ if batch_size % latents.shape[0] == 0:
+ # expand image_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_latents_per_image = batch_size // latents.shape[0]
+ latents = torch.cat([latents] * additional_latents_per_image, dim=0)
+ else:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ latents = torch.cat([latents], dim=0)
+
+ return latents
+
+ def get_epsilon(self, model_output: torch.Tensor, sample: torch.Tensor, timestep: int):
+ pred_type = self.inverse_scheduler.config.prediction_type
+ alpha_prod_t = self.inverse_scheduler.alphas_cumprod[timestep]
+
+ beta_prod_t = 1 - alpha_prod_t
+
+ if pred_type == "epsilon":
+ return model_output
+ elif pred_type == "sample":
+ return (sample - alpha_prod_t ** (0.5) * model_output) / beta_prod_t ** (0.5)
+ elif pred_type == "v_prediction":
+ return (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {pred_type} must be one of `epsilon`, `sample`, or `v_prediction`"
+ )
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def generate_mask(
+ self,
+ image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ target_prompt: Optional[Union[str, List[str]]] = None,
+ target_negative_prompt: Optional[Union[str, List[str]]] = None,
+ target_prompt_embeds: Optional[torch.FloatTensor] = None,
+ target_negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ source_prompt: Optional[Union[str, List[str]]] = None,
+ source_negative_prompt: Optional[Union[str, List[str]]] = None,
+ source_prompt_embeds: Optional[torch.FloatTensor] = None,
+ source_negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ num_maps_per_mask: Optional[int] = 10,
+ mask_encode_strength: Optional[float] = 0.5,
+ mask_thresholding_ratio: Optional[float] = 3.0,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ output_type: Optional[str] = "np",
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ ):
+ r"""
+ Generate a latent mask given a mask prompt, a target prompt, and an image.
+
+ Args:
+ image (`PIL.Image.Image`):
+ `Image` or tensor representing an image batch to be used for computing the mask.
+ target_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide semantic mask generation. If not defined, you need to pass
+ `prompt_embeds`.
+ target_negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ target_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ target_negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ source_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide semantic mask generation using DiffEdit. If not defined, you need to
+ pass `source_prompt_embeds` or `source_image` instead.
+ source_negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide semantic mask generation away from using DiffEdit. If not defined, you
+ need to pass `source_negative_prompt_embeds` or `source_image` instead.
+ source_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings to guide the semantic mask generation. Can be used to easily tweak text
+ inputs (prompt weighting). If not provided, text embeddings are generated from `source_prompt` input
+ argument.
+ source_negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings to negatively guide the semantic mask generation. Can be used to easily
+ tweak text inputs (prompt weighting). If not provided, text embeddings are generated from
+ `source_negative_prompt` input argument.
+ num_maps_per_mask (`int`, *optional*, defaults to 10):
+ The number of noise maps sampled to generate the semantic mask using DiffEdit.
+ mask_encode_strength (`float`, *optional*, defaults to 0.5):
+ The strength of the noise maps sampled to generate the semantic mask using DiffEdit. Must be between 0
+ and 1.
+ mask_thresholding_ratio (`float`, *optional*, defaults to 3.0):
+ The maximum multiple of the mean absolute difference used to clamp the semantic guidance map before
+ mask binarization.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the
+ [`~models.attention_processor.AttnProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+
+ Examples:
+
+ Returns:
+ `List[PIL.Image.Image]` or `np.array`:
+ When returning a `List[PIL.Image.Image]`, the list consists of a batch of single-channel binary images
+ with dimensions `(height // self.vae_scale_factor, width // self.vae_scale_factor)`. If it's
+ `np.array`, the shape is `(batch_size, height // self.vae_scale_factor, width //
+ self.vae_scale_factor)`.
+ """
+
+ # 1. Check inputs (Provide dummy argument for callback_steps)
+ self.check_inputs(
+ target_prompt,
+ mask_encode_strength,
+ 1,
+ target_negative_prompt,
+ target_prompt_embeds,
+ target_negative_prompt_embeds,
+ )
+
+ self.check_source_inputs(
+ source_prompt,
+ source_negative_prompt,
+ source_prompt_embeds,
+ source_negative_prompt_embeds,
+ )
+
+ if (num_maps_per_mask is None) or (
+ num_maps_per_mask is not None and (not isinstance(num_maps_per_mask, int) or num_maps_per_mask <= 0)
+ ):
+ raise ValueError(
+ f"`num_maps_per_mask` has to be a positive integer but is {num_maps_per_mask} of type"
+ f" {type(num_maps_per_mask)}."
+ )
+
+ if mask_thresholding_ratio is None or mask_thresholding_ratio <= 0:
+ raise ValueError(
+ f"`mask_thresholding_ratio` has to be positive but is {mask_thresholding_ratio} of type"
+ f" {type(mask_thresholding_ratio)}."
+ )
+
+ # 2. Define call parameters
+ if target_prompt is not None and isinstance(target_prompt, str):
+ batch_size = 1
+ elif target_prompt is not None and isinstance(target_prompt, list):
+ batch_size = len(target_prompt)
+ else:
+ batch_size = target_prompt_embeds.shape[0]
+ if cross_attention_kwargs is None:
+ cross_attention_kwargs = {}
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompts
+ (cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None)
+ target_negative_prompt_embeds, target_prompt_embeds = self.encode_prompt(
+ target_prompt,
+ device,
+ num_maps_per_mask,
+ do_classifier_free_guidance,
+ target_negative_prompt,
+ prompt_embeds=target_prompt_embeds,
+ negative_prompt_embeds=target_negative_prompt_embeds,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ target_prompt_embeds = torch.cat([target_negative_prompt_embeds, target_prompt_embeds])
+
+ source_negative_prompt_embeds, source_prompt_embeds = self.encode_prompt(
+ source_prompt,
+ device,
+ num_maps_per_mask,
+ do_classifier_free_guidance,
+ source_negative_prompt,
+ prompt_embeds=source_prompt_embeds,
+ negative_prompt_embeds=source_negative_prompt_embeds,
+ )
+ if do_classifier_free_guidance:
+ source_prompt_embeds = torch.cat([source_negative_prompt_embeds, source_prompt_embeds])
+
+ # 4. Preprocess image
+ image = self.image_processor.preprocess(image).repeat_interleave(num_maps_per_mask, dim=0)
+
+ # 5. Set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, _ = self.get_timesteps(num_inference_steps, mask_encode_strength, device)
+ encode_timestep = timesteps[0]
+
+ # 6. Prepare image latents and add noise with specified strength
+ image_latents = self.prepare_image_latents(
+ image, batch_size * num_maps_per_mask, self.vae.dtype, device, generator
+ )
+ noise = randn_tensor(image_latents.shape, generator=generator, device=device, dtype=self.vae.dtype)
+ image_latents = self.scheduler.add_noise(image_latents, noise, encode_timestep)
+
+ latent_model_input = torch.cat([image_latents] * (4 if do_classifier_free_guidance else 2))
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, encode_timestep)
+
+ # 7. Predict the noise residual
+ prompt_embeds = torch.cat([source_prompt_embeds, target_prompt_embeds])
+ noise_pred = self.unet(
+ latent_model_input,
+ encode_timestep,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ if do_classifier_free_guidance:
+ noise_pred_neg_src, noise_pred_source, noise_pred_uncond, noise_pred_target = noise_pred.chunk(4)
+ noise_pred_source = noise_pred_neg_src + guidance_scale * (noise_pred_source - noise_pred_neg_src)
+ noise_pred_target = noise_pred_uncond + guidance_scale * (noise_pred_target - noise_pred_uncond)
+ else:
+ noise_pred_source, noise_pred_target = noise_pred.chunk(2)
+
+ # 8. Compute the mask from the absolute difference of predicted noise residuals
+ # TODO: Consider smoothing mask guidance map
+ mask_guidance_map = (
+ torch.abs(noise_pred_target - noise_pred_source)
+ .reshape(batch_size, num_maps_per_mask, *noise_pred_target.shape[-3:])
+ .mean([1, 2])
+ )
+ clamp_magnitude = mask_guidance_map.mean() * mask_thresholding_ratio
+ semantic_mask_image = mask_guidance_map.clamp(0, clamp_magnitude) / clamp_magnitude
+ semantic_mask_image = torch.where(semantic_mask_image <= 0.5, 0, 1)
+ mask_image = semantic_mask_image.cpu().numpy()
+
+ # 9. Convert to Numpy array or PIL.
+ if output_type == "pil":
+ mask_image = self.image_processor.numpy_to_pil(mask_image)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ return mask_image
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_INVERT_DOC_STRING)
+ def invert(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ num_inference_steps: int = 50,
+ inpaint_strength: float = 0.8,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ decode_latents: bool = False,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ lambda_auto_corr: float = 20.0,
+ lambda_kl: float = 20.0,
+ num_reg_steps: int = 0,
+ num_auto_corr_rolls: int = 5,
+ ):
+ r"""
+ Generate inverted latents given a prompt and image.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`PIL.Image.Image`):
+ `Image` or tensor representing an image batch to produce the inverted latents guided by `prompt`.
+ inpaint_strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent of the noising process to run latent inversion. Must be between 0 and 1. When
+ `inpaint_strength` is 1, the inversion process is run for the full number of iterations specified in
+ `num_inference_steps`. `image` is used as a reference for the inversion process, and adding more noise
+ increases `inpaint_strength`. If `inpaint_strength` is 0, no inpainting occurs.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ decode_latents (`bool`, *optional*, defaults to `False`):
+ Whether or not to decode the inverted latents into a generated image. Setting this argument to `True`
+ decodes all inverted latents for each timestep into a list of generated images.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.DiffEditInversionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the
+ [`~models.attention_processor.AttnProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ lambda_auto_corr (`float`, *optional*, defaults to 20.0):
+ Lambda parameter to control auto correction.
+ lambda_kl (`float`, *optional*, defaults to 20.0):
+ Lambda parameter to control Kullback-Leibler divergence output.
+ num_reg_steps (`int`, *optional*, defaults to 0):
+ Number of regularization loss steps.
+ num_auto_corr_rolls (`int`, *optional*, defaults to 5):
+ Number of auto correction roll steps.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.pipeline_stable_diffusion_diffedit.DiffEditInversionPipelineOutput`] or
+ `tuple`:
+ If `return_dict` is `True`,
+ [`~pipelines.stable_diffusion.pipeline_stable_diffusion_diffedit.DiffEditInversionPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is the inverted latents tensors
+ ordered by increasing noise, and the second is the corresponding decoded images if `decode_latents` is
+ `True`, otherwise `None`.
+ """
+
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ inpaint_strength,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+ if cross_attention_kwargs is None:
+ cross_attention_kwargs = {}
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Preprocess image
+ image = self.image_processor.preprocess(image)
+
+ # 4. Prepare latent variables
+ num_images_per_prompt = 1
+ latents = self.prepare_image_latents(
+ image, batch_size * num_images_per_prompt, self.vae.dtype, device, generator
+ )
+
+ # 5. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 6. Prepare timesteps
+ self.inverse_scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_inverse_timesteps(num_inference_steps, inpaint_strength, device)
+
+ # 7. Noising loop where we obtain the intermediate noised latent image for each timestep.
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.inverse_scheduler.order
+ inverted_latents = []
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.inverse_scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # regularization of the noise prediction (not in original code or paper but borrowed from Pix2PixZero)
+ if num_reg_steps > 0:
+ with torch.enable_grad():
+ for _ in range(num_reg_steps):
+ if lambda_auto_corr > 0:
+ for _ in range(num_auto_corr_rolls):
+ var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True)
+
+ # Derive epsilon from model output before regularizing to IID standard normal
+ var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t)
+
+ l_ac = auto_corr_loss(var_epsilon, generator=generator)
+ l_ac.backward()
+
+ grad = var.grad.detach() / num_auto_corr_rolls
+ noise_pred = noise_pred - lambda_auto_corr * grad
+
+ if lambda_kl > 0:
+ var = torch.autograd.Variable(noise_pred.detach().clone(), requires_grad=True)
+
+ # Derive epsilon from model output before regularizing to IID standard normal
+ var_epsilon = self.get_epsilon(var, latent_model_input.detach(), t)
+
+ l_kld = kl_divergence(var_epsilon)
+ l_kld.backward()
+
+ grad = var.grad.detach()
+ noise_pred = noise_pred - lambda_kl * grad
+
+ noise_pred = noise_pred.detach()
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.inverse_scheduler.step(noise_pred, t, latents).prev_sample
+ inverted_latents.append(latents.detach().clone())
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or (
+ (i + 1) > num_warmup_steps and (i + 1) % self.inverse_scheduler.order == 0
+ ):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ assert len(inverted_latents) == len(timesteps)
+ latents = torch.stack(list(reversed(inverted_latents)), 1)
+
+ # 8. Post-processing
+ image = None
+ if decode_latents:
+ image = self.decode_latents(latents.flatten(0, 1))
+
+ # 9. Convert to PIL.
+ if decode_latents and output_type == "pil":
+ image = self.image_processor.numpy_to_pil(image)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (latents, image)
+
+ return DiffEditInversionPipelineOutput(latents=latents, images=image)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ image_latents: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ inpaint_strength: Optional[float] = 0.8,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_ckip: int = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ mask_image (`PIL.Image.Image`):
+ `Image` or tensor representing an image batch to mask the generated image. White pixels in the mask are
+ repainted, while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a
+ single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
+ instead of 3, so the expected shape would be `(B, 1, H, W)`.
+ image_latents (`PIL.Image.Image` or `torch.FloatTensor`):
+ Partially noised image latents from the inversion process to be used as inputs for image generation.
+ inpaint_strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to inpaint the masked area. Must be between 0 and 1. When `inpaint_strength` is 1, the
+ denoising process is run on the masked area for the full number of iterations specified in
+ `num_inference_steps`. `image_latents` is used as a reference for the masked area, and adding more
+ noise to a region increases `inpaint_strength`. If `inpaint_strength` is 0, no inpainting occurs.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ inpaint_strength,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ if mask_image is None:
+ raise ValueError(
+ "`mask_image` input cannot be undefined. Use `generate_mask()` to compute `mask_image` from text prompts."
+ )
+ if image_latents is None:
+ raise ValueError(
+ "`image_latents` input cannot be undefined. Use `invert()` to compute `image_latents` from input images."
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+ if cross_attention_kwargs is None:
+ cross_attention_kwargs = {}
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_ckip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Preprocess mask
+ mask_image = preprocess_mask(mask_image, batch_size)
+ latent_height, latent_width = mask_image.shape[-2:]
+ mask_image = torch.cat([mask_image] * num_images_per_prompt)
+ mask_image = mask_image.to(device=device, dtype=prompt_embeds.dtype)
+
+ # 5. Set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, inpaint_strength, device)
+
+ # 6. Preprocess image latents
+ if isinstance(image_latents, list) and any(isinstance(l, torch.Tensor) and l.ndim == 5 for l in image_latents):
+ image_latents = torch.cat(image_latents).detach()
+ elif isinstance(image_latents, torch.Tensor) and image_latents.ndim == 5:
+ image_latents = image_latents.detach()
+ else:
+ image_latents = self.image_processor.preprocess(image_latents).detach()
+
+ latent_shape = (self.vae.config.latent_channels, latent_height, latent_width)
+ if image_latents.shape[-3:] != latent_shape:
+ raise ValueError(
+ f"Each latent image in `image_latents` must have shape {latent_shape}, "
+ f"but has shape {image_latents.shape[-3:]}"
+ )
+ if image_latents.ndim == 4:
+ image_latents = image_latents.reshape(batch_size, len(timesteps), *latent_shape)
+ if image_latents.shape[:2] != (batch_size, len(timesteps)):
+ raise ValueError(
+ f"`image_latents` must have batch size {batch_size} with latent images from {len(timesteps)}"
+ f" timesteps, but has batch size {image_latents.shape[0]} with latent images from"
+ f" {image_latents.shape[1]} timesteps."
+ )
+ image_latents = image_latents.transpose(0, 1).repeat_interleave(num_images_per_prompt, dim=1)
+ image_latents = image_latents.to(device=device, dtype=prompt_embeds.dtype)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 8. Denoising loop
+ latents = image_latents[0].clone()
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # mask with inverted latents from appropriate timestep - use original image latent for last step
+ latents = latents * mask_image + image_latents[i] * (1 - mask_image)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_gligen/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_gligen/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..147980cbf9e5c3418fc1854787ae37b25e4fed56
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_gligen/__init__.py
@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_gligen"] = ["StableDiffusionGLIGENPipeline"]
+ _import_structure["pipeline_stable_diffusion_gligen_text_image"] = ["StableDiffusionGLIGENTextImagePipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_stable_diffusion_gligen import StableDiffusionGLIGENPipeline
+ from .pipeline_stable_diffusion_gligen_text_image import StableDiffusionGLIGENTextImagePipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..82a956a59ba3de6c5112d403158b2da285d40f7b
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4ad9b09fc026180a76c6f100b886cabec3035f8d
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen_text_image.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen_text_image.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d0446119749bb4fb2a5354da27b1cd10477b4ad7
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_gligen/__pycache__/pipeline_stable_diffusion_gligen_text_image.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py b/model/diffusers_c/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py
new file mode 100644
index 0000000000000000000000000000000000000000..9f0d1190fd878069875470c53c4e5cae379bc805
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen.py
@@ -0,0 +1,845 @@
+# Copyright 2024 The GLIGEN Authors and HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import warnings
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention import GatedSelfAttentionDense
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionGLIGENPipeline
+ >>> from diffusers.utils import load_image
+
+ >>> # Insert objects described by text at the region defined by bounding boxes
+ >>> pipe = StableDiffusionGLIGENPipeline.from_pretrained(
+ ... "masterful/gligen-1-4-inpainting-text-box", variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> input_image = load_image(
+ ... "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/livingroom_modern.png"
+ ... )
+ >>> prompt = "a birthday cake"
+ >>> boxes = [[0.2676, 0.6088, 0.4773, 0.7183]]
+ >>> phrases = ["a birthday cake"]
+
+ >>> images = pipe(
+ ... prompt=prompt,
+ ... gligen_phrases=phrases,
+ ... gligen_inpaint_image=input_image,
+ ... gligen_boxes=boxes,
+ ... gligen_scheduled_sampling_beta=1,
+ ... output_type="pil",
+ ... num_inference_steps=50,
+ ... ).images
+
+ >>> images[0].save("./gligen-1-4-inpainting-text-box.jpg")
+
+ >>> # Generate an image described by the prompt and
+ >>> # insert objects described by text at the region defined by bounding boxes
+ >>> pipe = StableDiffusionGLIGENPipeline.from_pretrained(
+ ... "masterful/gligen-1-4-generation-text-box", variant="fp16", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a waterfall and a modern high speed train running through the tunnel in a beautiful forest with fall foliage"
+ >>> boxes = [[0.1387, 0.2051, 0.4277, 0.7090], [0.4980, 0.4355, 0.8516, 0.7266]]
+ >>> phrases = ["a waterfall", "a modern high speed train running through the tunnel"]
+
+ >>> images = pipe(
+ ... prompt=prompt,
+ ... gligen_phrases=phrases,
+ ... gligen_boxes=boxes,
+ ... gligen_scheduled_sampling_beta=1,
+ ... output_type="pil",
+ ... num_inference_steps=50,
+ ... ).images
+
+ >>> images[0].save("./gligen-1-4-generation-text-box.jpg")
+ ```
+"""
+
+
+class StableDiffusionGLIGENPipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with Grounded-Language-to-Image Generation (GLIGEN).
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ _optional_components = ["safety_checker", "feature_extractor"]
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPFeatureExtractor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ gligen_phrases,
+ gligen_boxes,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if len(gligen_phrases) != len(gligen_boxes):
+ ValueError(
+ "length of `gligen_phrases` and `gligen_boxes` has to be same, but"
+ f" got: `gligen_phrases` {len(gligen_phrases)} != `gligen_boxes` {len(gligen_boxes)}"
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def enable_fuser(self, enabled=True):
+ for module in self.unet.modules():
+ if type(module) is GatedSelfAttentionDense:
+ module.enabled = enabled
+
+ def draw_inpaint_mask_from_boxes(self, boxes, size):
+ inpaint_mask = torch.ones(size[0], size[1])
+ for box in boxes:
+ x0, x1 = box[0] * size[0], box[2] * size[0]
+ y0, y1 = box[1] * size[1], box[3] * size[1]
+ inpaint_mask[int(y0) : int(y1), int(x0) : int(x1)] = 0
+ return inpaint_mask
+
+ def crop(self, im, new_width, new_height):
+ width, height = im.size
+ left = (width - new_width) / 2
+ top = (height - new_height) / 2
+ right = (width + new_width) / 2
+ bottom = (height + new_height) / 2
+ return im.crop((left, top, right, bottom))
+
+ def target_size_center_crop(self, im, new_hw):
+ width, height = im.size
+ if width != height:
+ im = self.crop(im, min(height, width), min(height, width))
+ return im.resize((new_hw, new_hw), PIL.Image.LANCZOS)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ gligen_scheduled_sampling_beta: float = 0.3,
+ gligen_phrases: List[str] = None,
+ gligen_boxes: List[List[float]] = None,
+ gligen_inpaint_image: Optional[PIL.Image.Image] = None,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ gligen_phrases (`List[str]`):
+ The phrases to guide what to include in each of the regions defined by the corresponding
+ `gligen_boxes`. There should only be one phrase per bounding box.
+ gligen_boxes (`List[List[float]]`):
+ The bounding boxes that identify rectangular regions of the image that are going to be filled with the
+ content described by the corresponding `gligen_phrases`. Each rectangular box is defined as a
+ `List[float]` of 4 elements `[xmin, ymin, xmax, ymax]` where each value is between [0,1].
+ gligen_inpaint_image (`PIL.Image.Image`, *optional*):
+ The input image, if provided, is inpainted with objects described by the `gligen_boxes` and
+ `gligen_phrases`. Otherwise, it is treated as a generation task on a blank input image.
+ gligen_scheduled_sampling_beta (`float`, defaults to 0.3):
+ Scheduled Sampling factor from [GLIGEN: Open-Set Grounded Text-to-Image
+ Generation](https://arxiv.org/pdf/2301.07093.pdf). Scheduled Sampling factor is only varied for
+ scheduled sampling during inference for improved quality and controllability.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ guidance_rescale (`float`, *optional*, defaults to 0.0):
+ Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
+ using zero terminal SNR.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ callback_steps,
+ gligen_phrases,
+ gligen_boxes,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 5.1 Prepare GLIGEN variables
+ max_objs = 30
+ if len(gligen_boxes) > max_objs:
+ warnings.warn(
+ f"More that {max_objs} objects found. Only first {max_objs} objects will be processed.",
+ FutureWarning,
+ )
+ gligen_phrases = gligen_phrases[:max_objs]
+ gligen_boxes = gligen_boxes[:max_objs]
+ # prepare batched input to the GLIGENTextBoundingboxProjection (boxes, phrases, mask)
+ # Get tokens for phrases from pre-trained CLIPTokenizer
+ tokenizer_inputs = self.tokenizer(gligen_phrases, padding=True, return_tensors="pt").to(device)
+ # For the token, we use the same pre-trained text encoder
+ # to obtain its text feature
+ _text_embeddings = self.text_encoder(**tokenizer_inputs).pooler_output
+ n_objs = len(gligen_boxes)
+ # For each entity, described in phrases, is denoted with a bounding box,
+ # we represent the location information as (xmin,ymin,xmax,ymax)
+ boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype)
+ boxes[:n_objs] = torch.tensor(gligen_boxes)
+ text_embeddings = torch.zeros(
+ max_objs, self.unet.cross_attention_dim, device=device, dtype=self.text_encoder.dtype
+ )
+ text_embeddings[:n_objs] = _text_embeddings
+ # Generate a mask for each object that is entity described by phrases
+ masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype)
+ masks[:n_objs] = 1
+
+ repeat_batch = batch_size * num_images_per_prompt
+ boxes = boxes.unsqueeze(0).expand(repeat_batch, -1, -1).clone()
+ text_embeddings = text_embeddings.unsqueeze(0).expand(repeat_batch, -1, -1).clone()
+ masks = masks.unsqueeze(0).expand(repeat_batch, -1).clone()
+ if do_classifier_free_guidance:
+ repeat_batch = repeat_batch * 2
+ boxes = torch.cat([boxes] * 2)
+ text_embeddings = torch.cat([text_embeddings] * 2)
+ masks = torch.cat([masks] * 2)
+ masks[: repeat_batch // 2] = 0
+ if cross_attention_kwargs is None:
+ cross_attention_kwargs = {}
+ cross_attention_kwargs["gligen"] = {"boxes": boxes, "positive_embeddings": text_embeddings, "masks": masks}
+
+ # Prepare latent variables for GLIGEN inpainting
+ if gligen_inpaint_image is not None:
+ # if the given input image is not of the same size as expected by VAE
+ # center crop and resize the input image to expected shape
+ if gligen_inpaint_image.size != (self.vae.sample_size, self.vae.sample_size):
+ gligen_inpaint_image = self.target_size_center_crop(gligen_inpaint_image, self.vae.sample_size)
+ # Convert a single image into a batch of images with a batch size of 1
+ # The resulting shape becomes (1, C, H, W), where C is the number of channels,
+ # and H and W are the height and width of the image.
+ # scales the pixel values to a range [-1, 1]
+ gligen_inpaint_image = self.image_processor.preprocess(gligen_inpaint_image)
+ gligen_inpaint_image = gligen_inpaint_image.to(dtype=self.vae.dtype, device=self.vae.device)
+ # Run AutoEncoder to get corresponding latents
+ gligen_inpaint_latent = self.vae.encode(gligen_inpaint_image).latent_dist.sample()
+ gligen_inpaint_latent = self.vae.config.scaling_factor * gligen_inpaint_latent
+ # Generate an inpainting mask
+ # pixel value = 0, where the object is present (defined by bounding boxes above)
+ # 1, everywhere else
+ gligen_inpaint_mask = self.draw_inpaint_mask_from_boxes(gligen_boxes, gligen_inpaint_latent.shape[2:])
+ gligen_inpaint_mask = gligen_inpaint_mask.to(
+ dtype=gligen_inpaint_latent.dtype, device=gligen_inpaint_latent.device
+ )
+ gligen_inpaint_mask = gligen_inpaint_mask[None, None]
+ gligen_inpaint_mask_addition = torch.cat(
+ (gligen_inpaint_latent * gligen_inpaint_mask, gligen_inpaint_mask), dim=1
+ )
+ # Convert a single mask into a batch of masks with a batch size of 1
+ gligen_inpaint_mask_addition = gligen_inpaint_mask_addition.expand(repeat_batch, -1, -1, -1).clone()
+
+ num_grounding_steps = int(gligen_scheduled_sampling_beta * len(timesteps))
+ self.enable_fuser(True)
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Scheduled sampling
+ if i == num_grounding_steps:
+ self.enable_fuser(False)
+
+ if latents.shape[1] != 4:
+ latents = torch.randn_like(latents[:, :4])
+
+ if gligen_inpaint_image is not None:
+ gligen_inpaint_latent_with_noise = (
+ self.scheduler.add_noise(
+ gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t])
+ )
+ .expand(latents.shape[0], -1, -1, -1)
+ .clone()
+ )
+ latents = gligen_inpaint_latent_with_noise * gligen_inpaint_mask + latents * (
+ 1 - gligen_inpaint_mask
+ )
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ if gligen_inpaint_image is not None:
+ latent_model_input = torch.cat((latent_model_input, gligen_inpaint_mask_addition), dim=1)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py b/model/diffusers_c/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py
new file mode 100644
index 0000000000000000000000000000000000000000..296ecae653e1128b89af69f2dfe0f9871eaf2f26
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_gligen/pipeline_stable_diffusion_gligen_text_image.py
@@ -0,0 +1,1017 @@
+# Copyright 2024 The GLIGEN Authors and HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+import warnings
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import PIL.Image
+import torch
+from transformers import (
+ CLIPFeatureExtractor,
+ CLIPProcessor,
+ CLIPTextModel,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention import GatedSelfAttentionDense
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import USE_PEFT_BACKEND, logging, replace_example_docstring, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.clip_image_project_model import CLIPImageProjection
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionGLIGENTextImagePipeline
+ >>> from diffusers.utils import load_image
+
+ >>> # Insert objects described by image at the region defined by bounding boxes
+ >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained(
+ ... "anhnct/Gligen_Inpainting_Text_Image", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> input_image = load_image(
+ ... "https://hf.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/livingroom_modern.png"
+ ... )
+ >>> prompt = "a backpack"
+ >>> boxes = [[0.2676, 0.4088, 0.4773, 0.7183]]
+ >>> phrases = None
+ >>> gligen_image = load_image(
+ ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/backpack.jpeg"
+ ... )
+
+ >>> images = pipe(
+ ... prompt=prompt,
+ ... gligen_phrases=phrases,
+ ... gligen_inpaint_image=input_image,
+ ... gligen_boxes=boxes,
+ ... gligen_images=[gligen_image],
+ ... gligen_scheduled_sampling_beta=1,
+ ... output_type="pil",
+ ... num_inference_steps=50,
+ ... ).images
+
+ >>> images[0].save("./gligen-inpainting-text-image-box.jpg")
+
+ >>> # Generate an image described by the prompt and
+ >>> # insert objects described by text and image at the region defined by bounding boxes
+ >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained(
+ ... "anhnct/Gligen_Text_Image", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a flower sitting on the beach"
+ >>> boxes = [[0.0, 0.09, 0.53, 0.76]]
+ >>> phrases = ["flower"]
+ >>> gligen_image = load_image(
+ ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/gligen/pexels-pixabay-60597.jpg"
+ ... )
+
+ >>> images = pipe(
+ ... prompt=prompt,
+ ... gligen_phrases=phrases,
+ ... gligen_images=[gligen_image],
+ ... gligen_boxes=boxes,
+ ... gligen_scheduled_sampling_beta=1,
+ ... output_type="pil",
+ ... num_inference_steps=50,
+ ... ).images
+
+ >>> images[0].save("./gligen-generation-text-image-box.jpg")
+
+ >>> # Generate an image described by the prompt and
+ >>> # transfer style described by image at the region defined by bounding boxes
+ >>> pipe = StableDiffusionGLIGENTextImagePipeline.from_pretrained(
+ ... "anhnct/Gligen_Text_Image", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a dragon flying on the sky"
+ >>> boxes = [[0.4, 0.2, 1.0, 0.8], [0.0, 1.0, 0.0, 1.0]] # Set `[0.0, 1.0, 0.0, 1.0]` for the style
+
+ >>> gligen_image = load_image(
+ ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/landscape.png"
+ ... )
+
+ >>> gligen_placeholder = load_image(
+ ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/landscape.png"
+ ... )
+
+ >>> images = pipe(
+ ... prompt=prompt,
+ ... gligen_phrases=[
+ ... "dragon",
+ ... "placeholder",
+ ... ], # Can use any text instead of `placeholder` token, because we will use mask here
+ ... gligen_images=[
+ ... gligen_placeholder,
+ ... gligen_image,
+ ... ], # Can use any image in gligen_placeholder, because we will use mask here
+ ... input_phrases_mask=[1, 0], # Set 0 for the placeholder token
+ ... input_images_mask=[0, 1], # Set 0 for the placeholder image
+ ... gligen_boxes=boxes,
+ ... gligen_scheduled_sampling_beta=1,
+ ... output_type="pil",
+ ... num_inference_steps=50,
+ ... ).images
+
+ >>> images[0].save("./gligen-generation-text-image-box-style-transfer.jpg")
+ ```
+"""
+
+
+class StableDiffusionGLIGENTextImagePipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with Grounded-Language-to-Image Generation (GLIGEN).
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ processor ([`~transformers.CLIPProcessor`]):
+ A `CLIPProcessor` to procces reference image.
+ image_encoder ([`~transformers.CLIPVisionModelWithProjection`]):
+ Frozen image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ image_project ([`CLIPImageProjection`]):
+ A `CLIPImageProjection` to project image embedding into phrases embedding space.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ processor: CLIPProcessor,
+ image_encoder: CLIPVisionModelWithProjection,
+ image_project: CLIPImageProjection,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPFeatureExtractor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ image_encoder=image_encoder,
+ processor=processor,
+ image_project=image_project,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def enable_fuser(self, enabled=True):
+ for module in self.unet.modules():
+ if type(module) is GatedSelfAttentionDense:
+ module.enabled = enabled
+
+ def draw_inpaint_mask_from_boxes(self, boxes, size):
+ """
+ Create an inpainting mask based on given boxes. This function generates an inpainting mask using the provided
+ boxes to mark regions that need to be inpainted.
+ """
+ inpaint_mask = torch.ones(size[0], size[1])
+ for box in boxes:
+ x0, x1 = box[0] * size[0], box[2] * size[0]
+ y0, y1 = box[1] * size[1], box[3] * size[1]
+ inpaint_mask[int(y0) : int(y1), int(x0) : int(x1)] = 0
+ return inpaint_mask
+
+ def crop(self, im, new_width, new_height):
+ """
+ Crop the input image to the specified dimensions.
+ """
+ width, height = im.size
+ left = (width - new_width) / 2
+ top = (height - new_height) / 2
+ right = (width + new_width) / 2
+ bottom = (height + new_height) / 2
+ return im.crop((left, top, right, bottom))
+
+ def target_size_center_crop(self, im, new_hw):
+ """
+ Crop and resize the image to the target size while keeping the center.
+ """
+ width, height = im.size
+ if width != height:
+ im = self.crop(im, min(height, width), min(height, width))
+ return im.resize((new_hw, new_hw), PIL.Image.LANCZOS)
+
+ def complete_mask(self, has_mask, max_objs, device):
+ """
+ Based on the input mask corresponding value `0 or 1` for each phrases and image, mask the features
+ corresponding to phrases and images.
+ """
+ mask = torch.ones(1, max_objs).type(self.text_encoder.dtype).to(device)
+ if has_mask is None:
+ return mask
+
+ if isinstance(has_mask, int):
+ return mask * has_mask
+ else:
+ for idx, value in enumerate(has_mask):
+ mask[0, idx] = value
+ return mask
+
+ def get_clip_feature(self, input, normalize_constant, device, is_image=False):
+ """
+ Get image and phrases embedding by using CLIP pretrain model. The image embedding is transformed into the
+ phrases embedding space through a projection.
+ """
+ if is_image:
+ if input is None:
+ return None
+ inputs = self.processor(images=[input], return_tensors="pt").to(device)
+ inputs["pixel_values"] = inputs["pixel_values"].to(self.image_encoder.dtype)
+
+ outputs = self.image_encoder(**inputs)
+ feature = outputs.image_embeds
+ feature = self.image_project(feature).squeeze(0)
+ feature = (feature / feature.norm()) * normalize_constant
+ feature = feature.unsqueeze(0)
+ else:
+ if input is None:
+ return None
+ inputs = self.tokenizer(input, return_tensors="pt", padding=True).to(device)
+ outputs = self.text_encoder(**inputs)
+ feature = outputs.pooler_output
+ return feature
+
+ def get_cross_attention_kwargs_with_grounded(
+ self,
+ hidden_size,
+ gligen_phrases,
+ gligen_images,
+ gligen_boxes,
+ input_phrases_mask,
+ input_images_mask,
+ repeat_batch,
+ normalize_constant,
+ max_objs,
+ device,
+ ):
+ """
+ Prepare the cross-attention kwargs containing information about the grounded input (boxes, mask, image
+ embedding, phrases embedding).
+ """
+ phrases, images = gligen_phrases, gligen_images
+ images = [None] * len(phrases) if images is None else images
+ phrases = [None] * len(images) if phrases is None else phrases
+
+ boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype)
+ masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype)
+ phrases_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype)
+ image_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype)
+ phrases_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype)
+ image_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype)
+
+ text_features = []
+ image_features = []
+ for phrase, image in zip(phrases, images):
+ text_features.append(self.get_clip_feature(phrase, normalize_constant, device, is_image=False))
+ image_features.append(self.get_clip_feature(image, normalize_constant, device, is_image=True))
+
+ for idx, (box, text_feature, image_feature) in enumerate(zip(gligen_boxes, text_features, image_features)):
+ boxes[idx] = torch.tensor(box)
+ masks[idx] = 1
+ if text_feature is not None:
+ phrases_embeddings[idx] = text_feature
+ phrases_masks[idx] = 1
+ if image_feature is not None:
+ image_embeddings[idx] = image_feature
+ image_masks[idx] = 1
+
+ input_phrases_mask = self.complete_mask(input_phrases_mask, max_objs, device)
+ phrases_masks = phrases_masks.unsqueeze(0).repeat(repeat_batch, 1) * input_phrases_mask
+ input_images_mask = self.complete_mask(input_images_mask, max_objs, device)
+ image_masks = image_masks.unsqueeze(0).repeat(repeat_batch, 1) * input_images_mask
+ boxes = boxes.unsqueeze(0).repeat(repeat_batch, 1, 1)
+ masks = masks.unsqueeze(0).repeat(repeat_batch, 1)
+ phrases_embeddings = phrases_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1)
+ image_embeddings = image_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1)
+
+ out = {
+ "boxes": boxes,
+ "masks": masks,
+ "phrases_masks": phrases_masks,
+ "image_masks": image_masks,
+ "phrases_embeddings": phrases_embeddings,
+ "image_embeddings": image_embeddings,
+ }
+
+ return out
+
+ def get_cross_attention_kwargs_without_grounded(self, hidden_size, repeat_batch, max_objs, device):
+ """
+ Prepare the cross-attention kwargs without information about the grounded input (boxes, mask, image embedding,
+ phrases embedding) (All are zero tensor).
+ """
+ boxes = torch.zeros(max_objs, 4, device=device, dtype=self.text_encoder.dtype)
+ masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype)
+ phrases_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype)
+ image_masks = torch.zeros(max_objs, device=device, dtype=self.text_encoder.dtype)
+ phrases_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype)
+ image_embeddings = torch.zeros(max_objs, hidden_size, device=device, dtype=self.text_encoder.dtype)
+
+ out = {
+ "boxes": boxes.unsqueeze(0).repeat(repeat_batch, 1, 1),
+ "masks": masks.unsqueeze(0).repeat(repeat_batch, 1),
+ "phrases_masks": phrases_masks.unsqueeze(0).repeat(repeat_batch, 1),
+ "image_masks": image_masks.unsqueeze(0).repeat(repeat_batch, 1),
+ "phrases_embeddings": phrases_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1),
+ "image_embeddings": image_embeddings.unsqueeze(0).repeat(repeat_batch, 1, 1),
+ }
+
+ return out
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ gligen_scheduled_sampling_beta: float = 0.3,
+ gligen_phrases: List[str] = None,
+ gligen_images: List[PIL.Image.Image] = None,
+ input_phrases_mask: Union[int, List[int]] = None,
+ input_images_mask: Union[int, List[int]] = None,
+ gligen_boxes: List[List[float]] = None,
+ gligen_inpaint_image: Optional[PIL.Image.Image] = None,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ gligen_normalize_constant: float = 28.7,
+ clip_skip: int = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ gligen_phrases (`List[str]`):
+ The phrases to guide what to include in each of the regions defined by the corresponding
+ `gligen_boxes`. There should only be one phrase per bounding box.
+ gligen_images (`List[PIL.Image.Image]`):
+ The images to guide what to include in each of the regions defined by the corresponding `gligen_boxes`.
+ There should only be one image per bounding box
+ input_phrases_mask (`int` or `List[int]`):
+ pre phrases mask input defined by the correspongding `input_phrases_mask`
+ input_images_mask (`int` or `List[int]`):
+ pre images mask input defined by the correspongding `input_images_mask`
+ gligen_boxes (`List[List[float]]`):
+ The bounding boxes that identify rectangular regions of the image that are going to be filled with the
+ content described by the corresponding `gligen_phrases`. Each rectangular box is defined as a
+ `List[float]` of 4 elements `[xmin, ymin, xmax, ymax]` where each value is between [0,1].
+ gligen_inpaint_image (`PIL.Image.Image`, *optional*):
+ The input image, if provided, is inpainted with objects described by the `gligen_boxes` and
+ `gligen_phrases`. Otherwise, it is treated as a generation task on a blank input image.
+ gligen_scheduled_sampling_beta (`float`, defaults to 0.3):
+ Scheduled Sampling factor from [GLIGEN: Open-Set Grounded Text-to-Image
+ Generation](https://arxiv.org/pdf/2301.07093.pdf). Scheduled Sampling factor is only varied for
+ scheduled sampling during inference for improved quality and controllability.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ gligen_normalize_constant (`float`, *optional*, defaults to 28.7):
+ The normalize value of the image embedding.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 5.1 Prepare GLIGEN variables
+ max_objs = 30
+ if len(gligen_boxes) > max_objs:
+ warnings.warn(
+ f"More that {max_objs} objects found. Only first {max_objs} objects will be processed.",
+ FutureWarning,
+ )
+ gligen_phrases = gligen_phrases[:max_objs]
+ gligen_boxes = gligen_boxes[:max_objs]
+ gligen_images = gligen_images[:max_objs]
+
+ repeat_batch = batch_size * num_images_per_prompt
+
+ if do_classifier_free_guidance:
+ repeat_batch = repeat_batch * 2
+
+ if cross_attention_kwargs is None:
+ cross_attention_kwargs = {}
+
+ hidden_size = prompt_embeds.shape[2]
+
+ cross_attention_kwargs["gligen"] = self.get_cross_attention_kwargs_with_grounded(
+ hidden_size=hidden_size,
+ gligen_phrases=gligen_phrases,
+ gligen_images=gligen_images,
+ gligen_boxes=gligen_boxes,
+ input_phrases_mask=input_phrases_mask,
+ input_images_mask=input_images_mask,
+ repeat_batch=repeat_batch,
+ normalize_constant=gligen_normalize_constant,
+ max_objs=max_objs,
+ device=device,
+ )
+
+ cross_attention_kwargs_without_grounded = {}
+ cross_attention_kwargs_without_grounded["gligen"] = self.get_cross_attention_kwargs_without_grounded(
+ hidden_size=hidden_size, repeat_batch=repeat_batch, max_objs=max_objs, device=device
+ )
+
+ # Prepare latent variables for GLIGEN inpainting
+ if gligen_inpaint_image is not None:
+ # if the given input image is not of the same size as expected by VAE
+ # center crop and resize the input image to expected shape
+ if gligen_inpaint_image.size != (self.vae.sample_size, self.vae.sample_size):
+ gligen_inpaint_image = self.target_size_center_crop(gligen_inpaint_image, self.vae.sample_size)
+ # Convert a single image into a batch of images with a batch size of 1
+ # The resulting shape becomes (1, C, H, W), where C is the number of channels,
+ # and H and W are the height and width of the image.
+ # scales the pixel values to a range [-1, 1]
+ gligen_inpaint_image = self.image_processor.preprocess(gligen_inpaint_image)
+ gligen_inpaint_image = gligen_inpaint_image.to(dtype=self.vae.dtype, device=self.vae.device)
+ # Run AutoEncoder to get corresponding latents
+ gligen_inpaint_latent = self.vae.encode(gligen_inpaint_image).latent_dist.sample()
+ gligen_inpaint_latent = self.vae.config.scaling_factor * gligen_inpaint_latent
+ # Generate an inpainting mask
+ # pixel value = 0, where the object is present (defined by bounding boxes above)
+ # 1, everywhere else
+ gligen_inpaint_mask = self.draw_inpaint_mask_from_boxes(gligen_boxes, gligen_inpaint_latent.shape[2:])
+ gligen_inpaint_mask = gligen_inpaint_mask.to(
+ dtype=gligen_inpaint_latent.dtype, device=gligen_inpaint_latent.device
+ )
+ gligen_inpaint_mask = gligen_inpaint_mask[None, None]
+ gligen_inpaint_mask_addition = torch.cat(
+ (gligen_inpaint_latent * gligen_inpaint_mask, gligen_inpaint_mask), dim=1
+ )
+ # Convert a single mask into a batch of masks with a batch size of 1
+ gligen_inpaint_mask_addition = gligen_inpaint_mask_addition.expand(repeat_batch, -1, -1, -1).clone()
+
+ int(gligen_scheduled_sampling_beta * len(timesteps))
+ self.enable_fuser(True)
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ if latents.shape[1] != 4:
+ latents = torch.randn_like(latents[:, :4])
+
+ if gligen_inpaint_image is not None:
+ gligen_inpaint_latent_with_noise = (
+ self.scheduler.add_noise(
+ gligen_inpaint_latent, torch.randn_like(gligen_inpaint_latent), torch.tensor([t])
+ )
+ .expand(latents.shape[0], -1, -1, -1)
+ .clone()
+ )
+ latents = gligen_inpaint_latent_with_noise * gligen_inpaint_mask + latents * (
+ 1 - gligen_inpaint_mask
+ )
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ if gligen_inpaint_image is not None:
+ latent_model_input = torch.cat((latent_model_input, gligen_inpaint_mask_addition), dim=1)
+
+ # predict the noise residual with grounded information
+ noise_pred_with_grounding = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ # predict the noise residual without grounded information
+ noise_pred_without_grounding = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs_without_grounded,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ # Using noise_pred_text from noise residual with grounded information and noise_pred_uncond from noise residual without grounded information
+ _, noise_pred_text = noise_pred_with_grounding.chunk(2)
+ noise_pred_uncond, _ = noise_pred_without_grounding.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ else:
+ noise_pred = noise_pred_with_grounding
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..7eb5bf8c229b584c934e33806efa73a3727dbb5b
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/__init__.py
@@ -0,0 +1,62 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_k_diffusion_available,
+ is_k_diffusion_version,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (
+ is_transformers_available()
+ and is_torch_available()
+ and is_k_diffusion_available()
+ and is_k_diffusion_version(">=", "0.0.12")
+ ):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_and_k_diffusion_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_and_k_diffusion_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_k_diffusion"] = ["StableDiffusionKDiffusionPipeline"]
+ _import_structure["pipeline_stable_diffusion_xl_k_diffusion"] = ["StableDiffusionXLKDiffusionPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (
+ is_transformers_available()
+ and is_torch_available()
+ and is_k_diffusion_available()
+ and is_k_diffusion_version(">=", "0.0.12")
+ ):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_and_k_diffusion_objects import *
+ else:
+ from .pipeline_stable_diffusion_k_diffusion import StableDiffusionKDiffusionPipeline
+ from .pipeline_stable_diffusion_xl_k_diffusion import StableDiffusionXLKDiffusionPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py b/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..bc565c938a30b4b5a8a849167a03dcbf4649dc39
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_k_diffusion.py
@@ -0,0 +1,664 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib
+import inspect
+from typing import Callable, List, Optional, Union
+
+import torch
+from k_diffusion.external import CompVisDenoiser, CompVisVDenoiser
+from k_diffusion.sampling import BrownianTreeNoiseSampler, get_sigmas_karras
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import LMSDiscreteScheduler
+from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class ModelWrapper:
+ def __init__(self, model, alphas_cumprod):
+ self.model = model
+ self.alphas_cumprod = alphas_cumprod
+
+ def apply_model(self, *args, **kwargs):
+ if len(args) == 3:
+ encoder_hidden_states = args[-1]
+ args = args[:2]
+ if kwargs.get("cond", None) is not None:
+ encoder_hidden_states = kwargs.pop("cond")
+ return self.model(*args, encoder_hidden_states=encoder_hidden_states, **kwargs).sample
+
+
+class StableDiffusionKDiffusionPipeline(
+ DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+
+
+ This is an experimental pipeline and is likely to change in the future.
+
+
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae,
+ text_encoder,
+ tokenizer,
+ unet,
+ scheduler,
+ safety_checker,
+ feature_extractor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ logger.info(
+ f"{self.__class__} is an experimntal pipeline and is likely to change in the future. We recommend to use"
+ " this pipeline for fast experimentation / iteration if needed, but advice to rely on existing pipelines"
+ " as defined in https://huggingface.co/docs/diffusers/api/schedulers#implemented-schedulers for"
+ " production settings."
+ )
+
+ # get correct sigmas from LMS
+ scheduler = LMSDiscreteScheduler.from_config(scheduler.config)
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ model = ModelWrapper(unet, scheduler.alphas_cumprod)
+ if scheduler.config.prediction_type == "v_prediction":
+ self.k_diffusion_model = CompVisVDenoiser(model)
+ else:
+ self.k_diffusion_model = CompVisDenoiser(model)
+
+ def set_scheduler(self, scheduler_type: str):
+ library = importlib.import_module("k_diffusion")
+ sampling = getattr(library, "sampling")
+ try:
+ self.sampler = getattr(sampling, scheduler_type)
+ except Exception:
+ valid_samplers = []
+ for s in dir(sampling):
+ if "sample_" in s:
+ valid_samplers.append(s)
+
+ raise ValueError(f"Invalid scheduler type {scheduler_type}. Please choose one of {valid_samplers}.")
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ use_karras_sigmas: Optional[bool] = False,
+ noise_sampler_seed: Optional[int] = None,
+ clip_skip: int = None,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. Ignored when not using guidance (i.e., ignored if `guidance_scale`
+ is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Use karras sigmas. For example, specifying `sample_dpmpp_2m` to `set_scheduler` will be equivalent to
+ `DPM++2M` in stable-diffusion-webui. On top of that, setting this option to True will make it `DPM++2M
+ Karras`.
+ noise_sampler_seed (`int`, *optional*, defaults to `None`):
+ The random seed to use for the noise sampler. If `None`, a random seed will be generated.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = True
+ if guidance_scale <= 1.0:
+ raise ValueError("has to use guidance_scale")
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=prompt_embeds.device)
+
+ # 5. Prepare sigmas
+ if use_karras_sigmas:
+ sigma_min: float = self.k_diffusion_model.sigmas[0].item()
+ sigma_max: float = self.k_diffusion_model.sigmas[-1].item()
+ sigmas = get_sigmas_karras(n=num_inference_steps, sigma_min=sigma_min, sigma_max=sigma_max)
+ sigmas = sigmas.to(device)
+ else:
+ sigmas = self.scheduler.sigmas
+ sigmas = sigmas.to(prompt_embeds.dtype)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+ latents = latents * sigmas[0]
+ self.k_diffusion_model.sigmas = self.k_diffusion_model.sigmas.to(latents.device)
+ self.k_diffusion_model.log_sigmas = self.k_diffusion_model.log_sigmas.to(latents.device)
+
+ # 7. Define model function
+ def model_fn(x, t):
+ latent_model_input = torch.cat([x] * 2)
+ t = torch.cat([t] * 2)
+
+ noise_pred = self.k_diffusion_model(latent_model_input, t, cond=prompt_embeds)
+
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ return noise_pred
+
+ # 8. Run k-diffusion solver
+ sampler_kwargs = {}
+
+ if "noise_sampler" in inspect.signature(self.sampler).parameters:
+ min_sigma, max_sigma = sigmas[sigmas > 0].min(), sigmas.max()
+ noise_sampler = BrownianTreeNoiseSampler(latents, min_sigma, max_sigma, noise_sampler_seed)
+ sampler_kwargs["noise_sampler"] = noise_sampler
+
+ if "generator" in inspect.signature(self.sampler).parameters:
+ sampler_kwargs["generator"] = generator
+
+ latents = self.sampler(model_fn, latents, sigmas, **sampler_kwargs)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py b/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..ed46a1e36b60b90b853ba488e2bd619ff79b3bbe
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_k_diffusion/pipeline_stable_diffusion_xl_k_diffusion.py
@@ -0,0 +1,891 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import importlib
+import inspect
+from typing import List, Optional, Tuple, Union
+
+import torch
+from k_diffusion.external import CompVisDenoiser, CompVisVDenoiser
+from k_diffusion.sampling import BrownianTreeNoiseSampler, get_sigmas_karras
+from transformers import (
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+)
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import (
+ FromSingleFileMixin,
+ IPAdapterMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers, LMSDiscreteScheduler
+from ...utils import (
+ USE_PEFT_BACKEND,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionXLKDiffusionPipeline
+
+ >>> pipe = StableDiffusionXLKDiffusionPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+ >>> pipe.set_scheduler("sample_dpmpp_2m_sde")
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> image = pipe(prompt).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.ModelWrapper
+class ModelWrapper:
+ def __init__(self, model, alphas_cumprod):
+ self.model = model
+ self.alphas_cumprod = alphas_cumprod
+
+ def apply_model(self, *args, **kwargs):
+ if len(args) == 3:
+ encoder_hidden_states = args[-1]
+ args = args[:2]
+ if kwargs.get("cond", None) is not None:
+ encoder_hidden_states = kwargs.pop("cond")
+ return self.model(*args, encoder_hidden_states=encoder_hidden_states, **kwargs).sample
+
+
+class StableDiffusionXLKDiffusionPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ FromSingleFileMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion XL and k-diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+ specifically the
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+ variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ tokenizer_2 (`CLIPTokenizer`):
+ Second Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+ Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+ `stabilityai/stable-diffusion-xl-base-1-0`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
+ _optional_components = [
+ "tokenizer",
+ "tokenizer_2",
+ "text_encoder",
+ "text_encoder_2",
+ "feature_extractor",
+ ]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ force_zeros_for_empty_prompt: bool = True,
+ ):
+ super().__init__()
+
+ # get correct sigmas from LMS
+ scheduler = LMSDiscreteScheduler.from_config(scheduler.config)
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ scheduler=scheduler,
+ )
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ self.default_sample_size = self.unet.config.sample_size
+
+ model = ModelWrapper(unet, scheduler.alphas_cumprod)
+ if scheduler.config.prediction_type == "v_prediction":
+ self.k_diffusion_model = CompVisVDenoiser(model)
+ else:
+ self.k_diffusion_model = CompVisDenoiser(model)
+
+ # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.set_scheduler
+ def set_scheduler(self, scheduler_type: str):
+ library = importlib.import_module("k_diffusion")
+ sampling = getattr(library, "sampling")
+ try:
+ self.sampler = getattr(sampling, scheduler_type)
+ except Exception:
+ valid_samplers = []
+ for s in dir(sampling):
+ if "sample_" in s:
+ valid_samplers.append(s)
+
+ raise ValueError(f"Invalid scheduler type {scheduler_type}. Please choose one of {valid_samplers}.")
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ height,
+ width,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ pooled_prompt_embeds=None,
+ negative_pooled_prompt_embeds=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+ )
+
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+ )
+
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ return latents
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
+ def _get_add_time_ids(
+ self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+ ):
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ return add_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ original_size: Optional[Tuple[int, int]] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Optional[Tuple[int, int]] = None,
+ negative_original_size: Optional[Tuple[int, int]] = None,
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+ negative_target_size: Optional[Tuple[int, int]] = None,
+ use_karras_sigmas: Optional[bool] = False,
+ noise_sampler_seed: Optional[int] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image. This is set to 1024 by default for the best results.
+ Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image. This is set to 1024 by default for the best results.
+ Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 5.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+ of a plain tuple.
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a target image resolution. It should be as same
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is a list with the generated images.
+ """
+
+ # 0. Default height and width to unet
+ height = height or self.default_sample_size * self.vae_scale_factor
+ width = width or self.default_sample_size * self.vae_scale_factor
+
+ original_size = original_size or (height, width)
+ target_size = target_size or (height, width)
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ height,
+ width,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ )
+
+ if guidance_scale <= 1.0:
+ raise ValueError("has to use guidance_scale")
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ lora_scale = None
+
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ prompt_2=prompt_2,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ negative_prompt_2=negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=prompt_embeds.device)
+
+ # 5. Prepare sigmas
+ if use_karras_sigmas:
+ sigma_min: float = self.k_diffusion_model.sigmas[0].item()
+ sigma_max: float = self.k_diffusion_model.sigmas[-1].item()
+ sigmas = get_sigmas_karras(n=num_inference_steps, sigma_min=sigma_min, sigma_max=sigma_max)
+ else:
+ sigmas = self.scheduler.sigmas
+ sigmas = sigmas.to(dtype=prompt_embeds.dtype, device=device)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+ latents = latents * sigmas[0]
+
+ self.k_diffusion_model.sigmas = self.k_diffusion_model.sigmas.to(latents.device)
+ self.k_diffusion_model.log_sigmas = self.k_diffusion_model.log_sigmas.to(latents.device)
+
+ # 7. Prepare added time ids & embeddings
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ if negative_original_size is not None and negative_target_size is not None:
+ negative_add_time_ids = self._get_add_time_ids(
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ else:
+ negative_add_time_ids = add_time_ids
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+
+ # 8. Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 9. Define model function
+ def model_fn(x, t):
+ latent_model_input = torch.cat([x] * 2)
+ t = torch.cat([t] * 2)
+
+ noise_pred = self.k_diffusion_model(
+ latent_model_input,
+ t,
+ cond=prompt_embeds,
+ timestep_cond=timestep_cond,
+ added_cond_kwargs=added_cond_kwargs,
+ )
+
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+ return noise_pred
+
+ # 10. Run k-diffusion solver
+ sampler_kwargs = {}
+
+ if "noise_sampler" in inspect.signature(self.sampler).parameters:
+ min_sigma, max_sigma = sigmas[sigmas > 0].min(), sigmas.max()
+ noise_sampler = BrownianTreeNoiseSampler(latents, min_sigma, max_sigma, noise_sampler_seed)
+ sampler_kwargs["noise_sampler"] = noise_sampler
+
+ if "generator" in inspect.signature(self.sampler).parameters:
+ sampler_kwargs["generator"] = generator
+
+ latents = self.sampler(model_fn, latents, sigmas, **sampler_kwargs)
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ image = latents
+
+ if not output_type == "latent":
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..dae2affddd1fd5952f454ed9cee906277dcceb16
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__init__.py
@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_ldm3d"] = ["StableDiffusionLDM3DPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_stable_diffusion_ldm3d import StableDiffusionLDM3DPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ba17ece93446f50f4572007d4ebbe867c1030cd5
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__pycache__/pipeline_stable_diffusion_ldm3d.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__pycache__/pipeline_stable_diffusion_ldm3d.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c95429cebbcfa214138cf6a048585122f5a8e1a5
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_ldm3d/__pycache__/pipeline_stable_diffusion_ldm3d.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py b/model/diffusers_c/pipelines/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py
new file mode 100644
index 0000000000000000000000000000000000000000..c7c05feaf0130ade69e65b25a02bb1aa36f482fd
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_ldm3d/pipeline_stable_diffusion_ldm3d.py
@@ -0,0 +1,985 @@
+# Copyright 2024 The Intel Labs Team Authors and the HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessorLDM3D
+from ...loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ BaseOutput,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```python
+ >>> from diffusers import StableDiffusionLDM3DPipeline
+
+ >>> pipe = StableDiffusionLDM3DPipeline.from_pretrained("Intel/ldm3d-4c")
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> output = pipe(prompt)
+ >>> rgb_image, depth_image = output.rgb, output.depth
+ >>> rgb_image[0].save("astronaut_ldm3d_rgb.jpg")
+ >>> depth_image[0].save("astronaut_ldm3d_depth.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+@dataclass
+class LDM3DPipelineOutput(BaseOutput):
+ """
+ Output class for Stable Diffusion pipelines.
+
+ Args:
+ rgb (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ depth (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ nsfw_content_detected (`List[bool]`)
+ List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or
+ `None` if safety checking could not be performed.
+ """
+
+ rgb: Union[List[PIL.Image.Image], np.ndarray]
+ depth: Union[List[PIL.Image.Image], np.ndarray]
+ nsfw_content_detected: Optional[List[bool]]
+
+
+class StableDiffusionLDM3DPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ LoraLoaderMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image and 3D generation using LDM3D.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: Optional[CLIPVisionModelWithProjection],
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessorLDM3D(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ rgb_feature_extractor_input = feature_extractor_input[0]
+ safety_checker_input = self.feature_extractor(rgb_feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def guidance_rescale(self):
+ return self._guidance_rescale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @property
+ def interrupt(self):
+ return self._interrupt
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 49,
+ timesteps: List[int] = None,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 5.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._guidance_rescale = guidance_rescale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+ self._interrupt = False
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 6.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
+ # 6.2 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ if self.interrupt:
+ continue
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ rgb, depth = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return ((rgb, depth), has_nsfw_concept)
+
+ return LDM3DPipelineOutput(rgb=rgb, depth=depth, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_panorama/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_panorama/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..f7572db7236cd6bcfd7dd032abcb29fd5f67cf1c
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_panorama/__init__.py
@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_panorama"] = ["StableDiffusionPanoramaPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_stable_diffusion_panorama import StableDiffusionPanoramaPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_panorama/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_panorama/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b4275d3b0271382d5d85599192c107e1f36e7284
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_panorama/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_panorama/__pycache__/pipeline_stable_diffusion_panorama.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_panorama/__pycache__/pipeline_stable_diffusion_panorama.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b1af7c06e7d5682aeabbb92ff02b89095845e47a
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_panorama/__pycache__/pipeline_stable_diffusion_panorama.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py b/model/diffusers_c/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py
new file mode 100644
index 0000000000000000000000000000000000000000..feda710e0049ebaab9b5293fa1649011179130a5
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_panorama/pipeline_stable_diffusion_panorama.py
@@ -0,0 +1,933 @@
+# Copyright 2024 MultiDiffusion Authors and The HuggingFace Team. All rights reserved."
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import DDIMScheduler
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionPanoramaPipeline, DDIMScheduler
+
+ >>> model_ckpt = "stabilityai/stable-diffusion-2-base"
+ >>> scheduler = DDIMScheduler.from_pretrained(model_ckpt, subfolder="scheduler")
+ >>> pipe = StableDiffusionPanoramaPipeline.from_pretrained(
+ ... model_ckpt, scheduler=scheduler, torch_dtype=torch.float16
+ ... )
+
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a photo of the dolomites"
+ >>> image = pipe(prompt).images[0]
+ ```
+"""
+
+
+class StableDiffusionPanoramaPipeline(
+ DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin
+):
+ r"""
+ Pipeline for text-to-image generation using MultiDiffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: DDIMScheduler,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ def decode_latents_with_padding(self, latents, padding=8):
+ # Add padding to latents for circular inference
+ # padding is the number of latents to add on each side
+ # it would slightly increase the memory usage, but remove the boundary artifacts
+ latents = 1 / self.vae.config.scaling_factor * latents
+ latents_left = latents[..., :padding]
+ latents_right = latents[..., -padding:]
+ latents = torch.cat((latents_right, latents, latents_left), axis=-1)
+ image = self.vae.decode(latents, return_dict=False)[0]
+ padding_pix = self.vae_scale_factor * padding
+ image = image[..., padding_pix:-padding_pix]
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def get_views(self, panorama_height, panorama_width, window_size=64, stride=8, circular_padding=False):
+ # Here, we define the mappings F_i (see Eq. 7 in the MultiDiffusion paper https://arxiv.org/abs/2302.08113)
+ # if panorama's height/width < window_size, num_blocks of height/width should return 1
+ panorama_height /= 8
+ panorama_width /= 8
+ num_blocks_height = (panorama_height - window_size) // stride + 1 if panorama_height > window_size else 1
+ if circular_padding:
+ num_blocks_width = panorama_width // stride if panorama_width > window_size else 1
+ else:
+ num_blocks_width = (panorama_width - window_size) // stride + 1 if panorama_width > window_size else 1
+ total_num_blocks = int(num_blocks_height * num_blocks_width)
+ views = []
+ for i in range(total_num_blocks):
+ h_start = int((i // num_blocks_width) * stride)
+ h_end = h_start + window_size
+ w_start = int((i % num_blocks_width) * stride)
+ w_end = w_start + window_size
+ views.append((h_start, h_end, w_start, w_end))
+ return views
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = 512,
+ width: Optional[int] = 2048,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ view_batch_size: int = 1,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ circular_padding: bool = False,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 2048):
+ The width in pixels of the generated image. The width is kept high because the pipeline is supposed
+ generate panorama-like images.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ view_batch_size (`int`, *optional*, defaults to 1):
+ The batch size to denoise split views. For some GPUs with high performance, higher view batch size can
+ speedup the generation and increase the VRAM usage.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ circular_padding (`bool`, *optional*, defaults to `False`):
+ If set to `True`, circular padding is applied to ensure there are no stitching artifacts. Circular
+ padding allows the model to seamlessly generate a transition from the rightmost part of the image to
+ the leftmost part, maintaining consistency in a 360-degree sense.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Define panorama grid and initialize views for synthesis.
+ # prepare batch grid
+ views = self.get_views(height, width, circular_padding=circular_padding)
+ views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)]
+ views_scheduler_status = [copy.deepcopy(self.scheduler.__dict__)] * len(views_batch)
+ count = torch.zeros_like(latents)
+ value = torch.zeros_like(latents)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 8. Denoising loop
+ # Each denoising step also includes refinement of the latents with respect to the
+ # views.
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ count.zero_()
+ value.zero_()
+
+ # generate views
+ # Here, we iterate through different spatial crops of the latents and denoise them. These
+ # denoised (latent) crops are then averaged to produce the final latent
+ # for the current timestep via MultiDiffusion. Please see Sec. 4.1 in the
+ # MultiDiffusion paper for more details: https://arxiv.org/abs/2302.08113
+ # Batch views denoise
+ for j, batch_view in enumerate(views_batch):
+ vb_size = len(batch_view)
+ # get the latents corresponding to the current view coordinates
+ if circular_padding:
+ latents_for_view = []
+ for h_start, h_end, w_start, w_end in batch_view:
+ if w_end > latents.shape[3]:
+ # Add circular horizontal padding
+ latent_view = torch.cat(
+ (
+ latents[:, :, h_start:h_end, w_start:],
+ latents[:, :, h_start:h_end, : w_end - latents.shape[3]],
+ ),
+ axis=-1,
+ )
+ else:
+ latent_view = latents[:, :, h_start:h_end, w_start:w_end]
+ latents_for_view.append(latent_view)
+ latents_for_view = torch.cat(latents_for_view)
+ else:
+ latents_for_view = torch.cat(
+ [
+ latents[:, :, h_start:h_end, w_start:w_end]
+ for h_start, h_end, w_start, w_end in batch_view
+ ]
+ )
+
+ # rematch block's scheduler status
+ self.scheduler.__dict__.update(views_scheduler_status[j])
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = (
+ latents_for_view.repeat_interleave(2, dim=0)
+ if do_classifier_free_guidance
+ else latents_for_view
+ )
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # repeat prompt_embeds for batch
+ prompt_embeds_input = torch.cat([prompt_embeds] * vb_size)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds_input,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::2]
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents_denoised_batch = self.scheduler.step(
+ noise_pred, t, latents_for_view, **extra_step_kwargs
+ ).prev_sample
+
+ # save views scheduler status after sample
+ views_scheduler_status[j] = copy.deepcopy(self.scheduler.__dict__)
+
+ # extract value from batch
+ for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip(
+ latents_denoised_batch.chunk(vb_size), batch_view
+ ):
+ if circular_padding and w_end > latents.shape[3]:
+ # Case for circular padding
+ value[:, :, h_start:h_end, w_start:] += latents_view_denoised[
+ :, :, h_start:h_end, : latents.shape[3] - w_start
+ ]
+ value[:, :, h_start:h_end, : w_end - latents.shape[3]] += latents_view_denoised[
+ :, :, h_start:h_end, latents.shape[3] - w_start :
+ ]
+ count[:, :, h_start:h_end, w_start:] += 1
+ count[:, :, h_start:h_end, : w_end - latents.shape[3]] += 1
+ else:
+ value[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised
+ count[:, :, h_start:h_end, w_start:w_end] += 1
+
+ # take the MultiDiffusion step. Eq. 5 in MultiDiffusion paper: https://arxiv.org/abs/2302.08113
+ latents = torch.where(count > 0, value / count, value)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ if circular_padding:
+ image = self.decode_latents_with_padding(latents)
+ else:
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_safe/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_safe/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..b432b9418c46257913d81c5bf56edc0f1fa74ed1
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_safe/__init__.py
@@ -0,0 +1,99 @@
+from dataclasses import dataclass
+from enum import Enum
+from typing import TYPE_CHECKING, List, Optional, Union
+
+import numpy as np
+import PIL
+from PIL import Image
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ BaseOutput,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+@dataclass
+class SafetyConfig(object):
+ WEAK = {
+ "sld_warmup_steps": 15,
+ "sld_guidance_scale": 20,
+ "sld_threshold": 0.0,
+ "sld_momentum_scale": 0.0,
+ "sld_mom_beta": 0.0,
+ }
+ MEDIUM = {
+ "sld_warmup_steps": 10,
+ "sld_guidance_scale": 1000,
+ "sld_threshold": 0.01,
+ "sld_momentum_scale": 0.3,
+ "sld_mom_beta": 0.4,
+ }
+ STRONG = {
+ "sld_warmup_steps": 7,
+ "sld_guidance_scale": 2000,
+ "sld_threshold": 0.025,
+ "sld_momentum_scale": 0.5,
+ "sld_mom_beta": 0.7,
+ }
+ MAX = {
+ "sld_warmup_steps": 0,
+ "sld_guidance_scale": 5000,
+ "sld_threshold": 1.0,
+ "sld_momentum_scale": 0.5,
+ "sld_mom_beta": 0.7,
+ }
+
+
+_dummy_objects = {}
+_additional_imports = {}
+_import_structure = {}
+
+_additional_imports.update({"SafetyConfig": SafetyConfig})
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure.update(
+ {
+ "pipeline_output": ["StableDiffusionSafePipelineOutput"],
+ "pipeline_stable_diffusion_safe": ["StableDiffusionPipelineSafe"],
+ "safety_checker": ["StableDiffusionSafetyChecker"],
+ }
+ )
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_output import StableDiffusionSafePipelineOutput
+ from .pipeline_stable_diffusion_safe import StableDiffusionPipelineSafe
+ from .safety_checker import SafeStableDiffusionSafetyChecker
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
+ for name, value in _additional_imports.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..70eb1d75c5f901bf3d66627d318b206959172a26
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/pipeline_output.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/pipeline_output.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5728e266660147d07524129a6817daf4f1840648
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/pipeline_output.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/pipeline_stable_diffusion_safe.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/pipeline_stable_diffusion_safe.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..fe663eac754faa706b9be296f911bb3a5f197b44
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/pipeline_stable_diffusion_safe.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/safety_checker.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/safety_checker.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5a882a1f3babf138086930825dcd9a52409bd75f
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_safe/__pycache__/safety_checker.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_safe/pipeline_output.py b/model/diffusers_c/pipelines/stable_diffusion_safe/pipeline_output.py
new file mode 100644
index 0000000000000000000000000000000000000000..69a064d6638df556d3007f59daf7e767ec7c298b
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_safe/pipeline_output.py
@@ -0,0 +1,34 @@
+from dataclasses import dataclass
+from typing import List, Optional, Union
+
+import numpy as np
+import PIL.Image
+
+from ...utils import (
+ BaseOutput,
+)
+
+
+@dataclass
+class StableDiffusionSafePipelineOutput(BaseOutput):
+ """
+ Output class for Safe Stable Diffusion pipelines.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+ num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+ nsfw_content_detected (`List[bool]`)
+ List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, or `None` if safety checking could not be performed.
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images that were flagged by the safety checker any may contain "not-safe-for-work"
+ (nsfw) content, or `None` if no safety check was performed or no images were flagged.
+ applied_safety_concept (`str`)
+ The safety concept that was applied for safety guidance, or `None` if safety guidance was disabled
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+ nsfw_content_detected: Optional[List[bool]]
+ unsafe_images: Optional[Union[List[PIL.Image.Image], np.ndarray]]
+ applied_safety_concept: Optional[str]
diff --git a/model/diffusers_c/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py b/model/diffusers_c/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py
new file mode 100644
index 0000000000000000000000000000000000000000..ae74e09678e3ad570b8f0d3e2b3359cbd937b5bb
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_safe/pipeline_stable_diffusion_safe.py
@@ -0,0 +1,764 @@
+import inspect
+import warnings
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...configuration_utils import FrozenDict
+from ...image_processor import PipelineImageInput
+from ...loaders import IPAdapterMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import deprecate, logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from . import StableDiffusionSafePipelineOutput
+from .safety_checker import SafeStableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class StableDiffusionPipelineSafe(DiffusionPipeline, StableDiffusionMixin, IPAdapterMixin):
+ r"""
+ Pipeline based on the [`StableDiffusionPipeline`] for text-to-image generation using Safe Latent Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: SafeStableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+ safety_concept: Optional[str] = (
+ "an image showing hate, harassment, violence, suffering, humiliation, harm, suicide, sexual, nudity,"
+ " bodily fluids, blood, obscene gestures, illegal activity, drug use, theft, vandalism, weapons, child"
+ " abuse, brutality, cruelty"
+ )
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+ " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+ " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
+ " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
+ " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
+ )
+ deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["clip_sample"] = False
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self._safety_text_concept = safety_concept
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ @property
+ def safety_concept(self):
+ r"""
+ Getter method for the safety concept used with SLD
+
+ Returns:
+ `str`: The text describing the safety concept
+ """
+ return self._safety_text_concept
+
+ @safety_concept.setter
+ def safety_concept(self, concept):
+ r"""
+ Setter method for the safety concept used with SLD
+
+ Args:
+ concept (`str`):
+ The text of the new safety concept
+ """
+ self._safety_text_concept = concept
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ enable_safety_guidance,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ """
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="max_length", return_tensors="pt").input_ids
+
+ if not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embeds = prompt_embeds[0]
+
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ max_length = text_input_ids.shape[-1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # Encode the safety concept text
+ if enable_safety_guidance:
+ safety_concept_input = self.tokenizer(
+ [self._safety_text_concept],
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ safety_embeddings = self.text_encoder(safety_concept_input.input_ids.to(self.device))[0]
+
+ # duplicate safety embeddings for each generation per prompt, using mps friendly method
+ seq_len = safety_embeddings.shape[1]
+ safety_embeddings = safety_embeddings.repeat(batch_size, num_images_per_prompt, 1)
+ safety_embeddings = safety_embeddings.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance + sld, we need to do three forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing three forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, safety_embeddings])
+
+ else:
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ def run_safety_checker(self, image, device, dtype, enable_safety_guidance):
+ if self.safety_checker is not None:
+ images = image.copy()
+ safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ flagged_images = np.zeros((2, *image.shape[1:]))
+ if any(has_nsfw_concept):
+ logger.warning(
+ "Potential NSFW content was detected in one or more images. A black image will be returned"
+ " instead."
+ f"{'You may look at this images in the `unsafe_images` variable of the output at your own discretion.' if enable_safety_guidance else 'Try again with a different prompt and/or seed.'}"
+ )
+ for idx, has_nsfw_concept in enumerate(has_nsfw_concept):
+ if has_nsfw_concept:
+ flagged_images[idx] = images[idx]
+ image[idx] = np.zeros(image[idx].shape) # black image
+ else:
+ has_nsfw_concept = None
+ flagged_images = None
+ return image, has_nsfw_concept, flagged_images
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def perform_safety_guidance(
+ self,
+ enable_safety_guidance,
+ safety_momentum,
+ noise_guidance,
+ noise_pred_out,
+ i,
+ sld_guidance_scale,
+ sld_warmup_steps,
+ sld_threshold,
+ sld_momentum_scale,
+ sld_mom_beta,
+ ):
+ # Perform SLD guidance
+ if enable_safety_guidance:
+ if safety_momentum is None:
+ safety_momentum = torch.zeros_like(noise_guidance)
+ noise_pred_text, noise_pred_uncond = noise_pred_out[0], noise_pred_out[1]
+ noise_pred_safety_concept = noise_pred_out[2]
+
+ # Equation 6
+ scale = torch.clamp(torch.abs((noise_pred_text - noise_pred_safety_concept)) * sld_guidance_scale, max=1.0)
+
+ # Equation 6
+ safety_concept_scale = torch.where(
+ (noise_pred_text - noise_pred_safety_concept) >= sld_threshold, torch.zeros_like(scale), scale
+ )
+
+ # Equation 4
+ noise_guidance_safety = torch.mul((noise_pred_safety_concept - noise_pred_uncond), safety_concept_scale)
+
+ # Equation 7
+ noise_guidance_safety = noise_guidance_safety + sld_momentum_scale * safety_momentum
+
+ # Equation 8
+ safety_momentum = sld_mom_beta * safety_momentum + (1 - sld_mom_beta) * noise_guidance_safety
+
+ if i >= sld_warmup_steps: # Warmup
+ # Equation 3
+ noise_guidance = noise_guidance - noise_guidance_safety
+ return noise_guidance, safety_momentum
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ sld_guidance_scale: Optional[float] = 1000,
+ sld_warmup_steps: Optional[int] = 10,
+ sld_threshold: Optional[float] = 0.01,
+ sld_momentum_scale: Optional[float] = 0.3,
+ sld_mom_beta: Optional[float] = 0.4,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ sld_guidance_scale (`float`, *optional*, defaults to 1000):
+ If `sld_guidance_scale < 1`, safety guidance is disabled.
+ sld_warmup_steps (`int`, *optional*, defaults to 10):
+ Number of warmup steps for safety guidance. SLD is only be applied for diffusion steps greater than
+ `sld_warmup_steps`.
+ sld_threshold (`float`, *optional*, defaults to 0.01):
+ Threshold that separates the hyperplane between appropriate and inappropriate images.
+ sld_momentum_scale (`float`, *optional*, defaults to 0.3):
+ Scale of the SLD momentum to be added to the safety guidance at each diffusion step. If set to 0.0,
+ momentum is disabled. Momentum is built up during warmup for diffusion steps smaller than
+ `sld_warmup_steps`.
+ sld_mom_beta (`float`, *optional*, defaults to 0.4):
+ Defines how safety guidance momentum builds up. `sld_mom_beta` indicates how much of the previous
+ momentum is kept. Momentum is built up during warmup for diffusion steps smaller than
+ `sld_warmup_steps`.
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+
+ Examples:
+
+ ```py
+ import torch
+ from diffusers import StableDiffusionPipelineSafe
+ from diffusers.pipelines.stable_diffusion_safe import SafetyConfig
+
+ pipeline = StableDiffusionPipelineSafe.from_pretrained(
+ "AIML-TUDA/stable-diffusion-safe", torch_dtype=torch.float16
+ ).to("cuda")
+ prompt = "the four horsewomen of the apocalypse, painting by tom of finland, gaston bussiere, craig mullins, j. c. leyendecker"
+ image = pipeline(prompt=prompt, **SafetyConfig.MEDIUM).images[0]
+ ```
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(prompt, height, width, callback_steps)
+
+ # 2. Define call parameters
+ batch_size = 1 if isinstance(prompt, str) else len(prompt)
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ enable_safety_guidance = sld_guidance_scale > 1.0 and do_classifier_free_guidance
+ if not enable_safety_guidance:
+ warnings.warn("Safety checker disabled!")
+
+ if ip_adapter_image is not None:
+ output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+ image_embeds, negative_image_embeds = self.encode_image(
+ ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+ )
+ if do_classifier_free_guidance:
+ if enable_safety_guidance:
+ image_embeds = torch.cat([negative_image_embeds, image_embeds, image_embeds])
+ else:
+ image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
+ # 3. Encode input prompt
+ prompt_embeds = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt, enable_safety_guidance
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 6.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+
+ safety_momentum = None
+
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = (
+ torch.cat([latents] * (3 if enable_safety_guidance else 2))
+ if do_classifier_free_guidance
+ else latents
+ )
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input, t, encoder_hidden_states=prompt_embeds, added_cond_kwargs=added_cond_kwargs
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_out = noise_pred.chunk((3 if enable_safety_guidance else 2))
+ noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
+
+ # default classifier free guidance
+ noise_guidance = noise_pred_text - noise_pred_uncond
+
+ # Perform SLD guidance
+ if enable_safety_guidance:
+ if safety_momentum is None:
+ safety_momentum = torch.zeros_like(noise_guidance)
+ noise_pred_safety_concept = noise_pred_out[2]
+
+ # Equation 6
+ scale = torch.clamp(
+ torch.abs((noise_pred_text - noise_pred_safety_concept)) * sld_guidance_scale, max=1.0
+ )
+
+ # Equation 6
+ safety_concept_scale = torch.where(
+ (noise_pred_text - noise_pred_safety_concept) >= sld_threshold,
+ torch.zeros_like(scale),
+ scale,
+ )
+
+ # Equation 4
+ noise_guidance_safety = torch.mul(
+ (noise_pred_safety_concept - noise_pred_uncond), safety_concept_scale
+ )
+
+ # Equation 7
+ noise_guidance_safety = noise_guidance_safety + sld_momentum_scale * safety_momentum
+
+ # Equation 8
+ safety_momentum = sld_mom_beta * safety_momentum + (1 - sld_mom_beta) * noise_guidance_safety
+
+ if i >= sld_warmup_steps: # Warmup
+ # Equation 3
+ noise_guidance = noise_guidance - noise_guidance_safety
+
+ noise_pred = noise_pred_uncond + guidance_scale * noise_guidance
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 8. Post-processing
+ image = self.decode_latents(latents)
+
+ # 9. Run safety checker
+ image, has_nsfw_concept, flagged_images = self.run_safety_checker(
+ image, device, prompt_embeds.dtype, enable_safety_guidance
+ )
+
+ # 10. Convert to PIL
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+ if flagged_images is not None:
+ flagged_images = self.numpy_to_pil(flagged_images)
+
+ if not return_dict:
+ return (
+ image,
+ has_nsfw_concept,
+ self._safety_text_concept if enable_safety_guidance else None,
+ flagged_images,
+ )
+
+ return StableDiffusionSafePipelineOutput(
+ images=image,
+ nsfw_content_detected=has_nsfw_concept,
+ applied_safety_concept=self._safety_text_concept if enable_safety_guidance else None,
+ unsafe_images=flagged_images,
+ )
diff --git a/model/diffusers_c/pipelines/stable_diffusion_safe/safety_checker.py b/model/diffusers_c/pipelines/stable_diffusion_safe/safety_checker.py
new file mode 100644
index 0000000000000000000000000000000000000000..549747e971466cfa31ad8504c925d747a1394c67
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_safe/safety_checker.py
@@ -0,0 +1,109 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from transformers import CLIPConfig, CLIPVisionModel, PreTrainedModel
+
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+def cosine_distance(image_embeds, text_embeds):
+ normalized_image_embeds = nn.functional.normalize(image_embeds)
+ normalized_text_embeds = nn.functional.normalize(text_embeds)
+ return torch.mm(normalized_image_embeds, normalized_text_embeds.t())
+
+
+class SafeStableDiffusionSafetyChecker(PreTrainedModel):
+ config_class = CLIPConfig
+
+ _no_split_modules = ["CLIPEncoderLayer"]
+
+ def __init__(self, config: CLIPConfig):
+ super().__init__(config)
+
+ self.vision_model = CLIPVisionModel(config.vision_config)
+ self.visual_projection = nn.Linear(config.vision_config.hidden_size, config.projection_dim, bias=False)
+
+ self.concept_embeds = nn.Parameter(torch.ones(17, config.projection_dim), requires_grad=False)
+ self.special_care_embeds = nn.Parameter(torch.ones(3, config.projection_dim), requires_grad=False)
+
+ self.concept_embeds_weights = nn.Parameter(torch.ones(17), requires_grad=False)
+ self.special_care_embeds_weights = nn.Parameter(torch.ones(3), requires_grad=False)
+
+ @torch.no_grad()
+ def forward(self, clip_input, images):
+ pooled_output = self.vision_model(clip_input)[1] # pooled_output
+ image_embeds = self.visual_projection(pooled_output)
+
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds).cpu().float().numpy()
+ cos_dist = cosine_distance(image_embeds, self.concept_embeds).cpu().float().numpy()
+
+ result = []
+ batch_size = image_embeds.shape[0]
+ for i in range(batch_size):
+ result_img = {"special_scores": {}, "special_care": [], "concept_scores": {}, "bad_concepts": []}
+
+ # increase this value to create a stronger `nfsw` filter
+ # at the cost of increasing the possibility of filtering benign images
+ adjustment = 0.0
+
+ for concept_idx in range(len(special_cos_dist[0])):
+ concept_cos = special_cos_dist[i][concept_idx]
+ concept_threshold = self.special_care_embeds_weights[concept_idx].item()
+ result_img["special_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3)
+ if result_img["special_scores"][concept_idx] > 0:
+ result_img["special_care"].append({concept_idx, result_img["special_scores"][concept_idx]})
+ adjustment = 0.01
+
+ for concept_idx in range(len(cos_dist[0])):
+ concept_cos = cos_dist[i][concept_idx]
+ concept_threshold = self.concept_embeds_weights[concept_idx].item()
+ result_img["concept_scores"][concept_idx] = round(concept_cos - concept_threshold + adjustment, 3)
+ if result_img["concept_scores"][concept_idx] > 0:
+ result_img["bad_concepts"].append(concept_idx)
+
+ result.append(result_img)
+
+ has_nsfw_concepts = [len(res["bad_concepts"]) > 0 for res in result]
+
+ return images, has_nsfw_concepts
+
+ @torch.no_grad()
+ def forward_onnx(self, clip_input: torch.FloatTensor, images: torch.FloatTensor):
+ pooled_output = self.vision_model(clip_input)[1] # pooled_output
+ image_embeds = self.visual_projection(pooled_output)
+
+ special_cos_dist = cosine_distance(image_embeds, self.special_care_embeds)
+ cos_dist = cosine_distance(image_embeds, self.concept_embeds)
+
+ # increase this value to create a stronger `nsfw` filter
+ # at the cost of increasing the possibility of filtering benign images
+ adjustment = 0.0
+
+ special_scores = special_cos_dist - self.special_care_embeds_weights + adjustment
+ # special_scores = special_scores.round(decimals=3)
+ special_care = torch.any(special_scores > 0, dim=1)
+ special_adjustment = special_care * 0.01
+ special_adjustment = special_adjustment.unsqueeze(1).expand(-1, cos_dist.shape[1])
+
+ concept_scores = (cos_dist - self.concept_embeds_weights) + special_adjustment
+ # concept_scores = concept_scores.round(decimals=3)
+ has_nsfw_concepts = torch.any(concept_scores > 0, dim=1)
+
+ return images, has_nsfw_concepts
diff --git a/model/diffusers_c/pipelines/stable_diffusion_sag/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_sag/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..378e0e57817f58a0a28afed5d6110f6ee3effb3a
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_sag/__init__.py
@@ -0,0 +1,48 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_sag"] = ["StableDiffusionSAGPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_stable_diffusion_sag import StableDiffusionSAGPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_sag/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_sag/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..951a933adb745f3d5a8dbea901c888cd8740c9b4
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_sag/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_sag/__pycache__/pipeline_stable_diffusion_sag.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_sag/__pycache__/pipeline_stable_diffusion_sag.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5c067a396ce6715afe6a20d7208f41baf9c1347c
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_sag/__pycache__/pipeline_stable_diffusion_sag.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py b/model/diffusers_c/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py
new file mode 100644
index 0000000000000000000000000000000000000000..96aa006d2ab38aaaa30d6ff1c5765baee1677897
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_sag/pipeline_stable_diffusion_sag.py
@@ -0,0 +1,886 @@
+# Copyright 2024 Susung Hong and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionPipelineOutput
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionSAGPipeline
+
+ >>> pipe = StableDiffusionSAGPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> image = pipe(prompt, sag_scale=0.75).images[0]
+ ```
+"""
+
+
+# processes and stores attention probabilities
+class CrossAttnStoreProcessor:
+ def __init__(self):
+ self.attention_probs = None
+
+ def __call__(
+ self,
+ attn,
+ hidden_states,
+ encoder_hidden_states=None,
+ attention_mask=None,
+ ):
+ batch_size, sequence_length, _ = hidden_states.shape
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ query = attn.to_q(hidden_states)
+
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ query = attn.head_to_batch_dim(query)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ self.attention_probs = attn.get_attention_scores(query, key, attention_mask)
+ hidden_states = torch.bmm(self.attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ return hidden_states
+
+
+# Modified to get self-attention guidance scale in this paper (https://arxiv.org/pdf/2210.00939.pdf) as an input
+class StableDiffusionSAGPipeline(DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, IPAdapterMixin):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: Optional[CLIPVisionModelWithProjection] = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ sag_scale: float = 0.75,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ sag_scale (`float`, *optional*, defaults to 0.75):
+ Chosen between [0, 1.0] for better quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+ # and `sag_scale` is` `s` of equation (16)
+ # of the self-attentnion guidance paper: https://arxiv.org/pdf/2210.00939.pdf
+ # `sag_scale = 0` means no self-attention guidance
+ do_self_attention_guidance = sag_scale > 0.0
+
+ if ip_adapter_image is not None:
+ output_hidden_state = False if isinstance(self.unet.encoder_hid_proj, ImageProjection) else True
+ image_embeds, negative_image_embeds = self.encode_image(
+ ip_adapter_image, device, num_images_per_prompt, output_hidden_state
+ )
+ if do_classifier_free_guidance:
+ image_embeds = torch.cat([negative_image_embeds, image_embeds])
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ if timesteps.dtype not in [torch.int16, torch.int32, torch.int64]:
+ raise ValueError(
+ f"{self.__class__.__name__} does not support using a scheduler of type {self.scheduler.__class__.__name__}. Please make sure to use one of 'DDIMScheduler, PNDMScheduler, DDPMScheduler, DEISMultistepScheduler, UniPCMultistepScheduler, DPMSolverMultistepScheduler, DPMSolverSinlgestepScheduler'."
+ )
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 6.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None
+ added_uncond_kwargs = {"image_embeds": negative_image_embeds} if ip_adapter_image is not None else None
+
+ # 7. Denoising loop
+ store_processor = CrossAttnStoreProcessor()
+ self.unet.mid_block.attentions[0].transformer_blocks[0].attn1.processor = store_processor
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+
+ map_size = None
+
+ def get_map_size(module, input, output):
+ nonlocal map_size
+ map_size = output[0].shape[-2:]
+
+ with self.unet.mid_block.attentions[0].register_forward_hook(get_map_size):
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # perform self-attention guidance with the stored self-attentnion map
+ if do_self_attention_guidance:
+ # classifier-free guidance produces two chunks of attention map
+ # and we only use unconditional one according to equation (25)
+ # in https://arxiv.org/pdf/2210.00939.pdf
+ if do_classifier_free_guidance:
+ # DDIM-like prediction of x0
+ pred_x0 = self.pred_x0(latents, noise_pred_uncond, t)
+ # get the stored attention maps
+ uncond_attn, cond_attn = store_processor.attention_probs.chunk(2)
+ # self-attention-based degrading of latents
+ degraded_latents = self.sag_masking(
+ pred_x0, uncond_attn, map_size, t, self.pred_epsilon(latents, noise_pred_uncond, t)
+ )
+ uncond_emb, _ = prompt_embeds.chunk(2)
+ # forward and give guidance
+ degraded_pred = self.unet(
+ degraded_latents,
+ t,
+ encoder_hidden_states=uncond_emb,
+ added_cond_kwargs=added_uncond_kwargs,
+ ).sample
+ noise_pred += sag_scale * (noise_pred_uncond - degraded_pred)
+ else:
+ # DDIM-like prediction of x0
+ pred_x0 = self.pred_x0(latents, noise_pred, t)
+ # get the stored attention maps
+ cond_attn = store_processor.attention_probs
+ # self-attention-based degrading of latents
+ degraded_latents = self.sag_masking(
+ pred_x0, cond_attn, map_size, t, self.pred_epsilon(latents, noise_pred, t)
+ )
+ # forward and give guidance
+ degraded_pred = self.unet(
+ degraded_latents,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ added_cond_kwargs=added_cond_kwargs,
+ ).sample
+ noise_pred += sag_scale * (noise_pred - degraded_pred)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+ def sag_masking(self, original_latents, attn_map, map_size, t, eps):
+ # Same masking process as in SAG paper: https://arxiv.org/pdf/2210.00939.pdf
+ bh, hw1, hw2 = attn_map.shape
+ b, latent_channel, latent_h, latent_w = original_latents.shape
+ h = self.unet.config.attention_head_dim
+ if isinstance(h, list):
+ h = h[-1]
+
+ # Produce attention mask
+ attn_map = attn_map.reshape(b, h, hw1, hw2)
+ attn_mask = attn_map.mean(1, keepdim=False).sum(1, keepdim=False) > 1.0
+ attn_mask = (
+ attn_mask.reshape(b, map_size[0], map_size[1])
+ .unsqueeze(1)
+ .repeat(1, latent_channel, 1, 1)
+ .type(attn_map.dtype)
+ )
+ attn_mask = F.interpolate(attn_mask, (latent_h, latent_w))
+
+ # Blur according to the self-attention mask
+ degraded_latents = gaussian_blur_2d(original_latents, kernel_size=9, sigma=1.0)
+ degraded_latents = degraded_latents * attn_mask + original_latents * (1 - attn_mask)
+
+ # Noise it again to match the noise level
+ degraded_latents = self.scheduler.add_noise(degraded_latents, noise=eps, timesteps=t[None])
+
+ return degraded_latents
+
+ # Modified from diffusers.schedulers.scheduling_ddim.DDIMScheduler.step
+ # Note: there are some schedulers that clip or do not return x_0 (PNDMScheduler, DDIMScheduler, etc.)
+ def pred_x0(self, sample, model_output, timestep):
+ alpha_prod_t = self.scheduler.alphas_cumprod[timestep].to(sample.device)
+
+ beta_prod_t = 1 - alpha_prod_t
+ if self.scheduler.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ elif self.scheduler.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ elif self.scheduler.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ # predict V
+ model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.scheduler.config.prediction_type} must be one of `epsilon`, `sample`,"
+ " or `v_prediction`"
+ )
+
+ return pred_original_sample
+
+ def pred_epsilon(self, sample, model_output, timestep):
+ alpha_prod_t = self.scheduler.alphas_cumprod[timestep]
+
+ beta_prod_t = 1 - alpha_prod_t
+ if self.scheduler.config.prediction_type == "epsilon":
+ pred_eps = model_output
+ elif self.scheduler.config.prediction_type == "sample":
+ pred_eps = (sample - (alpha_prod_t**0.5) * model_output) / (beta_prod_t**0.5)
+ elif self.scheduler.config.prediction_type == "v_prediction":
+ pred_eps = (beta_prod_t**0.5) * sample + (alpha_prod_t**0.5) * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.scheduler.config.prediction_type} must be one of `epsilon`, `sample`,"
+ " or `v_prediction`"
+ )
+
+ return pred_eps
+
+
+# Gaussian blur
+def gaussian_blur_2d(img, kernel_size, sigma):
+ ksize_half = (kernel_size - 1) * 0.5
+
+ x = torch.linspace(-ksize_half, ksize_half, steps=kernel_size)
+
+ pdf = torch.exp(-0.5 * (x / sigma).pow(2))
+
+ x_kernel = pdf / pdf.sum()
+ x_kernel = x_kernel.to(device=img.device, dtype=img.dtype)
+
+ kernel2d = torch.mm(x_kernel[:, None], x_kernel[None, :])
+ kernel2d = kernel2d.expand(img.shape[-3], 1, kernel2d.shape[0], kernel2d.shape[1])
+
+ padding = [kernel_size // 2, kernel_size // 2, kernel_size // 2, kernel_size // 2]
+
+ img = F.pad(img, padding, mode="reflect")
+ img = F.conv2d(img, kernel2d, groups=img.shape[-3])
+
+ return img
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/__init__.py b/model/diffusers_c/pipelines/stable_diffusion_xl/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8088fbcfceba205b9b908613f4ca3fdc579120e8
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_xl/__init__.py
@@ -0,0 +1,76 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_flax_available,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_additional_imports = {}
+_import_structure = {"pipeline_output": ["StableDiffusionXLPipelineOutput"]}
+
+if is_transformers_available() and is_flax_available():
+ _import_structure["pipeline_output"].extend(["FlaxStableDiffusionXLPipelineOutput"])
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_xl"] = ["StableDiffusionXLPipeline"]
+ _import_structure["pipeline_stable_diffusion_xl_img2img"] = ["StableDiffusionXLImg2ImgPipeline"]
+ _import_structure["pipeline_stable_diffusion_xl_inpaint"] = ["StableDiffusionXLInpaintPipeline"]
+ _import_structure["pipeline_stable_diffusion_xl_instruct_pix2pix"] = ["StableDiffusionXLInstructPix2PixPipeline"]
+
+if is_transformers_available() and is_flax_available():
+ from ...schedulers.scheduling_pndm_flax import PNDMSchedulerState
+
+ _additional_imports.update({"PNDMSchedulerState": PNDMSchedulerState})
+ _import_structure["pipeline_flax_stable_diffusion_xl"] = ["FlaxStableDiffusionXLPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import * # noqa F403
+ else:
+ from .pipeline_stable_diffusion_xl import StableDiffusionXLPipeline
+ from .pipeline_stable_diffusion_xl_img2img import StableDiffusionXLImg2ImgPipeline
+ from .pipeline_stable_diffusion_xl_inpaint import StableDiffusionXLInpaintPipeline
+ from .pipeline_stable_diffusion_xl_instruct_pix2pix import StableDiffusionXLInstructPix2PixPipeline
+
+ try:
+ if not (is_transformers_available() and is_flax_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_flax_objects import *
+ else:
+ from .pipeline_flax_stable_diffusion_xl import (
+ FlaxStableDiffusionXLPipeline,
+ )
+ from .pipeline_output import FlaxStableDiffusionXLPipelineOutput
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
+ for name, value in _additional_imports.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..41f2748d992f19bc39d6e4f54f042a8126868224
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_output.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_output.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..32023ffe4352f2eb98f2346c2c13ec9621773bd7
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_output.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..92dd6e04f985cc35c70d540d7d4c3e280df635c1
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0416068f729b27835e8bc55ecca3daa940838b7c
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_inpaint.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_inpaint.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..36b3564b7bf1c6259ec5c59bee6ee618bc556e35
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_inpaint.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_instruct_pix2pix.cpython-39.pyc b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_instruct_pix2pix.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7bd94c7610404e0418a61fdd2c5b297e3edc5fad
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_diffusion_xl/__pycache__/pipeline_stable_diffusion_xl_instruct_pix2pix.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_flax_stable_diffusion_xl.py b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_flax_stable_diffusion_xl.py
new file mode 100644
index 0000000000000000000000000000000000000000..77363b2546d72836618d4fad3d24ca572d1c8f12
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_flax_stable_diffusion_xl.py
@@ -0,0 +1,308 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import partial
+from typing import Dict, List, Optional, Union
+
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict
+from transformers import CLIPTokenizer, FlaxCLIPTextModel
+
+from diffusers.utils import logging
+
+from ...models import FlaxAutoencoderKL, FlaxUNet2DConditionModel
+from ...schedulers import (
+ FlaxDDIMScheduler,
+ FlaxDPMSolverMultistepScheduler,
+ FlaxLMSDiscreteScheduler,
+ FlaxPNDMScheduler,
+)
+from ..pipeline_flax_utils import FlaxDiffusionPipeline
+from .pipeline_output import FlaxStableDiffusionXLPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+# Set to True to use python for loop instead of jax.fori_loop for easier debugging
+DEBUG = False
+
+
+class FlaxStableDiffusionXLPipeline(FlaxDiffusionPipeline):
+ def __init__(
+ self,
+ text_encoder: FlaxCLIPTextModel,
+ text_encoder_2: FlaxCLIPTextModel,
+ vae: FlaxAutoencoderKL,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: FlaxUNet2DConditionModel,
+ scheduler: Union[
+ FlaxDDIMScheduler, FlaxPNDMScheduler, FlaxLMSDiscreteScheduler, FlaxDPMSolverMultistepScheduler
+ ],
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ super().__init__()
+ self.dtype = dtype
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+
+ def prepare_inputs(self, prompt: Union[str, List[str]]):
+ if not isinstance(prompt, (str, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ # Assume we have the two encoders
+ inputs = []
+ for tokenizer in [self.tokenizer, self.tokenizer_2]:
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="np",
+ )
+ inputs.append(text_inputs.input_ids)
+ inputs = jnp.stack(inputs, axis=1)
+ return inputs
+
+ def __call__(
+ self,
+ prompt_ids: jax.Array,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ num_inference_steps: int = 50,
+ guidance_scale: Union[float, jax.Array] = 7.5,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ latents: jnp.array = None,
+ neg_prompt_ids: jnp.array = None,
+ return_dict: bool = True,
+ output_type: str = None,
+ jit: bool = False,
+ ):
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ if isinstance(guidance_scale, float) and jit:
+ # Convert to a tensor so each device gets a copy.
+ guidance_scale = jnp.array([guidance_scale] * prompt_ids.shape[0])
+ guidance_scale = guidance_scale[:, None]
+
+ return_latents = output_type == "latent"
+
+ if jit:
+ images = _p_generate(
+ self,
+ prompt_ids,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ return_latents,
+ )
+ else:
+ images = self._generate(
+ prompt_ids,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ return_latents,
+ )
+
+ if not return_dict:
+ return (images,)
+
+ return FlaxStableDiffusionXLPipelineOutput(images=images)
+
+ def get_embeddings(self, prompt_ids: jnp.array, params):
+ # We assume we have the two encoders
+
+ # bs, encoder_input, seq_length
+ te_1_inputs = prompt_ids[:, 0, :]
+ te_2_inputs = prompt_ids[:, 1, :]
+
+ prompt_embeds = self.text_encoder(te_1_inputs, params=params["text_encoder"], output_hidden_states=True)
+ prompt_embeds = prompt_embeds["hidden_states"][-2]
+ prompt_embeds_2_out = self.text_encoder_2(
+ te_2_inputs, params=params["text_encoder_2"], output_hidden_states=True
+ )
+ prompt_embeds_2 = prompt_embeds_2_out["hidden_states"][-2]
+ text_embeds = prompt_embeds_2_out["text_embeds"]
+ prompt_embeds = jnp.concatenate([prompt_embeds, prompt_embeds_2], axis=-1)
+ return prompt_embeds, text_embeds
+
+ def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, bs, dtype):
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+ add_time_ids = jnp.array([add_time_ids] * bs, dtype=dtype)
+ return add_time_ids
+
+ def _generate(
+ self,
+ prompt_ids: jnp.array,
+ params: Union[Dict, FrozenDict],
+ prng_seed: jax.Array,
+ num_inference_steps: int,
+ height: int,
+ width: int,
+ guidance_scale: float,
+ latents: Optional[jnp.array] = None,
+ neg_prompt_ids: Optional[jnp.array] = None,
+ return_latents=False,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ # Encode input prompt
+ prompt_embeds, pooled_embeds = self.get_embeddings(prompt_ids, params)
+
+ # Get unconditional embeddings
+ batch_size = prompt_embeds.shape[0]
+ if neg_prompt_ids is None:
+ neg_prompt_embeds = jnp.zeros_like(prompt_embeds)
+ negative_pooled_embeds = jnp.zeros_like(pooled_embeds)
+ else:
+ neg_prompt_embeds, negative_pooled_embeds = self.get_embeddings(neg_prompt_ids, params)
+
+ add_time_ids = self._get_add_time_ids(
+ (height, width), (0, 0), (height, width), prompt_embeds.shape[0], dtype=prompt_embeds.dtype
+ )
+
+ prompt_embeds = jnp.concatenate([neg_prompt_embeds, prompt_embeds], axis=0) # (2, 77, 2048)
+ add_text_embeds = jnp.concatenate([negative_pooled_embeds, pooled_embeds], axis=0)
+ add_time_ids = jnp.concatenate([add_time_ids, add_time_ids], axis=0)
+
+ # Ensure model output will be `float32` before going into the scheduler
+ guidance_scale = jnp.array([guidance_scale], dtype=jnp.float32)
+
+ # Create random latents
+ latents_shape = (
+ batch_size,
+ self.unet.config.in_channels,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if latents is None:
+ latents = jax.random.normal(prng_seed, shape=latents_shape, dtype=jnp.float32)
+ else:
+ if latents.shape != latents_shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {latents_shape}")
+
+ # Prepare scheduler state
+ scheduler_state = self.scheduler.set_timesteps(
+ params["scheduler"], num_inference_steps=num_inference_steps, shape=latents.shape
+ )
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * scheduler_state.init_noise_sigma
+
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+
+ # Denoising loop
+ def loop_body(step, args):
+ latents, scheduler_state = args
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ latents_input = jnp.concatenate([latents] * 2)
+
+ t = jnp.array(scheduler_state.timesteps, dtype=jnp.int32)[step]
+ timestep = jnp.broadcast_to(t, latents_input.shape[0])
+
+ latents_input = self.scheduler.scale_model_input(scheduler_state, latents_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet.apply(
+ {"params": params["unet"]},
+ jnp.array(latents_input),
+ jnp.array(timestep, dtype=jnp.int32),
+ encoder_hidden_states=prompt_embeds,
+ added_cond_kwargs=added_cond_kwargs,
+ ).sample
+ # perform guidance
+ noise_pred_uncond, noise_prediction_text = jnp.split(noise_pred, 2, axis=0)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents, scheduler_state = self.scheduler.step(scheduler_state, noise_pred, t, latents).to_tuple()
+ return latents, scheduler_state
+
+ if DEBUG:
+ # run with python for loop
+ for i in range(num_inference_steps):
+ latents, scheduler_state = loop_body(i, (latents, scheduler_state))
+ else:
+ latents, _ = jax.lax.fori_loop(0, num_inference_steps, loop_body, (latents, scheduler_state))
+
+ if return_latents:
+ return latents
+
+ # Decode latents
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.apply({"params": params["vae"]}, latents, method=self.vae.decode).sample
+
+ image = (image / 2 + 0.5).clip(0, 1).transpose(0, 2, 3, 1)
+ return image
+
+
+# Static argnums are pipe, num_inference_steps, height, width, return_latents. A change would trigger recompilation.
+# Non-static args are (sharded) input tensors mapped over their first dimension (hence, `0`).
+@partial(
+ jax.pmap,
+ in_axes=(None, 0, 0, 0, None, None, None, 0, 0, 0, None),
+ static_broadcasted_argnums=(0, 4, 5, 6, 10),
+)
+def _p_generate(
+ pipe,
+ prompt_ids,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ return_latents,
+):
+ return pipe._generate(
+ prompt_ids,
+ params,
+ prng_seed,
+ num_inference_steps,
+ height,
+ width,
+ guidance_scale,
+ latents,
+ neg_prompt_ids,
+ return_latents,
+ )
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_output.py b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_output.py
new file mode 100644
index 0000000000000000000000000000000000000000..0783f44486ee1448bd15529f745af381ee7fa69f
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_output.py
@@ -0,0 +1,37 @@
+from dataclasses import dataclass
+from typing import List, Union
+
+import numpy as np
+import PIL.Image
+
+from ...utils import BaseOutput, is_flax_available
+
+
+@dataclass
+class StableDiffusionXLPipelineOutput(BaseOutput):
+ """
+ Output class for Stable Diffusion pipelines.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+ num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+if is_flax_available():
+ import flax
+
+ @flax.struct.dataclass
+ class FlaxStableDiffusionXLPipelineOutput(BaseOutput):
+ """
+ Output class for Flax Stable Diffusion XL pipelines.
+
+ Args:
+ images (`np.ndarray`)
+ Array of shape `(batch_size, height, width, num_channels)` with images from the diffusion pipeline.
+ """
+
+ images: np.ndarray
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py
new file mode 100644
index 0000000000000000000000000000000000000000..776696e9d4862011332fe6b36e2fd036abd6efaa
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py
@@ -0,0 +1,1266 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import (
+ FromSingleFileMixin,
+ IPAdapterMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ is_invisible_watermark_available,
+ is_torch_xla_available,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+ from .watermark import StableDiffusionXLWatermarker
+
+if is_torch_xla_available():
+ import torch_xla.core.xla_model as xm
+
+ XLA_AVAILABLE = True
+else:
+ XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionXLPipeline
+
+ >>> pipe = StableDiffusionXLPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> image = pipe(prompt).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionXLPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ FromSingleFileMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion XL.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+ specifically the
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+ variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ tokenizer_2 (`CLIPTokenizer`):
+ Second Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+ Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+ `stabilityai/stable-diffusion-xl-base-1-0`.
+ add_watermarker (`bool`, *optional*):
+ Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
+ watermark output images. If not defined, it will default to True if the package is installed, otherwise no
+ watermarker will be used.
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+ _optional_components = [
+ "tokenizer",
+ "tokenizer_2",
+ "text_encoder",
+ "text_encoder_2",
+ "image_encoder",
+ "feature_extractor",
+ ]
+ _callback_tensor_inputs = [
+ "latents",
+ "prompt_embeds",
+ "negative_prompt_embeds",
+ "add_text_embeds",
+ "add_time_ids",
+ "negative_pooled_prompt_embeds",
+ "negative_add_time_ids",
+ ]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ feature_extractor: CLIPImageProcessor = None,
+ force_zeros_for_empty_prompt: bool = True,
+ add_watermarker: Optional[bool] = None,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ scheduler=scheduler,
+ image_encoder=image_encoder,
+ feature_extractor=feature_extractor,
+ )
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ self.default_sample_size = self.unet.config.sample_size
+
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+ if add_watermarker:
+ self.watermark = StableDiffusionXLWatermarker()
+ else:
+ self.watermark = None
+
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ pooled_prompt_embeds=None,
+ negative_pooled_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+ )
+
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def _get_add_time_ids(
+ self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+ ):
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ return add_time_ids
+
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def guidance_rescale(self):
+ return self._guidance_rescale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def denoising_end(self):
+ return self._denoising_end
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @property
+ def interrupt(self):
+ return self._interrupt
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ denoising_end: Optional[float] = None,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ original_size: Optional[Tuple[int, int]] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Optional[Tuple[int, int]] = None,
+ negative_original_size: Optional[Tuple[int, int]] = None,
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+ negative_target_size: Optional[Tuple[int, int]] = None,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image. This is set to 1024 by default for the best results.
+ Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image. This is set to 1024 by default for the best results.
+ Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ denoising_end (`float`, *optional*):
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
+ still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+ scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+ "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+ guidance_scale (`float`, *optional*, defaults to 5.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+ of a plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ guidance_rescale (`float`, *optional*, defaults to 0.0):
+ Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+ [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+ Guidance rescale factor should fix overexposure when using zero terminal SNR.
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a target image resolution. It should be as same
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is a list with the generated images.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 0. Default height and width to unet
+ height = height or self.default_sample_size * self.vae_scale_factor
+ width = width or self.default_sample_size * self.vae_scale_factor
+
+ original_size = original_size or (height, width)
+ target_size = target_size or (height, width)
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._guidance_rescale = guidance_rescale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+ self._denoising_end = denoising_end
+ self._interrupt = False
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ prompt_2=prompt_2,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ negative_prompt_2=negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Prepare added time ids & embeddings
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ if negative_original_size is not None and negative_target_size is not None:
+ negative_add_time_ids = self._get_add_time_ids(
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ else:
+ negative_add_time_ids = add_time_ids
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 8. Denoising loop
+ num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+ # 8.1 Apply denoising_end
+ if (
+ self.denoising_end is not None
+ and isinstance(self.denoising_end, float)
+ and self.denoising_end > 0
+ and self.denoising_end < 1
+ ):
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+ )
+ )
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+ timesteps = timesteps[:num_inference_steps]
+
+ # 9. Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ if self.interrupt:
+ continue
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ added_cond_kwargs["image_embeds"] = image_embeds
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+ add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
+ negative_pooled_prompt_embeds = callback_outputs.pop(
+ "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+ )
+ add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+ negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if XLA_AVAILABLE:
+ xm.mark_step()
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ # unscale/denormalize the latents
+ # denormalize with the mean and std if available and not None
+ has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+ has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+ if has_latents_mean and has_latents_std:
+ latents_mean = (
+ torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents_std = (
+ torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+ else:
+ latents = latents / self.vae.config.scaling_factor
+
+ image = self.vae.decode(latents, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ image = latents
+
+ if not output_type == "latent":
+ # apply watermark if available
+ if self.watermark is not None:
+ image = self.watermark.apply_watermark(image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..fd4c412f48cbdd326d9900cea442934ae4111227
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_img2img.py
@@ -0,0 +1,1442 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import PIL.Image
+import torch
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import (
+ FromSingleFileMixin,
+ IPAdapterMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ is_invisible_watermark_available,
+ is_torch_xla_available,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+ from .watermark import StableDiffusionXLWatermarker
+
+if is_torch_xla_available():
+ import torch_xla.core.xla_model as xm
+
+ XLA_AVAILABLE = True
+else:
+ XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionXLImg2ImgPipeline
+ >>> from diffusers.utils import load_image
+
+ >>> pipe = StableDiffusionXLImg2ImgPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-xl-refiner-1.0", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+ >>> url = "https://huggingface.co/datasets/patrickvonplaten/images/resolve/main/aa_xl/000000009.png"
+
+ >>> init_image = load_image(url).convert("RGB")
+ >>> prompt = "a photo of an astronaut riding a horse on mars"
+ >>> image = pipe(prompt, image=init_image).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionXLImg2ImgPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ FromSingleFileMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ IPAdapterMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion XL.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+ specifically the
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+ variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ tokenizer_2 (`CLIPTokenizer`):
+ Second Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`):
+ Whether the `unet` requires an `aesthetic_score` condition to be passed during inference. Also see the
+ config of `stabilityai/stable-diffusion-xl-refiner-1-0`.
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+ Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+ `stabilityai/stable-diffusion-xl-base-1-0`.
+ add_watermarker (`bool`, *optional*):
+ Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
+ watermark output images. If not defined, it will default to True if the package is installed, otherwise no
+ watermarker will be used.
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+ _optional_components = [
+ "tokenizer",
+ "tokenizer_2",
+ "text_encoder",
+ "text_encoder_2",
+ "image_encoder",
+ "feature_extractor",
+ ]
+ _callback_tensor_inputs = [
+ "latents",
+ "prompt_embeds",
+ "negative_prompt_embeds",
+ "add_text_embeds",
+ "add_time_ids",
+ "negative_pooled_prompt_embeds",
+ "add_neg_time_ids",
+ ]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ feature_extractor: CLIPImageProcessor = None,
+ requires_aesthetics_score: bool = False,
+ force_zeros_for_empty_prompt: bool = True,
+ add_watermarker: Optional[bool] = None,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ image_encoder=image_encoder,
+ feature_extractor=feature_extractor,
+ scheduler=scheduler,
+ )
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+ if add_watermarker:
+ self.watermark = StableDiffusionXLWatermarker()
+ else:
+ self.watermark = None
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ strength,
+ num_inference_steps,
+ callback_steps,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+ if num_inference_steps is None:
+ raise ValueError("`num_inference_steps` cannot be None.")
+ elif not isinstance(num_inference_steps, int) or num_inference_steps <= 0:
+ raise ValueError(
+ f"`num_inference_steps` has to be a positive integer but is {num_inference_steps} of type"
+ f" {type(num_inference_steps)}."
+ )
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
+ # get the original timestep using init_timestep
+ if denoising_start is None:
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+ t_start = max(num_inference_steps - init_timestep, 0)
+ else:
+ t_start = 0
+
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+ # Strength is irrelevant if we directly request a timestep to start at;
+ # that is, strength is determined by the denoising_start instead.
+ if denoising_start is not None:
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (denoising_start * self.scheduler.config.num_train_timesteps)
+ )
+ )
+
+ num_inference_steps = (timesteps < discrete_timestep_cutoff).sum().item()
+ if self.scheduler.order == 2 and num_inference_steps % 2 == 0:
+ # if the scheduler is a 2nd order scheduler we might have to do +1
+ # because `num_inference_steps` might be even given that every timestep
+ # (except the highest one) is duplicated. If `num_inference_steps` is even it would
+ # mean that we cut the timesteps in the middle of the denoising step
+ # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1
+ # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
+ num_inference_steps = num_inference_steps + 1
+
+ # because t_n+1 >= t_n, we slice the timesteps starting from the end
+ timesteps = timesteps[-num_inference_steps:]
+ return timesteps, num_inference_steps
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(
+ self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None, add_noise=True
+ ):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ # Offload text encoder if `enable_model_cpu_offload` was enabled
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.text_encoder_2.to("cpu")
+ torch.cuda.empty_cache()
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ init_latents = image
+
+ else:
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ if self.vae.config.force_upcast:
+ image = image.float()
+ self.vae.to(dtype=torch.float32)
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ elif isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ if self.vae.config.force_upcast:
+ self.vae.to(dtype)
+
+ init_latents = init_latents.to(dtype)
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
+ # expand init_latents for batch_size
+ additional_image_per_prompt = batch_size // init_latents.shape[0]
+ init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ if add_noise:
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+
+ latents = init_latents
+
+ return latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ def _get_add_time_ids(
+ self,
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ aesthetic_score,
+ negative_aesthetic_score,
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype,
+ text_encoder_projection_dim=None,
+ ):
+ if self.config.requires_aesthetics_score:
+ add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,))
+ add_neg_time_ids = list(
+ negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,)
+ )
+ else:
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+ add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if (
+ expected_add_embed_dim > passed_add_embed_dim
+ and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim
+ ):
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model."
+ )
+ elif (
+ expected_add_embed_dim < passed_add_embed_dim
+ and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim
+ ):
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model."
+ )
+ elif expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype)
+
+ return add_time_ids, add_neg_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def guidance_rescale(self):
+ return self._guidance_rescale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def denoising_end(self):
+ return self._denoising_end
+
+ @property
+ def denoising_start(self):
+ return self._denoising_start
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @property
+ def interrupt(self):
+ return self._interrupt
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ image: PipelineImageInput = None,
+ strength: float = 0.3,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ denoising_start: Optional[float] = None,
+ denoising_end: Optional[float] = None,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ original_size: Tuple[int, int] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Tuple[int, int] = None,
+ negative_original_size: Optional[Tuple[int, int]] = None,
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+ negative_target_size: Optional[Tuple[int, int]] = None,
+ aesthetic_score: float = 6.0,
+ negative_aesthetic_score: float = 2.5,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ image (`torch.FloatTensor` or `PIL.Image.Image` or `np.ndarray` or `List[torch.FloatTensor]` or `List[PIL.Image.Image]` or `List[np.ndarray]`):
+ The image(s) to modify with the pipeline.
+ strength (`float`, *optional*, defaults to 0.3):
+ Conceptually, indicates how much to transform the reference `image`. Must be between 0 and 1. `image`
+ will be used as a starting point, adding more noise to it the larger the `strength`. The number of
+ denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will
+ be maximum and the denoising process will run for the full number of iterations specified in
+ `num_inference_steps`. A value of 1, therefore, essentially ignores `image`. Note that in the case of
+ `denoising_start` being declared as an integer, the value of `strength` will be ignored.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ denoising_start (`float`, *optional*):
+ When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be
+ bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and
+ it is assumed that the passed `image` is a partly denoised image. Note that when this is specified,
+ strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline
+ is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refine Image
+ Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality).
+ denoising_end (`float`, *optional*):
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
+ still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be
+ denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the
+ final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline
+ forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refine Image
+ Quality**](https://huggingface.co/docs/diffusers/using-diffusers/sdxl#refine-image-quality).
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ guidance_rescale (`float`, *optional*, defaults to 0.0):
+ Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+ [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+ Guidance rescale factor should fix overexposure when using zero terminal SNR.
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a target image resolution. It should be as same
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ aesthetic_score (`float`, *optional*, defaults to 6.0):
+ Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
+ simulate an aesthetic score of the generated image by influencing the negative text condition.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+ `tuple. When returning a tuple, the first element is a list with the generated images.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ strength,
+ num_inference_steps,
+ callback_steps,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._guidance_rescale = guidance_rescale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+ self._denoising_end = denoising_end
+ self._denoising_start = denoising_start
+ self._interrupt = False
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ prompt_2=prompt_2,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ negative_prompt_2=negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 4. Preprocess image
+ image = self.image_processor.preprocess(image)
+
+ # 5. Prepare timesteps
+ def denoising_value_valid(dnv):
+ return isinstance(dnv, float) and 0 < dnv < 1
+
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ timesteps, num_inference_steps = self.get_timesteps(
+ num_inference_steps,
+ strength,
+ device,
+ denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None,
+ )
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+ add_noise = True if self.denoising_start is None else False
+ # 6. Prepare latent variables
+ latents = self.prepare_latents(
+ image,
+ latent_timestep,
+ batch_size,
+ num_images_per_prompt,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ add_noise,
+ )
+ # 7. Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ height, width = latents.shape[-2:]
+ height = height * self.vae_scale_factor
+ width = width * self.vae_scale_factor
+
+ original_size = original_size or (height, width)
+ target_size = target_size or (height, width)
+
+ # 8. Prepare added time ids & embeddings
+ if negative_original_size is None:
+ negative_original_size = original_size
+ if negative_target_size is None:
+ negative_target_size = target_size
+
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids, add_neg_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ aesthetic_score,
+ negative_aesthetic_score,
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+ add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device)
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 9. Denoising loop
+ num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+ # 9.1 Apply denoising_end
+ if (
+ self.denoising_end is not None
+ and self.denoising_start is not None
+ and denoising_value_valid(self.denoising_end)
+ and denoising_value_valid(self.denoising_start)
+ and self.denoising_start >= self.denoising_end
+ ):
+ raise ValueError(
+ f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: "
+ + f" {self.denoising_end} when using type float."
+ )
+ elif self.denoising_end is not None and denoising_value_valid(self.denoising_end):
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+ )
+ )
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+ timesteps = timesteps[:num_inference_steps]
+
+ # 9.2 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ if self.interrupt:
+ continue
+
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ added_cond_kwargs["image_embeds"] = image_embeds
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+ add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
+ negative_pooled_prompt_embeds = callback_outputs.pop(
+ "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+ )
+ add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+ add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if XLA_AVAILABLE:
+ xm.mark_step()
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ # unscale/denormalize the latents
+ # denormalize with the mean and std if available and not None
+ has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+ has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+ if has_latents_mean and has_latents_std:
+ latents_mean = (
+ torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents_std = (
+ torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+ else:
+ latents = latents / self.vae.config.scaling_factor
+
+ image = self.vae.decode(latents, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ image = latents
+
+ # apply watermark if available
+ if self.watermark is not None:
+ image = self.watermark.apply_watermark(image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..c25628c22c7b519d6bad8d0b7c8303a2ea6aed62
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_inpaint.py
@@ -0,0 +1,1812 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import (
+ FromSingleFileMixin,
+ IPAdapterMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ is_invisible_watermark_available,
+ is_torch_xla_available,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+ from .watermark import StableDiffusionXLWatermarker
+
+if is_torch_xla_available():
+ import torch_xla.core.xla_model as xm
+
+ XLA_AVAILABLE = True
+else:
+ XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionXLInpaintPipeline
+ >>> from diffusers.utils import load_image
+
+ >>> pipe = StableDiffusionXLInpaintPipeline.from_pretrained(
+ ... "stabilityai/stable-diffusion-xl-base-1.0",
+ ... torch_dtype=torch.float16,
+ ... variant="fp16",
+ ... use_safetensors=True,
+ ... )
+ >>> pipe.to("cuda")
+
+ >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+ >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+ >>> init_image = load_image(img_url).convert("RGB")
+ >>> mask_image = load_image(mask_url).convert("RGB")
+
+ >>> prompt = "A majestic tiger sitting on a bench"
+ >>> image = pipe(
+ ... prompt=prompt, image=init_image, mask_image=mask_image, num_inference_steps=50, strength=0.80
+ ... ).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+def mask_pil_to_torch(mask, height, width):
+ # preprocess mask
+ if isinstance(mask, (PIL.Image.Image, np.ndarray)):
+ mask = [mask]
+
+ if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
+ mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
+ mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+ mask = mask.astype(np.float32) / 255.0
+ elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+ mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+ mask = torch.from_numpy(mask)
+ return mask
+
+
+def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool = False):
+ """
+ Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
+ converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
+ ``image`` and ``1`` for the ``mask``.
+
+ The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+ binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+ Args:
+ image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+ ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+ mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+ ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+
+
+ Raises:
+ ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+ should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+ TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+ (ot the other way around).
+
+ Returns:
+ tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
+ dimensions: ``batch x channels x height x width``.
+ """
+
+ # checkpoint. TOD(Yiyi) - need to clean this up later
+ deprecation_message = "The prepare_mask_and_masked_image method is deprecated and will be removed in a future version. Please use VaeImageProcessor.preprocess instead"
+ deprecate(
+ "prepare_mask_and_masked_image",
+ "0.30.0",
+ deprecation_message,
+ )
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ if mask is None:
+ raise ValueError("`mask_image` input cannot be undefined.")
+
+ if isinstance(image, torch.Tensor):
+ if not isinstance(mask, torch.Tensor):
+ mask = mask_pil_to_torch(mask, height, width)
+
+ if image.ndim == 3:
+ image = image.unsqueeze(0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Single batched mask, no channel dim or single mask not batched but channel dim
+ if mask.shape[0] == 1:
+ mask = mask.unsqueeze(0)
+
+ # Batched masks no channel dim
+ else:
+ mask = mask.unsqueeze(1)
+
+ assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+ # assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+ assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+
+ # Check image is in [-1, 1]
+ # if image.min() < -1 or image.max() > 1:
+ # raise ValueError("Image should be in [-1, 1] range")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("Mask should be in [0, 1] range")
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ # Image as float32
+ image = image.to(dtype=torch.float32)
+ elif isinstance(mask, torch.Tensor):
+ raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+ else:
+ # preprocess image
+ if isinstance(image, (PIL.Image.Image, np.ndarray)):
+ image = [image]
+ if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+ # resize all images w.r.t passed height an width
+ image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
+ image = [np.array(i.convert("RGB"))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+ image = np.concatenate([i[None, :] for i in image], axis=0)
+
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ mask = mask_pil_to_torch(mask, height, width)
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ if image.shape[1] == 4:
+ # images are in latent space and thus can't
+ # be masked set masked_image to None
+ # we assume that the checkpoint is not an inpainting
+ # checkpoint. TOD(Yiyi) - need to clean this up later
+ masked_image = None
+ else:
+ masked_image = image * (mask < 0.5)
+
+ # n.b. ensure backwards compatibility as old function does not return image
+ if return_image:
+ return mask, masked_image, image
+
+ return mask, masked_image
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionXLInpaintPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ FromSingleFileMixin,
+ IPAdapterMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion XL.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+ specifically the
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+ variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ tokenizer_2 (`CLIPTokenizer`):
+ Second Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`):
+ Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config
+ of `stabilityai/stable-diffusion-xl-refiner-1-0`.
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+ Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+ `stabilityai/stable-diffusion-xl-base-1-0`.
+ add_watermarker (`bool`, *optional*):
+ Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
+ watermark output images. If not defined, it will default to True if the package is installed, otherwise no
+ watermarker will be used.
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+
+ _optional_components = [
+ "tokenizer",
+ "tokenizer_2",
+ "text_encoder",
+ "text_encoder_2",
+ "image_encoder",
+ "feature_extractor",
+ ]
+ _callback_tensor_inputs = [
+ "latents",
+ "prompt_embeds",
+ "negative_prompt_embeds",
+ "add_text_embeds",
+ "add_time_ids",
+ "negative_pooled_prompt_embeds",
+ "add_neg_time_ids",
+ "mask",
+ "masked_image_latents",
+ ]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ feature_extractor: CLIPImageProcessor = None,
+ requires_aesthetics_score: bool = False,
+ force_zeros_for_empty_prompt: bool = True,
+ add_watermarker: Optional[bool] = None,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ image_encoder=image_encoder,
+ feature_extractor=feature_extractor,
+ scheduler=scheduler,
+ )
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.register_to_config(requires_aesthetics_score=requires_aesthetics_score)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.mask_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_normalize=False, do_binarize=True, do_convert_grayscale=True
+ )
+
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+ if add_watermarker:
+ self.watermark = StableDiffusionXLWatermarker()
+ else:
+ self.watermark = None
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ image,
+ mask_image,
+ height,
+ width,
+ strength,
+ callback_steps,
+ output_type,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ padding_mask_crop=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+ if padding_mask_crop is not None:
+ if not isinstance(image, PIL.Image.Image):
+ raise ValueError(
+ f"The image should be a PIL image when inpainting mask crop, but is of type" f" {type(image)}."
+ )
+ if not isinstance(mask_image, PIL.Image.Image):
+ raise ValueError(
+ f"The mask image should be a PIL image when inpainting mask crop, but is of type"
+ f" {type(mask_image)}."
+ )
+ if output_type != "pil":
+ raise ValueError(f"The output type should be PIL when inpainting mask crop, but is" f" {output_type}.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def prepare_latents(
+ self,
+ batch_size,
+ num_channels_latents,
+ height,
+ width,
+ dtype,
+ device,
+ generator,
+ latents=None,
+ image=None,
+ timestep=None,
+ is_strength_max=True,
+ add_noise=True,
+ return_noise=False,
+ return_image_latents=False,
+ ):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if (image is None or timestep is None) and not is_strength_max:
+ raise ValueError(
+ "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
+ "However, either the image or the noise timestep has not been provided."
+ )
+
+ if image.shape[1] == 4:
+ image_latents = image.to(device=device, dtype=dtype)
+ image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
+ elif return_image_latents or (latents is None and not is_strength_max):
+ image = image.to(device=device, dtype=dtype)
+ image_latents = self._encode_vae_image(image=image, generator=generator)
+ image_latents = image_latents.repeat(batch_size // image_latents.shape[0], 1, 1, 1)
+
+ if latents is None and add_noise:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ # if strength is 1. then initialise the latents to noise, else initial to image + noise
+ latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
+ # if pure noise then scale the initial latents by the Scheduler's init sigma
+ latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+ elif add_noise:
+ noise = latents.to(device)
+ latents = noise * self.scheduler.init_noise_sigma
+ else:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ latents = image_latents.to(device)
+
+ outputs = (latents,)
+
+ if return_noise:
+ outputs += (noise,)
+
+ if return_image_latents:
+ outputs += (image_latents,)
+
+ return outputs
+
+ def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+ dtype = image.dtype
+ if self.vae.config.force_upcast:
+ image = image.float()
+ self.vae.to(dtype=torch.float32)
+
+ if isinstance(generator, list):
+ image_latents = [
+ retrieve_latents(self.vae.encode(image[i : i + 1]), generator=generator[i])
+ for i in range(image.shape[0])
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = retrieve_latents(self.vae.encode(image), generator=generator)
+
+ if self.vae.config.force_upcast:
+ self.vae.to(dtype)
+
+ image_latents = image_latents.to(dtype)
+ image_latents = self.vae.config.scaling_factor * image_latents
+
+ return image_latents
+
+ def prepare_mask_latents(
+ self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+ ):
+ # resize the mask to latents shape as we concatenate the mask to the latents
+ # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+ # and half precision
+ mask = torch.nn.functional.interpolate(
+ mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+ )
+ mask = mask.to(device=device, dtype=dtype)
+
+ # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+ if mask.shape[0] < batch_size:
+ if not batch_size % mask.shape[0] == 0:
+ raise ValueError(
+ "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+ f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+ " of masks that you pass is divisible by the total requested batch size."
+ )
+ mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+
+ mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+
+ if masked_image is not None and masked_image.shape[1] == 4:
+ masked_image_latents = masked_image
+ else:
+ masked_image_latents = None
+
+ if masked_image is not None:
+ if masked_image_latents is None:
+ masked_image = masked_image.to(device=device, dtype=dtype)
+ masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+
+ if masked_image_latents.shape[0] < batch_size:
+ if not batch_size % masked_image_latents.shape[0] == 0:
+ raise ValueError(
+ "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+ f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+ " Make sure the number of images that you pass is divisible by the total requested batch size."
+ )
+ masked_image_latents = masked_image_latents.repeat(
+ batch_size // masked_image_latents.shape[0], 1, 1, 1
+ )
+
+ masked_image_latents = (
+ torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+ )
+
+ # aligning device to prevent device errors when concating it with the latent model input
+ masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+
+ return mask, masked_image_latents
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None):
+ # get the original timestep using init_timestep
+ if denoising_start is None:
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+ t_start = max(num_inference_steps - init_timestep, 0)
+ else:
+ t_start = 0
+
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+ # Strength is irrelevant if we directly request a timestep to start at;
+ # that is, strength is determined by the denoising_start instead.
+ if denoising_start is not None:
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (denoising_start * self.scheduler.config.num_train_timesteps)
+ )
+ )
+
+ num_inference_steps = (timesteps < discrete_timestep_cutoff).sum().item()
+ if self.scheduler.order == 2 and num_inference_steps % 2 == 0:
+ # if the scheduler is a 2nd order scheduler we might have to do +1
+ # because `num_inference_steps` might be even given that every timestep
+ # (except the highest one) is duplicated. If `num_inference_steps` is even it would
+ # mean that we cut the timesteps in the middle of the denoising step
+ # (between 1st and 2nd devirative) which leads to incorrect results. By adding 1
+ # we ensure that the denoising process always ends after the 2nd derivate step of the scheduler
+ num_inference_steps = num_inference_steps + 1
+
+ # because t_n+1 >= t_n, we slice the timesteps starting from the end
+ timesteps = timesteps[-num_inference_steps:]
+ return timesteps, num_inference_steps
+
+ return timesteps, num_inference_steps - t_start
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl_img2img.StableDiffusionXLImg2ImgPipeline._get_add_time_ids
+ def _get_add_time_ids(
+ self,
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ aesthetic_score,
+ negative_aesthetic_score,
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype,
+ text_encoder_projection_dim=None,
+ ):
+ if self.config.requires_aesthetics_score:
+ add_time_ids = list(original_size + crops_coords_top_left + (aesthetic_score,))
+ add_neg_time_ids = list(
+ negative_original_size + negative_crops_coords_top_left + (negative_aesthetic_score,)
+ )
+ else:
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+ add_neg_time_ids = list(negative_original_size + crops_coords_top_left + negative_target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if (
+ expected_add_embed_dim > passed_add_embed_dim
+ and (expected_add_embed_dim - passed_add_embed_dim) == self.unet.config.addition_time_embed_dim
+ ):
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to enable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=True)` to make sure `aesthetic_score` {aesthetic_score} and `negative_aesthetic_score` {negative_aesthetic_score} is correctly used by the model."
+ )
+ elif (
+ expected_add_embed_dim < passed_add_embed_dim
+ and (passed_add_embed_dim - expected_add_embed_dim) == self.unet.config.addition_time_embed_dim
+ ):
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. Please make sure to disable `requires_aesthetics_score` with `pipe.register_to_config(requires_aesthetics_score=False)` to make sure `target_size` {target_size} is correctly used by the model."
+ )
+ elif expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ add_neg_time_ids = torch.tensor([add_neg_time_ids], dtype=dtype)
+
+ return add_time_ids, add_neg_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def guidance_rescale(self):
+ return self._guidance_rescale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def denoising_end(self):
+ return self._denoising_end
+
+ @property
+ def denoising_start(self):
+ return self._denoising_start
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @property
+ def interrupt(self):
+ return self._interrupt
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ image: PipelineImageInput = None,
+ mask_image: PipelineImageInput = None,
+ masked_image_latents: torch.FloatTensor = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ padding_mask_crop: Optional[int] = None,
+ strength: float = 0.9999,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ denoising_start: Optional[float] = None,
+ denoising_end: Optional[float] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ original_size: Tuple[int, int] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Tuple[int, int] = None,
+ negative_original_size: Optional[Tuple[int, int]] = None,
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+ negative_target_size: Optional[Tuple[int, int]] = None,
+ aesthetic_score: float = 6.0,
+ negative_aesthetic_score: float = 2.5,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will
+ be masked out with `mask_image` and repainted according to `prompt`.
+ mask_image (`PIL.Image.Image`):
+ `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be
+ repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted
+ to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L)
+ instead of 3, so the expected shape would be `(B, H, W, 1)`.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image. This is set to 1024 by default for the best results.
+ Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image. This is set to 1024 by default for the best results.
+ Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ padding_mask_crop (`int`, *optional*, defaults to `None`):
+ The size of margin in the crop to be applied to the image and masking. If `None`, no crop is applied to image and mask_image. If
+ `padding_mask_crop` is not `None`, it will first find a rectangular region with the same aspect ration of the image and
+ contains all masked area, and then expand that area based on `padding_mask_crop`. The image and mask_image will then be cropped based on
+ the expanded area before resizing to the original image size for inpainting. This is useful when the masked area is small while the image is large
+ and contain information inreleant for inpainging, such as background.
+ strength (`float`, *optional*, defaults to 0.9999):
+ Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be
+ between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the
+ `strength`. The number of denoising steps depends on the amount of noise initially added. When
+ `strength` is 1, added noise will be maximum and the denoising process will run for the full number of
+ iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked
+ portion of the reference `image`. Note that in the case of `denoising_start` being declared as an
+ integer, the value of `strength` will be ignored.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ denoising_start (`float`, *optional*):
+ When specified, indicates the fraction (between 0.0 and 1.0) of the total denoising process to be
+ bypassed before it is initiated. Consequently, the initial part of the denoising process is skipped and
+ it is assumed that the passed `image` is a partly denoised image. Note that when this is specified,
+ strength will be ignored. The `denoising_start` parameter is particularly beneficial when this pipeline
+ is integrated into a "Mixture of Denoisers" multi-pipeline setup, as detailed in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
+ denoising_end (`float`, *optional*):
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
+ still retain a substantial amount of noise (ca. final 20% of timesteps still needed) and should be
+ denoised by a successor pipeline that has `denoising_start` set to 0.8 so that it only denoises the
+ final 20% of the scheduler. The denoising_end parameter should ideally be utilized when this pipeline
+ forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output).
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a target image resolution. It should be as same
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ aesthetic_score (`float`, *optional*, defaults to 6.0):
+ Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
+ simulate an aesthetic score of the generated image by influencing the negative text condition.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+ `tuple. `tuple. When returning a tuple, the first element is a list with the generated images.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ image,
+ mask_image,
+ height,
+ width,
+ strength,
+ callback_steps,
+ output_type,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs,
+ padding_mask_crop,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._guidance_rescale = guidance_rescale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+ self._denoising_end = denoising_end
+ self._denoising_start = denoising_start
+ self._interrupt = False
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ prompt_2=prompt_2,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ negative_prompt_2=negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+
+ # 4. set timesteps
+ def denoising_value_valid(dnv):
+ return isinstance(dnv, float) and 0 < dnv < 1
+
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ timesteps, num_inference_steps = self.get_timesteps(
+ num_inference_steps,
+ strength,
+ device,
+ denoising_start=self.denoising_start if denoising_value_valid(self.denoising_start) else None,
+ )
+ # check that number of inference steps is not < 1 - as this doesn't make sense
+ if num_inference_steps < 1:
+ raise ValueError(
+ f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+ f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+ )
+ # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+ is_strength_max = strength == 1.0
+
+ # 5. Preprocess mask and image
+ if padding_mask_crop is not None:
+ crops_coords = self.mask_processor.get_crop_region(mask_image, width, height, pad=padding_mask_crop)
+ resize_mode = "fill"
+ else:
+ crops_coords = None
+ resize_mode = "default"
+
+ original_image = image
+ init_image = self.image_processor.preprocess(
+ image, height=height, width=width, crops_coords=crops_coords, resize_mode=resize_mode
+ )
+ init_image = init_image.to(dtype=torch.float32)
+
+ mask = self.mask_processor.preprocess(
+ mask_image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
+ )
+
+ if masked_image_latents is not None:
+ masked_image = masked_image_latents
+ elif init_image.shape[1] == 4:
+ # if images are in latent space, we can't mask it
+ masked_image = None
+ else:
+ masked_image = init_image * (mask < 0.5)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ num_channels_unet = self.unet.config.in_channels
+ return_image_latents = num_channels_unet == 4
+
+ add_noise = True if self.denoising_start is None else False
+ latents_outputs = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ image=init_image,
+ timestep=latent_timestep,
+ is_strength_max=is_strength_max,
+ add_noise=add_noise,
+ return_noise=True,
+ return_image_latents=return_image_latents,
+ )
+
+ if return_image_latents:
+ latents, noise, image_latents = latents_outputs
+ else:
+ latents, noise = latents_outputs
+
+ # 7. Prepare mask latent variables
+ mask, masked_image_latents = self.prepare_mask_latents(
+ mask,
+ masked_image,
+ batch_size * num_images_per_prompt,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ self.do_classifier_free_guidance,
+ )
+
+ # 8. Check that sizes of mask, masked image and latents match
+ if num_channels_unet == 9:
+ # default case for runwayml/stable-diffusion-inpainting
+ num_channels_mask = mask.shape[1]
+ num_channels_masked_image = masked_image_latents.shape[1]
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+ elif num_channels_unet != 4:
+ raise ValueError(
+ f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
+ )
+ # 8.1 Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ height, width = latents.shape[-2:]
+ height = height * self.vae_scale_factor
+ width = width * self.vae_scale_factor
+
+ original_size = original_size or (height, width)
+ target_size = target_size or (height, width)
+
+ # 10. Prepare added time ids & embeddings
+ if negative_original_size is None:
+ negative_original_size = original_size
+ if negative_target_size is None:
+ negative_target_size = target_size
+
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids, add_neg_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ aesthetic_score,
+ negative_aesthetic_score,
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ add_time_ids = add_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_neg_time_ids = add_neg_time_ids.repeat(batch_size * num_images_per_prompt, 1)
+ add_time_ids = torch.cat([add_neg_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device)
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 11. Denoising loop
+ num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+
+ if (
+ self.denoising_end is not None
+ and self.denoising_start is not None
+ and denoising_value_valid(self.denoising_end)
+ and denoising_value_valid(self.denoising_start)
+ and self.denoising_start >= self.denoising_end
+ ):
+ raise ValueError(
+ f"`denoising_start`: {self.denoising_start} cannot be larger than or equal to `denoising_end`: "
+ + f" {self.denoising_end} when using type float."
+ )
+ elif self.denoising_end is not None and denoising_value_valid(self.denoising_end):
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (self.denoising_end * self.scheduler.config.num_train_timesteps)
+ )
+ )
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+ timesteps = timesteps[:num_inference_steps]
+
+ # 11.1 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ if self.interrupt:
+ continue
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ if num_channels_unet == 9:
+ latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+ # predict the noise residual
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ added_cond_kwargs["image_embeds"] = image_embeds
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if num_channels_unet == 4:
+ init_latents_proper = image_latents
+ if self.do_classifier_free_guidance:
+ init_mask, _ = mask.chunk(2)
+ else:
+ init_mask = mask
+
+ if i < len(timesteps) - 1:
+ noise_timestep = timesteps[i + 1]
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_proper, noise, torch.tensor([noise_timestep])
+ )
+
+ latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+ add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
+ negative_pooled_prompt_embeds = callback_outputs.pop(
+ "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
+ )
+ add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
+ add_neg_time_ids = callback_outputs.pop("add_neg_time_ids", add_neg_time_ids)
+ mask = callback_outputs.pop("mask", mask)
+ masked_image_latents = callback_outputs.pop("masked_image_latents", masked_image_latents)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if XLA_AVAILABLE:
+ xm.mark_step()
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ # unscale/denormalize the latents
+ # denormalize with the mean and std if available and not None
+ has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+ has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+ if has_latents_mean and has_latents_std:
+ latents_mean = (
+ torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents_std = (
+ torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+ else:
+ latents = latents / self.vae.config.scaling_factor
+
+ image = self.vae.decode(latents, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ return StableDiffusionXLPipelineOutput(images=latents)
+
+ # apply watermark if available
+ if self.watermark is not None:
+ image = self.watermark.apply_watermark(image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ if padding_mask_crop is not None:
+ image = [self.image_processor.apply_overlay(mask_image, original_image, i, crops_coords) for i in image]
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py
new file mode 100644
index 0000000000000000000000000000000000000000..8bf07d2f808cbfe3823ef09b9bb4e2f7fc15a03c
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl_instruct_pix2pix.py
@@ -0,0 +1,996 @@
+# Copyright 2024 Harutatsu Akiyama and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import PIL.Image
+import torch
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ is_invisible_watermark_available,
+ is_torch_xla_available,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import StableDiffusionXLPipelineOutput
+
+
+if is_invisible_watermark_available():
+ from .watermark import StableDiffusionXLWatermarker
+
+if is_torch_xla_available():
+ import torch_xla.core.xla_model as xm
+
+ XLA_AVAILABLE = True
+else:
+ XLA_AVAILABLE = False
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import StableDiffusionXLInstructPix2PixPipeline
+ >>> from diffusers.utils import load_image
+
+ >>> resolution = 768
+ >>> image = load_image(
+ ... "https://hf.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png"
+ ... ).resize((resolution, resolution))
+ >>> edit_instruction = "Turn sky into a cloudy one"
+
+ >>> pipe = StableDiffusionXLInstructPix2PixPipeline.from_pretrained(
+ ... "diffusers/sdxl-instructpix2pix-768", torch_dtype=torch.float16
+ ... ).to("cuda")
+
+ >>> edited_image = pipe(
+ ... prompt=edit_instruction,
+ ... image=image,
+ ... height=resolution,
+ ... width=resolution,
+ ... guidance_scale=3.0,
+ ... image_guidance_scale=1.5,
+ ... num_inference_steps=30,
+ ... ).images[0]
+ >>> edited_image
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+class StableDiffusionXLInstructPix2PixPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ FromSingleFileMixin,
+ StableDiffusionXLLoraLoaderMixin,
+):
+ r"""
+ Pipeline for pixel-level image editing by following text instructions. Based on Stable Diffusion XL.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+ specifically the
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+ variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ tokenizer_2 (`CLIPTokenizer`):
+ Second Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ requires_aesthetics_score (`bool`, *optional*, defaults to `"False"`):
+ Whether the `unet` requires a aesthetic_score condition to be passed during inference. Also see the config
+ of `stabilityai/stable-diffusion-xl-refiner-1-0`.
+ force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
+ Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
+ `stabilityai/stable-diffusion-xl-base-1-0`.
+ add_watermarker (`bool`, *optional*):
+ Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
+ watermark output images. If not defined, it will default to True if the package is installed, otherwise no
+ watermarker will be used.
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
+ _optional_components = ["tokenizer", "tokenizer_2", "text_encoder", "text_encoder_2"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ force_zeros_for_empty_prompt: bool = True,
+ add_watermarker: Optional[bool] = None,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ scheduler=scheduler,
+ )
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.default_sample_size = self.unet.config.sample_size
+
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+ if add_watermarker:
+ self.watermark = StableDiffusionXLWatermarker()
+ else:
+ self.watermark = None
+
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(
+ text_input_ids.to(device),
+ output_hidden_states=True,
+ )
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ prompt_embeds_dtype = self.text_encoder_2.dtype if self.text_encoder_2 is not None else self.unet.dtype
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_instruct_pix2pix.StableDiffusionInstructPix2PixPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def prepare_image_latents(
+ self, image, batch_size, num_images_per_prompt, dtype, device, do_classifier_free_guidance, generator=None
+ ):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_images_per_prompt
+
+ if image.shape[1] == 4:
+ image_latents = image
+ else:
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+ if needs_upcasting:
+ self.upcast_vae()
+ image = image.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax")
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+
+ if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+ # expand image_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // image_latents.shape[0]
+ image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ image_latents = torch.cat([image_latents], dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_image_latents = torch.zeros_like(image_latents)
+ image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0)
+
+ if image_latents.dtype != self.vae.dtype:
+ image_latents = image_latents.to(dtype=self.vae.dtype)
+
+ return image_latents
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
+ def _get_add_time_ids(
+ self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+ ):
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ return add_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ image: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 100,
+ denoising_end: Optional[float] = None,
+ guidance_scale: float = 5.0,
+ image_guidance_scale: float = 1.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ original_size: Tuple[int, int] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Tuple[int, int] = None,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ image (`torch.FloatTensor` or `PIL.Image.Image` or `np.ndarray` or `List[torch.FloatTensor]` or `List[PIL.Image.Image]` or `List[np.ndarray]`):
+ The image(s) to modify with the pipeline.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ denoising_end (`float`, *optional*):
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
+ still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+ scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+ "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+ guidance_scale (`float`, *optional*, defaults to 5.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ image_guidance_scale (`float`, *optional*, defaults to 1.5):
+ Image guidance scale is to push the generated image towards the inital image `image`. Image guidance
+ scale is enabled by setting `image_guidance_scale > 1`. Higher image guidance scale encourages to
+ generate images that are closely linked to the source image `image`, usually at the expense of lower
+ image quality. This pipeline requires a value of at least `1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionXLPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ guidance_rescale (`float`, *optional*, defaults to 0.0):
+ Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+ [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+ Guidance rescale factor should fix overexposure when using zero terminal SNR.
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ aesthetic_score (`float`, *optional*, defaults to 6.0):
+ Used to simulate an aesthetic score of the generated image by influencing the positive text condition.
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_aesthetic_score (`float`, *optional*, defaults to 2.5):
+ Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). Can be used to
+ simulate an aesthetic score of the generated image by influencing the negative text condition.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is a list with the generated images.
+ """
+ # 0. Default height and width to unet
+ height = height or self.default_sample_size * self.vae_scale_factor
+ width = width or self.default_sample_size * self.vae_scale_factor
+
+ original_size = original_size or (height, width)
+ target_size = target_size or (height, width)
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(prompt, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0 and image_guidance_scale >= 1.0
+ # check if scheduler is in sigmas space
+ scheduler_is_in_sigma_space = hasattr(self.scheduler, "sigmas")
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ prompt_2=prompt_2,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ negative_prompt_2=negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ )
+
+ # 4. Preprocess image
+ image = self.image_processor.preprocess(image, height=height, width=width).to(device)
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 6. Prepare Image latents
+ image_latents = self.prepare_image_latents(
+ image,
+ batch_size,
+ num_images_per_prompt,
+ prompt_embeds.dtype,
+ device,
+ do_classifier_free_guidance,
+ )
+
+ # 7. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 8. Check that shapes of latents and image match the UNet channels
+ num_channels_image = image_latents.shape[1]
+ if num_channels_latents + num_channels_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_image`: {num_channels_image} "
+ f" = {num_channels_latents + num_channels_image}. Please verify the config of"
+ " `pipeline.unet` or your `image` input."
+ )
+
+ # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 10. Prepare added time ids & embeddings
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+
+ if do_classifier_free_guidance:
+ # The extra concat similar to how it's done in SD InstructPix2Pix.
+ prompt_embeds = torch.cat([prompt_embeds, negative_prompt_embeds, negative_prompt_embeds], dim=0)
+ add_text_embeds = torch.cat(
+ [add_text_embeds, negative_pooled_prompt_embeds, negative_pooled_prompt_embeds], dim=0
+ )
+ add_time_ids = torch.cat([add_time_ids, add_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+ # 11. Denoising loop
+ num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+ if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (denoising_end * self.scheduler.config.num_train_timesteps)
+ )
+ )
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+ timesteps = timesteps[:num_inference_steps]
+
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Expand the latents if we are doing classifier free guidance.
+ # The latents are expanded 3 times because for pix2pix the guidance
+ # is applied for both the text and the input image.
+ latent_model_input = torch.cat([latents] * 3) if do_classifier_free_guidance else latents
+
+ # concat latents, image_latents in the channel dimension
+ scaled_latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+ scaled_latent_model_input = torch.cat([scaled_latent_model_input, image_latents], dim=1)
+
+ # predict the noise residual
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+ noise_pred = self.unet(
+ scaled_latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # Hack:
+ # For karras style schedulers the model does classifer free guidance using the
+ # predicted_original_sample instead of the noise_pred. So we need to compute the
+ # predicted_original_sample here if we are using a karras style scheduler.
+ if scheduler_is_in_sigma_space:
+ step_index = (self.scheduler.timesteps == t).nonzero()[0].item()
+ sigma = self.scheduler.sigmas[step_index]
+ noise_pred = latent_model_input - sigma * noise_pred
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_text, noise_pred_image, noise_pred_uncond = noise_pred.chunk(3)
+ noise_pred = (
+ noise_pred_uncond
+ + guidance_scale * (noise_pred_text - noise_pred_image)
+ + image_guidance_scale * (noise_pred_image - noise_pred_uncond)
+ )
+
+ if do_classifier_free_guidance and guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+ # Hack:
+ # For karras style schedulers the model does classifer free guidance using the
+ # predicted_original_sample instead of the noise_pred. But the scheduler.step function
+ # expects the noise_pred and computes the predicted_original_sample internally. So we
+ # need to overwrite the noise_pred here such that the value of the computed
+ # predicted_original_sample is correct.
+ if scheduler_is_in_sigma_space:
+ noise_pred = (noise_pred - latents) / (-sigma)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if XLA_AVAILABLE:
+ xm.mark_step()
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ # unscale/denormalize the latents
+ # denormalize with the mean and std if available and not None
+ has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None
+ has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None
+ if has_latents_mean and has_latents_std:
+ latents_mean = (
+ torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents_std = (
+ torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype)
+ )
+ latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean
+ else:
+ latents = latents / self.vae.config.scaling_factor
+
+ image = self.vae.decode(latents, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ return StableDiffusionXLPipelineOutput(images=latents)
+
+ # apply watermark if available
+ if self.watermark is not None:
+ image = self.watermark.apply_watermark(image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/stable_diffusion_xl/watermark.py b/model/diffusers_c/pipelines/stable_diffusion_xl/watermark.py
new file mode 100644
index 0000000000000000000000000000000000000000..5b6e36d9f44756da494cee0b996b1871721872e7
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_diffusion_xl/watermark.py
@@ -0,0 +1,36 @@
+import numpy as np
+import torch
+
+from ...utils import is_invisible_watermark_available
+
+
+if is_invisible_watermark_available():
+ from imwatermark import WatermarkEncoder
+
+
+# Copied from https://github.com/Stability-AI/generative-models/blob/613af104c6b85184091d42d374fef420eddb356d/scripts/demo/streamlit_helpers.py#L66
+WATERMARK_MESSAGE = 0b101100111110110010010000011110111011000110011110
+# bin(x)[2:] gives bits of x as str, use int to convert them to 0/1
+WATERMARK_BITS = [int(bit) for bit in bin(WATERMARK_MESSAGE)[2:]]
+
+
+class StableDiffusionXLWatermarker:
+ def __init__(self):
+ self.watermark = WATERMARK_BITS
+ self.encoder = WatermarkEncoder()
+
+ self.encoder.set_watermark("bits", self.watermark)
+
+ def apply_watermark(self, images: torch.FloatTensor):
+ # can't encode images that are smaller than 256
+ if images.shape[-1] < 256:
+ return images
+
+ images = (255 * (images / 2 + 0.5)).cpu().permute(0, 2, 3, 1).float().numpy()
+
+ images = [self.encoder.encode(image, "dwtDct") for image in images]
+
+ images = torch.from_numpy(np.array(images)).permute(0, 3, 1, 2)
+
+ images = torch.clamp(2 * (images / 255 - 0.5), min=-1.0, max=1.0)
+ return images
diff --git a/model/diffusers_c/pipelines/stable_video_diffusion/__init__.py b/model/diffusers_c/pipelines/stable_video_diffusion/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..3bd4dc78966e217d85769691b98ed8fb0b6ac05c
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_video_diffusion/__init__.py
@@ -0,0 +1,58 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ BaseOutput,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure.update(
+ {
+ "pipeline_stable_video_diffusion": [
+ "StableVideoDiffusionPipeline",
+ "StableVideoDiffusionPipelineOutput",
+ ],
+ }
+ )
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import *
+ else:
+ from .pipeline_stable_video_diffusion import (
+ StableVideoDiffusionPipeline,
+ StableVideoDiffusionPipelineOutput,
+ )
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/stable_video_diffusion/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/stable_video_diffusion/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..966cb2c15095433c87e07b54615095c70bf68349
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_video_diffusion/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_video_diffusion/__pycache__/pipeline_stable_video_diffusion.cpython-39.pyc b/model/diffusers_c/pipelines/stable_video_diffusion/__pycache__/pipeline_stable_video_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c164c9bbe08e89af00eb096210f01083c8d2fe56
Binary files /dev/null and b/model/diffusers_c/pipelines/stable_video_diffusion/__pycache__/pipeline_stable_video_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py b/model/diffusers_c/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..1342fe429145497ffca34b341de064e22fd9b40c
--- /dev/null
+++ b/model/diffusers_c/pipelines/stable_video_diffusion/pipeline_stable_video_diffusion.py
@@ -0,0 +1,673 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...models import AutoencoderKLTemporalDecoder, UNetSpatioTemporalConditionModel
+from ...schedulers import EulerDiscreteScheduler
+from ...utils import BaseOutput, logging, replace_example_docstring
+from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusers import StableVideoDiffusionPipeline
+ >>> from diffusers.utils import load_image, export_to_video
+
+ >>> pipe = StableVideoDiffusionPipeline.from_pretrained("stabilityai/stable-video-diffusion-img2vid-xt", torch_dtype=torch.float16, variant="fp16")
+ >>> pipe.to("cuda")
+
+ >>> image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd-docstring-example.jpeg")
+ >>> image = image.resize((1024, 576))
+
+ >>> frames = pipe(image, num_frames=25, decode_chunk_size=8).frames[0]
+ >>> export_to_video(frames, "generated.mp4", fps=7)
+ ```
+"""
+
+
+def _append_dims(x, target_dims):
+ """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+ dims_to_append = target_dims - x.ndim
+ if dims_to_append < 0:
+ raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
+ return x[(...,) + (None,) * dims_to_append]
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor: VaeImageProcessor, output_type: str = "np"):
+ batch_size, channels, num_frames, height, width = video.shape
+ outputs = []
+ for batch_idx in range(batch_size):
+ batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+ batch_output = processor.postprocess(batch_vid, output_type)
+
+ outputs.append(batch_output)
+
+ if output_type == "np":
+ outputs = np.stack(outputs)
+
+ elif output_type == "pt":
+ outputs = torch.stack(outputs)
+
+ elif not output_type == "pil":
+ raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+ return outputs
+
+
+@dataclass
+class StableVideoDiffusionPipelineOutput(BaseOutput):
+ r"""
+ Output class for Stable Video Diffusion pipeline.
+
+ Args:
+ frames (`[List[List[PIL.Image.Image]]`, `np.ndarray`, `torch.FloatTensor`]):
+ List of denoised PIL images of length `batch_size` or numpy array or torch tensor
+ of shape `(batch_size, num_frames, height, width, num_channels)`.
+ """
+
+ frames: Union[List[List[PIL.Image.Image]], np.ndarray, torch.FloatTensor]
+
+
+class StableVideoDiffusionPipeline(DiffusionPipeline):
+ r"""
+ Pipeline to generate video from an input image using Stable Video Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKLTemporalDecoder`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ image_encoder ([`~transformers.CLIPVisionModelWithProjection`]):
+ Frozen CLIP image-encoder ([laion/CLIP-ViT-H-14-laion2B-s32B-b79K](https://huggingface.co/laion/CLIP-ViT-H-14-laion2B-s32B-b79K)).
+ unet ([`UNetSpatioTemporalConditionModel`]):
+ A `UNetSpatioTemporalConditionModel` to denoise the encoded image latents.
+ scheduler ([`EulerDiscreteScheduler`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images.
+ """
+
+ model_cpu_offload_seq = "image_encoder->unet->vae"
+ _callback_tensor_inputs = ["latents"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKLTemporalDecoder,
+ image_encoder: CLIPVisionModelWithProjection,
+ unet: UNetSpatioTemporalConditionModel,
+ scheduler: EulerDiscreteScheduler,
+ feature_extractor: CLIPImageProcessor,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ image_encoder=image_encoder,
+ unet=unet,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ def _encode_image(
+ self,
+ image: PipelineImageInput,
+ device: Union[str, torch.device],
+ num_videos_per_prompt: int,
+ do_classifier_free_guidance: bool,
+ ) -> torch.FloatTensor:
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.image_processor.pil_to_numpy(image)
+ image = self.image_processor.numpy_to_pt(image)
+
+ # We normalize the image before resizing to match with the original implementation.
+ # Then we unnormalize it after resizing.
+ image = image * 2.0 - 1.0
+ image = _resize_with_antialiasing(image, (224, 224))
+ image = (image + 1.0) / 2.0
+
+ # Normalize the image with for CLIP input
+ image = self.feature_extractor(
+ images=image,
+ do_normalize=True,
+ do_center_crop=False,
+ do_resize=False,
+ do_rescale=False,
+ return_tensors="pt",
+ ).pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ image_embeddings = self.image_encoder(image).image_embeds
+ image_embeddings = image_embeddings.unsqueeze(1)
+
+ # duplicate image embeddings for each generation per prompt, using mps friendly method
+ bs_embed, seq_len, _ = image_embeddings.shape
+ image_embeddings = image_embeddings.repeat(1, num_videos_per_prompt, 1)
+ image_embeddings = image_embeddings.view(bs_embed * num_videos_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ negative_image_embeddings = torch.zeros_like(image_embeddings)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ image_embeddings = torch.cat([negative_image_embeddings, image_embeddings])
+
+ return image_embeddings
+
+ def _encode_vae_image(
+ self,
+ image: torch.Tensor,
+ device: Union[str, torch.device],
+ num_videos_per_prompt: int,
+ do_classifier_free_guidance: bool,
+ ):
+ image = image.to(device=device)
+ image_latents = self.vae.encode(image).latent_dist.mode()
+
+ if do_classifier_free_guidance:
+ negative_image_latents = torch.zeros_like(image_latents)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ image_latents = torch.cat([negative_image_latents, image_latents])
+
+ # duplicate image_latents for each generation per prompt, using mps friendly method
+ image_latents = image_latents.repeat(num_videos_per_prompt, 1, 1, 1)
+
+ return image_latents
+
+ def _get_add_time_ids(
+ self,
+ fps: int,
+ motion_bucket_id: int,
+ noise_aug_strength: float,
+ dtype: torch.dtype,
+ batch_size: int,
+ num_videos_per_prompt: int,
+ do_classifier_free_guidance: bool,
+ ):
+ add_time_ids = [fps, motion_bucket_id, noise_aug_strength]
+
+ passed_add_embed_dim = self.unet.config.addition_time_embed_dim * len(add_time_ids)
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ add_time_ids = add_time_ids.repeat(batch_size * num_videos_per_prompt, 1)
+
+ if do_classifier_free_guidance:
+ add_time_ids = torch.cat([add_time_ids, add_time_ids])
+
+ return add_time_ids
+
+ def decode_latents(self, latents: torch.FloatTensor, num_frames: int, decode_chunk_size: int = 14):
+ # [batch, frames, channels, height, width] -> [batch*frames, channels, height, width]
+ latents = latents.flatten(0, 1)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+
+ forward_vae_fn = self.vae._orig_mod.forward if is_compiled_module(self.vae) else self.vae.forward
+ accepts_num_frames = "num_frames" in set(inspect.signature(forward_vae_fn).parameters.keys())
+
+ # decode decode_chunk_size frames at a time to avoid OOM
+ frames = []
+ for i in range(0, latents.shape[0], decode_chunk_size):
+ num_frames_in = latents[i : i + decode_chunk_size].shape[0]
+ decode_kwargs = {}
+ if accepts_num_frames:
+ # we only pass num_frames_in if it's expected
+ decode_kwargs["num_frames"] = num_frames_in
+
+ frame = self.vae.decode(latents[i : i + decode_chunk_size], **decode_kwargs).sample
+ frames.append(frame)
+ frames = torch.cat(frames, dim=0)
+
+ # [batch*frames, channels, height, width] -> [batch, channels, frames, height, width]
+ frames = frames.reshape(-1, num_frames, *frames.shape[1:]).permute(0, 2, 1, 3, 4)
+
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ frames = frames.float()
+ return frames
+
+ def check_inputs(self, image, height, width):
+ if (
+ not isinstance(image, torch.Tensor)
+ and not isinstance(image, PIL.Image.Image)
+ and not isinstance(image, list)
+ ):
+ raise ValueError(
+ "`image` has to be of type `torch.FloatTensor` or `PIL.Image.Image` or `List[PIL.Image.Image]` but is"
+ f" {type(image)}"
+ )
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ def prepare_latents(
+ self,
+ batch_size: int,
+ num_frames: int,
+ num_channels_latents: int,
+ height: int,
+ width: int,
+ dtype: torch.dtype,
+ device: Union[str, torch.device],
+ generator: torch.Generator,
+ latents: Optional[torch.FloatTensor] = None,
+ ):
+ shape = (
+ batch_size,
+ num_frames,
+ num_channels_latents // 2,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ if isinstance(self.guidance_scale, (int, float)):
+ return self.guidance_scale > 1
+ return self.guidance_scale.max() > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image: Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor],
+ height: int = 576,
+ width: int = 1024,
+ num_frames: Optional[int] = None,
+ num_inference_steps: int = 25,
+ min_guidance_scale: float = 1.0,
+ max_guidance_scale: float = 3.0,
+ fps: int = 7,
+ motion_bucket_id: int = 127,
+ noise_aug_strength: float = 0.02,
+ decode_chunk_size: Optional[int] = None,
+ num_videos_per_prompt: Optional[int] = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ return_dict: bool = True,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+ Image(s) to guide image generation. If you provide a tensor, the expected value range is between `[0, 1]`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_frames (`int`, *optional*):
+ The number of video frames to generate. Defaults to `self.unet.config.num_frames`
+ (14 for `stable-video-diffusion-img2vid` and to 25 for `stable-video-diffusion-img2vid-xt`).
+ num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps. More denoising steps usually lead to a higher quality video at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ min_guidance_scale (`float`, *optional*, defaults to 1.0):
+ The minimum guidance scale. Used for the classifier free guidance with first frame.
+ max_guidance_scale (`float`, *optional*, defaults to 3.0):
+ The maximum guidance scale. Used for the classifier free guidance with last frame.
+ fps (`int`, *optional*, defaults to 7):
+ Frames per second. The rate at which the generated images shall be exported to a video after generation.
+ Note that Stable Diffusion Video's UNet was micro-conditioned on fps-1 during training.
+ motion_bucket_id (`int`, *optional*, defaults to 127):
+ Used for conditioning the amount of motion for the generation. The higher the number the more motion
+ will be in the video.
+ noise_aug_strength (`float`, *optional*, defaults to 0.02):
+ The amount of noise added to the init image, the higher it is the less the video will look like the init image. Increase it for more motion.
+ decode_chunk_size (`int`, *optional*):
+ The number of frames to decode at a time. Higher chunk size leads to better temporal consistency at the expense of more memory usage. By default, the decoder decodes all frames at once for maximal
+ quality. For lower memory usage, reduce `decode_chunk_size`.
+ num_videos_per_prompt (`int`, *optional*, defaults to 1):
+ The number of videos to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `pil`, `np` or `pt`.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that is called at the end of each denoising step during inference. The function is called
+ with the following arguments:
+ `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`.
+ `callback_kwargs` will include a list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableVideoDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` of (`List[List[PIL.Image.Image]]` or `np.ndarray` or `torch.FloatTensor`) is returned.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ num_frames = num_frames if num_frames is not None else self.unet.config.num_frames
+ decode_chunk_size = decode_chunk_size if decode_chunk_size is not None else num_frames
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(image, height, width)
+
+ # 2. Define call parameters
+ if isinstance(image, PIL.Image.Image):
+ batch_size = 1
+ elif isinstance(image, list):
+ batch_size = len(image)
+ else:
+ batch_size = image.shape[0]
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ self._guidance_scale = max_guidance_scale
+
+ # 3. Encode input image
+ image_embeddings = self._encode_image(image, device, num_videos_per_prompt, self.do_classifier_free_guidance)
+
+ # NOTE: Stable Video Diffusion was conditioned on fps - 1, which is why it is reduced here.
+ # See: https://github.com/Stability-AI/generative-models/blob/ed0997173f98eaf8f4edf7ba5fe8f15c6b877fd3/scripts/sampling/simple_video_sample.py#L188
+ fps = fps - 1
+
+ # 4. Encode input image using VAE
+ image = self.image_processor.preprocess(image, height=height, width=width).to(device)
+ noise = randn_tensor(image.shape, generator=generator, device=device, dtype=image.dtype)
+ image = image + noise_aug_strength * noise
+
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float32)
+
+ image_latents = self._encode_vae_image(
+ image,
+ device=device,
+ num_videos_per_prompt=num_videos_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ )
+ image_latents = image_latents.to(image_embeddings.dtype)
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+
+ # Repeat the image latents for each frame so we can concatenate them with the noise
+ # image_latents [batch, channels, height, width] ->[batch, num_frames, channels, height, width]
+ image_latents = image_latents.unsqueeze(1).repeat(1, num_frames, 1, 1, 1)
+
+ # 5. Get Added Time IDs
+ added_time_ids = self._get_add_time_ids(
+ fps,
+ motion_bucket_id,
+ noise_aug_strength,
+ image_embeddings.dtype,
+ batch_size,
+ num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+ added_time_ids = added_time_ids.to(device)
+
+ # 6. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 7. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_videos_per_prompt,
+ num_frames,
+ num_channels_latents,
+ height,
+ width,
+ image_embeddings.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 8. Prepare guidance scale
+ guidance_scale = torch.linspace(min_guidance_scale, max_guidance_scale, num_frames).unsqueeze(0)
+ guidance_scale = guidance_scale.to(device, latents.dtype)
+ guidance_scale = guidance_scale.repeat(batch_size * num_videos_per_prompt, 1)
+ guidance_scale = _append_dims(guidance_scale, latents.ndim)
+
+ self._guidance_scale = guidance_scale
+
+ # 9. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ self._num_timesteps = len(timesteps)
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # Concatenate image_latents over channels dimension
+ latent_model_input = torch.cat([latent_model_input, image_latents], dim=2)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=image_embeddings,
+ added_time_ids=added_time_ids,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_cond - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+
+ if not output_type == "latent":
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ frames = self.decode_latents(latents, num_frames, decode_chunk_size)
+ frames = tensor2vid(frames, self.image_processor, output_type=output_type)
+ else:
+ frames = latents
+
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return frames
+
+ return StableVideoDiffusionPipelineOutput(frames=frames)
+
+
+# resizing utils
+# TODO: clean up later
+def _resize_with_antialiasing(input, size, interpolation="bicubic", align_corners=True):
+ h, w = input.shape[-2:]
+ factors = (h / size[0], w / size[1])
+
+ # First, we have to determine sigma
+ # Taken from skimage: https://github.com/scikit-image/scikit-image/blob/v0.19.2/skimage/transform/_warps.py#L171
+ sigmas = (
+ max((factors[0] - 1.0) / 2.0, 0.001),
+ max((factors[1] - 1.0) / 2.0, 0.001),
+ )
+
+ # Now kernel size. Good results are for 3 sigma, but that is kind of slow. Pillow uses 1 sigma
+ # https://github.com/python-pillow/Pillow/blob/master/src/libImaging/Resample.c#L206
+ # But they do it in the 2 passes, which gives better results. Let's try 2 sigmas for now
+ ks = int(max(2.0 * 2 * sigmas[0], 3)), int(max(2.0 * 2 * sigmas[1], 3))
+
+ # Make sure it is odd
+ if (ks[0] % 2) == 0:
+ ks = ks[0] + 1, ks[1]
+
+ if (ks[1] % 2) == 0:
+ ks = ks[0], ks[1] + 1
+
+ input = _gaussian_blur2d(input, ks, sigmas)
+
+ output = torch.nn.functional.interpolate(input, size=size, mode=interpolation, align_corners=align_corners)
+ return output
+
+
+def _compute_padding(kernel_size):
+ """Compute padding tuple."""
+ # 4 or 6 ints: (padding_left, padding_right,padding_top,padding_bottom)
+ # https://pytorch.org/docs/stable/nn.html#torch.nn.functional.pad
+ if len(kernel_size) < 2:
+ raise AssertionError(kernel_size)
+ computed = [k - 1 for k in kernel_size]
+
+ # for even kernels we need to do asymmetric padding :(
+ out_padding = 2 * len(kernel_size) * [0]
+
+ for i in range(len(kernel_size)):
+ computed_tmp = computed[-(i + 1)]
+
+ pad_front = computed_tmp // 2
+ pad_rear = computed_tmp - pad_front
+
+ out_padding[2 * i + 0] = pad_front
+ out_padding[2 * i + 1] = pad_rear
+
+ return out_padding
+
+
+def _filter2d(input, kernel):
+ # prepare kernel
+ b, c, h, w = input.shape
+ tmp_kernel = kernel[:, None, ...].to(device=input.device, dtype=input.dtype)
+
+ tmp_kernel = tmp_kernel.expand(-1, c, -1, -1)
+
+ height, width = tmp_kernel.shape[-2:]
+
+ padding_shape: list[int] = _compute_padding([height, width])
+ input = torch.nn.functional.pad(input, padding_shape, mode="reflect")
+
+ # kernel and input tensor reshape to align element-wise or batch-wise params
+ tmp_kernel = tmp_kernel.reshape(-1, 1, height, width)
+ input = input.view(-1, tmp_kernel.size(0), input.size(-2), input.size(-1))
+
+ # convolve the tensor with the kernel.
+ output = torch.nn.functional.conv2d(input, tmp_kernel, groups=tmp_kernel.size(0), padding=0, stride=1)
+
+ out = output.view(b, c, h, w)
+ return out
+
+
+def _gaussian(window_size: int, sigma):
+ if isinstance(sigma, float):
+ sigma = torch.tensor([[sigma]])
+
+ batch_size = sigma.shape[0]
+
+ x = (torch.arange(window_size, device=sigma.device, dtype=sigma.dtype) - window_size // 2).expand(batch_size, -1)
+
+ if window_size % 2 == 0:
+ x = x + 0.5
+
+ gauss = torch.exp(-x.pow(2.0) / (2 * sigma.pow(2.0)))
+
+ return gauss / gauss.sum(-1, keepdim=True)
+
+
+def _gaussian_blur2d(input, kernel_size, sigma):
+ if isinstance(sigma, tuple):
+ sigma = torch.tensor([sigma], dtype=input.dtype)
+ else:
+ sigma = sigma.to(dtype=input.dtype)
+
+ ky, kx = int(kernel_size[0]), int(kernel_size[1])
+ bs = sigma.shape[0]
+ kernel_x = _gaussian(kx, sigma[:, 1].view(bs, 1))
+ kernel_y = _gaussian(ky, sigma[:, 0].view(bs, 1))
+ out_x = _filter2d(input, kernel_x[..., None, :])
+ out = _filter2d(out_x, kernel_y[..., None])
+
+ return out
diff --git a/model/diffusers_c/pipelines/t2i_adapter/__init__.py b/model/diffusers_c/pipelines/t2i_adapter/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..08c22a2707fe55770a519db481954881c1cad26e
--- /dev/null
+++ b/model/diffusers_c/pipelines/t2i_adapter/__init__.py
@@ -0,0 +1,47 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_stable_diffusion_adapter"] = ["StableDiffusionAdapterPipeline"]
+ _import_structure["pipeline_stable_diffusion_xl_adapter"] = ["StableDiffusionXLAdapterPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import * # noqa F403
+ else:
+ from .pipeline_stable_diffusion_adapter import StableDiffusionAdapterPipeline
+ from .pipeline_stable_diffusion_xl_adapter import StableDiffusionXLAdapterPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/t2i_adapter/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/t2i_adapter/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b09e9e3d2d3df3cba278dfa6cc0b55f598b6b087
Binary files /dev/null and b/model/diffusers_c/pipelines/t2i_adapter/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_adapter.cpython-39.pyc b/model/diffusers_c/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_adapter.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3176d80e4e702d3b1c869d3d4a102def8effe91c
Binary files /dev/null and b/model/diffusers_c/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_adapter.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_xl_adapter.cpython-39.pyc b/model/diffusers_c/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_xl_adapter.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e6de03bebf0b4811a4b7fd849de0c0fddaac4f27
Binary files /dev/null and b/model/diffusers_c/pipelines/t2i_adapter/__pycache__/pipeline_stable_diffusion_xl_adapter.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py b/model/diffusers_c/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py
new file mode 100644
index 0000000000000000000000000000000000000000..0b55bb38b5eba83a60eb4655a41ee74b3a7eee41
--- /dev/null
+++ b/model/diffusers_c/pipelines/t2i_adapter/pipeline_stable_diffusion_adapter.py
@@ -0,0 +1,912 @@
+# Copyright 2024 TencentARC and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, MultiAdapter, T2IAdapter, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ PIL_INTERPOLATION,
+ USE_PEFT_BACKEND,
+ BaseOutput,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+@dataclass
+class StableDiffusionAdapterPipelineOutput(BaseOutput):
+ """
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+ num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+ nsfw_content_detected (`List[bool]`)
+ List of flags denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, or `None` if safety checking could not be performed.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+ nsfw_content_detected: Optional[List[bool]]
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from PIL import Image
+ >>> from diffusers.utils import load_image
+ >>> import torch
+ >>> from diffusers import StableDiffusionAdapterPipeline, T2IAdapter
+
+ >>> image = load_image(
+ ... "https://huggingface.co/datasets/diffusers/docs-images/resolve/main/t2i-adapter/color_ref.png"
+ ... )
+
+ >>> color_palette = image.resize((8, 8))
+ >>> color_palette = color_palette.resize((512, 512), resample=Image.Resampling.NEAREST)
+
+ >>> adapter = T2IAdapter.from_pretrained("TencentARC/t2iadapter_color_sd14v1", torch_dtype=torch.float16)
+ >>> pipe = StableDiffusionAdapterPipeline.from_pretrained(
+ ... "CompVis/stable-diffusion-v1-4",
+ ... adapter=adapter,
+ ... torch_dtype=torch.float16,
+ ... )
+
+ >>> pipe.to("cuda")
+
+ >>> out_image = pipe(
+ ... "At night, glowing cubes in front of the beach",
+ ... image=color_palette,
+ ... ).images[0]
+ ```
+"""
+
+
+def _preprocess_adapter_image(image, height, width):
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ image = [np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])) for i in image]
+ image = [
+ i[None, ..., None] if i.ndim == 2 else i[None, ...] for i in image
+ ] # expand [h, w] or [h, w, c] to [b, h, w, c]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ if image[0].ndim == 3:
+ image = torch.stack(image, dim=0)
+ elif image[0].ndim == 4:
+ image = torch.cat(image, dim=0)
+ else:
+ raise ValueError(
+ f"Invalid image tensor! Expecting image tensor with 3 or 4 dimension, but recive: {image[0].ndim}"
+ )
+ return image
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionAdapterPipeline(DiffusionPipeline, StableDiffusionMixin):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter
+ https://arxiv.org/abs/2302.08453
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ adapter ([`T2IAdapter`] or [`MultiAdapter`] or `List[T2IAdapter]`):
+ Provides additional conditioning to the unet during the denoising process. If you set multiple Adapter as a
+ list, the outputs from each Adapter are added together to create one combined additional conditioning.
+ adapter_weights (`List[float]`, *optional*, defaults to None):
+ List of floats representing the weight which will be multiply to each adapter's output before adding them
+ together.
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPFeatureExtractor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->adapter->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ adapter: Union[T2IAdapter, MultiAdapter, List[T2IAdapter]],
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPFeatureExtractor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ if isinstance(adapter, (list, tuple)):
+ adapter = MultiAdapter(adapter)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ adapter=adapter,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ image,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if isinstance(self.adapter, MultiAdapter):
+ if not isinstance(image, list):
+ raise ValueError(
+ "MultiAdapter is enabled, but `image` is not a list. Please pass a list of images to `image`."
+ )
+
+ if len(image) != len(self.adapter.adapters):
+ raise ValueError(
+ f"MultiAdapter requires passing the same number of images as adapters. Given {len(image)} images and {len(self.adapter.adapters)} adapters."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def _default_height_width(self, height, width, image):
+ # NOTE: It is possible that a list of images have different
+ # dimensions for each image, so just checking the first image
+ # is not _exactly_ correct, but it is simple.
+ while isinstance(image, list):
+ image = image[0]
+
+ if height is None:
+ if isinstance(image, PIL.Image.Image):
+ height = image.height
+ elif isinstance(image, torch.Tensor):
+ height = image.shape[-2]
+
+ # round down to nearest multiple of `self.adapter.downscale_factor`
+ height = (height // self.adapter.downscale_factor) * self.adapter.downscale_factor
+
+ if width is None:
+ if isinstance(image, PIL.Image.Image):
+ width = image.width
+ elif isinstance(image, torch.Tensor):
+ width = image.shape[-1]
+
+ # round down to nearest multiple of `self.adapter.downscale_factor`
+ width = (width // self.adapter.downscale_factor) * self.adapter.downscale_factor
+
+ return height, width
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ image: Union[torch.Tensor, PIL.Image.Image, List[PIL.Image.Image]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ adapter_conditioning_scale: Union[float, List[float]] = 1.0,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]` or `List[PIL.Image.Image]` or `List[List[PIL.Image.Image]]`):
+ The Adapter input condition. Adapter uses this input condition to generate guidance to Unet. If the
+ type is specified as `Torch.FloatTensor`, it is passed to Adapter as is. PIL.Image.Image` can also be
+ accepted as an image. The control image is automatically resized to fit the output image.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+ Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] instead
+ of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ adapter_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the adapter are multiplied by `adapter_conditioning_scale` before they are added to the
+ residual in the original unet. If multiple adapters are specified in init, you can set the
+ corresponding scale as a list.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] if `return_dict` is True, otherwise a
+ `tuple. When returning a tuple, the first element is a list with the generated images, and the second
+ element is a list of `bool`s denoting whether the corresponding generated image likely represents
+ "not-safe-for-work" (nsfw) content, according to the `safety_checker`.
+ """
+ # 0. Default height and width to unet
+ height, width = self._default_height_width(height, width, image)
+ device = self._execution_device
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt, height, width, callback_steps, image, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ self._guidance_scale = guidance_scale
+
+ if isinstance(self.adapter, MultiAdapter):
+ adapter_input = []
+
+ for one_image in image:
+ one_image = _preprocess_adapter_image(one_image, height, width)
+ one_image = one_image.to(device=device, dtype=self.adapter.dtype)
+ adapter_input.append(one_image)
+ else:
+ adapter_input = _preprocess_adapter_image(image, height, width)
+ adapter_input = adapter_input.to(device=device, dtype=self.adapter.dtype)
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # 3. Encode input prompt
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 6.5 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Denoising loop
+ if isinstance(self.adapter, MultiAdapter):
+ adapter_state = self.adapter(adapter_input, adapter_conditioning_scale)
+ for k, v in enumerate(adapter_state):
+ adapter_state[k] = v
+ else:
+ adapter_state = self.adapter(adapter_input)
+ for k, v in enumerate(adapter_state):
+ adapter_state[k] = v * adapter_conditioning_scale
+ if num_images_per_prompt > 1:
+ for k, v in enumerate(adapter_state):
+ adapter_state[k] = v.repeat(num_images_per_prompt, 1, 1, 1)
+ if self.do_classifier_free_guidance:
+ for k, v in enumerate(adapter_state):
+ adapter_state[k] = torch.cat([v] * 2, dim=0)
+
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=cross_attention_kwargs,
+ down_intrablock_additional_residuals=[state.clone() for state in adapter_state],
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if output_type == "latent":
+ image = latents
+ has_nsfw_concept = None
+ elif output_type == "pil":
+ # 8. Post-processing
+ image = self.decode_latents(latents)
+
+ # 9. Run safety checker
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # 10. Convert to PIL
+ image = self.numpy_to_pil(image)
+ else:
+ # 8. Post-processing
+ image = self.decode_latents(latents)
+
+ # 9. Run safety checker
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionAdapterPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/model/diffusers_c/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py b/model/diffusers_c/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py
new file mode 100644
index 0000000000000000000000000000000000000000..4e0cc61f5c1d4f81aea6bce71c48ca30b78dd8fe
--- /dev/null
+++ b/model/diffusers_c/pipelines/t2i_adapter/pipeline_stable_diffusion_xl_adapter.py
@@ -0,0 +1,1258 @@
+# Copyright 2024 TencentARC and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import (
+ FromSingleFileMixin,
+ IPAdapterMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+)
+from ...models import AutoencoderKL, ImageProjection, MultiAdapter, T2IAdapter, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ PIL_INTERPOLATION,
+ USE_PEFT_BACKEND,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import T2IAdapter, StableDiffusionXLAdapterPipeline, DDPMScheduler
+ >>> from diffusers.utils import load_image
+
+ >>> sketch_image = load_image("https://huggingface.co/Adapter/t2iadapter/resolve/main/sketch.png").convert("L")
+
+ >>> model_id = "stabilityai/stable-diffusion-xl-base-1.0"
+
+ >>> adapter = T2IAdapter.from_pretrained(
+ ... "Adapter/t2iadapter",
+ ... subfolder="sketch_sdxl_1.0",
+ ... torch_dtype=torch.float16,
+ ... adapter_type="full_adapter_xl",
+ ... )
+ >>> scheduler = DDPMScheduler.from_pretrained(model_id, subfolder="scheduler")
+
+ >>> pipe = StableDiffusionXLAdapterPipeline.from_pretrained(
+ ... model_id, adapter=adapter, torch_dtype=torch.float16, variant="fp16", scheduler=scheduler
+ ... ).to("cuda")
+
+ >>> generator = torch.manual_seed(42)
+ >>> sketch_image_out = pipe(
+ ... prompt="a photo of a dog in real world, high quality",
+ ... negative_prompt="extra digit, fewer digits, cropped, worst quality, low quality",
+ ... image=sketch_image,
+ ... generator=generator,
+ ... guidance_scale=7.5,
+ ... ).images[0]
+ ```
+"""
+
+
+def _preprocess_adapter_image(image, height, width):
+ if isinstance(image, torch.Tensor):
+ return image
+ elif isinstance(image, PIL.Image.Image):
+ image = [image]
+
+ if isinstance(image[0], PIL.Image.Image):
+ image = [np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"])) for i in image]
+ image = [
+ i[None, ..., None] if i.ndim == 2 else i[None, ...] for i in image
+ ] # expand [h, w] or [h, w, c] to [b, h, w, c]
+ image = np.concatenate(image, axis=0)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image)
+ elif isinstance(image[0], torch.Tensor):
+ if image[0].ndim == 3:
+ image = torch.stack(image, dim=0)
+ elif image[0].ndim == 4:
+ image = torch.cat(image, dim=0)
+ else:
+ raise ValueError(
+ f"Invalid image tensor! Expecting image tensor with 3 or 4 dimension, but recive: {image[0].ndim}"
+ )
+ return image
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionXLAdapterPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion augmented with T2I-Adapter
+ https://arxiv.org/abs/2302.08453
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ adapter ([`T2IAdapter`] or [`MultiAdapter`] or `List[T2IAdapter]`):
+ Provides additional conditioning to the unet during the denoising process. If you set multiple Adapter as a
+ list, the outputs from each Adapter are added together to create one combined additional conditioning.
+ adapter_weights (`List[float]`, *optional*, defaults to None):
+ List of floats representing the weight which will be multiply to each adapter's output before adding them
+ together.
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPFeatureExtractor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae"
+ _optional_components = [
+ "tokenizer",
+ "tokenizer_2",
+ "text_encoder",
+ "text_encoder_2",
+ "feature_extractor",
+ "image_encoder",
+ ]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ adapter: Union[T2IAdapter, MultiAdapter, List[T2IAdapter]],
+ scheduler: KarrasDiffusionSchedulers,
+ force_zeros_for_empty_prompt: bool = True,
+ feature_extractor: CLIPImageProcessor = None,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ adapter=adapter,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.default_sample_size = self.unet.config.sample_size
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ pooled_prompt_embeds=None,
+ negative_pooled_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+ )
+
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
+ def _get_add_time_ids(
+ self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+ ):
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ return add_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ # Copied from diffusers.pipelines.t2i_adapter.pipeline_stable_diffusion_adapter.StableDiffusionAdapterPipeline._default_height_width
+ def _default_height_width(self, height, width, image):
+ # NOTE: It is possible that a list of images have different
+ # dimensions for each image, so just checking the first image
+ # is not _exactly_ correct, but it is simple.
+ while isinstance(image, list):
+ image = image[0]
+
+ if height is None:
+ if isinstance(image, PIL.Image.Image):
+ height = image.height
+ elif isinstance(image, torch.Tensor):
+ height = image.shape[-2]
+
+ # round down to nearest multiple of `self.adapter.downscale_factor`
+ height = (height // self.adapter.downscale_factor) * self.adapter.downscale_factor
+
+ if width is None:
+ if isinstance(image, PIL.Image.Image):
+ width = image.width
+ elif isinstance(image, torch.Tensor):
+ width = image.shape[-1]
+
+ # round down to nearest multiple of `self.adapter.downscale_factor`
+ width = (width // self.adapter.downscale_factor) * self.adapter.downscale_factor
+
+ return height, width
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ image: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ denoising_end: Optional[float] = None,
+ guidance_scale: float = 5.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ original_size: Optional[Tuple[int, int]] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Optional[Tuple[int, int]] = None,
+ negative_original_size: Optional[Tuple[int, int]] = None,
+ negative_crops_coords_top_left: Tuple[int, int] = (0, 0),
+ negative_target_size: Optional[Tuple[int, int]] = None,
+ adapter_conditioning_scale: Union[float, List[float]] = 1.0,
+ adapter_conditioning_factor: float = 1.0,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]` or `List[PIL.Image.Image]` or `List[List[PIL.Image.Image]]`):
+ The Adapter input condition. Adapter uses this input condition to generate guidance to Unet. If the
+ type is specified as `Torch.FloatTensor`, it is passed to Adapter as is. PIL.Image.Image` can also be
+ accepted as an image. The control image is automatically resized to fit the output image.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image. Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image. Anything below 512 pixels won't work well for
+ [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0)
+ and checkpoints that are not specifically fine-tuned on low resolutions.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ denoising_end (`float`, *optional*):
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
+ still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+ scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+ "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+ guidance_scale (`float`, *optional*, defaults to 5.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionAdapterPipelineOutput`]
+ instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ guidance_rescale (`float`, *optional*, defaults to 0.0):
+ Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+ [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+ Guidance rescale factor should fix overexposure when using zero terminal SNR.
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a specific image resolution. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's
+ micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ To negatively condition the generation process based on a target image resolution. It should be as same
+ as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more
+ information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208.
+ adapter_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the adapter are multiplied by `adapter_conditioning_scale` before they are added to the
+ residual in the original unet. If multiple adapters are specified in init, you can set the
+ corresponding scale as a list.
+ adapter_conditioning_factor (`float`, *optional*, defaults to 1.0):
+ The fraction of timesteps for which adapter should be applied. If `adapter_conditioning_factor` is
+ `0.0`, adapter is not applied at all. If `adapter_conditioning_factor` is `1.0`, adapter is applied for
+ all timesteps. If `adapter_conditioning_factor` is `0.5`, adapter is applied for half of the timesteps.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionAdapterPipelineOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is a list with the generated images.
+ """
+ # 0. Default height and width to unet
+
+ height, width = self._default_height_width(height, width, image)
+ device = self._execution_device
+
+ if isinstance(self.adapter, MultiAdapter):
+ adapter_input = []
+
+ for one_image in image:
+ one_image = _preprocess_adapter_image(one_image, height, width)
+ one_image = one_image.to(device=device, dtype=self.adapter.dtype)
+ adapter_input.append(one_image)
+ else:
+ adapter_input = _preprocess_adapter_image(image, height, width)
+ adapter_input = adapter_input.to(device=device, dtype=self.adapter.dtype)
+ original_size = original_size or (height, width)
+ target_size = target_size or (height, width)
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ )
+
+ self._guidance_scale = guidance_scale
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ # 3.1 Encode input prompt
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ prompt_2=prompt_2,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ negative_prompt_2=negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ clip_skip=clip_skip,
+ )
+
+ # 3.2 Encode ip_adapter_image
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6.1 Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 6.2 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Prepare added time ids & embeddings & adapter features
+ if isinstance(self.adapter, MultiAdapter):
+ adapter_state = self.adapter(adapter_input, adapter_conditioning_scale)
+ for k, v in enumerate(adapter_state):
+ adapter_state[k] = v
+ else:
+ adapter_state = self.adapter(adapter_input)
+ for k, v in enumerate(adapter_state):
+ adapter_state[k] = v * adapter_conditioning_scale
+ if num_images_per_prompt > 1:
+ for k, v in enumerate(adapter_state):
+ adapter_state[k] = v.repeat(num_images_per_prompt, 1, 1, 1)
+ if self.do_classifier_free_guidance:
+ for k, v in enumerate(adapter_state):
+ adapter_state[k] = torch.cat([v] * 2, dim=0)
+
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ if negative_original_size is not None and negative_target_size is not None:
+ negative_add_time_ids = self._get_add_time_ids(
+ negative_original_size,
+ negative_crops_coords_top_left,
+ negative_target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+ else:
+ negative_add_time_ids = add_time_ids
+
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1)
+
+ # 8. Denoising loop
+ num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+ # Apply denoising_end
+ if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (denoising_end * self.scheduler.config.num_train_timesteps)
+ )
+ )
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+ timesteps = timesteps[:num_inference_steps]
+
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ added_cond_kwargs["image_embeds"] = image_embeds
+
+ # predict the noise residual
+ if i < int(num_inference_steps * adapter_conditioning_factor):
+ down_intrablock_additional_residuals = [state.clone() for state in adapter_state]
+ else:
+ down_intrablock_additional_residuals = None
+
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=cross_attention_kwargs,
+ down_intrablock_additional_residuals=down_intrablock_additional_residuals,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if self.do_classifier_free_guidance and guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ image = latents
+ return StableDiffusionXLPipelineOutput(images=image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image,)
+
+ return StableDiffusionXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/__init__.py b/model/diffusers_c/pipelines/text_to_video_synthesis/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8d8fdb92769bb3dcb2dd7696115b16f197062262
--- /dev/null
+++ b/model/diffusers_c/pipelines/text_to_video_synthesis/__init__.py
@@ -0,0 +1,54 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["pipeline_output"] = ["TextToVideoSDPipelineOutput"]
+ _import_structure["pipeline_text_to_video_synth"] = ["TextToVideoSDPipeline"]
+ _import_structure["pipeline_text_to_video_synth_img2img"] = ["VideoToVideoSDPipeline"]
+ _import_structure["pipeline_text_to_video_zero"] = ["TextToVideoZeroPipeline"]
+ _import_structure["pipeline_text_to_video_zero_sdxl"] = ["TextToVideoZeroSDXLPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import * # noqa F403
+ else:
+ from .pipeline_output import TextToVideoSDPipelineOutput
+ from .pipeline_text_to_video_synth import TextToVideoSDPipeline
+ from .pipeline_text_to_video_synth_img2img import VideoToVideoSDPipeline
+ from .pipeline_text_to_video_zero import TextToVideoZeroPipeline
+ from .pipeline_text_to_video_zero_sdxl import TextToVideoZeroSDXLPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..097d5bed7f74c9f7bbeacda6eef5941568d6d382
Binary files /dev/null and b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_output.cpython-39.pyc b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_output.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1946551ad9b5d85c22b5f34f30ddd311d03b5371
Binary files /dev/null and b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_output.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth.cpython-39.pyc b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d39cda0cbda2152c304ad56a364936cad19123eb
Binary files /dev/null and b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth_img2img.cpython-39.pyc b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth_img2img.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f0cf2e785021afa1dc2a163a806fb48a561f7dc4
Binary files /dev/null and b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_synth_img2img.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero.cpython-39.pyc b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ce2b5a1ee0a60cebf5c2c1c8a40cfa63d8772bc2
Binary files /dev/null and b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero_sdxl.cpython-39.pyc b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero_sdxl.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b9a27391303545d2d0d86bea348d5e4c1b695442
Binary files /dev/null and b/model/diffusers_c/pipelines/text_to_video_synthesis/__pycache__/pipeline_text_to_video_zero_sdxl.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_output.py b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_output.py
new file mode 100644
index 0000000000000000000000000000000000000000..c155386cf1733556ce19af64293dc06c7c972c4a
--- /dev/null
+++ b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_output.py
@@ -0,0 +1,25 @@
+from dataclasses import dataclass
+from typing import List, Union
+
+import numpy as np
+import PIL
+import torch
+
+from ...utils import (
+ BaseOutput,
+)
+
+
+@dataclass
+class TextToVideoSDPipelineOutput(BaseOutput):
+ """
+ Output class for text-to-video pipelines.
+
+ Args:
+ frames (`torch.Tensor`, `np.ndarray`, or List[List[PIL.Image.Image]]):
+ List of video outputs - It can be a nested list of length `batch_size,` with each sub-list containing denoised
+ PIL image sequences of length `num_frames.` It can also be a NumPy array or Torch tensor of shape
+ `(batch_size, num_frames, channels, height, width)`
+ """
+
+ frames: Union[torch.Tensor, np.ndarray, List[List[PIL.Image.Image]]]
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py
new file mode 100644
index 0000000000000000000000000000000000000000..6c33836e60da8c0e152bb7b0ed32a5a2786b97eb
--- /dev/null
+++ b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth.py
@@ -0,0 +1,663 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet3DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from . import TextToVideoSDPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import TextToVideoSDPipeline
+ >>> from diffusers.utils import export_to_video
+
+ >>> pipe = TextToVideoSDPipeline.from_pretrained(
+ ... "damo-vilab/text-to-video-ms-1.7b", torch_dtype=torch.float16, variant="fp16"
+ ... )
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> prompt = "Spiderman is surfing"
+ >>> video_frames = pipe(prompt).frames[0]
+ >>> video_path = export_to_video(video_frames)
+ >>> video_path
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor: "VaeImageProcessor", output_type: str = "np"):
+ batch_size, channels, num_frames, height, width = video.shape
+ outputs = []
+ for batch_idx in range(batch_size):
+ batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+ batch_output = processor.postprocess(batch_vid, output_type)
+
+ outputs.append(batch_output)
+
+ if output_type == "np":
+ outputs = np.stack(outputs)
+
+ elif output_type == "pt":
+ outputs = torch.stack(outputs)
+
+ elif not output_type == "pil":
+ raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+ return outputs
+
+
+class TextToVideoSDPipeline(DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin):
+ r"""
+ Pipeline for text-to-video generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`] to tokenize text.
+ unet ([`UNet3DConditionModel`]):
+ A [`UNet3DConditionModel`] to denoise the encoded video latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet3DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def decode_latents(self, latents):
+ latents = 1 / self.vae.config.scaling_factor * latents
+
+ batch_size, channels, num_frames, height, width = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+ image = self.vae.decode(latents).sample
+ video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ video = video.float()
+ return video
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def prepare_latents(
+ self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+ ):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ num_frames,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_frames: int = 16,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 9.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "np",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated video.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated video.
+ num_frames (`int`, *optional*, defaults to 16):
+ The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds
+ amounts to 2 seconds of video.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`. Latents should be of shape
+ `(batch_size, num_channel, num_frames, height, width)`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"np"`):
+ The output format of the generated video. Choose between `torch.FloatTensor` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead
+ of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ num_images_per_prompt = 1
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ num_frames,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # reshape latents
+ bsz, channel, frames, width, height = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+ noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # reshape latents back
+ latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 8. Post processing
+ if output_type == "latent":
+ video = latents
+ else:
+ video_tensor = self.decode_latents(latents)
+ video = tensor2vid(video_tensor, self.image_processor, output_type)
+
+ # 9. Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (video,)
+
+ return TextToVideoSDPipelineOutput(frames=video)
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py
new file mode 100644
index 0000000000000000000000000000000000000000..3901946afe465f5d8af629e45ae6044574ad69d4
--- /dev/null
+++ b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_synth_img2img.py
@@ -0,0 +1,760 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet3DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from . import TextToVideoSDPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import DiffusionPipeline, DPMSolverMultistepScheduler
+ >>> from diffusers.utils import export_to_video
+
+ >>> pipe = DiffusionPipeline.from_pretrained("cerspense/zeroscope_v2_576w", torch_dtype=torch.float16)
+ >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+ >>> pipe.to("cuda")
+
+ >>> prompt = "spiderman running in the desert"
+ >>> video_frames = pipe(prompt, num_inference_steps=40, height=320, width=576, num_frames=24).frames[0]
+ >>> # safe low-res video
+ >>> video_path = export_to_video(video_frames, output_video_path="./video_576_spiderman.mp4")
+
+ >>> # let's offload the text-to-image model
+ >>> pipe.to("cpu")
+
+ >>> # and load the image-to-image model
+ >>> pipe = DiffusionPipeline.from_pretrained(
+ ... "cerspense/zeroscope_v2_XL", torch_dtype=torch.float16, revision="refs/pr/15"
+ ... )
+ >>> pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> # The VAE consumes A LOT of memory, let's make sure we run it in sliced mode
+ >>> pipe.vae.enable_slicing()
+
+ >>> # now let's upscale it
+ >>> video = [Image.fromarray(frame).resize((1024, 576)) for frame in video_frames]
+
+ >>> # and denoise it
+ >>> video_frames = pipe(prompt, video=video, strength=0.6).frames[0]
+ >>> video_path = export_to_video(video_frames, output_video_path="./video_1024_spiderman.mp4")
+ >>> video_path
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+# Copied from diffusers.pipelines.animatediff.pipeline_animatediff.tensor2vid
+def tensor2vid(video: torch.Tensor, processor: "VaeImageProcessor", output_type: str = "np"):
+ batch_size, channels, num_frames, height, width = video.shape
+ outputs = []
+ for batch_idx in range(batch_size):
+ batch_vid = video[batch_idx].permute(1, 0, 2, 3)
+ batch_output = processor.postprocess(batch_vid, output_type)
+
+ outputs.append(batch_output)
+
+ if output_type == "np":
+ outputs = np.stack(outputs)
+
+ elif output_type == "pt":
+ outputs = torch.stack(outputs)
+
+ elif not output_type == "pil":
+ raise ValueError(f"{output_type} does not exist. Please choose one of ['np', 'pt', 'pil']")
+
+ return outputs
+
+
+def preprocess_video(video):
+ supported_formats = (np.ndarray, torch.Tensor, PIL.Image.Image)
+
+ if isinstance(video, supported_formats):
+ video = [video]
+ elif not (isinstance(video, list) and all(isinstance(i, supported_formats) for i in video)):
+ raise ValueError(
+ f"Input is in incorrect format: {[type(i) for i in video]}. Currently, we only support {', '.join(supported_formats)}"
+ )
+
+ if isinstance(video[0], PIL.Image.Image):
+ video = [np.array(frame) for frame in video]
+
+ if isinstance(video[0], np.ndarray):
+ video = np.concatenate(video, axis=0) if video[0].ndim == 5 else np.stack(video, axis=0)
+
+ if video.dtype == np.uint8:
+ video = np.array(video).astype(np.float32) / 255.0
+
+ if video.ndim == 4:
+ video = video[None, ...]
+
+ video = torch.from_numpy(video.transpose(0, 4, 1, 2, 3))
+
+ elif isinstance(video[0], torch.Tensor):
+ video = torch.cat(video, axis=0) if video[0].ndim == 5 else torch.stack(video, axis=0)
+
+ # don't need any preprocess if the video is latents
+ channel = video.shape[1]
+ if channel == 4:
+ return video
+
+ # move channels before num_frames
+ video = video.permute(0, 2, 1, 3, 4)
+
+ # normalize video
+ video = 2.0 * video - 1.0
+
+ return video
+
+
+class VideoToVideoSDPipeline(DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin):
+ r"""
+ Pipeline for text-guided video-to-video generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`] to tokenize text.
+ unet ([`UNet3DConditionModel`]):
+ A [`UNet3DConditionModel`] to denoise the encoded video latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->unet->vae"
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet3DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
+ def decode_latents(self, latents):
+ latents = 1 / self.vae.config.scaling_factor * latents
+
+ batch_size, channels, num_frames, height, width = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+ image = self.vae.decode(latents).sample
+ video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ video = video.float()
+ return video
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ strength,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+ if hasattr(self.scheduler, "set_begin_index"):
+ self.scheduler.set_begin_index(t_start * self.scheduler.order)
+
+ return timesteps, num_inference_steps - t_start
+
+ def prepare_latents(self, video, timestep, batch_size, dtype, device, generator=None):
+ video = video.to(device=device, dtype=dtype)
+
+ # change from (b, c, f, h, w) -> (b * f, c, w, h)
+ bsz, channel, frames, width, height = video.shape
+ video = video.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+
+ if video.shape[1] == 4:
+ init_latents = video
+ else:
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+ elif isinstance(generator, list):
+ init_latents = [
+ retrieve_latents(self.vae.encode(video[i : i + 1]), generator=generator[i])
+ for i in range(batch_size)
+ ]
+ init_latents = torch.cat(init_latents, dim=0)
+ else:
+ init_latents = retrieve_latents(self.vae.encode(video), generator=generator)
+
+ init_latents = self.vae.config.scaling_factor * init_latents
+
+ if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `video` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ init_latents = torch.cat([init_latents], dim=0)
+
+ shape = init_latents.shape
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+
+ # get latents
+ init_latents = self.scheduler.add_noise(init_latents, noise, timestep)
+ latents = init_latents
+
+ latents = latents[None, :].reshape((bsz, frames, latents.shape[1]) + latents.shape[2:]).permute(0, 2, 1, 3, 4)
+
+ return latents
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Union[str, List[str]] = None,
+ video: Union[List[np.ndarray], torch.FloatTensor] = None,
+ strength: float = 0.6,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 15.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "np",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ video (`List[np.ndarray]` or `torch.FloatTensor`):
+ `video` frames or tensor representing a video batch to be used as the starting point for the process.
+ Can also accept video latents as `image`, if passing latents directly, it will not be encoded again.
+ strength (`float`, *optional*, defaults to 0.8):
+ Indicates extent to transform the reference `video`. Must be between 0 and 1. `video` is used as a
+ starting point, adding more noise to it the larger the `strength`. The number of denoising steps
+ depends on the amount of noise initially added. When `strength` is 1, added noise is maximum and the
+ denoising process runs for the full number of iterations specified in `num_inference_steps`. A value of
+ 1 essentially ignores `video`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in video generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`. Latents should be of shape
+ `(batch_size, num_channel, num_frames, height, width)`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"np"`):
+ The output format of the generated video. Choose between `torch.FloatTensor` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead
+ of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ Examples:
+
+ Returns:
+ [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
+ """
+ # 0. Default height and width to unet
+ num_images_per_prompt = 1
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(prompt, strength, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds)
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Preprocess video
+ video = preprocess_video(video)
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+
+ # 6. Prepare latent variables
+ latents = self.prepare_latents(video, latent_timestep, batch_size, prompt_embeds.dtype, device, generator)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # reshape latents
+ bsz, channel, frames, width, height = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+ noise_pred = noise_pred.permute(0, 2, 1, 3, 4).reshape(bsz * frames, channel, width, height)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # reshape latents back
+ latents = latents[None, :].reshape(bsz, frames, channel, width, height).permute(0, 2, 1, 3, 4)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+
+ # 9. Post processing
+ if output_type == "latent":
+ video = latents
+ else:
+ video_tensor = self.decode_latents(latents)
+ video = tensor2vid(video_tensor, self.image_processor, output_type)
+
+ # 10. Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (video,)
+
+ return TextToVideoSDPipelineOutput(frames=video)
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py
new file mode 100644
index 0000000000000000000000000000000000000000..d3ff3728c22ef91853411abf513b3900506fa229
--- /dev/null
+++ b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py
@@ -0,0 +1,969 @@
+import copy
+import inspect
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from torch.nn.functional import grid_sample
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import USE_PEFT_BACKEND, BaseOutput, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from ..stable_diffusion import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def rearrange_0(tensor, f):
+ F, C, H, W = tensor.size()
+ tensor = torch.permute(torch.reshape(tensor, (F // f, f, C, H, W)), (0, 2, 1, 3, 4))
+ return tensor
+
+
+def rearrange_1(tensor):
+ B, C, F, H, W = tensor.size()
+ return torch.reshape(torch.permute(tensor, (0, 2, 1, 3, 4)), (B * F, C, H, W))
+
+
+def rearrange_3(tensor, f):
+ F, D, C = tensor.size()
+ return torch.reshape(tensor, (F // f, f, D, C))
+
+
+def rearrange_4(tensor):
+ B, F, D, C = tensor.size()
+ return torch.reshape(tensor, (B * F, D, C))
+
+
+class CrossFrameAttnProcessor:
+ """
+ Cross frame attention processor. Each frame attends the first frame.
+
+ Args:
+ batch_size: The number that represents actual batch size, other than the frames.
+ For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to
+ 2, due to classifier-free guidance.
+ """
+
+ def __init__(self, batch_size=2):
+ self.batch_size = batch_size
+
+ def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None):
+ batch_size, sequence_length, _ = hidden_states.shape
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ query = attn.to_q(hidden_states)
+
+ is_cross_attention = encoder_hidden_states is not None
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ # Cross Frame Attention
+ if not is_cross_attention:
+ video_length = key.size()[0] // self.batch_size
+ first_frame_index = [0] * video_length
+
+ # rearrange keys to have batch and frames in the 1st and 2nd dims respectively
+ key = rearrange_3(key, video_length)
+ key = key[:, first_frame_index]
+ # rearrange values to have batch and frames in the 1st and 2nd dims respectively
+ value = rearrange_3(value, video_length)
+ value = value[:, first_frame_index]
+
+ # rearrange back to original shape
+ key = rearrange_4(key)
+ value = rearrange_4(value)
+
+ query = attn.head_to_batch_dim(query)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ attention_probs = attn.get_attention_scores(query, key, attention_mask)
+ hidden_states = torch.bmm(attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ return hidden_states
+
+
+class CrossFrameAttnProcessor2_0:
+ """
+ Cross frame attention processor with scaled_dot_product attention of Pytorch 2.0.
+
+ Args:
+ batch_size: The number that represents actual batch size, other than the frames.
+ For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to
+ 2, due to classifier-free guidance.
+ """
+
+ def __init__(self, batch_size=2):
+ if not hasattr(F, "scaled_dot_product_attention"):
+ raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+ self.batch_size = batch_size
+
+ def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None):
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+ inner_dim = hidden_states.shape[-1]
+
+ if attention_mask is not None:
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ # scaled_dot_product_attention expects attention_mask shape to be
+ # (batch, heads, source_length, target_length)
+ attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+ query = attn.to_q(hidden_states)
+
+ is_cross_attention = encoder_hidden_states is not None
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ # Cross Frame Attention
+ if not is_cross_attention:
+ video_length = max(1, key.size()[0] // self.batch_size)
+ first_frame_index = [0] * video_length
+
+ # rearrange keys to have batch and frames in the 1st and 2nd dims respectively
+ key = rearrange_3(key, video_length)
+ key = key[:, first_frame_index]
+ # rearrange values to have batch and frames in the 1st and 2nd dims respectively
+ value = rearrange_3(value, video_length)
+ value = value[:, first_frame_index]
+
+ # rearrange back to original shape
+ key = rearrange_4(key)
+ value = rearrange_4(value)
+
+ head_dim = inner_dim // attn.heads
+ query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ # the output of sdp = (batch, num_heads, seq_len, head_dim)
+ # TODO: add support for attn.scale when we move to Torch 2.1
+ hidden_states = F.scaled_dot_product_attention(
+ query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+ )
+
+ hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+ hidden_states = hidden_states.to(query.dtype)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+ return hidden_states
+
+
+@dataclass
+class TextToVideoPipelineOutput(BaseOutput):
+ r"""
+ Output class for zero-shot text-to-video pipeline.
+
+ Args:
+ images (`[List[PIL.Image.Image]`, `np.ndarray`]):
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ nsfw_content_detected (`[List[bool]]`):
+ List indicating whether the corresponding generated image contains "not-safe-for-work" (nsfw) content or
+ `None` if safety checking could not be performed.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+ nsfw_content_detected: Optional[List[bool]]
+
+
+def coords_grid(batch, ht, wd, device):
+ # Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py
+ coords = torch.meshgrid(torch.arange(ht, device=device), torch.arange(wd, device=device))
+ coords = torch.stack(coords[::-1], dim=0).float()
+ return coords[None].repeat(batch, 1, 1, 1)
+
+
+def warp_single_latent(latent, reference_flow):
+ """
+ Warp latent of a single frame with given flow
+
+ Args:
+ latent: latent code of a single frame
+ reference_flow: flow which to warp the latent with
+
+ Returns:
+ warped: warped latent
+ """
+ _, _, H, W = reference_flow.size()
+ _, _, h, w = latent.size()
+ coords0 = coords_grid(1, H, W, device=latent.device).to(latent.dtype)
+
+ coords_t0 = coords0 + reference_flow
+ coords_t0[:, 0] /= W
+ coords_t0[:, 1] /= H
+
+ coords_t0 = coords_t0 * 2.0 - 1.0
+ coords_t0 = F.interpolate(coords_t0, size=(h, w), mode="bilinear")
+ coords_t0 = torch.permute(coords_t0, (0, 2, 3, 1))
+
+ warped = grid_sample(latent, coords_t0, mode="nearest", padding_mode="reflection")
+ return warped
+
+
+def create_motion_field(motion_field_strength_x, motion_field_strength_y, frame_ids, device, dtype):
+ """
+ Create translation motion field
+
+ Args:
+ motion_field_strength_x: motion strength along x-axis
+ motion_field_strength_y: motion strength along y-axis
+ frame_ids: indexes of the frames the latents of which are being processed.
+ This is needed when we perform chunk-by-chunk inference
+ device: device
+ dtype: dtype
+
+ Returns:
+
+ """
+ seq_length = len(frame_ids)
+ reference_flow = torch.zeros((seq_length, 2, 512, 512), device=device, dtype=dtype)
+ for fr_idx in range(seq_length):
+ reference_flow[fr_idx, 0, :, :] = motion_field_strength_x * (frame_ids[fr_idx])
+ reference_flow[fr_idx, 1, :, :] = motion_field_strength_y * (frame_ids[fr_idx])
+ return reference_flow
+
+
+def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_strength_y, frame_ids, latents):
+ """
+ Creates translation motion and warps the latents accordingly
+
+ Args:
+ motion_field_strength_x: motion strength along x-axis
+ motion_field_strength_y: motion strength along y-axis
+ frame_ids: indexes of the frames the latents of which are being processed.
+ This is needed when we perform chunk-by-chunk inference
+ latents: latent codes of frames
+
+ Returns:
+ warped_latents: warped latents
+ """
+ motion_field = create_motion_field(
+ motion_field_strength_x=motion_field_strength_x,
+ motion_field_strength_y=motion_field_strength_y,
+ frame_ids=frame_ids,
+ device=latents.device,
+ dtype=latents.dtype,
+ )
+ warped_latents = latents.clone().detach()
+ for i in range(len(warped_latents)):
+ warped_latents[i] = warp_single_latent(latents[i][None], motion_field[i][None])
+ return warped_latents
+
+
+class TextToVideoZeroPipeline(DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, LoraLoaderMixin):
+ r"""
+ Pipeline for zero-shot text-to-video generation using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`] to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A [`UNet3DConditionModel`] to denoise the encoded video latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`CLIPImageProcessor`]):
+ A [`CLIPImageProcessor`] to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ processor = (
+ CrossFrameAttnProcessor2_0(batch_size=2)
+ if hasattr(F, "scaled_dot_product_attention")
+ else CrossFrameAttnProcessor(batch_size=2)
+ )
+ self.unet.set_attn_processor(processor)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ def forward_loop(self, x_t0, t0, t1, generator):
+ """
+ Perform DDPM forward process from time t0 to t1. This is the same as adding noise with corresponding variance.
+
+ Args:
+ x_t0:
+ Latent code at time t0.
+ t0:
+ Timestep at t0.
+ t1:
+ Timestamp at t1.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+
+ Returns:
+ x_t1:
+ Forward process applied to x_t0 from time t0 to t1.
+ """
+ eps = randn_tensor(x_t0.size(), generator=generator, dtype=x_t0.dtype, device=x_t0.device)
+ alpha_vec = torch.prod(self.scheduler.alphas[t0:t1])
+ x_t1 = torch.sqrt(alpha_vec) * x_t0 + torch.sqrt(1 - alpha_vec) * eps
+ return x_t1
+
+ def backward_loop(
+ self,
+ latents,
+ timesteps,
+ prompt_embeds,
+ guidance_scale,
+ callback,
+ callback_steps,
+ num_warmup_steps,
+ extra_step_kwargs,
+ cross_attention_kwargs=None,
+ ):
+ """
+ Perform backward process given list of time steps.
+
+ Args:
+ latents:
+ Latents at time timesteps[0].
+ timesteps:
+ Time steps along which to perform backward process.
+ prompt_embeds:
+ Pre-generated text embeddings.
+ guidance_scale:
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ extra_step_kwargs:
+ Extra_step_kwargs.
+ cross_attention_kwargs:
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ num_warmup_steps:
+ number of warmup steps.
+
+ Returns:
+ latents:
+ Latents of backward process output at time timesteps[-1].
+ """
+ do_classifier_free_guidance = guidance_scale > 1.0
+ num_steps = (len(timesteps) - num_warmup_steps) // self.scheduler.order
+ with self.progress_bar(total=num_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ ).sample
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+ return latents.clone().detach()
+
+ # Copied from diffusers.pipelines.stable_diffusion_k_diffusion.pipeline_stable_diffusion_k_diffusion.StableDiffusionKDiffusionPipeline.check_inputs
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ video_length: Optional[int] = 8,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_videos_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ motion_field_strength_x: float = 12,
+ motion_field_strength_y: float = 12,
+ output_type: Optional[str] = "tensor",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: Optional[int] = 1,
+ t0: int = 44,
+ t1: int = 47,
+ frame_ids: Optional[List[int]] = None,
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ video_length (`int`, *optional*, defaults to 8):
+ The number of generated video frames.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in video generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_videos_per_prompt (`int`, *optional*, defaults to 1):
+ The number of videos to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"numpy"`):
+ The output format of the generated video. Choose between `"latent"` and `"numpy"`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a
+ [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoPipelineOutput`] instead of
+ a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ motion_field_strength_x (`float`, *optional*, defaults to 12):
+ Strength of motion in generated video along x-axis. See the [paper](https://arxiv.org/abs/2303.13439),
+ Sect. 3.3.1.
+ motion_field_strength_y (`float`, *optional*, defaults to 12):
+ Strength of motion in generated video along y-axis. See the [paper](https://arxiv.org/abs/2303.13439),
+ Sect. 3.3.1.
+ t0 (`int`, *optional*, defaults to 44):
+ Timestep t0. Should be in the range [0, num_inference_steps - 1]. See the
+ [paper](https://arxiv.org/abs/2303.13439), Sect. 3.3.1.
+ t1 (`int`, *optional*, defaults to 47):
+ Timestep t0. Should be in the range [t0 + 1, num_inference_steps - 1]. See the
+ [paper](https://arxiv.org/abs/2303.13439), Sect. 3.3.1.
+ frame_ids (`List[int]`, *optional*):
+ Indexes of the frames that are being generated. This is used when generating longer videos
+ chunk-by-chunk.
+
+ Returns:
+ [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoPipelineOutput`]:
+ The output contains a `ndarray` of the generated video, when `output_type` != `"latent"`, otherwise a
+ latent code of generated videos and a list of `bool`s indicating whether the corresponding generated
+ video contains "not-safe-for-work" (nsfw) content..
+ """
+ assert video_length > 0
+ if frame_ids is None:
+ frame_ids = list(range(video_length))
+ assert len(frame_ids) == video_length
+
+ assert num_videos_per_prompt == 1
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+
+ # Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+ # Check inputs. Raise error if not correct
+ self.check_inputs(prompt, height, width, callback_steps)
+
+ # Define call parameters
+ batch_size = 1 if isinstance(prompt, str) else len(prompt)
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # Encode input prompt
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt, device, num_videos_per_prompt, do_classifier_free_guidance, negative_prompt
+ )
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ # Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_videos_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+ # Prepare extra step kwargs.
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+
+ # Perform the first backward process up to time T_1
+ x_1_t1 = self.backward_loop(
+ timesteps=timesteps[: -t1 - 1],
+ prompt_embeds=prompt_embeds,
+ latents=latents,
+ guidance_scale=guidance_scale,
+ callback=callback,
+ callback_steps=callback_steps,
+ extra_step_kwargs=extra_step_kwargs,
+ num_warmup_steps=num_warmup_steps,
+ )
+ scheduler_copy = copy.deepcopy(self.scheduler)
+
+ # Perform the second backward process up to time T_0
+ x_1_t0 = self.backward_loop(
+ timesteps=timesteps[-t1 - 1 : -t0 - 1],
+ prompt_embeds=prompt_embeds,
+ latents=x_1_t1,
+ guidance_scale=guidance_scale,
+ callback=callback,
+ callback_steps=callback_steps,
+ extra_step_kwargs=extra_step_kwargs,
+ num_warmup_steps=0,
+ )
+
+ # Propagate first frame latents at time T_0 to remaining frames
+ x_2k_t0 = x_1_t0.repeat(video_length - 1, 1, 1, 1)
+
+ # Add motion in latents at time T_0
+ x_2k_t0 = create_motion_field_and_warp_latents(
+ motion_field_strength_x=motion_field_strength_x,
+ motion_field_strength_y=motion_field_strength_y,
+ latents=x_2k_t0,
+ frame_ids=frame_ids[1:],
+ )
+
+ # Perform forward process up to time T_1
+ x_2k_t1 = self.forward_loop(
+ x_t0=x_2k_t0,
+ t0=timesteps[-t0 - 1].item(),
+ t1=timesteps[-t1 - 1].item(),
+ generator=generator,
+ )
+
+ # Perform backward process from time T_1 to 0
+ x_1k_t1 = torch.cat([x_1_t1, x_2k_t1])
+ b, l, d = prompt_embeds.size()
+ prompt_embeds = prompt_embeds[:, None].repeat(1, video_length, 1, 1).reshape(b * video_length, l, d)
+
+ self.scheduler = scheduler_copy
+ x_1k_0 = self.backward_loop(
+ timesteps=timesteps[-t1 - 1 :],
+ prompt_embeds=prompt_embeds,
+ latents=x_1k_t1,
+ guidance_scale=guidance_scale,
+ callback=callback,
+ callback_steps=callback_steps,
+ extra_step_kwargs=extra_step_kwargs,
+ num_warmup_steps=0,
+ )
+ latents = x_1k_0
+
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if output_type == "latent":
+ image = latents
+ has_nsfw_concept = None
+ else:
+ image = self.decode_latents(latents)
+ # Run safety checker
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return TextToVideoPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def decode_latents(self, latents):
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
diff --git a/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py
new file mode 100644
index 0000000000000000000000000000000000000000..eaa2760363a93ff3c0603250a2c4bd2c3acacde9
--- /dev/null
+++ b/model/diffusers_c/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero_sdxl.py
@@ -0,0 +1,1315 @@
+import copy
+import inspect
+from dataclasses import dataclass
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL
+import torch
+import torch.nn.functional as F
+from torch.nn.functional import grid_sample
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, UNet2DConditionModel
+from ...models.attention_processor import (
+ AttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ LoRAAttnProcessor2_0,
+ LoRAXFormersAttnProcessor,
+ XFormersAttnProcessor,
+)
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ BaseOutput,
+ is_invisible_watermark_available,
+ logging,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+
+
+if is_invisible_watermark_available():
+ from ..stable_diffusion_xl.watermark import StableDiffusionXLWatermarker
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_0
+def rearrange_0(tensor, f):
+ F, C, H, W = tensor.size()
+ tensor = torch.permute(torch.reshape(tensor, (F // f, f, C, H, W)), (0, 2, 1, 3, 4))
+ return tensor
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_1
+def rearrange_1(tensor):
+ B, C, F, H, W = tensor.size()
+ return torch.reshape(torch.permute(tensor, (0, 2, 1, 3, 4)), (B * F, C, H, W))
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_3
+def rearrange_3(tensor, f):
+ F, D, C = tensor.size()
+ return torch.reshape(tensor, (F // f, f, D, C))
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.rearrange_4
+def rearrange_4(tensor):
+ B, F, D, C = tensor.size()
+ return torch.reshape(tensor, (B * F, D, C))
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor
+class CrossFrameAttnProcessor:
+ """
+ Cross frame attention processor. Each frame attends the first frame.
+
+ Args:
+ batch_size: The number that represents actual batch size, other than the frames.
+ For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to
+ 2, due to classifier-free guidance.
+ """
+
+ def __init__(self, batch_size=2):
+ self.batch_size = batch_size
+
+ def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None):
+ batch_size, sequence_length, _ = hidden_states.shape
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ query = attn.to_q(hidden_states)
+
+ is_cross_attention = encoder_hidden_states is not None
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ # Cross Frame Attention
+ if not is_cross_attention:
+ video_length = key.size()[0] // self.batch_size
+ first_frame_index = [0] * video_length
+
+ # rearrange keys to have batch and frames in the 1st and 2nd dims respectively
+ key = rearrange_3(key, video_length)
+ key = key[:, first_frame_index]
+ # rearrange values to have batch and frames in the 1st and 2nd dims respectively
+ value = rearrange_3(value, video_length)
+ value = value[:, first_frame_index]
+
+ # rearrange back to original shape
+ key = rearrange_4(key)
+ value = rearrange_4(value)
+
+ query = attn.head_to_batch_dim(query)
+ key = attn.head_to_batch_dim(key)
+ value = attn.head_to_batch_dim(value)
+
+ attention_probs = attn.get_attention_scores(query, key, attention_mask)
+ hidden_states = torch.bmm(attention_probs, value)
+ hidden_states = attn.batch_to_head_dim(hidden_states)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+
+ return hidden_states
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.CrossFrameAttnProcessor2_0
+class CrossFrameAttnProcessor2_0:
+ """
+ Cross frame attention processor with scaled_dot_product attention of Pytorch 2.0.
+
+ Args:
+ batch_size: The number that represents actual batch size, other than the frames.
+ For example, calling unet with a single prompt and num_images_per_prompt=1, batch_size should be equal to
+ 2, due to classifier-free guidance.
+ """
+
+ def __init__(self, batch_size=2):
+ if not hasattr(F, "scaled_dot_product_attention"):
+ raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+ self.batch_size = batch_size
+
+ def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None):
+ batch_size, sequence_length, _ = (
+ hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+ )
+ inner_dim = hidden_states.shape[-1]
+
+ if attention_mask is not None:
+ attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+ # scaled_dot_product_attention expects attention_mask shape to be
+ # (batch, heads, source_length, target_length)
+ attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+ query = attn.to_q(hidden_states)
+
+ is_cross_attention = encoder_hidden_states is not None
+ if encoder_hidden_states is None:
+ encoder_hidden_states = hidden_states
+ elif attn.norm_cross:
+ encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+ key = attn.to_k(encoder_hidden_states)
+ value = attn.to_v(encoder_hidden_states)
+
+ # Cross Frame Attention
+ if not is_cross_attention:
+ video_length = max(1, key.size()[0] // self.batch_size)
+ first_frame_index = [0] * video_length
+
+ # rearrange keys to have batch and frames in the 1st and 2nd dims respectively
+ key = rearrange_3(key, video_length)
+ key = key[:, first_frame_index]
+ # rearrange values to have batch and frames in the 1st and 2nd dims respectively
+ value = rearrange_3(value, video_length)
+ value = value[:, first_frame_index]
+
+ # rearrange back to original shape
+ key = rearrange_4(key)
+ value = rearrange_4(value)
+
+ head_dim = inner_dim // attn.heads
+ query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+ value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+ # the output of sdp = (batch, num_heads, seq_len, head_dim)
+ # TODO: add support for attn.scale when we move to Torch 2.1
+ hidden_states = F.scaled_dot_product_attention(
+ query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+ )
+
+ hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+ hidden_states = hidden_states.to(query.dtype)
+
+ # linear proj
+ hidden_states = attn.to_out[0](hidden_states)
+ # dropout
+ hidden_states = attn.to_out[1](hidden_states)
+ return hidden_states
+
+
+@dataclass
+class TextToVideoSDXLPipelineOutput(BaseOutput):
+ """
+ Output class for zero-shot text-to-video pipeline.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+ num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+ """
+
+ images: Union[List[PIL.Image.Image], np.ndarray]
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.coords_grid
+def coords_grid(batch, ht, wd, device):
+ # Adapted from https://github.com/princeton-vl/RAFT/blob/master/core/utils/utils.py
+ coords = torch.meshgrid(torch.arange(ht, device=device), torch.arange(wd, device=device))
+ coords = torch.stack(coords[::-1], dim=0).float()
+ return coords[None].repeat(batch, 1, 1, 1)
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.warp_single_latent
+def warp_single_latent(latent, reference_flow):
+ """
+ Warp latent of a single frame with given flow
+
+ Args:
+ latent: latent code of a single frame
+ reference_flow: flow which to warp the latent with
+
+ Returns:
+ warped: warped latent
+ """
+ _, _, H, W = reference_flow.size()
+ _, _, h, w = latent.size()
+ coords0 = coords_grid(1, H, W, device=latent.device).to(latent.dtype)
+
+ coords_t0 = coords0 + reference_flow
+ coords_t0[:, 0] /= W
+ coords_t0[:, 1] /= H
+
+ coords_t0 = coords_t0 * 2.0 - 1.0
+ coords_t0 = F.interpolate(coords_t0, size=(h, w), mode="bilinear")
+ coords_t0 = torch.permute(coords_t0, (0, 2, 3, 1))
+
+ warped = grid_sample(latent, coords_t0, mode="nearest", padding_mode="reflection")
+ return warped
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.create_motion_field
+def create_motion_field(motion_field_strength_x, motion_field_strength_y, frame_ids, device, dtype):
+ """
+ Create translation motion field
+
+ Args:
+ motion_field_strength_x: motion strength along x-axis
+ motion_field_strength_y: motion strength along y-axis
+ frame_ids: indexes of the frames the latents of which are being processed.
+ This is needed when we perform chunk-by-chunk inference
+ device: device
+ dtype: dtype
+
+ Returns:
+
+ """
+ seq_length = len(frame_ids)
+ reference_flow = torch.zeros((seq_length, 2, 512, 512), device=device, dtype=dtype)
+ for fr_idx in range(seq_length):
+ reference_flow[fr_idx, 0, :, :] = motion_field_strength_x * (frame_ids[fr_idx])
+ reference_flow[fr_idx, 1, :, :] = motion_field_strength_y * (frame_ids[fr_idx])
+ return reference_flow
+
+
+# Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.create_motion_field_and_warp_latents
+def create_motion_field_and_warp_latents(motion_field_strength_x, motion_field_strength_y, frame_ids, latents):
+ """
+ Creates translation motion and warps the latents accordingly
+
+ Args:
+ motion_field_strength_x: motion strength along x-axis
+ motion_field_strength_y: motion strength along y-axis
+ frame_ids: indexes of the frames the latents of which are being processed.
+ This is needed when we perform chunk-by-chunk inference
+ latents: latent codes of frames
+
+ Returns:
+ warped_latents: warped latents
+ """
+ motion_field = create_motion_field(
+ motion_field_strength_x=motion_field_strength_x,
+ motion_field_strength_y=motion_field_strength_y,
+ frame_ids=frame_ids,
+ device=latents.device,
+ dtype=latents.dtype,
+ )
+ warped_latents = latents.clone().detach()
+ for i in range(len(warped_latents)):
+ warped_latents[i] = warp_single_latent(latents[i][None], motion_field[i][None])
+ return warped_latents
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+ """
+ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+ """
+ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+ std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+ # rescale the results from guidance (fixes overexposure)
+ noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+ # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+ noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+ return noise_cfg
+
+
+class TextToVideoZeroSDXLPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ StableDiffusionXLLoraLoaderMixin,
+ TextualInversionLoaderMixin,
+):
+ r"""
+ Pipeline for zero-shot text-to-video generation using Stable Diffusion XL.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion XL uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ text_encoder_2 ([` CLIPTextModelWithProjection`]):
+ Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
+ specifically the
+ [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
+ variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ tokenizer_2 (`CLIPTokenizer`):
+ Second Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
+ _optional_components = [
+ "tokenizer",
+ "tokenizer_2",
+ "text_encoder",
+ "text_encoder_2",
+ "image_encoder",
+ "feature_extractor",
+ ]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_2: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ tokenizer_2: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ feature_extractor: CLIPImageProcessor = None,
+ force_zeros_for_empty_prompt: bool = True,
+ add_watermarker: Optional[bool] = None,
+ ):
+ super().__init__()
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_2=text_encoder_2,
+ tokenizer=tokenizer,
+ tokenizer_2=tokenizer_2,
+ unet=unet,
+ scheduler=scheduler,
+ image_encoder=image_encoder,
+ feature_extractor=feature_extractor,
+ )
+ self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ self.default_sample_size = self.unet.config.sample_size
+
+ add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()
+
+ if add_watermarker:
+ self.watermark = StableDiffusionXLWatermarker()
+ else:
+ self.watermark = None
+
+ processor = (
+ CrossFrameAttnProcessor2_0(batch_size=2)
+ if hasattr(F, "scaled_dot_product_attention")
+ else CrossFrameAttnProcessor(batch_size=2)
+ )
+
+ self.unet.set_attn_processor(processor)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.upcast_vae
+ def upcast_vae(self):
+ dtype = self.vae.dtype
+ self.vae.to(dtype=torch.float32)
+ use_torch_2_0_or_xformers = isinstance(
+ self.vae.decoder.mid_block.attentions[0].processor,
+ (
+ AttnProcessor2_0,
+ XFormersAttnProcessor,
+ LoRAXFormersAttnProcessor,
+ LoRAAttnProcessor2_0,
+ FusedAttnProcessor2_0,
+ ),
+ )
+ # if xformers or torch_2_0 is used attention block does not need
+ # to be in float32 which can save lots of memory
+ if use_torch_2_0_or_xformers:
+ self.vae.post_quant_conv.to(dtype)
+ self.vae.decoder.conv_in.to(dtype)
+ self.vae.decoder.mid_block.to(dtype)
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids
+ def _get_add_time_ids(
+ self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
+ ):
+ add_time_ids = list(original_size + crops_coords_top_left + target_size)
+
+ passed_add_embed_dim = (
+ self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
+ )
+ expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features
+
+ if expected_add_embed_dim != passed_add_embed_dim:
+ raise ValueError(
+ f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."
+ )
+
+ add_time_ids = torch.tensor([add_time_ids], dtype=dtype)
+ return add_time_ids
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def check_inputs(
+ self,
+ prompt,
+ prompt_2,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ negative_prompt_2=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ pooled_prompt_embeds=None,
+ negative_pooled_prompt_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt_2 is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
+ raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+ elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if prompt_embeds is not None and pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
+ )
+
+ if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
+ raise ValueError(
+ "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt: str,
+ prompt_2: Optional[str] = None,
+ device: Optional[torch.device] = None,
+ num_images_per_prompt: int = 1,
+ do_classifier_free_guidance: bool = True,
+ negative_prompt: Optional[str] = None,
+ negative_prompt_2: Optional[str] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ device = device or self._execution_device
+
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if self.text_encoder is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder_2, lora_scale)
+
+ prompt = [prompt] if isinstance(prompt, str) else prompt
+
+ if prompt is not None:
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # Define tokenizers and text encoders
+ tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
+ text_encoders = (
+ [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
+ )
+
+ if prompt_embeds is None:
+ prompt_2 = prompt_2 or prompt
+ prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+
+ # textual inversion: process multi-vector tokens if necessary
+ prompt_embeds_list = []
+ prompts = [prompt, prompt_2]
+ for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, tokenizer)
+
+ text_inputs = tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True)
+
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ pooled_prompt_embeds = prompt_embeds[0]
+ if clip_skip is None:
+ prompt_embeds = prompt_embeds.hidden_states[-2]
+ else:
+ # "2" because SDXL always indexes from the penultimate layer.
+ prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)]
+
+ prompt_embeds_list.append(prompt_embeds)
+
+ prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+
+ # get unconditional embeddings for classifier free guidance
+ zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt
+ if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt:
+ negative_prompt_embeds = torch.zeros_like(prompt_embeds)
+ negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds)
+ elif do_classifier_free_guidance and negative_prompt_embeds is None:
+ negative_prompt = negative_prompt or ""
+ negative_prompt_2 = negative_prompt_2 or negative_prompt
+
+ # normalize str to list
+ negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
+ negative_prompt_2 = (
+ batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
+ )
+
+ uncond_tokens: List[str]
+ if prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = [negative_prompt, negative_prompt_2]
+
+ negative_prompt_embeds_list = []
+ for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
+ if isinstance(self, TextualInversionLoaderMixin):
+ negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = tokenizer(
+ negative_prompt,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ negative_prompt_embeds = text_encoder(
+ uncond_input.input_ids.to(device),
+ output_hidden_states=True,
+ )
+ # We are only ALWAYS interested in the pooled output of the final text encoder
+ negative_pooled_prompt_embeds = negative_prompt_embeds[0]
+ negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2]
+
+ negative_prompt_embeds_list.append(negative_prompt_embeds)
+
+ negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1)
+
+ if self.text_encoder_2 is not None:
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ if self.text_encoder_2 is not None:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device)
+ else:
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+ if do_classifier_free_guidance:
+ negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view(
+ bs_embed * num_images_per_prompt, -1
+ )
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ if self.text_encoder_2 is not None:
+ if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder_2, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoZeroPipeline.forward_loop
+ def forward_loop(self, x_t0, t0, t1, generator):
+ """
+ Perform DDPM forward process from time t0 to t1. This is the same as adding noise with corresponding variance.
+
+ Args:
+ x_t0:
+ Latent code at time t0.
+ t0:
+ Timestep at t0.
+ t1:
+ Timestamp at t1.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+
+ Returns:
+ x_t1:
+ Forward process applied to x_t0 from time t0 to t1.
+ """
+ eps = randn_tensor(x_t0.size(), generator=generator, dtype=x_t0.dtype, device=x_t0.device)
+ alpha_vec = torch.prod(self.scheduler.alphas[t0:t1])
+ x_t1 = torch.sqrt(alpha_vec) * x_t0 + torch.sqrt(1 - alpha_vec) * eps
+ return x_t1
+
+ def backward_loop(
+ self,
+ latents,
+ timesteps,
+ prompt_embeds,
+ guidance_scale,
+ callback,
+ callback_steps,
+ num_warmup_steps,
+ extra_step_kwargs,
+ add_text_embeds,
+ add_time_ids,
+ cross_attention_kwargs=None,
+ guidance_rescale: float = 0.0,
+ ):
+ """
+ Perform backward process given list of time steps
+
+ Args:
+ latents:
+ Latents at time timesteps[0].
+ timesteps:
+ Time steps along which to perform backward process.
+ prompt_embeds:
+ Pre-generated text embeddings.
+ guidance_scale:
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ extra_step_kwargs:
+ Extra_step_kwargs.
+ cross_attention_kwargs:
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ num_warmup_steps:
+ number of warmup steps.
+
+ Returns:
+ latents: latents of backward process output at time timesteps[-1]
+ """
+
+ do_classifier_free_guidance = guidance_scale > 1.0
+ num_steps = (len(timesteps) - num_warmup_steps) // self.scheduler.order
+
+ with self.progress_bar(total=num_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # predict the noise residual
+ added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ if do_classifier_free_guidance and guidance_rescale > 0.0:
+ # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+ noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, latents)
+ return latents.clone().detach()
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Union[str, List[str]],
+ prompt_2: Optional[Union[str, List[str]]] = None,
+ video_length: Optional[int] = 8,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ denoising_end: Optional[float] = None,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ negative_prompt_2: Optional[Union[str, List[str]]] = None,
+ num_videos_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ frame_ids: Optional[List[int]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ motion_field_strength_x: float = 12,
+ motion_field_strength_y: float = 12,
+ output_type: Optional[str] = "tensor",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ guidance_rescale: float = 0.0,
+ original_size: Optional[Tuple[int, int]] = None,
+ crops_coords_top_left: Tuple[int, int] = (0, 0),
+ target_size: Optional[Tuple[int, int]] = None,
+ t0: int = 44,
+ t1: int = 47,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+ used in both text-encoders
+ video_length (`int`, *optional*, defaults to 8):
+ The number of generated video frames.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ denoising_end (`float`, *optional*):
+ When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
+ completed before it is intentionally prematurely terminated. As a result, the returned sample will
+ still retain a substantial amount of noise as determined by the discrete timesteps selected by the
+ scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
+ "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
+ Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output)
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ negative_prompt_2 (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
+ `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
+ num_videos_per_prompt (`int`, *optional*, defaults to 1):
+ The number of videos to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ frame_ids (`List[int]`, *optional*):
+ Indexes of the frames that are being generated. This is used when generating longer videos
+ chunk-by-chunk.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+ If not provided, pooled text embeddings will be generated from `prompt` input argument.
+ negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ motion_field_strength_x (`float`, *optional*, defaults to 12):
+ Strength of motion in generated video along x-axis. See the [paper](https://arxiv.org/abs/2303.13439),
+ Sect. 3.3.1.
+ motion_field_strength_y (`float`, *optional*, defaults to 12):
+ Strength of motion in generated video along y-axis. See the [paper](https://arxiv.org/abs/2303.13439),
+ Sect. 3.3.1.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
+ of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+ guidance_rescale (`float`, *optional*, defaults to 0.7):
+ Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
+ Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
+ [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
+ Guidance rescale factor should fix overexposure when using zero terminal SNR.
+ original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled.
+ `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as
+ explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
+ `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
+ `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
+ `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
+ [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
+ For most cases, `target_size` should be set to the desired height and width of the generated image. If
+ not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
+ section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
+ t0 (`int`, *optional*, defaults to 44):
+ Timestep t0. Should be in the range [0, num_inference_steps - 1]. See the
+ [paper](https://arxiv.org/abs/2303.13439), Sect. 3.3.1.
+ t1 (`int`, *optional*, defaults to 47):
+ Timestep t0. Should be in the range [t0 + 1, num_inference_steps - 1]. See the
+ [paper](https://arxiv.org/abs/2303.13439), Sect. 3.3.1.
+
+ Returns:
+ [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoSDXLPipelineOutput`] or
+ `tuple`: [`~pipelines.text_to_video_synthesis.pipeline_text_to_video_zero.TextToVideoSDXLPipelineOutput`]
+ if `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
+ generated images.
+ """
+ assert video_length > 0
+ if frame_ids is None:
+ frame_ids = list(range(video_length))
+ assert len(frame_ids) == video_length
+
+ assert num_videos_per_prompt == 1
+
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+
+ # 0. Default height and width to unet
+ height = height or self.default_sample_size * self.vae_scale_factor
+ width = width or self.default_sample_size * self.vae_scale_factor
+
+ original_size = original_size or (height, width)
+ target_size = target_size or (height, width)
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ prompt_2,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ negative_prompt_2,
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ batch_size = (
+ 1 if isinstance(prompt, str) else len(prompt) if isinstance(prompt, list) else prompt_embeds.shape[0]
+ )
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ (
+ prompt_embeds,
+ negative_prompt_embeds,
+ pooled_prompt_embeds,
+ negative_pooled_prompt_embeds,
+ ) = self.encode_prompt(
+ prompt=prompt,
+ prompt_2=prompt_2,
+ device=device,
+ num_images_per_prompt=num_videos_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ negative_prompt_2=negative_prompt_2,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ pooled_prompt_embeds=pooled_prompt_embeds,
+ negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ )
+
+ # 4. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+
+ latents = self.prepare_latents(
+ batch_size * num_videos_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7. Prepare added time ids & embeddings
+ add_text_embeds = pooled_prompt_embeds
+ if self.text_encoder_2 is None:
+ text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1])
+ else:
+ text_encoder_projection_dim = self.text_encoder_2.config.projection_dim
+
+ add_time_ids = self._get_add_time_ids(
+ original_size,
+ crops_coords_top_left,
+ target_size,
+ dtype=prompt_embeds.dtype,
+ text_encoder_projection_dim=text_encoder_projection_dim,
+ )
+
+ if do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
+ add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0)
+ add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0)
+
+ prompt_embeds = prompt_embeds.to(device)
+ add_text_embeds = add_text_embeds.to(device)
+ add_time_ids = add_time_ids.to(device).repeat(batch_size * num_videos_per_prompt, 1)
+
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+
+ # Perform the first backward process up to time T_1
+ x_1_t1 = self.backward_loop(
+ timesteps=timesteps[: -t1 - 1],
+ prompt_embeds=prompt_embeds,
+ latents=latents,
+ guidance_scale=guidance_scale,
+ callback=callback,
+ callback_steps=callback_steps,
+ extra_step_kwargs=extra_step_kwargs,
+ num_warmup_steps=num_warmup_steps,
+ add_text_embeds=add_text_embeds,
+ add_time_ids=add_time_ids,
+ )
+
+ scheduler_copy = copy.deepcopy(self.scheduler)
+
+ # Perform the second backward process up to time T_0
+ x_1_t0 = self.backward_loop(
+ timesteps=timesteps[-t1 - 1 : -t0 - 1],
+ prompt_embeds=prompt_embeds,
+ latents=x_1_t1,
+ guidance_scale=guidance_scale,
+ callback=callback,
+ callback_steps=callback_steps,
+ extra_step_kwargs=extra_step_kwargs,
+ num_warmup_steps=0,
+ add_text_embeds=add_text_embeds,
+ add_time_ids=add_time_ids,
+ )
+
+ # Propagate first frame latents at time T_0 to remaining frames
+ x_2k_t0 = x_1_t0.repeat(video_length - 1, 1, 1, 1)
+
+ # Add motion in latents at time T_0
+ x_2k_t0 = create_motion_field_and_warp_latents(
+ motion_field_strength_x=motion_field_strength_x,
+ motion_field_strength_y=motion_field_strength_y,
+ latents=x_2k_t0,
+ frame_ids=frame_ids[1:],
+ )
+
+ # Perform forward process up to time T_1
+ x_2k_t1 = self.forward_loop(
+ x_t0=x_2k_t0,
+ t0=timesteps[-t0 - 1].to(torch.long),
+ t1=timesteps[-t1 - 1].to(torch.long),
+ generator=generator,
+ )
+
+ # Perform backward process from time T_1 to 0
+ latents = torch.cat([x_1_t1, x_2k_t1])
+
+ self.scheduler = scheduler_copy
+ timesteps = timesteps[-t1 - 1 :]
+
+ b, l, d = prompt_embeds.size()
+ prompt_embeds = prompt_embeds[:, None].repeat(1, video_length, 1, 1).reshape(b * video_length, l, d)
+
+ b, k = add_text_embeds.size()
+ add_text_embeds = add_text_embeds[:, None].repeat(1, video_length, 1).reshape(b * video_length, k)
+
+ b, k = add_time_ids.size()
+ add_time_ids = add_time_ids[:, None].repeat(1, video_length, 1).reshape(b * video_length, k)
+
+ # 7.1 Apply denoising_end
+ if denoising_end is not None and isinstance(denoising_end, float) and denoising_end > 0 and denoising_end < 1:
+ discrete_timestep_cutoff = int(
+ round(
+ self.scheduler.config.num_train_timesteps
+ - (denoising_end * self.scheduler.config.num_train_timesteps)
+ )
+ )
+ num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
+ timesteps = timesteps[:num_inference_steps]
+
+ x_1k_0 = self.backward_loop(
+ timesteps=timesteps,
+ prompt_embeds=prompt_embeds,
+ latents=latents,
+ guidance_scale=guidance_scale,
+ callback=callback,
+ callback_steps=callback_steps,
+ extra_step_kwargs=extra_step_kwargs,
+ num_warmup_steps=0,
+ add_text_embeds=add_text_embeds,
+ add_time_ids=add_time_ids,
+ )
+
+ latents = x_1k_0
+
+ if not output_type == "latent":
+ # make sure the VAE is in float32 mode, as it overflows in float16
+ needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast
+
+ if needs_upcasting:
+ self.upcast_vae()
+ latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype)
+
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+
+ # cast back to fp16 if needed
+ if needs_upcasting:
+ self.vae.to(dtype=torch.float16)
+ else:
+ image = latents
+ return TextToVideoSDXLPipelineOutput(images=image)
+
+ # apply watermark if available
+ if self.watermark is not None:
+ image = self.watermark.apply_watermark(image)
+
+ image = self.image_processor.postprocess(image, output_type=output_type)
+
+ # Offload last model to CPU manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.final_offload_hook.offload()
+
+ if not return_dict:
+ return (image,)
+
+ return TextToVideoSDXLPipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/unclip/__init__.py b/model/diffusers_c/pipelines/unclip/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..c89e899463beede59b8ccf02688f6168b8ee3d77
--- /dev/null
+++ b/model/diffusers_c/pipelines/unclip/__init__.py
@@ -0,0 +1,52 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ is_torch_available,
+ is_transformers_available,
+ is_transformers_version,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import UnCLIPImageVariationPipeline, UnCLIPPipeline
+
+ _dummy_objects.update(
+ {"UnCLIPImageVariationPipeline": UnCLIPImageVariationPipeline, "UnCLIPPipeline": UnCLIPPipeline}
+ )
+else:
+ _import_structure["pipeline_unclip"] = ["UnCLIPPipeline"]
+ _import_structure["pipeline_unclip_image_variation"] = ["UnCLIPImageVariationPipeline"]
+ _import_structure["text_proj"] = ["UnCLIPTextProjModel"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available() and is_transformers_version(">=", "4.25.0")):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import * # noqa F403
+ else:
+ from .pipeline_unclip import UnCLIPPipeline
+ from .pipeline_unclip_image_variation import UnCLIPImageVariationPipeline
+ from .text_proj import UnCLIPTextProjModel
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/unclip/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/unclip/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..bf287572e92adad569a5049c4f2fdf10d9e1df6a
Binary files /dev/null and b/model/diffusers_c/pipelines/unclip/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/unclip/__pycache__/pipeline_unclip.cpython-39.pyc b/model/diffusers_c/pipelines/unclip/__pycache__/pipeline_unclip.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..16b211bb760fd6abd52accbbe10d336651d11398
Binary files /dev/null and b/model/diffusers_c/pipelines/unclip/__pycache__/pipeline_unclip.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/unclip/__pycache__/pipeline_unclip_image_variation.cpython-39.pyc b/model/diffusers_c/pipelines/unclip/__pycache__/pipeline_unclip_image_variation.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..70b3713dc88822fa1f3f02a4445dd2be94b69e95
Binary files /dev/null and b/model/diffusers_c/pipelines/unclip/__pycache__/pipeline_unclip_image_variation.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/unclip/__pycache__/text_proj.cpython-39.pyc b/model/diffusers_c/pipelines/unclip/__pycache__/text_proj.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d03fb9fb764c6fc20f15d4e2b055a82225340038
Binary files /dev/null and b/model/diffusers_c/pipelines/unclip/__pycache__/text_proj.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/unclip/pipeline_unclip.py b/model/diffusers_c/pipelines/unclip/pipeline_unclip.py
new file mode 100644
index 0000000000000000000000000000000000000000..72e5b311394e954e75080370e4206bd254386115
--- /dev/null
+++ b/model/diffusers_c/pipelines/unclip/pipeline_unclip.py
@@ -0,0 +1,493 @@
+# Copyright 2024 Kakao Brain and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import List, Optional, Tuple, Union
+
+import torch
+from torch.nn import functional as F
+from transformers import CLIPTextModelWithProjection, CLIPTokenizer
+from transformers.models.clip.modeling_clip import CLIPTextModelOutput
+
+from ...models import PriorTransformer, UNet2DConditionModel, UNet2DModel
+from ...schedulers import UnCLIPScheduler
+from ...utils import logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .text_proj import UnCLIPTextProjModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class UnCLIPPipeline(DiffusionPipeline):
+ """
+ Pipeline for text-to-image generation using unCLIP.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ text_encoder ([`~transformers.CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ prior ([`PriorTransformer`]):
+ The canonical unCLIP prior to approximate the image embedding from the text embedding.
+ text_proj ([`UnCLIPTextProjModel`]):
+ Utility class to prepare and combine the embeddings before they are passed to the decoder.
+ decoder ([`UNet2DConditionModel`]):
+ The decoder to invert the image embedding into an image.
+ super_res_first ([`UNet2DModel`]):
+ Super resolution UNet. Used in all but the last step of the super resolution diffusion process.
+ super_res_last ([`UNet2DModel`]):
+ Super resolution UNet. Used in the last step of the super resolution diffusion process.
+ prior_scheduler ([`UnCLIPScheduler`]):
+ Scheduler used in the prior denoising process (a modified [`DDPMScheduler`]).
+ decoder_scheduler ([`UnCLIPScheduler`]):
+ Scheduler used in the decoder denoising process (a modified [`DDPMScheduler`]).
+ super_res_scheduler ([`UnCLIPScheduler`]):
+ Scheduler used in the super resolution denoising process (a modified [`DDPMScheduler`]).
+
+ """
+
+ _exclude_from_cpu_offload = ["prior"]
+
+ prior: PriorTransformer
+ decoder: UNet2DConditionModel
+ text_proj: UnCLIPTextProjModel
+ text_encoder: CLIPTextModelWithProjection
+ tokenizer: CLIPTokenizer
+ super_res_first: UNet2DModel
+ super_res_last: UNet2DModel
+
+ prior_scheduler: UnCLIPScheduler
+ decoder_scheduler: UnCLIPScheduler
+ super_res_scheduler: UnCLIPScheduler
+
+ model_cpu_offload_seq = "text_encoder->text_proj->decoder->super_res_first->super_res_last"
+
+ def __init__(
+ self,
+ prior: PriorTransformer,
+ decoder: UNet2DConditionModel,
+ text_encoder: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ text_proj: UnCLIPTextProjModel,
+ super_res_first: UNet2DModel,
+ super_res_last: UNet2DModel,
+ prior_scheduler: UnCLIPScheduler,
+ decoder_scheduler: UnCLIPScheduler,
+ super_res_scheduler: UnCLIPScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ prior=prior,
+ decoder=decoder,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ text_proj=text_proj,
+ super_res_first=super_res_first,
+ super_res_last=super_res_last,
+ prior_scheduler=prior_scheduler,
+ decoder_scheduler=decoder_scheduler,
+ super_res_scheduler=super_res_scheduler,
+ )
+
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None,
+ text_attention_mask: Optional[torch.Tensor] = None,
+ ):
+ if text_model_output is None:
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ text_mask = text_inputs.attention_mask.bool().to(device)
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+
+ text_encoder_output = self.text_encoder(text_input_ids.to(device))
+
+ prompt_embeds = text_encoder_output.text_embeds
+ text_enc_hid_states = text_encoder_output.last_hidden_state
+
+ else:
+ batch_size = text_model_output[0].shape[0]
+ prompt_embeds, text_enc_hid_states = text_model_output[0], text_model_output[1]
+ text_mask = text_attention_mask
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_enc_hid_states = text_enc_hid_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens = [""] * batch_size
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_text_mask = uncond_input.attention_mask.bool().to(device)
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device))
+
+ negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds
+ uncond_text_enc_hid_states = negative_prompt_embeds_text_encoder_output.last_hidden_state
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_enc_hid_states.shape[1]
+ uncond_text_enc_hid_states = uncond_text_enc_hid_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_enc_hid_states = uncond_text_enc_hid_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_enc_hid_states = torch.cat([uncond_text_enc_hid_states, text_enc_hid_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_enc_hid_states, text_mask
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: int = 1,
+ prior_num_inference_steps: int = 25,
+ decoder_num_inference_steps: int = 25,
+ super_res_num_inference_steps: int = 7,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ prior_latents: Optional[torch.FloatTensor] = None,
+ decoder_latents: Optional[torch.FloatTensor] = None,
+ super_res_latents: Optional[torch.FloatTensor] = None,
+ text_model_output: Optional[Union[CLIPTextModelOutput, Tuple]] = None,
+ text_attention_mask: Optional[torch.Tensor] = None,
+ prior_guidance_scale: float = 4.0,
+ decoder_guidance_scale: float = 8.0,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide image generation. This can only be left undefined if `text_model_output`
+ and `text_attention_mask` is passed.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ prior_num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps for the prior. More denoising steps usually lead to a higher quality
+ image at the expense of slower inference.
+ decoder_num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps for the decoder. More denoising steps usually lead to a higher quality
+ image at the expense of slower inference.
+ super_res_num_inference_steps (`int`, *optional*, defaults to 7):
+ The number of denoising steps for super resolution. More denoising steps usually lead to a higher
+ quality image at the expense of slower inference.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ prior_latents (`torch.FloatTensor` of shape (batch size, embeddings dimension), *optional*):
+ Pre-generated noisy latents to be used as inputs for the prior.
+ decoder_latents (`torch.FloatTensor` of shape (batch size, channels, height, width), *optional*):
+ Pre-generated noisy latents to be used as inputs for the decoder.
+ super_res_latents (`torch.FloatTensor` of shape (batch size, channels, super res height, super res width), *optional*):
+ Pre-generated noisy latents to be used as inputs for the decoder.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ decoder_guidance_scale (`float`, *optional*, defaults to 4.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ text_model_output (`CLIPTextModelOutput`, *optional*):
+ Pre-defined [`CLIPTextModel`] outputs that can be derived from the text encoder. Pre-defined text
+ outputs can be passed for tasks like text embedding interpolations. Make sure to also pass
+ `text_attention_mask` in this case. `prompt` can the be left `None`.
+ text_attention_mask (`torch.Tensor`, *optional*):
+ Pre-defined CLIP text attention mask that can be derived from the tokenizer. Pre-defined text attention
+ masks are necessary when passing `text_model_output`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+ if prompt is not None:
+ if isinstance(prompt, str):
+ batch_size = 1
+ elif isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+ else:
+ batch_size = text_model_output[0].shape[0]
+
+ device = self._execution_device
+
+ batch_size = batch_size * num_images_per_prompt
+
+ do_classifier_free_guidance = prior_guidance_scale > 1.0 or decoder_guidance_scale > 1.0
+
+ prompt_embeds, text_enc_hid_states, text_mask = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance, text_model_output, text_attention_mask
+ )
+
+ # prior
+
+ self.prior_scheduler.set_timesteps(prior_num_inference_steps, device=device)
+ prior_timesteps_tensor = self.prior_scheduler.timesteps
+
+ embedding_dim = self.prior.config.embedding_dim
+
+ prior_latents = self.prepare_latents(
+ (batch_size, embedding_dim),
+ prompt_embeds.dtype,
+ device,
+ generator,
+ prior_latents,
+ self.prior_scheduler,
+ )
+
+ for i, t in enumerate(self.progress_bar(prior_timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([prior_latents] * 2) if do_classifier_free_guidance else prior_latents
+
+ predicted_image_embedding = self.prior(
+ latent_model_input,
+ timestep=t,
+ proj_embedding=prompt_embeds,
+ encoder_hidden_states=text_enc_hid_states,
+ attention_mask=text_mask,
+ ).predicted_image_embedding
+
+ if do_classifier_free_guidance:
+ predicted_image_embedding_uncond, predicted_image_embedding_text = predicted_image_embedding.chunk(2)
+ predicted_image_embedding = predicted_image_embedding_uncond + prior_guidance_scale * (
+ predicted_image_embedding_text - predicted_image_embedding_uncond
+ )
+
+ if i + 1 == prior_timesteps_tensor.shape[0]:
+ prev_timestep = None
+ else:
+ prev_timestep = prior_timesteps_tensor[i + 1]
+
+ prior_latents = self.prior_scheduler.step(
+ predicted_image_embedding,
+ timestep=t,
+ sample=prior_latents,
+ generator=generator,
+ prev_timestep=prev_timestep,
+ ).prev_sample
+
+ prior_latents = self.prior.post_process_latents(prior_latents)
+
+ image_embeddings = prior_latents
+
+ # done prior
+
+ # decoder
+
+ text_enc_hid_states, additive_clip_time_embeddings = self.text_proj(
+ image_embeddings=image_embeddings,
+ prompt_embeds=prompt_embeds,
+ text_encoder_hidden_states=text_enc_hid_states,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ )
+
+ if device.type == "mps":
+ # HACK: MPS: There is a panic when padding bool tensors,
+ # so cast to int tensor for the pad and back to bool afterwards
+ text_mask = text_mask.type(torch.int)
+ decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=1)
+ decoder_text_mask = decoder_text_mask.type(torch.bool)
+ else:
+ decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=True)
+
+ self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device)
+ decoder_timesteps_tensor = self.decoder_scheduler.timesteps
+
+ num_channels_latents = self.decoder.config.in_channels
+ height = self.decoder.config.sample_size
+ width = self.decoder.config.sample_size
+
+ decoder_latents = self.prepare_latents(
+ (batch_size, num_channels_latents, height, width),
+ text_enc_hid_states.dtype,
+ device,
+ generator,
+ decoder_latents,
+ self.decoder_scheduler,
+ )
+
+ for i, t in enumerate(self.progress_bar(decoder_timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([decoder_latents] * 2) if do_classifier_free_guidance else decoder_latents
+
+ noise_pred = self.decoder(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=text_enc_hid_states,
+ class_labels=additive_clip_time_embeddings,
+ attention_mask=decoder_text_mask,
+ ).sample
+
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred_uncond, _ = noise_pred_uncond.split(latent_model_input.shape[1], dim=1)
+ noise_pred_text, predicted_variance = noise_pred_text.split(latent_model_input.shape[1], dim=1)
+ noise_pred = noise_pred_uncond + decoder_guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
+
+ if i + 1 == decoder_timesteps_tensor.shape[0]:
+ prev_timestep = None
+ else:
+ prev_timestep = decoder_timesteps_tensor[i + 1]
+
+ # compute the previous noisy sample x_t -> x_t-1
+ decoder_latents = self.decoder_scheduler.step(
+ noise_pred, t, decoder_latents, prev_timestep=prev_timestep, generator=generator
+ ).prev_sample
+
+ decoder_latents = decoder_latents.clamp(-1, 1)
+
+ image_small = decoder_latents
+
+ # done decoder
+
+ # super res
+
+ self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device)
+ super_res_timesteps_tensor = self.super_res_scheduler.timesteps
+
+ channels = self.super_res_first.config.in_channels // 2
+ height = self.super_res_first.config.sample_size
+ width = self.super_res_first.config.sample_size
+
+ super_res_latents = self.prepare_latents(
+ (batch_size, channels, height, width),
+ image_small.dtype,
+ device,
+ generator,
+ super_res_latents,
+ self.super_res_scheduler,
+ )
+
+ if device.type == "mps":
+ # MPS does not support many interpolations
+ image_upscaled = F.interpolate(image_small, size=[height, width])
+ else:
+ interpolate_antialias = {}
+ if "antialias" in inspect.signature(F.interpolate).parameters:
+ interpolate_antialias["antialias"] = True
+
+ image_upscaled = F.interpolate(
+ image_small, size=[height, width], mode="bicubic", align_corners=False, **interpolate_antialias
+ )
+
+ for i, t in enumerate(self.progress_bar(super_res_timesteps_tensor)):
+ # no classifier free guidance
+
+ if i == super_res_timesteps_tensor.shape[0] - 1:
+ unet = self.super_res_last
+ else:
+ unet = self.super_res_first
+
+ latent_model_input = torch.cat([super_res_latents, image_upscaled], dim=1)
+
+ noise_pred = unet(
+ sample=latent_model_input,
+ timestep=t,
+ ).sample
+
+ if i + 1 == super_res_timesteps_tensor.shape[0]:
+ prev_timestep = None
+ else:
+ prev_timestep = super_res_timesteps_tensor[i + 1]
+
+ # compute the previous noisy sample x_t -> x_t-1
+ super_res_latents = self.super_res_scheduler.step(
+ noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator
+ ).prev_sample
+
+ image = super_res_latents
+ # done super res
+
+ self.maybe_free_model_hooks()
+
+ # post processing
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/unclip/pipeline_unclip_image_variation.py b/model/diffusers_c/pipelines/unclip/pipeline_unclip_image_variation.py
new file mode 100644
index 0000000000000000000000000000000000000000..6c646a7df3629881d7d24a1c4f9985684bf4a241
--- /dev/null
+++ b/model/diffusers_c/pipelines/unclip/pipeline_unclip_image_variation.py
@@ -0,0 +1,420 @@
+# Copyright 2024 Kakao Brain and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import List, Optional, Union
+
+import PIL.Image
+import torch
+from torch.nn import functional as F
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModelWithProjection,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+)
+
+from ...models import UNet2DConditionModel, UNet2DModel
+from ...schedulers import UnCLIPScheduler
+from ...utils import logging
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .text_proj import UnCLIPTextProjModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+class UnCLIPImageVariationPipeline(DiffusionPipeline):
+ """
+ Pipeline to generate image variations from an input image using UnCLIP.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ text_encoder ([`~transformers.CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `image_encoder`.
+ image_encoder ([`~transformers.CLIPVisionModelWithProjection`]):
+ Frozen CLIP image-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ text_proj ([`UnCLIPTextProjModel`]):
+ Utility class to prepare and combine the embeddings before they are passed to the decoder.
+ decoder ([`UNet2DConditionModel`]):
+ The decoder to invert the image embedding into an image.
+ super_res_first ([`UNet2DModel`]):
+ Super resolution UNet. Used in all but the last step of the super resolution diffusion process.
+ super_res_last ([`UNet2DModel`]):
+ Super resolution UNet. Used in the last step of the super resolution diffusion process.
+ decoder_scheduler ([`UnCLIPScheduler`]):
+ Scheduler used in the decoder denoising process (a modified [`DDPMScheduler`]).
+ super_res_scheduler ([`UnCLIPScheduler`]):
+ Scheduler used in the super resolution denoising process (a modified [`DDPMScheduler`]).
+ """
+
+ decoder: UNet2DConditionModel
+ text_proj: UnCLIPTextProjModel
+ text_encoder: CLIPTextModelWithProjection
+ tokenizer: CLIPTokenizer
+ feature_extractor: CLIPImageProcessor
+ image_encoder: CLIPVisionModelWithProjection
+ super_res_first: UNet2DModel
+ super_res_last: UNet2DModel
+
+ decoder_scheduler: UnCLIPScheduler
+ super_res_scheduler: UnCLIPScheduler
+ model_cpu_offload_seq = "text_encoder->image_encoder->text_proj->decoder->super_res_first->super_res_last"
+
+ def __init__(
+ self,
+ decoder: UNet2DConditionModel,
+ text_encoder: CLIPTextModelWithProjection,
+ tokenizer: CLIPTokenizer,
+ text_proj: UnCLIPTextProjModel,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection,
+ super_res_first: UNet2DModel,
+ super_res_last: UNet2DModel,
+ decoder_scheduler: UnCLIPScheduler,
+ super_res_scheduler: UnCLIPScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ decoder=decoder,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ text_proj=text_proj,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ super_res_first=super_res_first,
+ super_res_last=super_res_last,
+ decoder_scheduler=decoder_scheduler,
+ super_res_scheduler=super_res_scheduler,
+ )
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def _encode_prompt(self, prompt, device, num_images_per_prompt, do_classifier_free_guidance):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ text_mask = text_inputs.attention_mask.bool().to(device)
+ text_encoder_output = self.text_encoder(text_input_ids.to(device))
+
+ prompt_embeds = text_encoder_output.text_embeds
+ text_encoder_hidden_states = text_encoder_output.last_hidden_state
+
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ text_mask = text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_tokens = [""] * batch_size
+
+ max_length = text_input_ids.shape[-1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_text_mask = uncond_input.attention_mask.bool().to(device)
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(uncond_input.input_ids.to(device))
+
+ negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.text_embeds
+ uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len)
+
+ seq_len = uncond_text_encoder_hidden_states.shape[1]
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ uncond_text_mask = uncond_text_mask.repeat_interleave(num_images_per_prompt, dim=0)
+
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ text_encoder_hidden_states = torch.cat([uncond_text_encoder_hidden_states, text_encoder_hidden_states])
+
+ text_mask = torch.cat([uncond_text_mask, text_mask])
+
+ return prompt_embeds, text_encoder_hidden_states, text_mask
+
+ def _encode_image(self, image, device, num_images_per_prompt, image_embeddings: Optional[torch.Tensor] = None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if image_embeddings is None:
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(images=image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ image_embeddings = self.image_encoder(image).image_embeds
+
+ image_embeddings = image_embeddings.repeat_interleave(num_images_per_prompt, dim=0)
+
+ return image_embeddings
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ image: Optional[Union[PIL.Image.Image, List[PIL.Image.Image], torch.FloatTensor]] = None,
+ num_images_per_prompt: int = 1,
+ decoder_num_inference_steps: int = 25,
+ super_res_num_inference_steps: int = 7,
+ generator: Optional[torch.Generator] = None,
+ decoder_latents: Optional[torch.FloatTensor] = None,
+ super_res_latents: Optional[torch.FloatTensor] = None,
+ image_embeddings: Optional[torch.Tensor] = None,
+ decoder_guidance_scale: float = 8.0,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ ):
+ """
+ The call function to the pipeline for generation.
+
+ Args:
+ image (`PIL.Image.Image` or `List[PIL.Image.Image]` or `torch.FloatTensor`):
+ `Image` or tensor representing an image batch to be used as the starting point. If you provide a
+ tensor, it needs to be compatible with the [`CLIPImageProcessor`]
+ [configuration](https://huggingface.co/fusing/karlo-image-variations-diffusers/blob/main/feature_extractor/preprocessor_config.json).
+ Can be left as `None` only when `image_embeddings` are passed.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ decoder_num_inference_steps (`int`, *optional*, defaults to 25):
+ The number of denoising steps for the decoder. More denoising steps usually lead to a higher quality
+ image at the expense of slower inference.
+ super_res_num_inference_steps (`int`, *optional*, defaults to 7):
+ The number of denoising steps for super resolution. More denoising steps usually lead to a higher
+ quality image at the expense of slower inference.
+ generator (`torch.Generator`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ decoder_latents (`torch.FloatTensor` of shape (batch size, channels, height, width), *optional*):
+ Pre-generated noisy latents to be used as inputs for the decoder.
+ super_res_latents (`torch.FloatTensor` of shape (batch size, channels, super res height, super res width), *optional*):
+ Pre-generated noisy latents to be used as inputs for the decoder.
+ decoder_guidance_scale (`float`, *optional*, defaults to 4.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ image_embeddings (`torch.Tensor`, *optional*):
+ Pre-defined image embeddings that can be derived from the image encoder. Pre-defined image embeddings
+ can be passed for tasks like image interpolations. `image` can be left as `None`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
+ returned where the first element is a list with the generated images.
+ """
+ if image is not None:
+ if isinstance(image, PIL.Image.Image):
+ batch_size = 1
+ elif isinstance(image, list):
+ batch_size = len(image)
+ else:
+ batch_size = image.shape[0]
+ else:
+ batch_size = image_embeddings.shape[0]
+
+ prompt = [""] * batch_size
+
+ device = self._execution_device
+
+ batch_size = batch_size * num_images_per_prompt
+
+ do_classifier_free_guidance = decoder_guidance_scale > 1.0
+
+ prompt_embeds, text_encoder_hidden_states, text_mask = self._encode_prompt(
+ prompt, device, num_images_per_prompt, do_classifier_free_guidance
+ )
+
+ image_embeddings = self._encode_image(image, device, num_images_per_prompt, image_embeddings)
+
+ # decoder
+ text_encoder_hidden_states, additive_clip_time_embeddings = self.text_proj(
+ image_embeddings=image_embeddings,
+ prompt_embeds=prompt_embeds,
+ text_encoder_hidden_states=text_encoder_hidden_states,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ )
+
+ if device.type == "mps":
+ # HACK: MPS: There is a panic when padding bool tensors,
+ # so cast to int tensor for the pad and back to bool afterwards
+ text_mask = text_mask.type(torch.int)
+ decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=1)
+ decoder_text_mask = decoder_text_mask.type(torch.bool)
+ else:
+ decoder_text_mask = F.pad(text_mask, (self.text_proj.clip_extra_context_tokens, 0), value=True)
+
+ self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device)
+ decoder_timesteps_tensor = self.decoder_scheduler.timesteps
+
+ num_channels_latents = self.decoder.config.in_channels
+ height = self.decoder.config.sample_size
+ width = self.decoder.config.sample_size
+
+ if decoder_latents is None:
+ decoder_latents = self.prepare_latents(
+ (batch_size, num_channels_latents, height, width),
+ text_encoder_hidden_states.dtype,
+ device,
+ generator,
+ decoder_latents,
+ self.decoder_scheduler,
+ )
+
+ for i, t in enumerate(self.progress_bar(decoder_timesteps_tensor)):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([decoder_latents] * 2) if do_classifier_free_guidance else decoder_latents
+
+ noise_pred = self.decoder(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=text_encoder_hidden_states,
+ class_labels=additive_clip_time_embeddings,
+ attention_mask=decoder_text_mask,
+ ).sample
+
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred_uncond, _ = noise_pred_uncond.split(latent_model_input.shape[1], dim=1)
+ noise_pred_text, predicted_variance = noise_pred_text.split(latent_model_input.shape[1], dim=1)
+ noise_pred = noise_pred_uncond + decoder_guidance_scale * (noise_pred_text - noise_pred_uncond)
+ noise_pred = torch.cat([noise_pred, predicted_variance], dim=1)
+
+ if i + 1 == decoder_timesteps_tensor.shape[0]:
+ prev_timestep = None
+ else:
+ prev_timestep = decoder_timesteps_tensor[i + 1]
+
+ # compute the previous noisy sample x_t -> x_t-1
+ decoder_latents = self.decoder_scheduler.step(
+ noise_pred, t, decoder_latents, prev_timestep=prev_timestep, generator=generator
+ ).prev_sample
+
+ decoder_latents = decoder_latents.clamp(-1, 1)
+
+ image_small = decoder_latents
+
+ # done decoder
+
+ # super res
+
+ self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device)
+ super_res_timesteps_tensor = self.super_res_scheduler.timesteps
+
+ channels = self.super_res_first.config.in_channels // 2
+ height = self.super_res_first.config.sample_size
+ width = self.super_res_first.config.sample_size
+
+ if super_res_latents is None:
+ super_res_latents = self.prepare_latents(
+ (batch_size, channels, height, width),
+ image_small.dtype,
+ device,
+ generator,
+ super_res_latents,
+ self.super_res_scheduler,
+ )
+
+ if device.type == "mps":
+ # MPS does not support many interpolations
+ image_upscaled = F.interpolate(image_small, size=[height, width])
+ else:
+ interpolate_antialias = {}
+ if "antialias" in inspect.signature(F.interpolate).parameters:
+ interpolate_antialias["antialias"] = True
+
+ image_upscaled = F.interpolate(
+ image_small, size=[height, width], mode="bicubic", align_corners=False, **interpolate_antialias
+ )
+
+ for i, t in enumerate(self.progress_bar(super_res_timesteps_tensor)):
+ # no classifier free guidance
+
+ if i == super_res_timesteps_tensor.shape[0] - 1:
+ unet = self.super_res_last
+ else:
+ unet = self.super_res_first
+
+ latent_model_input = torch.cat([super_res_latents, image_upscaled], dim=1)
+
+ noise_pred = unet(
+ sample=latent_model_input,
+ timestep=t,
+ ).sample
+
+ if i + 1 == super_res_timesteps_tensor.shape[0]:
+ prev_timestep = None
+ else:
+ prev_timestep = super_res_timesteps_tensor[i + 1]
+
+ # compute the previous noisy sample x_t -> x_t-1
+ super_res_latents = self.super_res_scheduler.step(
+ noise_pred, t, super_res_latents, prev_timestep=prev_timestep, generator=generator
+ ).prev_sample
+
+ image = super_res_latents
+
+ # done super res
+ self.maybe_free_model_hooks()
+
+ # post processing
+
+ image = image * 0.5 + 0.5
+ image = image.clamp(0, 1)
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+
+ if output_type == "pil":
+ image = self.numpy_to_pil(image)
+
+ if not return_dict:
+ return (image,)
+
+ return ImagePipelineOutput(images=image)
diff --git a/model/diffusers_c/pipelines/unclip/text_proj.py b/model/diffusers_c/pipelines/unclip/text_proj.py
new file mode 100644
index 0000000000000000000000000000000000000000..5a86d0c08a8da90ff27505c8868b030cfa10068f
--- /dev/null
+++ b/model/diffusers_c/pipelines/unclip/text_proj.py
@@ -0,0 +1,86 @@
+# Copyright 2024 Kakao Brain and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models import ModelMixin
+
+
+class UnCLIPTextProjModel(ModelMixin, ConfigMixin):
+ """
+ Utility class for CLIP embeddings. Used to combine the image and text embeddings into a format usable by the
+ decoder.
+
+ For more details, see the original paper: https://arxiv.org/abs/2204.06125 section 2.1
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ *,
+ clip_extra_context_tokens: int = 4,
+ clip_embeddings_dim: int = 768,
+ time_embed_dim: int,
+ cross_attention_dim,
+ ):
+ super().__init__()
+
+ self.learned_classifier_free_guidance_embeddings = nn.Parameter(torch.zeros(clip_embeddings_dim))
+
+ # parameters for additional clip time embeddings
+ self.embedding_proj = nn.Linear(clip_embeddings_dim, time_embed_dim)
+ self.clip_image_embeddings_project_to_time_embeddings = nn.Linear(clip_embeddings_dim, time_embed_dim)
+
+ # parameters for encoder hidden states
+ self.clip_extra_context_tokens = clip_extra_context_tokens
+ self.clip_extra_context_tokens_proj = nn.Linear(
+ clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim
+ )
+ self.encoder_hidden_states_proj = nn.Linear(clip_embeddings_dim, cross_attention_dim)
+ self.text_encoder_hidden_states_norm = nn.LayerNorm(cross_attention_dim)
+
+ def forward(self, *, image_embeddings, prompt_embeds, text_encoder_hidden_states, do_classifier_free_guidance):
+ if do_classifier_free_guidance:
+ # Add the classifier free guidance embeddings to the image embeddings
+ image_embeddings_batch_size = image_embeddings.shape[0]
+ classifier_free_guidance_embeddings = self.learned_classifier_free_guidance_embeddings.unsqueeze(0)
+ classifier_free_guidance_embeddings = classifier_free_guidance_embeddings.expand(
+ image_embeddings_batch_size, -1
+ )
+ image_embeddings = torch.cat([classifier_free_guidance_embeddings, image_embeddings], dim=0)
+
+ # The image embeddings batch size and the text embeddings batch size are equal
+ assert image_embeddings.shape[0] == prompt_embeds.shape[0]
+
+ batch_size = prompt_embeds.shape[0]
+
+ # "Specifically, we modify the architecture described in Nichol et al. (2021) by projecting and
+ # adding CLIP embeddings to the existing timestep embedding, ...
+ time_projected_prompt_embeds = self.embedding_proj(prompt_embeds)
+ time_projected_image_embeddings = self.clip_image_embeddings_project_to_time_embeddings(image_embeddings)
+ additive_clip_time_embeddings = time_projected_image_embeddings + time_projected_prompt_embeds
+
+ # ... and by projecting CLIP embeddings into four
+ # extra tokens of context that are concatenated to the sequence of outputs from the GLIDE text encoder"
+ clip_extra_context_tokens = self.clip_extra_context_tokens_proj(image_embeddings)
+ clip_extra_context_tokens = clip_extra_context_tokens.reshape(batch_size, -1, self.clip_extra_context_tokens)
+ clip_extra_context_tokens = clip_extra_context_tokens.permute(0, 2, 1)
+
+ text_encoder_hidden_states = self.encoder_hidden_states_proj(text_encoder_hidden_states)
+ text_encoder_hidden_states = self.text_encoder_hidden_states_norm(text_encoder_hidden_states)
+ text_encoder_hidden_states = torch.cat([clip_extra_context_tokens, text_encoder_hidden_states], dim=1)
+
+ return text_encoder_hidden_states, additive_clip_time_embeddings
diff --git a/model/diffusers_c/pipelines/unidiffuser/__init__.py b/model/diffusers_c/pipelines/unidiffuser/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..1ac2b09a6e570087c80bc11bf1a8102dd4970b8f
--- /dev/null
+++ b/model/diffusers_c/pipelines/unidiffuser/__init__.py
@@ -0,0 +1,58 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ ImageTextPipelineOutput,
+ UniDiffuserPipeline,
+ )
+
+ _dummy_objects.update(
+ {"ImageTextPipelineOutput": ImageTextPipelineOutput, "UniDiffuserPipeline": UniDiffuserPipeline}
+ )
+else:
+ _import_structure["modeling_text_decoder"] = ["UniDiffuserTextDecoder"]
+ _import_structure["modeling_uvit"] = ["UniDiffuserModel", "UTransformer2DModel"]
+ _import_structure["pipeline_unidiffuser"] = ["ImageTextPipelineOutput", "UniDiffuserPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import (
+ ImageTextPipelineOutput,
+ UniDiffuserPipeline,
+ )
+ else:
+ from .modeling_text_decoder import UniDiffuserTextDecoder
+ from .modeling_uvit import UniDiffuserModel, UTransformer2DModel
+ from .pipeline_unidiffuser import ImageTextPipelineOutput, UniDiffuserPipeline
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/unidiffuser/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/unidiffuser/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..741c304ef39e3b251d716cffbfc24f4cab7402eb
Binary files /dev/null and b/model/diffusers_c/pipelines/unidiffuser/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/unidiffuser/__pycache__/modeling_text_decoder.cpython-39.pyc b/model/diffusers_c/pipelines/unidiffuser/__pycache__/modeling_text_decoder.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d33dd882ae3c0e602b765c3ca200bf0f4c5019e3
Binary files /dev/null and b/model/diffusers_c/pipelines/unidiffuser/__pycache__/modeling_text_decoder.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/unidiffuser/__pycache__/modeling_uvit.cpython-39.pyc b/model/diffusers_c/pipelines/unidiffuser/__pycache__/modeling_uvit.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..029738e307de613d2a6bd84280e5cdeeeb25cd21
Binary files /dev/null and b/model/diffusers_c/pipelines/unidiffuser/__pycache__/modeling_uvit.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/unidiffuser/__pycache__/pipeline_unidiffuser.cpython-39.pyc b/model/diffusers_c/pipelines/unidiffuser/__pycache__/pipeline_unidiffuser.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..84c3d06d0ea4173191e89851ce35712d63d38142
Binary files /dev/null and b/model/diffusers_c/pipelines/unidiffuser/__pycache__/pipeline_unidiffuser.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/unidiffuser/modeling_text_decoder.py b/model/diffusers_c/pipelines/unidiffuser/modeling_text_decoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..bf0a4eb475c0e8cda0a62a8b4bb83a8a02733903
--- /dev/null
+++ b/model/diffusers_c/pipelines/unidiffuser/modeling_text_decoder.py
@@ -0,0 +1,296 @@
+from typing import Optional
+
+import numpy as np
+import torch
+from torch import nn
+from transformers import GPT2Config, GPT2LMHeadModel
+from transformers.modeling_utils import ModuleUtilsMixin
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models import ModelMixin
+
+
+# Modified from ClipCaptionModel in https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py
+class UniDiffuserTextDecoder(ModelMixin, ConfigMixin, ModuleUtilsMixin):
+ """
+ Text decoder model for a image-text [UniDiffuser](https://arxiv.org/pdf/2303.06555.pdf) model. This is used to
+ generate text from the UniDiffuser image-text embedding.
+
+ Parameters:
+ prefix_length (`int`):
+ Max number of prefix tokens that will be supplied to the model.
+ prefix_inner_dim (`int`):
+ The hidden size of the incoming prefix embeddings. For UniDiffuser, this would be the hidden dim of the
+ CLIP text encoder.
+ prefix_hidden_dim (`int`, *optional*):
+ Hidden dim of the MLP if we encode the prefix.
+ vocab_size (`int`, *optional*, defaults to 50257):
+ Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the
+ `inputs_ids` passed when calling [`GPT2Model`] or [`TFGPT2Model`].
+ n_positions (`int`, *optional*, defaults to 1024):
+ The maximum sequence length that this model might ever be used with. Typically set this to something large
+ just in case (e.g., 512 or 1024 or 2048).
+ n_embd (`int`, *optional*, defaults to 768):
+ Dimensionality of the embeddings and hidden states.
+ n_layer (`int`, *optional*, defaults to 12):
+ Number of hidden layers in the Transformer encoder.
+ n_head (`int`, *optional*, defaults to 12):
+ Number of attention heads for each attention layer in the Transformer encoder.
+ n_inner (`int`, *optional*, defaults to None):
+ Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
+ activation_function (`str`, *optional*, defaults to `"gelu"`):
+ Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new"]`.
+ resid_pdrop (`float`, *optional*, defaults to 0.1):
+ The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+ embd_pdrop (`float`, *optional*, defaults to 0.1):
+ The dropout ratio for the embeddings.
+ attn_pdrop (`float`, *optional*, defaults to 0.1):
+ The dropout ratio for the attention.
+ layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
+ The epsilon to use in the layer normalization layers.
+ initializer_range (`float`, *optional*, defaults to 0.02):
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+ scale_attn_weights (`bool`, *optional*, defaults to `True`):
+ Scale attention weights by dividing by sqrt(hidden_size)..
+ use_cache (`bool`, *optional*, defaults to `True`):
+ Whether or not the model should return the last key/values attentions (not used by all models).
+ scale_attn_by_inverse_layer_idx (`bool`, *optional*, defaults to `False`):
+ Whether to additionally scale attention weights by `1 / layer_idx + 1`.
+ reorder_and_upcast_attn (`bool`, *optional*, defaults to `False`):
+ Whether to scale keys (K) prior to computing attention (dot-product) and upcast attention
+ dot-product/softmax to float() when training with mixed precision.
+ """
+
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.bias", r"h\.\d+\.attn\.masked_bias"]
+
+ @register_to_config
+ def __init__(
+ self,
+ prefix_length: int,
+ prefix_inner_dim: int,
+ prefix_hidden_dim: Optional[int] = None,
+ vocab_size: int = 50257, # Start of GPT2 config args
+ n_positions: int = 1024,
+ n_embd: int = 768,
+ n_layer: int = 12,
+ n_head: int = 12,
+ n_inner: Optional[int] = None,
+ activation_function: str = "gelu_new",
+ resid_pdrop: float = 0.1,
+ embd_pdrop: float = 0.1,
+ attn_pdrop: float = 0.1,
+ layer_norm_epsilon: float = 1e-5,
+ initializer_range: float = 0.02,
+ scale_attn_weights: bool = True,
+ use_cache: bool = True,
+ scale_attn_by_inverse_layer_idx: bool = False,
+ reorder_and_upcast_attn: bool = False,
+ ):
+ super().__init__()
+
+ self.prefix_length = prefix_length
+
+ if prefix_inner_dim != n_embd and prefix_hidden_dim is None:
+ raise ValueError(
+ f"`prefix_hidden_dim` cannot be `None` when `prefix_inner_dim`: {prefix_hidden_dim} and"
+ f" `n_embd`: {n_embd} are not equal."
+ )
+
+ self.prefix_inner_dim = prefix_inner_dim
+ self.prefix_hidden_dim = prefix_hidden_dim
+
+ self.encode_prefix = (
+ nn.Linear(self.prefix_inner_dim, self.prefix_hidden_dim)
+ if self.prefix_hidden_dim is not None
+ else nn.Identity()
+ )
+ self.decode_prefix = (
+ nn.Linear(self.prefix_hidden_dim, n_embd) if self.prefix_hidden_dim is not None else nn.Identity()
+ )
+
+ gpt_config = GPT2Config(
+ vocab_size=vocab_size,
+ n_positions=n_positions,
+ n_embd=n_embd,
+ n_layer=n_layer,
+ n_head=n_head,
+ n_inner=n_inner,
+ activation_function=activation_function,
+ resid_pdrop=resid_pdrop,
+ embd_pdrop=embd_pdrop,
+ attn_pdrop=attn_pdrop,
+ layer_norm_epsilon=layer_norm_epsilon,
+ initializer_range=initializer_range,
+ scale_attn_weights=scale_attn_weights,
+ use_cache=use_cache,
+ scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
+ reorder_and_upcast_attn=reorder_and_upcast_attn,
+ )
+ self.transformer = GPT2LMHeadModel(gpt_config)
+
+ def forward(
+ self,
+ input_ids: torch.Tensor,
+ prefix_embeds: torch.Tensor,
+ attention_mask: Optional[torch.Tensor] = None,
+ labels: Optional[torch.Tensor] = None,
+ ):
+ """
+ Args:
+ input_ids (`torch.Tensor` of shape `(N, max_seq_len)`):
+ Text tokens to use for inference.
+ prefix_embeds (`torch.Tensor` of shape `(N, prefix_length, 768)`):
+ Prefix embedding to preprend to the embedded tokens.
+ attention_mask (`torch.Tensor` of shape `(N, prefix_length + max_seq_len, 768)`, *optional*):
+ Attention mask for the prefix embedding.
+ labels (`torch.Tensor`, *optional*):
+ Labels to use for language modeling.
+ """
+ embedding_text = self.transformer.transformer.wte(input_ids)
+ hidden = self.encode_prefix(prefix_embeds)
+ prefix_embeds = self.decode_prefix(hidden)
+ embedding_cat = torch.cat((prefix_embeds, embedding_text), dim=1)
+
+ if labels is not None:
+ dummy_token = self.get_dummy_token(input_ids.shape[0], input_ids.device)
+ labels = torch.cat((dummy_token, input_ids), dim=1)
+ out = self.transformer(inputs_embeds=embedding_cat, labels=labels, attention_mask=attention_mask)
+ if self.prefix_hidden_dim is not None:
+ return out, hidden
+ else:
+ return out
+
+ def get_dummy_token(self, batch_size: int, device: torch.device) -> torch.Tensor:
+ return torch.zeros(batch_size, self.prefix_length, dtype=torch.int64, device=device)
+
+ def encode(self, prefix):
+ return self.encode_prefix(prefix)
+
+ @torch.no_grad()
+ def generate_captions(self, features, eos_token_id, device):
+ """
+ Generate captions given text embedding features. Returns list[L].
+
+ Args:
+ features (`torch.Tensor` of shape `(B, L, D)`):
+ Text embedding features to generate captions from.
+ eos_token_id (`int`):
+ The token ID of the EOS token for the text decoder model.
+ device:
+ Device to perform text generation on.
+
+ Returns:
+ `List[str]`: A list of strings generated from the decoder model.
+ """
+
+ features = torch.split(features, 1, dim=0)
+ generated_tokens = []
+ generated_seq_lengths = []
+ for feature in features:
+ feature = self.decode_prefix(feature.to(device)) # back to the clip feature
+ # Only support beam search for now
+ output_tokens, seq_lengths = self.generate_beam(
+ input_embeds=feature, device=device, eos_token_id=eos_token_id
+ )
+ generated_tokens.append(output_tokens[0])
+ generated_seq_lengths.append(seq_lengths[0])
+ generated_tokens = torch.stack(generated_tokens)
+ generated_seq_lengths = torch.stack(generated_seq_lengths)
+ return generated_tokens, generated_seq_lengths
+
+ @torch.no_grad()
+ def generate_beam(
+ self,
+ input_ids=None,
+ input_embeds=None,
+ device=None,
+ beam_size: int = 5,
+ entry_length: int = 67,
+ temperature: float = 1.0,
+ eos_token_id: Optional[int] = None,
+ ):
+ """
+ Generates text using the given tokenizer and text prompt or token embedding via beam search. This
+ implementation is based on the beam search implementation from the [original UniDiffuser
+ code](https://github.com/thu-ml/unidiffuser/blob/main/libs/caption_decoder.py#L89).
+
+ Args:
+ eos_token_id (`int`, *optional*):
+ The token ID of the EOS token for the text decoder model.
+ input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`, *optional*):
+ Tokenizer indices of input sequence tokens in the vocabulary. One of `input_ids` and `input_embeds`
+ must be supplied.
+ input_embeds (`torch.FloatTensor` of shape `(batch_size, seq_len, hidden_size)`, *optional*):
+ An embedded representation to directly pass to the transformer as a prefix for beam search. One of
+ `input_ids` and `input_embeds` must be supplied.
+ device:
+ The device to perform beam search on.
+ beam_size (`int`, *optional*, defaults to `5`):
+ The number of best states to store during beam search.
+ entry_length (`int`, *optional*, defaults to `67`):
+ The number of iterations to run beam search.
+ temperature (`float`, *optional*, defaults to 1.0):
+ The temperature to use when performing the softmax over logits from the decoding model.
+
+ Returns:
+ `Tuple(torch.Tensor, torch.Tensor)`: A tuple of tensors where the first element is a tensor of generated
+ token sequences sorted by score in descending order, and the second element is the sequence lengths
+ corresponding to those sequences.
+ """
+ # Generates text until stop_token is reached using beam search with the desired beam size.
+ stop_token_index = eos_token_id
+ tokens = None
+ scores = None
+ seq_lengths = torch.ones(beam_size, device=device, dtype=torch.int)
+ is_stopped = torch.zeros(beam_size, device=device, dtype=torch.bool)
+
+ if input_embeds is not None:
+ generated = input_embeds
+ else:
+ generated = self.transformer.transformer.wte(input_ids)
+
+ for i in range(entry_length):
+ outputs = self.transformer(inputs_embeds=generated)
+ logits = outputs.logits
+ logits = logits[:, -1, :] / (temperature if temperature > 0 else 1.0)
+ logits = logits.softmax(-1).log()
+
+ if scores is None:
+ scores, next_tokens = logits.topk(beam_size, -1)
+ generated = generated.expand(beam_size, *generated.shape[1:])
+ next_tokens, scores = next_tokens.permute(1, 0), scores.squeeze(0)
+ if tokens is None:
+ tokens = next_tokens
+ else:
+ tokens = tokens.expand(beam_size, *tokens.shape[1:])
+ tokens = torch.cat((tokens, next_tokens), dim=1)
+ else:
+ logits[is_stopped] = -float(np.inf)
+ logits[is_stopped, 0] = 0
+ scores_sum = scores[:, None] + logits
+ seq_lengths[~is_stopped] += 1
+ scores_sum_average = scores_sum / seq_lengths[:, None]
+ scores_sum_average, next_tokens = scores_sum_average.view(-1).topk(beam_size, -1)
+ next_tokens_source = next_tokens // scores_sum.shape[1]
+ seq_lengths = seq_lengths[next_tokens_source]
+ next_tokens = next_tokens % scores_sum.shape[1]
+ next_tokens = next_tokens.unsqueeze(1)
+ tokens = tokens[next_tokens_source]
+ tokens = torch.cat((tokens, next_tokens), dim=1)
+ generated = generated[next_tokens_source]
+ scores = scores_sum_average * seq_lengths
+ is_stopped = is_stopped[next_tokens_source]
+
+ next_token_embed = self.transformer.transformer.wte(next_tokens.squeeze()).view(generated.shape[0], 1, -1)
+ generated = torch.cat((generated, next_token_embed), dim=1)
+ is_stopped = is_stopped + next_tokens.eq(stop_token_index).squeeze()
+ if is_stopped.all():
+ break
+
+ scores = scores / seq_lengths
+ order = scores.argsort(descending=True)
+ # tokens tensors are already padded to max_seq_length
+ output_texts = [tokens[i] for i in order]
+ output_texts = torch.stack(output_texts, dim=0)
+ seq_lengths = torch.tensor([seq_lengths[i] for i in order], dtype=seq_lengths.dtype)
+ return output_texts, seq_lengths
diff --git a/model/diffusers_c/pipelines/unidiffuser/modeling_uvit.py b/model/diffusers_c/pipelines/unidiffuser/modeling_uvit.py
new file mode 100644
index 0000000000000000000000000000000000000000..c074b9916301f1b673580d2974f5f3bc0f802535
--- /dev/null
+++ b/model/diffusers_c/pipelines/unidiffuser/modeling_uvit.py
@@ -0,0 +1,1197 @@
+import math
+from typing import Optional, Union
+
+import torch
+from torch import nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models import ModelMixin
+from ...models.attention import FeedForward
+from ...models.attention_processor import Attention
+from ...models.embeddings import TimestepEmbedding, Timesteps, get_2d_sincos_pos_embed
+from ...models.normalization import AdaLayerNorm
+from ...models.transformers.transformer_2d import Transformer2DModelOutput
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+ # Cut & paste from PyTorch official master until it's in a few official releases - RW
+ # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+ def norm_cdf(x):
+ # Computes standard normal cumulative distribution function
+ return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+ if (mean < a - 2 * std) or (mean > b + 2 * std):
+ logger.warning(
+ "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+ "The distribution of values may be incorrect."
+ )
+
+ with torch.no_grad():
+ # Values are generated by using a truncated uniform distribution and
+ # then using the inverse CDF for the normal distribution.
+ # Get upper and lower cdf values
+ l = norm_cdf((a - mean) / std)
+ u = norm_cdf((b - mean) / std)
+
+ # Uniformly fill tensor with values from [l, u], then translate to
+ # [2l-1, 2u-1].
+ tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+ # Use inverse cdf transform for normal distribution to get truncated
+ # standard normal
+ tensor.erfinv_()
+
+ # Transform to proper mean, std
+ tensor.mul_(std * math.sqrt(2.0))
+ tensor.add_(mean)
+
+ # Clamp to ensure it's in the proper range
+ tensor.clamp_(min=a, max=b)
+ return tensor
+
+
+def trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):
+ # type: (torch.Tensor, float, float, float, float) -> torch.Tensor
+ r"""Fills the input Tensor with values drawn from a truncated
+ normal distribution. The values are effectively drawn from the normal distribution :math:`\mathcal{N}(\text{mean},
+ \text{std}^2)` with values outside :math:`[a, b]` redrawn until they are within the bounds. The method used for
+ generating the random values works best when :math:`a \leq \text{mean} \leq b`.
+
+ Args:
+ tensor: an n-dimensional `torch.Tensor`
+ mean: the mean of the normal distribution
+ std: the standard deviation of the normal distribution
+ a: the minimum cutoff value
+ b: the maximum cutoff value
+ Examples:
+ >>> w = torch.empty(3, 5) >>> nn.init.trunc_normal_(w)
+ """
+ return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+
+
+class PatchEmbed(nn.Module):
+ """2D Image to Patch Embedding"""
+
+ def __init__(
+ self,
+ height=224,
+ width=224,
+ patch_size=16,
+ in_channels=3,
+ embed_dim=768,
+ layer_norm=False,
+ flatten=True,
+ bias=True,
+ use_pos_embed=True,
+ ):
+ super().__init__()
+
+ num_patches = (height // patch_size) * (width // patch_size)
+ self.flatten = flatten
+ self.layer_norm = layer_norm
+
+ self.proj = nn.Conv2d(
+ in_channels, embed_dim, kernel_size=(patch_size, patch_size), stride=patch_size, bias=bias
+ )
+ if layer_norm:
+ self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
+ else:
+ self.norm = None
+
+ self.use_pos_embed = use_pos_embed
+ if self.use_pos_embed:
+ pos_embed = get_2d_sincos_pos_embed(embed_dim, int(num_patches**0.5))
+ self.register_buffer("pos_embed", torch.from_numpy(pos_embed).float().unsqueeze(0), persistent=False)
+
+ def forward(self, latent):
+ latent = self.proj(latent)
+ if self.flatten:
+ latent = latent.flatten(2).transpose(1, 2) # BCHW -> BNC
+ if self.layer_norm:
+ latent = self.norm(latent)
+ if self.use_pos_embed:
+ return latent + self.pos_embed
+ else:
+ return latent
+
+
+class SkipBlock(nn.Module):
+ def __init__(self, dim: int):
+ super().__init__()
+
+ self.skip_linear = nn.Linear(2 * dim, dim)
+
+ # Use torch.nn.LayerNorm for now, following the original code
+ self.norm = nn.LayerNorm(dim)
+
+ def forward(self, x, skip):
+ x = self.skip_linear(torch.cat([x, skip], dim=-1))
+ x = self.norm(x)
+
+ return x
+
+
+# Modified to support both pre-LayerNorm and post-LayerNorm configurations
+# Don't support AdaLayerNormZero for now
+# Modified from diffusers.models.attention.BasicTransformerBlock
+class UTransformerBlock(nn.Module):
+ r"""
+ A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations.
+
+ Parameters:
+ dim (`int`): The number of channels in the input and output.
+ num_attention_heads (`int`): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`): The number of channels in each head.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`):
+ Activation function to be used in feed-forward.
+ num_embeds_ada_norm (:obj: `int`, *optional*):
+ The number of diffusion steps used during training. See `Transformer2DModel`.
+ attention_bias (:obj: `bool`, *optional*, defaults to `False`):
+ Configure if the attentions should contain a bias parameter.
+ only_cross_attention (`bool`, *optional*):
+ Whether to use only cross-attention layers. In this case two cross attention layers are used.
+ double_self_attention (`bool`, *optional*):
+ Whether to use two self-attention layers. In this case no cross attention layers are used.
+ upcast_attention (`bool`, *optional*):
+ Whether to upcast the query and key to float32 when performing the attention calculation.
+ norm_elementwise_affine (`bool`, *optional*):
+ Whether to use learnable per-element affine parameters during layer normalization.
+ norm_type (`str`, defaults to `"layer_norm"`):
+ The layer norm implementation to use.
+ pre_layer_norm (`bool`, *optional*):
+ Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
+ as opposed to after ("post-LayerNorm"). Note that `BasicTransformerBlock` uses pre-LayerNorm, e.g.
+ `pre_layer_norm = True`.
+ final_dropout (`bool`, *optional*):
+ Whether to use a final Dropout layer after the feedforward network.
+ """
+
+ def __init__(
+ self,
+ dim: int,
+ num_attention_heads: int,
+ attention_head_dim: int,
+ dropout=0.0,
+ cross_attention_dim: Optional[int] = None,
+ activation_fn: str = "geglu",
+ num_embeds_ada_norm: Optional[int] = None,
+ attention_bias: bool = False,
+ only_cross_attention: bool = False,
+ double_self_attention: bool = False,
+ upcast_attention: bool = False,
+ norm_elementwise_affine: bool = True,
+ norm_type: str = "layer_norm",
+ pre_layer_norm: bool = True,
+ final_dropout: bool = False,
+ ):
+ super().__init__()
+ self.only_cross_attention = only_cross_attention
+
+ self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+
+ self.pre_layer_norm = pre_layer_norm
+
+ if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+ raise ValueError(
+ f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+ f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+ )
+
+ # 1. Self-Attn
+ self.attn1 = Attention(
+ query_dim=dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+ upcast_attention=upcast_attention,
+ )
+
+ # 2. Cross-Attn
+ if cross_attention_dim is not None or double_self_attention:
+ self.attn2 = Attention(
+ query_dim=dim,
+ cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ upcast_attention=upcast_attention,
+ ) # is self-attn if encoder_hidden_states is none
+ else:
+ self.attn2 = None
+
+ if self.use_ada_layer_norm:
+ self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+ else:
+ self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+
+ if cross_attention_dim is not None or double_self_attention:
+ # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+ # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+ # the second cross attention block.
+ self.norm2 = (
+ AdaLayerNorm(dim, num_embeds_ada_norm)
+ if self.use_ada_layer_norm
+ else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+ )
+ else:
+ self.norm2 = None
+
+ # 3. Feed-forward
+ self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+ self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+
+ def forward(
+ self,
+ hidden_states,
+ attention_mask=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ timestep=None,
+ cross_attention_kwargs=None,
+ class_labels=None,
+ ):
+ # Pre-LayerNorm
+ if self.pre_layer_norm:
+ if self.use_ada_layer_norm:
+ norm_hidden_states = self.norm1(hidden_states, timestep)
+ else:
+ norm_hidden_states = self.norm1(hidden_states)
+ else:
+ norm_hidden_states = hidden_states
+
+ # 1. Self-Attention
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+ attn_output = self.attn1(
+ norm_hidden_states,
+ encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+ attention_mask=attention_mask,
+ **cross_attention_kwargs,
+ )
+
+ # Post-LayerNorm
+ if not self.pre_layer_norm:
+ if self.use_ada_layer_norm:
+ attn_output = self.norm1(attn_output, timestep)
+ else:
+ attn_output = self.norm1(attn_output)
+
+ hidden_states = attn_output + hidden_states
+
+ if self.attn2 is not None:
+ # Pre-LayerNorm
+ if self.pre_layer_norm:
+ norm_hidden_states = (
+ self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+ )
+ else:
+ norm_hidden_states = hidden_states
+ # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly
+ # prepare attention mask here
+
+ # 2. Cross-Attention
+ attn_output = self.attn2(
+ norm_hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=encoder_attention_mask,
+ **cross_attention_kwargs,
+ )
+
+ # Post-LayerNorm
+ if not self.pre_layer_norm:
+ attn_output = self.norm2(attn_output, timestep) if self.use_ada_layer_norm else self.norm2(attn_output)
+
+ hidden_states = attn_output + hidden_states
+
+ # 3. Feed-forward
+ # Pre-LayerNorm
+ if self.pre_layer_norm:
+ norm_hidden_states = self.norm3(hidden_states)
+ else:
+ norm_hidden_states = hidden_states
+
+ ff_output = self.ff(norm_hidden_states)
+
+ # Post-LayerNorm
+ if not self.pre_layer_norm:
+ ff_output = self.norm3(ff_output)
+
+ hidden_states = ff_output + hidden_states
+
+ return hidden_states
+
+
+# Like UTransformerBlock except with LayerNorms on the residual backbone of the block
+# Modified from diffusers.models.attention.BasicTransformerBlock
+class UniDiffuserBlock(nn.Module):
+ r"""
+ A modification of BasicTransformerBlock which supports pre-LayerNorm and post-LayerNorm configurations and puts the
+ LayerNorms on the residual backbone of the block. This matches the transformer block in the [original UniDiffuser
+ implementation](https://github.com/thu-ml/unidiffuser/blob/main/libs/uvit_multi_post_ln_v1.py#L104).
+
+ Parameters:
+ dim (`int`): The number of channels in the input and output.
+ num_attention_heads (`int`): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`): The number of channels in each head.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`):
+ Activation function to be used in feed-forward.
+ num_embeds_ada_norm (:obj: `int`, *optional*):
+ The number of diffusion steps used during training. See `Transformer2DModel`.
+ attention_bias (:obj: `bool`, *optional*, defaults to `False`):
+ Configure if the attentions should contain a bias parameter.
+ only_cross_attention (`bool`, *optional*):
+ Whether to use only cross-attention layers. In this case two cross attention layers are used.
+ double_self_attention (`bool`, *optional*):
+ Whether to use two self-attention layers. In this case no cross attention layers are used.
+ upcast_attention (`bool`, *optional*):
+ Whether to upcast the query and key to float() when performing the attention calculation.
+ norm_elementwise_affine (`bool`, *optional*):
+ Whether to use learnable per-element affine parameters during layer normalization.
+ norm_type (`str`, defaults to `"layer_norm"`):
+ The layer norm implementation to use.
+ pre_layer_norm (`bool`, *optional*):
+ Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
+ as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
+ (`pre_layer_norm = False`).
+ final_dropout (`bool`, *optional*):
+ Whether to use a final Dropout layer after the feedforward network.
+ """
+
+ def __init__(
+ self,
+ dim: int,
+ num_attention_heads: int,
+ attention_head_dim: int,
+ dropout=0.0,
+ cross_attention_dim: Optional[int] = None,
+ activation_fn: str = "geglu",
+ num_embeds_ada_norm: Optional[int] = None,
+ attention_bias: bool = False,
+ only_cross_attention: bool = False,
+ double_self_attention: bool = False,
+ upcast_attention: bool = False,
+ norm_elementwise_affine: bool = True,
+ norm_type: str = "layer_norm",
+ pre_layer_norm: bool = False,
+ final_dropout: bool = True,
+ ):
+ super().__init__()
+ self.only_cross_attention = only_cross_attention
+
+ self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+
+ self.pre_layer_norm = pre_layer_norm
+
+ if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+ raise ValueError(
+ f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+ f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+ )
+
+ # 1. Self-Attn
+ self.attn1 = Attention(
+ query_dim=dim,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+ upcast_attention=upcast_attention,
+ )
+
+ # 2. Cross-Attn
+ if cross_attention_dim is not None or double_self_attention:
+ self.attn2 = Attention(
+ query_dim=dim,
+ cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+ heads=num_attention_heads,
+ dim_head=attention_head_dim,
+ dropout=dropout,
+ bias=attention_bias,
+ upcast_attention=upcast_attention,
+ ) # is self-attn if encoder_hidden_states is none
+ else:
+ self.attn2 = None
+
+ if self.use_ada_layer_norm:
+ self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+ else:
+ self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+
+ if cross_attention_dim is not None or double_self_attention:
+ # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+ # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+ # the second cross attention block.
+ self.norm2 = (
+ AdaLayerNorm(dim, num_embeds_ada_norm)
+ if self.use_ada_layer_norm
+ else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+ )
+ else:
+ self.norm2 = None
+
+ # 3. Feed-forward
+ self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+ self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+
+ def forward(
+ self,
+ hidden_states,
+ attention_mask=None,
+ encoder_hidden_states=None,
+ encoder_attention_mask=None,
+ timestep=None,
+ cross_attention_kwargs=None,
+ class_labels=None,
+ ):
+ # Following the diffusers transformer block implementation, put the LayerNorm on the
+ # residual backbone
+ # Pre-LayerNorm
+ if self.pre_layer_norm:
+ if self.use_ada_layer_norm:
+ hidden_states = self.norm1(hidden_states, timestep)
+ else:
+ hidden_states = self.norm1(hidden_states)
+
+ # 1. Self-Attention
+ cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+ attn_output = self.attn1(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+ attention_mask=attention_mask,
+ **cross_attention_kwargs,
+ )
+
+ hidden_states = attn_output + hidden_states
+
+ # Following the diffusers transformer block implementation, put the LayerNorm on the
+ # residual backbone
+ # Post-LayerNorm
+ if not self.pre_layer_norm:
+ if self.use_ada_layer_norm:
+ hidden_states = self.norm1(hidden_states, timestep)
+ else:
+ hidden_states = self.norm1(hidden_states)
+
+ if self.attn2 is not None:
+ # Pre-LayerNorm
+ if self.pre_layer_norm:
+ hidden_states = (
+ self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+ )
+ # TODO (Birch-San): Here we should prepare the encoder_attention mask correctly
+ # prepare attention mask here
+
+ # 2. Cross-Attention
+ attn_output = self.attn2(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=encoder_attention_mask,
+ **cross_attention_kwargs,
+ )
+
+ hidden_states = attn_output + hidden_states
+
+ # Post-LayerNorm
+ if not self.pre_layer_norm:
+ hidden_states = (
+ self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+ )
+
+ # 3. Feed-forward
+ # Pre-LayerNorm
+ if self.pre_layer_norm:
+ hidden_states = self.norm3(hidden_states)
+
+ ff_output = self.ff(hidden_states)
+
+ hidden_states = ff_output + hidden_states
+
+ # Post-LayerNorm
+ if not self.pre_layer_norm:
+ hidden_states = self.norm3(hidden_states)
+
+ return hidden_states
+
+
+# Modified from diffusers.models.transformer_2d.Transformer2DModel
+# Modify the transformer block structure to be U-Net like following U-ViT
+# Only supports patch-style input and torch.nn.LayerNorm currently
+# https://github.com/baofff/U-ViT
+class UTransformer2DModel(ModelMixin, ConfigMixin):
+ """
+ Transformer model based on the [U-ViT](https://github.com/baofff/U-ViT) architecture for image-like data. Compared
+ to [`Transformer2DModel`], this model has skip connections between transformer blocks in a "U"-shaped fashion,
+ similar to a U-Net. Supports only continuous (actual embeddings) inputs, which are embedded via a [`PatchEmbed`]
+ layer and then reshaped to (b, t, d).
+
+ Parameters:
+ num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+ in_channels (`int`, *optional*):
+ Pass if the input is continuous. The number of channels in the input.
+ out_channels (`int`, *optional*):
+ The number of output channels; if `None`, defaults to `in_channels`.
+ num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ norm_num_groups (`int`, *optional*, defaults to `32`):
+ The number of groups to use when performing Group Normalization.
+ cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
+ attention_bias (`bool`, *optional*):
+ Configure if the TransformerBlocks' attention should contain a bias parameter.
+ sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
+ Note that this is fixed at training time as it is used for learning a number of position embeddings. See
+ `ImagePositionalEmbeddings`.
+ num_vector_embeds (`int`, *optional*):
+ Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
+ Includes the class for the masked latent pixel.
+ patch_size (`int`, *optional*, defaults to 2):
+ The patch size to use in the patch embedding.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+ num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
+ The number of diffusion steps used during training. Note that this is fixed at training time as it is used
+ to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
+ up to but not more than steps than `num_embeds_ada_norm`.
+ use_linear_projection (int, *optional*): TODO: Not used
+ only_cross_attention (`bool`, *optional*):
+ Whether to use only cross-attention layers. In this case two cross attention layers are used in each
+ transformer block.
+ upcast_attention (`bool`, *optional*):
+ Whether to upcast the query and key to float() when performing the attention calculation.
+ norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+ The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`.
+ block_type (`str`, *optional*, defaults to `"unidiffuser"`):
+ The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual
+ backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard
+ behavior in `diffusers`.)
+ pre_layer_norm (`bool`, *optional*):
+ Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
+ as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
+ (`pre_layer_norm = False`).
+ norm_elementwise_affine (`bool`, *optional*):
+ Whether to use learnable per-element affine parameters during layer normalization.
+ use_patch_pos_embed (`bool`, *optional*):
+ Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`).
+ final_dropout (`bool`, *optional*):
+ Whether to use a final Dropout layer after the feedforward network.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ num_attention_heads: int = 16,
+ attention_head_dim: int = 88,
+ in_channels: Optional[int] = None,
+ out_channels: Optional[int] = None,
+ num_layers: int = 1,
+ dropout: float = 0.0,
+ norm_num_groups: int = 32,
+ cross_attention_dim: Optional[int] = None,
+ attention_bias: bool = False,
+ sample_size: Optional[int] = None,
+ num_vector_embeds: Optional[int] = None,
+ patch_size: Optional[int] = 2,
+ activation_fn: str = "geglu",
+ num_embeds_ada_norm: Optional[int] = None,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ norm_type: str = "layer_norm",
+ block_type: str = "unidiffuser",
+ pre_layer_norm: bool = False,
+ norm_elementwise_affine: bool = True,
+ use_patch_pos_embed=False,
+ ff_final_dropout: bool = False,
+ ):
+ super().__init__()
+ self.use_linear_projection = use_linear_projection
+ self.num_attention_heads = num_attention_heads
+ self.attention_head_dim = attention_head_dim
+ inner_dim = num_attention_heads * attention_head_dim
+
+ # 1. Input
+ # Only support patch input of shape (batch_size, num_channels, height, width) for now
+ assert in_channels is not None and patch_size is not None, "Patch input requires in_channels and patch_size."
+
+ assert sample_size is not None, "UTransformer2DModel over patched input must provide sample_size"
+
+ # 2. Define input layers
+ self.height = sample_size
+ self.width = sample_size
+
+ self.patch_size = patch_size
+ self.pos_embed = PatchEmbed(
+ height=sample_size,
+ width=sample_size,
+ patch_size=patch_size,
+ in_channels=in_channels,
+ embed_dim=inner_dim,
+ use_pos_embed=use_patch_pos_embed,
+ )
+
+ # 3. Define transformers blocks
+ # Modify this to have in_blocks ("downsample" blocks, even though we don't actually downsample), a mid_block,
+ # and out_blocks ("upsample" blocks). Like a U-Net, there are skip connections from in_blocks to out_blocks in
+ # a "U"-shaped fashion (e.g. first in_block to last out_block, etc.).
+ # Quick hack to make the transformer block type configurable
+ if block_type == "unidiffuser":
+ block_cls = UniDiffuserBlock
+ else:
+ block_cls = UTransformerBlock
+ self.transformer_in_blocks = nn.ModuleList(
+ [
+ block_cls(
+ inner_dim,
+ num_attention_heads,
+ attention_head_dim,
+ dropout=dropout,
+ cross_attention_dim=cross_attention_dim,
+ activation_fn=activation_fn,
+ num_embeds_ada_norm=num_embeds_ada_norm,
+ attention_bias=attention_bias,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ norm_type=norm_type,
+ pre_layer_norm=pre_layer_norm,
+ norm_elementwise_affine=norm_elementwise_affine,
+ final_dropout=ff_final_dropout,
+ )
+ for d in range(num_layers // 2)
+ ]
+ )
+
+ self.transformer_mid_block = block_cls(
+ inner_dim,
+ num_attention_heads,
+ attention_head_dim,
+ dropout=dropout,
+ cross_attention_dim=cross_attention_dim,
+ activation_fn=activation_fn,
+ num_embeds_ada_norm=num_embeds_ada_norm,
+ attention_bias=attention_bias,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ norm_type=norm_type,
+ pre_layer_norm=pre_layer_norm,
+ norm_elementwise_affine=norm_elementwise_affine,
+ final_dropout=ff_final_dropout,
+ )
+
+ # For each skip connection, we use a SkipBlock (concatenation + Linear + LayerNorm) to process the inputs
+ # before each transformer out_block.
+ self.transformer_out_blocks = nn.ModuleList(
+ [
+ nn.ModuleDict(
+ {
+ "skip": SkipBlock(
+ inner_dim,
+ ),
+ "block": block_cls(
+ inner_dim,
+ num_attention_heads,
+ attention_head_dim,
+ dropout=dropout,
+ cross_attention_dim=cross_attention_dim,
+ activation_fn=activation_fn,
+ num_embeds_ada_norm=num_embeds_ada_norm,
+ attention_bias=attention_bias,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ norm_type=norm_type,
+ pre_layer_norm=pre_layer_norm,
+ norm_elementwise_affine=norm_elementwise_affine,
+ final_dropout=ff_final_dropout,
+ ),
+ }
+ )
+ for d in range(num_layers // 2)
+ ]
+ )
+
+ # 4. Define output layers
+ self.out_channels = in_channels if out_channels is None else out_channels
+
+ # Following the UniDiffuser U-ViT implementation, we process the transformer output with
+ # a LayerNorm layer with per-element affine params
+ self.norm_out = nn.LayerNorm(inner_dim)
+
+ def forward(
+ self,
+ hidden_states,
+ encoder_hidden_states=None,
+ timestep=None,
+ class_labels=None,
+ cross_attention_kwargs=None,
+ return_dict: bool = True,
+ hidden_states_is_embedding: bool = False,
+ unpatchify: bool = True,
+ ):
+ """
+ Args:
+ hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`.
+ When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input
+ hidden_states
+ encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+ Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+ self-attention.
+ timestep ( `torch.long`, *optional*):
+ Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step.
+ class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+ Optional class labels to be applied as an embedding in AdaLayerZeroNorm. Used to indicate class labels
+ conditioning.
+ cross_attention_kwargs (*optional*):
+ Keyword arguments to supply to the cross attention layers, if used.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
+ hidden_states_is_embedding (`bool`, *optional*, defaults to `False`):
+ Whether or not hidden_states is an embedding directly usable by the transformer. In this case we will
+ ignore input handling (e.g. continuous, vectorized, etc.) and directly feed hidden_states into the
+ transformer blocks.
+ unpatchify (`bool`, *optional*, defaults to `True`):
+ Whether to unpatchify the transformer output.
+
+ Returns:
+ [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`:
+ [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When
+ returning a tuple, the first element is the sample tensor.
+ """
+ # 0. Check inputs
+
+ if not unpatchify and return_dict:
+ raise ValueError(
+ f"Cannot both define `unpatchify`: {unpatchify} and `return_dict`: {return_dict} since when"
+ f" `unpatchify` is {unpatchify} the returned output is of shape (batch_size, seq_len, hidden_dim)"
+ " rather than (batch_size, num_channels, height, width)."
+ )
+
+ # 1. Input
+ if not hidden_states_is_embedding:
+ hidden_states = self.pos_embed(hidden_states)
+
+ # 2. Blocks
+
+ # In ("downsample") blocks
+ skips = []
+ for in_block in self.transformer_in_blocks:
+ hidden_states = in_block(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ timestep=timestep,
+ cross_attention_kwargs=cross_attention_kwargs,
+ class_labels=class_labels,
+ )
+ skips.append(hidden_states)
+
+ # Mid block
+ hidden_states = self.transformer_mid_block(hidden_states)
+
+ # Out ("upsample") blocks
+ for out_block in self.transformer_out_blocks:
+ hidden_states = out_block["skip"](hidden_states, skips.pop())
+ hidden_states = out_block["block"](
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ timestep=timestep,
+ cross_attention_kwargs=cross_attention_kwargs,
+ class_labels=class_labels,
+ )
+
+ # 3. Output
+ # Don't support AdaLayerNorm for now, so no conditioning/scale/shift logic
+ hidden_states = self.norm_out(hidden_states)
+ # hidden_states = self.proj_out(hidden_states)
+
+ if unpatchify:
+ # unpatchify
+ height = width = int(hidden_states.shape[1] ** 0.5)
+ hidden_states = hidden_states.reshape(
+ shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+ )
+ hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+ output = hidden_states.reshape(
+ shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+ )
+ else:
+ output = hidden_states
+
+ if not return_dict:
+ return (output,)
+
+ return Transformer2DModelOutput(sample=output)
+
+
+class UniDiffuserModel(ModelMixin, ConfigMixin):
+ """
+ Transformer model for a image-text [UniDiffuser](https://arxiv.org/pdf/2303.06555.pdf) model. This is a
+ modification of [`UTransformer2DModel`] with input and output heads for the VAE-embedded latent image, the
+ CLIP-embedded image, and the CLIP-embedded prompt (see paper for more details).
+
+ Parameters:
+ text_dim (`int`): The hidden dimension of the CLIP text model used to embed images.
+ clip_img_dim (`int`): The hidden dimension of the CLIP vision model used to embed prompts.
+ num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+ attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+ in_channels (`int`, *optional*):
+ Pass if the input is continuous. The number of channels in the input.
+ out_channels (`int`, *optional*):
+ The number of output channels; if `None`, defaults to `in_channels`.
+ num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+ dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+ norm_num_groups (`int`, *optional*, defaults to `32`):
+ The number of groups to use when performing Group Normalization.
+ cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use.
+ attention_bias (`bool`, *optional*):
+ Configure if the TransformerBlocks' attention should contain a bias parameter.
+ sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images.
+ Note that this is fixed at training time as it is used for learning a number of position embeddings. See
+ `ImagePositionalEmbeddings`.
+ num_vector_embeds (`int`, *optional*):
+ Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels.
+ Includes the class for the masked latent pixel.
+ patch_size (`int`, *optional*, defaults to 2):
+ The patch size to use in the patch embedding.
+ activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+ num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`.
+ The number of diffusion steps used during training. Note that this is fixed at training time as it is used
+ to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for
+ up to but not more than steps than `num_embeds_ada_norm`.
+ use_linear_projection (int, *optional*): TODO: Not used
+ only_cross_attention (`bool`, *optional*):
+ Whether to use only cross-attention layers. In this case two cross attention layers are used in each
+ transformer block.
+ upcast_attention (`bool`, *optional*):
+ Whether to upcast the query and key to float32 when performing the attention calculation.
+ norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+ The Layer Normalization implementation to use. Defaults to `torch.nn.LayerNorm`.
+ block_type (`str`, *optional*, defaults to `"unidiffuser"`):
+ The transformer block implementation to use. If `"unidiffuser"`, has the LayerNorms on the residual
+ backbone of each transformer block; otherwise has them in the attention/feedforward branches (the standard
+ behavior in `diffusers`.)
+ pre_layer_norm (`bool`, *optional*):
+ Whether to perform layer normalization before the attention and feedforward operations ("pre-LayerNorm"),
+ as opposed to after ("post-LayerNorm"). The original UniDiffuser implementation is post-LayerNorm
+ (`pre_layer_norm = False`).
+ norm_elementwise_affine (`bool`, *optional*):
+ Whether to use learnable per-element affine parameters during layer normalization.
+ use_patch_pos_embed (`bool`, *optional*):
+ Whether to use position embeddings inside the patch embedding layer (`PatchEmbed`).
+ ff_final_dropout (`bool`, *optional*):
+ Whether to use a final Dropout layer after the feedforward network.
+ use_data_type_embedding (`bool`, *optional*):
+ Whether to use a data type embedding. This is only relevant for UniDiffuser-v1 style models; UniDiffuser-v1
+ is continue-trained from UniDiffuser-v0 on non-publically-available data and accepts a `data_type`
+ argument, which can either be `1` to use the weights trained on non-publically-available data or `0`
+ otherwise. This argument is subsequently embedded by the data type embedding, if used.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ text_dim: int = 768,
+ clip_img_dim: int = 512,
+ num_text_tokens: int = 77,
+ num_attention_heads: int = 16,
+ attention_head_dim: int = 88,
+ in_channels: Optional[int] = None,
+ out_channels: Optional[int] = None,
+ num_layers: int = 1,
+ dropout: float = 0.0,
+ norm_num_groups: int = 32,
+ cross_attention_dim: Optional[int] = None,
+ attention_bias: bool = False,
+ sample_size: Optional[int] = None,
+ num_vector_embeds: Optional[int] = None,
+ patch_size: Optional[int] = None,
+ activation_fn: str = "geglu",
+ num_embeds_ada_norm: Optional[int] = None,
+ use_linear_projection: bool = False,
+ only_cross_attention: bool = False,
+ upcast_attention: bool = False,
+ norm_type: str = "layer_norm",
+ block_type: str = "unidiffuser",
+ pre_layer_norm: bool = False,
+ use_timestep_embedding=False,
+ norm_elementwise_affine: bool = True,
+ use_patch_pos_embed=False,
+ ff_final_dropout: bool = True,
+ use_data_type_embedding: bool = False,
+ ):
+ super().__init__()
+
+ # 0. Handle dimensions
+ self.inner_dim = num_attention_heads * attention_head_dim
+
+ assert sample_size is not None, "UniDiffuserModel over patched input must provide sample_size"
+ self.sample_size = sample_size
+ self.in_channels = in_channels
+ self.out_channels = in_channels if out_channels is None else out_channels
+
+ self.patch_size = patch_size
+ # Assume image is square...
+ self.num_patches = (self.sample_size // patch_size) * (self.sample_size // patch_size)
+
+ # 1. Define input layers
+ # 1.1 Input layers for text and image input
+ # For now, only support patch input for VAE latent image input
+ self.vae_img_in = PatchEmbed(
+ height=sample_size,
+ width=sample_size,
+ patch_size=patch_size,
+ in_channels=in_channels,
+ embed_dim=self.inner_dim,
+ use_pos_embed=use_patch_pos_embed,
+ )
+ self.clip_img_in = nn.Linear(clip_img_dim, self.inner_dim)
+ self.text_in = nn.Linear(text_dim, self.inner_dim)
+
+ # 1.2. Timestep embeddings for t_img, t_text
+ self.timestep_img_proj = Timesteps(
+ self.inner_dim,
+ flip_sin_to_cos=True,
+ downscale_freq_shift=0,
+ )
+ self.timestep_img_embed = (
+ TimestepEmbedding(
+ self.inner_dim,
+ 4 * self.inner_dim,
+ out_dim=self.inner_dim,
+ )
+ if use_timestep_embedding
+ else nn.Identity()
+ )
+
+ self.timestep_text_proj = Timesteps(
+ self.inner_dim,
+ flip_sin_to_cos=True,
+ downscale_freq_shift=0,
+ )
+ self.timestep_text_embed = (
+ TimestepEmbedding(
+ self.inner_dim,
+ 4 * self.inner_dim,
+ out_dim=self.inner_dim,
+ )
+ if use_timestep_embedding
+ else nn.Identity()
+ )
+
+ # 1.3. Positional embedding
+ self.num_text_tokens = num_text_tokens
+ self.num_tokens = 1 + 1 + num_text_tokens + 1 + self.num_patches
+ self.pos_embed = nn.Parameter(torch.zeros(1, self.num_tokens, self.inner_dim))
+ self.pos_embed_drop = nn.Dropout(p=dropout)
+ trunc_normal_(self.pos_embed, std=0.02)
+
+ # 1.4. Handle data type token embeddings for UniDiffuser-V1, if necessary
+ self.use_data_type_embedding = use_data_type_embedding
+ if self.use_data_type_embedding:
+ self.data_type_token_embedding = nn.Embedding(2, self.inner_dim)
+ self.data_type_pos_embed_token = nn.Parameter(torch.zeros(1, 1, self.inner_dim))
+
+ # 2. Define transformer blocks
+ self.transformer = UTransformer2DModel(
+ num_attention_heads=num_attention_heads,
+ attention_head_dim=attention_head_dim,
+ in_channels=in_channels,
+ out_channels=out_channels,
+ num_layers=num_layers,
+ dropout=dropout,
+ norm_num_groups=norm_num_groups,
+ cross_attention_dim=cross_attention_dim,
+ attention_bias=attention_bias,
+ sample_size=sample_size,
+ num_vector_embeds=num_vector_embeds,
+ patch_size=patch_size,
+ activation_fn=activation_fn,
+ num_embeds_ada_norm=num_embeds_ada_norm,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention,
+ upcast_attention=upcast_attention,
+ norm_type=norm_type,
+ block_type=block_type,
+ pre_layer_norm=pre_layer_norm,
+ norm_elementwise_affine=norm_elementwise_affine,
+ use_patch_pos_embed=use_patch_pos_embed,
+ ff_final_dropout=ff_final_dropout,
+ )
+
+ # 3. Define output layers
+ patch_dim = (patch_size**2) * out_channels
+ self.vae_img_out = nn.Linear(self.inner_dim, patch_dim)
+ self.clip_img_out = nn.Linear(self.inner_dim, clip_img_dim)
+ self.text_out = nn.Linear(self.inner_dim, text_dim)
+
+ @torch.jit.ignore
+ def no_weight_decay(self):
+ return {"pos_embed"}
+
+ def forward(
+ self,
+ latent_image_embeds: torch.FloatTensor,
+ image_embeds: torch.FloatTensor,
+ prompt_embeds: torch.FloatTensor,
+ timestep_img: Union[torch.Tensor, float, int],
+ timestep_text: Union[torch.Tensor, float, int],
+ data_type: Optional[Union[torch.Tensor, float, int]] = 1,
+ encoder_hidden_states=None,
+ cross_attention_kwargs=None,
+ ):
+ """
+ Args:
+ latent_image_embeds (`torch.FloatTensor` of shape `(batch size, latent channels, height, width)`):
+ Latent image representation from the VAE encoder.
+ image_embeds (`torch.FloatTensor` of shape `(batch size, 1, clip_img_dim)`):
+ CLIP-embedded image representation (unsqueezed in the first dimension).
+ prompt_embeds (`torch.FloatTensor` of shape `(batch size, seq_len, text_dim)`):
+ CLIP-embedded text representation.
+ timestep_img (`torch.long` or `float` or `int`):
+ Current denoising step for the image.
+ timestep_text (`torch.long` or `float` or `int`):
+ Current denoising step for the text.
+ data_type: (`torch.int` or `float` or `int`, *optional*, defaults to `1`):
+ Only used in UniDiffuser-v1-style models. Can be either `1`, to use weights trained on nonpublic data,
+ or `0` otherwise.
+ encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
+ Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+ self-attention.
+ cross_attention_kwargs (*optional*):
+ Keyword arguments to supply to the cross attention layers, if used.
+
+
+ Returns:
+ `tuple`: Returns relevant parts of the model's noise prediction: the first element of the tuple is tbe VAE
+ image embedding, the second element is the CLIP image embedding, and the third element is the CLIP text
+ embedding.
+ """
+ batch_size = latent_image_embeds.shape[0]
+
+ # 1. Input
+ # 1.1. Map inputs to shape (B, N, inner_dim)
+ vae_hidden_states = self.vae_img_in(latent_image_embeds)
+ clip_hidden_states = self.clip_img_in(image_embeds)
+ text_hidden_states = self.text_in(prompt_embeds)
+
+ num_text_tokens, num_img_tokens = text_hidden_states.size(1), vae_hidden_states.size(1)
+
+ # 1.2. Encode image timesteps to single token (B, 1, inner_dim)
+ if not torch.is_tensor(timestep_img):
+ timestep_img = torch.tensor([timestep_img], dtype=torch.long, device=vae_hidden_states.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timestep_img = timestep_img * torch.ones(batch_size, dtype=timestep_img.dtype, device=timestep_img.device)
+
+ timestep_img_token = self.timestep_img_proj(timestep_img)
+ # t_img_token does not contain any weights and will always return f32 tensors
+ # but time_embedding might be fp16, so we need to cast here.
+ timestep_img_token = timestep_img_token.to(dtype=self.dtype)
+ timestep_img_token = self.timestep_img_embed(timestep_img_token)
+ timestep_img_token = timestep_img_token.unsqueeze(dim=1)
+
+ # 1.3. Encode text timesteps to single token (B, 1, inner_dim)
+ if not torch.is_tensor(timestep_text):
+ timestep_text = torch.tensor([timestep_text], dtype=torch.long, device=vae_hidden_states.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timestep_text = timestep_text * torch.ones(batch_size, dtype=timestep_text.dtype, device=timestep_text.device)
+
+ timestep_text_token = self.timestep_text_proj(timestep_text)
+ # t_text_token does not contain any weights and will always return f32 tensors
+ # but time_embedding might be fp16, so we need to cast here.
+ timestep_text_token = timestep_text_token.to(dtype=self.dtype)
+ timestep_text_token = self.timestep_text_embed(timestep_text_token)
+ timestep_text_token = timestep_text_token.unsqueeze(dim=1)
+
+ # 1.4. Concatenate all of the embeddings together.
+ if self.use_data_type_embedding:
+ assert data_type is not None, "data_type must be supplied if the model uses a data type embedding"
+ if not torch.is_tensor(data_type):
+ data_type = torch.tensor([data_type], dtype=torch.int, device=vae_hidden_states.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ data_type = data_type * torch.ones(batch_size, dtype=data_type.dtype, device=data_type.device)
+
+ data_type_token = self.data_type_token_embedding(data_type).unsqueeze(dim=1)
+ hidden_states = torch.cat(
+ [
+ timestep_img_token,
+ timestep_text_token,
+ data_type_token,
+ text_hidden_states,
+ clip_hidden_states,
+ vae_hidden_states,
+ ],
+ dim=1,
+ )
+ else:
+ hidden_states = torch.cat(
+ [timestep_img_token, timestep_text_token, text_hidden_states, clip_hidden_states, vae_hidden_states],
+ dim=1,
+ )
+
+ # 1.5. Prepare the positional embeddings and add to hidden states
+ # Note: I think img_vae should always have the proper shape, so there's no need to interpolate
+ # the position embeddings.
+ if self.use_data_type_embedding:
+ pos_embed = torch.cat(
+ [self.pos_embed[:, : 1 + 1, :], self.data_type_pos_embed_token, self.pos_embed[:, 1 + 1 :, :]], dim=1
+ )
+ else:
+ pos_embed = self.pos_embed
+ hidden_states = hidden_states + pos_embed
+ hidden_states = self.pos_embed_drop(hidden_states)
+
+ # 2. Blocks
+ hidden_states = self.transformer(
+ hidden_states,
+ encoder_hidden_states=encoder_hidden_states,
+ timestep=None,
+ class_labels=None,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ hidden_states_is_embedding=True,
+ unpatchify=False,
+ )[0]
+
+ # 3. Output
+ # Split out the predicted noise representation.
+ if self.use_data_type_embedding:
+ (
+ t_img_token_out,
+ t_text_token_out,
+ data_type_token_out,
+ text_out,
+ img_clip_out,
+ img_vae_out,
+ ) = hidden_states.split((1, 1, 1, num_text_tokens, 1, num_img_tokens), dim=1)
+ else:
+ t_img_token_out, t_text_token_out, text_out, img_clip_out, img_vae_out = hidden_states.split(
+ (1, 1, num_text_tokens, 1, num_img_tokens), dim=1
+ )
+
+ img_vae_out = self.vae_img_out(img_vae_out)
+
+ # unpatchify
+ height = width = int(img_vae_out.shape[1] ** 0.5)
+ img_vae_out = img_vae_out.reshape(
+ shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+ )
+ img_vae_out = torch.einsum("nhwpqc->nchpwq", img_vae_out)
+ img_vae_out = img_vae_out.reshape(
+ shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+ )
+
+ img_clip_out = self.clip_img_out(img_clip_out)
+
+ text_out = self.text_out(text_out)
+
+ return img_vae_out, img_clip_out, text_out
diff --git a/model/diffusers_c/pipelines/unidiffuser/pipeline_unidiffuser.py b/model/diffusers_c/pipelines/unidiffuser/pipeline_unidiffuser.py
new file mode 100644
index 0000000000000000000000000000000000000000..5d61b1054e1caaada47ee87ef414723c014c5db2
--- /dev/null
+++ b/model/diffusers_c/pipelines/unidiffuser/pipeline_unidiffuser.py
@@ -0,0 +1,1419 @@
+import inspect
+from dataclasses import dataclass
+from typing import Callable, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+from transformers import (
+ CLIPImageProcessor,
+ CLIPTextModel,
+ CLIPTokenizer,
+ CLIPVisionModelWithProjection,
+ GPT2Tokenizer,
+)
+
+from ...image_processor import VaeImageProcessor
+from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import USE_PEFT_BACKEND, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from ...utils.outputs import BaseOutput
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .modeling_text_decoder import UniDiffuserTextDecoder
+from .modeling_uvit import UniDiffuserModel
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# New BaseOutput child class for joint image-text output
+@dataclass
+class ImageTextPipelineOutput(BaseOutput):
+ """
+ Output class for joint image-text pipelines.
+
+ Args:
+ images (`List[PIL.Image.Image]` or `np.ndarray`)
+ List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
+ num_channels)`.
+ text (`List[str]` or `List[List[str]]`)
+ List of generated text strings of length `batch_size` or a list of list of strings whose outer list has
+ length `batch_size`.
+ """
+
+ images: Optional[Union[List[PIL.Image.Image], np.ndarray]]
+ text: Optional[Union[List[str], List[List[str]]]]
+
+
+class UniDiffuserPipeline(DiffusionPipeline):
+ r"""
+ Pipeline for a bimodal image-text model which supports unconditional text and image generation, text-conditioned
+ image generation, image-conditioned text generation, and joint image-text generation.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. This
+ is part of the UniDiffuser image representation along with the CLIP vision encoding.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ image_encoder ([`CLIPVisionModel`]):
+ A [`~transformers.CLIPVisionModel`] to encode images as part of its image representation along with the VAE
+ latent representation.
+ image_processor ([`CLIPImageProcessor`]):
+ [`~transformers.CLIPImageProcessor`] to preprocess an image before CLIP encoding it with `image_encoder`.
+ clip_tokenizer ([`CLIPTokenizer`]):
+ A [`~transformers.CLIPTokenizer`] to tokenize the prompt before encoding it with `text_encoder`.
+ text_decoder ([`UniDiffuserTextDecoder`]):
+ Frozen text decoder. This is a GPT-style model which is used to generate text from the UniDiffuser
+ embedding.
+ text_tokenizer ([`GPT2Tokenizer`]):
+ A [`~transformers.GPT2Tokenizer`] to decode text for text generation; used along with the `text_decoder`.
+ unet ([`UniDiffuserModel`]):
+ A [U-ViT](https://github.com/baofff/U-ViT) model with UNNet-style skip connections between transformer
+ layers to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image and/or text latents. The
+ original UniDiffuser paper uses the [`DPMSolverMultistepScheduler`] scheduler.
+ """
+
+ # TODO: support for moving submodules for components with enable_model_cpu_offload
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae->text_decoder"
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ image_encoder: CLIPVisionModelWithProjection,
+ clip_image_processor: CLIPImageProcessor,
+ clip_tokenizer: CLIPTokenizer,
+ text_decoder: UniDiffuserTextDecoder,
+ text_tokenizer: GPT2Tokenizer,
+ unet: UniDiffuserModel,
+ scheduler: KarrasDiffusionSchedulers,
+ ):
+ super().__init__()
+
+ if text_encoder.config.hidden_size != text_decoder.prefix_inner_dim:
+ raise ValueError(
+ f"The text encoder hidden size and text decoder prefix inner dim must be the same, but"
+ f" `text_encoder.config.hidden_size`: {text_encoder.config.hidden_size} and `text_decoder.prefix_inner_dim`: {text_decoder.prefix_inner_dim}"
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ image_encoder=image_encoder,
+ clip_image_processor=clip_image_processor,
+ clip_tokenizer=clip_tokenizer,
+ text_decoder=text_decoder,
+ text_tokenizer=text_tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ )
+
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+
+ self.num_channels_latents = vae.config.latent_channels
+ self.text_encoder_seq_len = text_encoder.config.max_position_embeddings
+ self.text_encoder_hidden_size = text_encoder.config.hidden_size
+ self.image_encoder_projection_dim = image_encoder.config.projection_dim
+ self.unet_resolution = unet.config.sample_size
+
+ self.text_intermediate_dim = self.text_encoder_hidden_size
+ if self.text_decoder.prefix_hidden_dim is not None:
+ self.text_intermediate_dim = self.text_decoder.prefix_hidden_dim
+
+ self.mode = None
+
+ # TODO: handle safety checking?
+ self.safety_checker = None
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def _infer_mode(self, prompt, prompt_embeds, image, latents, prompt_latents, vae_latents, clip_latents):
+ r"""
+ Infer the generation task ('mode') from the inputs to `__call__`. If the mode has been manually set, the set
+ mode will be used.
+ """
+ prompt_available = (prompt is not None) or (prompt_embeds is not None)
+ image_available = image is not None
+ input_available = prompt_available or image_available
+
+ prompt_latents_available = prompt_latents is not None
+ vae_latents_available = vae_latents is not None
+ clip_latents_available = clip_latents is not None
+ full_latents_available = latents is not None
+ image_latents_available = vae_latents_available and clip_latents_available
+ all_indv_latents_available = prompt_latents_available and image_latents_available
+
+ if self.mode is not None:
+ # Preferentially use the mode set by the user
+ mode = self.mode
+ elif prompt_available:
+ mode = "text2img"
+ elif image_available:
+ mode = "img2text"
+ else:
+ # Neither prompt nor image supplied, infer based on availability of latents
+ if full_latents_available or all_indv_latents_available:
+ mode = "joint"
+ elif prompt_latents_available:
+ mode = "text"
+ elif image_latents_available:
+ mode = "img"
+ else:
+ # No inputs or latents available
+ mode = "joint"
+
+ # Give warnings for ambiguous cases
+ if self.mode is None and prompt_available and image_available:
+ logger.warning(
+ f"You have supplied both a text prompt and image to the pipeline and mode has not been set manually,"
+ f" defaulting to mode '{mode}'."
+ )
+
+ if self.mode is None and not input_available:
+ if vae_latents_available != clip_latents_available:
+ # Exactly one of vae_latents and clip_latents is supplied
+ logger.warning(
+ f"You have supplied exactly one of `vae_latents` and `clip_latents`, whereas either both or none"
+ f" are expected to be supplied. Defaulting to mode '{mode}'."
+ )
+ elif not prompt_latents_available and not vae_latents_available and not clip_latents_available:
+ # No inputs or latents supplied
+ logger.warning(
+ f"No inputs or latents have been supplied, and mode has not been manually set,"
+ f" defaulting to mode '{mode}'."
+ )
+
+ return mode
+
+ # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.enable_vae_slicing
+ def enable_vae_slicing(self):
+ r"""
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+ """
+ self.vae.enable_slicing()
+
+ # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.disable_vae_slicing
+ def disable_vae_slicing(self):
+ r"""
+ Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+ computing decoding in one step.
+ """
+ self.vae.disable_slicing()
+
+ # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.enable_vae_tiling
+ def enable_vae_tiling(self):
+ r"""
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+ processing larger images.
+ """
+ self.vae.enable_tiling()
+
+ # Copied from diffusers.pipelines.pipeline_utils.StableDiffusionMixin.disable_vae_tiling
+ def disable_vae_tiling(self):
+ r"""
+ Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+ computing decoding in one step.
+ """
+ self.vae.disable_tiling()
+
+ # Functions to manually set the mode
+ def set_text_mode(self):
+ r"""Manually set the generation mode to unconditional ("marginal") text generation."""
+ self.mode = "text"
+
+ def set_image_mode(self):
+ r"""Manually set the generation mode to unconditional ("marginal") image generation."""
+ self.mode = "img"
+
+ def set_text_to_image_mode(self):
+ r"""Manually set the generation mode to text-conditioned image generation."""
+ self.mode = "text2img"
+
+ def set_image_to_text_mode(self):
+ r"""Manually set the generation mode to image-conditioned text generation."""
+ self.mode = "img2text"
+
+ def set_joint_mode(self):
+ r"""Manually set the generation mode to unconditional joint image-text generation."""
+ self.mode = "joint"
+
+ def reset_mode(self):
+ r"""Removes a manually set mode; after calling this, the pipeline will infer the mode from inputs."""
+ self.mode = None
+
+ def _infer_batch_size(
+ self,
+ mode,
+ prompt,
+ prompt_embeds,
+ image,
+ num_images_per_prompt,
+ num_prompts_per_image,
+ latents,
+ prompt_latents,
+ vae_latents,
+ clip_latents,
+ ):
+ r"""Infers the batch size and multiplier depending on mode and supplied arguments to `__call__`."""
+ if num_images_per_prompt is None:
+ num_images_per_prompt = 1
+ if num_prompts_per_image is None:
+ num_prompts_per_image = 1
+
+ assert num_images_per_prompt > 0, "num_images_per_prompt must be a positive integer"
+ assert num_prompts_per_image > 0, "num_prompts_per_image must be a positive integer"
+
+ if mode in ["text2img"]:
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ # Either prompt or prompt_embeds must be present for text2img.
+ batch_size = prompt_embeds.shape[0]
+ multiplier = num_images_per_prompt
+ elif mode in ["img2text"]:
+ if isinstance(image, PIL.Image.Image):
+ batch_size = 1
+ else:
+ # Image must be available and type either PIL.Image.Image or torch.FloatTensor.
+ # Not currently supporting something like image_embeds.
+ batch_size = image.shape[0]
+ multiplier = num_prompts_per_image
+ elif mode in ["img"]:
+ if vae_latents is not None:
+ batch_size = vae_latents.shape[0]
+ elif clip_latents is not None:
+ batch_size = clip_latents.shape[0]
+ else:
+ batch_size = 1
+ multiplier = num_images_per_prompt
+ elif mode in ["text"]:
+ if prompt_latents is not None:
+ batch_size = prompt_latents.shape[0]
+ else:
+ batch_size = 1
+ multiplier = num_prompts_per_image
+ elif mode in ["joint"]:
+ if latents is not None:
+ batch_size = latents.shape[0]
+ elif prompt_latents is not None:
+ batch_size = prompt_latents.shape[0]
+ elif vae_latents is not None:
+ batch_size = vae_latents.shape[0]
+ elif clip_latents is not None:
+ batch_size = clip_latents.shape[0]
+ else:
+ batch_size = 1
+
+ if num_images_per_prompt == num_prompts_per_image:
+ multiplier = num_images_per_prompt
+ else:
+ multiplier = min(num_images_per_prompt, num_prompts_per_image)
+ logger.warning(
+ f"You are using mode `{mode}` and `num_images_per_prompt`: {num_images_per_prompt} and"
+ f" num_prompts_per_image: {num_prompts_per_image} are not equal. Using batch size equal to"
+ f" `min(num_images_per_prompt, num_prompts_per_image) = {batch_size}."
+ )
+ return batch_size, multiplier
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ **kwargs,
+ ):
+ deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+ deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+ prompt_embeds_tuple = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=lora_scale,
+ **kwargs,
+ )
+
+ # concatenate for backwards comp
+ prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with self.tokenizer->self.clip_tokenizer
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.clip_tokenizer)
+
+ text_inputs = self.clip_tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.clip_tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.clip_tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.clip_tokenizer.batch_decode(
+ untruncated_ids[:, self.clip_tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.clip_tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.clip_tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.clip_tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_instruct_pix2pix.StableDiffusionInstructPix2PixPipeline.prepare_image_latents
+ # Add num_prompts_per_image argument, sample from autoencoder moment distribution
+ def encode_image_vae_latents(
+ self,
+ image,
+ batch_size,
+ num_prompts_per_image,
+ dtype,
+ device,
+ do_classifier_free_guidance,
+ generator=None,
+ ):
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ image = image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_prompts_per_image
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if isinstance(generator, list):
+ image_latents = [
+ self.vae.encode(image[i : i + 1]).latent_dist.sample(generator=generator[i])
+ * self.vae.config.scaling_factor
+ for i in range(batch_size)
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = self.vae.encode(image).latent_dist.sample(generator=generator)
+ # Scale image_latents by the VAE's scaling factor
+ image_latents = image_latents * self.vae.config.scaling_factor
+
+ if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+ # expand image_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // image_latents.shape[0]
+ image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ image_latents = torch.cat([image_latents], dim=0)
+
+ if do_classifier_free_guidance:
+ uncond_image_latents = torch.zeros_like(image_latents)
+ image_latents = torch.cat([image_latents, image_latents, uncond_image_latents], dim=0)
+
+ return image_latents
+
+ def encode_image_clip_latents(
+ self,
+ image,
+ batch_size,
+ num_prompts_per_image,
+ dtype,
+ device,
+ generator=None,
+ ):
+ # Map image to CLIP embedding.
+ if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)):
+ raise ValueError(
+ f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}"
+ )
+
+ preprocessed_image = self.clip_image_processor.preprocess(
+ image,
+ return_tensors="pt",
+ )
+ preprocessed_image = preprocessed_image.to(device=device, dtype=dtype)
+
+ batch_size = batch_size * num_prompts_per_image
+ if isinstance(generator, list):
+ image_latents = [
+ self.image_encoder(**preprocessed_image[i : i + 1]).image_embeds for i in range(batch_size)
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = self.image_encoder(**preprocessed_image).image_embeds
+
+ if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
+ # expand image_latents for batch_size
+ deprecation_message = (
+ f"You have passed {batch_size} text prompts (`prompt`), but only {image_latents.shape[0]} initial"
+ " images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
+ " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
+ " your script to pass as many initial images as text prompts to suppress this warning."
+ )
+ deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
+ additional_image_per_prompt = batch_size // image_latents.shape[0]
+ image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
+ elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
+ raise ValueError(
+ f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
+ )
+ else:
+ image_latents = torch.cat([image_latents], dim=0)
+
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ return image_latents
+
+ def prepare_text_latents(
+ self, batch_size, num_images_per_prompt, seq_len, hidden_size, dtype, device, generator, latents=None
+ ):
+ # Prepare latents for the CLIP embedded prompt.
+ shape = (batch_size * num_images_per_prompt, seq_len, hidden_size)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ # latents is assumed to have shace (B, L, D)
+ latents = latents.repeat(num_images_per_prompt, 1, 1)
+ latents = latents.to(device=device, dtype=dtype)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ # Rename prepare_latents -> prepare_image_vae_latents and add num_prompts_per_image argument.
+ def prepare_image_vae_latents(
+ self,
+ batch_size,
+ num_prompts_per_image,
+ num_channels_latents,
+ height,
+ width,
+ dtype,
+ device,
+ generator,
+ latents=None,
+ ):
+ shape = (
+ batch_size * num_prompts_per_image,
+ num_channels_latents,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ # latents is assumed to have shape (B, C, H, W)
+ latents = latents.repeat(num_prompts_per_image, 1, 1, 1)
+ latents = latents.to(device=device, dtype=dtype)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def prepare_image_clip_latents(
+ self, batch_size, num_prompts_per_image, clip_img_dim, dtype, device, generator, latents=None
+ ):
+ # Prepare latents for the CLIP embedded image.
+ shape = (batch_size * num_prompts_per_image, 1, clip_img_dim)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ # latents is assumed to have shape (B, L, D)
+ latents = latents.repeat(num_prompts_per_image, 1, 1)
+ latents = latents.to(device=device, dtype=dtype)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def decode_text_latents(self, text_latents, device):
+ output_token_list, seq_lengths = self.text_decoder.generate_captions(
+ text_latents, self.text_tokenizer.eos_token_id, device=device
+ )
+ output_list = output_token_list.cpu().numpy()
+ generated_text = [
+ self.text_tokenizer.decode(output[: int(length)], skip_special_tokens=True)
+ for output, length in zip(output_list, seq_lengths)
+ ]
+ return generated_text
+
+ def _split(self, x, height, width):
+ r"""
+ Splits a flattened embedding x of shape (B, C * H * W + clip_img_dim) into two tensors of shape (B, C, H, W)
+ and (B, 1, clip_img_dim)
+ """
+ batch_size = x.shape[0]
+ latent_height = height // self.vae_scale_factor
+ latent_width = width // self.vae_scale_factor
+ img_vae_dim = self.num_channels_latents * latent_height * latent_width
+
+ img_vae, img_clip = x.split([img_vae_dim, self.image_encoder_projection_dim], dim=1)
+
+ img_vae = torch.reshape(img_vae, (batch_size, self.num_channels_latents, latent_height, latent_width))
+ img_clip = torch.reshape(img_clip, (batch_size, 1, self.image_encoder_projection_dim))
+ return img_vae, img_clip
+
+ def _combine(self, img_vae, img_clip):
+ r"""
+ Combines a latent iamge img_vae of shape (B, C, H, W) and a CLIP-embedded image img_clip of shape (B, 1,
+ clip_img_dim) into a single tensor of shape (B, C * H * W + clip_img_dim).
+ """
+ img_vae = torch.reshape(img_vae, (img_vae.shape[0], -1))
+ img_clip = torch.reshape(img_clip, (img_clip.shape[0], -1))
+ return torch.concat([img_vae, img_clip], dim=-1)
+
+ def _split_joint(self, x, height, width):
+ r"""
+ Splits a flattened embedding x of shape (B, C * H * W + clip_img_dim + text_seq_len * text_dim] into (img_vae,
+ img_clip, text) where img_vae is of shape (B, C, H, W), img_clip is of shape (B, 1, clip_img_dim), and text is
+ of shape (B, text_seq_len, text_dim).
+ """
+ batch_size = x.shape[0]
+ latent_height = height // self.vae_scale_factor
+ latent_width = width // self.vae_scale_factor
+ img_vae_dim = self.num_channels_latents * latent_height * latent_width
+ text_dim = self.text_encoder_seq_len * self.text_intermediate_dim
+
+ img_vae, img_clip, text = x.split([img_vae_dim, self.image_encoder_projection_dim, text_dim], dim=1)
+
+ img_vae = torch.reshape(img_vae, (batch_size, self.num_channels_latents, latent_height, latent_width))
+ img_clip = torch.reshape(img_clip, (batch_size, 1, self.image_encoder_projection_dim))
+ text = torch.reshape(text, (batch_size, self.text_encoder_seq_len, self.text_intermediate_dim))
+ return img_vae, img_clip, text
+
+ def _combine_joint(self, img_vae, img_clip, text):
+ r"""
+ Combines a latent image img_vae of shape (B, C, H, W), a CLIP-embedded image img_clip of shape (B, L_img,
+ clip_img_dim), and a text embedding text of shape (B, L_text, text_dim) into a single embedding x of shape (B,
+ C * H * W + L_img * clip_img_dim + L_text * text_dim).
+ """
+ img_vae = torch.reshape(img_vae, (img_vae.shape[0], -1))
+ img_clip = torch.reshape(img_clip, (img_clip.shape[0], -1))
+ text = torch.reshape(text, (text.shape[0], -1))
+ return torch.concat([img_vae, img_clip, text], dim=-1)
+
+ def _get_noise_pred(
+ self,
+ mode,
+ latents,
+ t,
+ prompt_embeds,
+ img_vae,
+ img_clip,
+ max_timestep,
+ data_type,
+ guidance_scale,
+ generator,
+ device,
+ height,
+ width,
+ ):
+ r"""
+ Gets the noise prediction using the `unet` and performs classifier-free guidance, if necessary.
+ """
+ if mode == "joint":
+ # Joint text-image generation
+ img_vae_latents, img_clip_latents, text_latents = self._split_joint(latents, height, width)
+
+ img_vae_out, img_clip_out, text_out = self.unet(
+ img_vae_latents, img_clip_latents, text_latents, timestep_img=t, timestep_text=t, data_type=data_type
+ )
+
+ x_out = self._combine_joint(img_vae_out, img_clip_out, text_out)
+
+ if guidance_scale <= 1.0:
+ return x_out
+
+ # Classifier-free guidance
+ img_vae_T = randn_tensor(img_vae.shape, generator=generator, device=device, dtype=img_vae.dtype)
+ img_clip_T = randn_tensor(img_clip.shape, generator=generator, device=device, dtype=img_clip.dtype)
+ text_T = randn_tensor(prompt_embeds.shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
+
+ _, _, text_out_uncond = self.unet(
+ img_vae_T, img_clip_T, text_latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type
+ )
+
+ img_vae_out_uncond, img_clip_out_uncond, _ = self.unet(
+ img_vae_latents,
+ img_clip_latents,
+ text_T,
+ timestep_img=t,
+ timestep_text=max_timestep,
+ data_type=data_type,
+ )
+
+ x_out_uncond = self._combine_joint(img_vae_out_uncond, img_clip_out_uncond, text_out_uncond)
+
+ return guidance_scale * x_out + (1.0 - guidance_scale) * x_out_uncond
+ elif mode == "text2img":
+ # Text-conditioned image generation
+ img_vae_latents, img_clip_latents = self._split(latents, height, width)
+
+ img_vae_out, img_clip_out, text_out = self.unet(
+ img_vae_latents, img_clip_latents, prompt_embeds, timestep_img=t, timestep_text=0, data_type=data_type
+ )
+
+ img_out = self._combine(img_vae_out, img_clip_out)
+
+ if guidance_scale <= 1.0:
+ return img_out
+
+ # Classifier-free guidance
+ text_T = randn_tensor(prompt_embeds.shape, generator=generator, device=device, dtype=prompt_embeds.dtype)
+
+ img_vae_out_uncond, img_clip_out_uncond, text_out_uncond = self.unet(
+ img_vae_latents,
+ img_clip_latents,
+ text_T,
+ timestep_img=t,
+ timestep_text=max_timestep,
+ data_type=data_type,
+ )
+
+ img_out_uncond = self._combine(img_vae_out_uncond, img_clip_out_uncond)
+
+ return guidance_scale * img_out + (1.0 - guidance_scale) * img_out_uncond
+ elif mode == "img2text":
+ # Image-conditioned text generation
+ img_vae_out, img_clip_out, text_out = self.unet(
+ img_vae, img_clip, latents, timestep_img=0, timestep_text=t, data_type=data_type
+ )
+
+ if guidance_scale <= 1.0:
+ return text_out
+
+ # Classifier-free guidance
+ img_vae_T = randn_tensor(img_vae.shape, generator=generator, device=device, dtype=img_vae.dtype)
+ img_clip_T = randn_tensor(img_clip.shape, generator=generator, device=device, dtype=img_clip.dtype)
+
+ img_vae_out_uncond, img_clip_out_uncond, text_out_uncond = self.unet(
+ img_vae_T, img_clip_T, latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type
+ )
+
+ return guidance_scale * text_out + (1.0 - guidance_scale) * text_out_uncond
+ elif mode == "text":
+ # Unconditional ("marginal") text generation (no CFG)
+ img_vae_out, img_clip_out, text_out = self.unet(
+ img_vae, img_clip, latents, timestep_img=max_timestep, timestep_text=t, data_type=data_type
+ )
+
+ return text_out
+ elif mode == "img":
+ # Unconditional ("marginal") image generation (no CFG)
+ img_vae_latents, img_clip_latents = self._split(latents, height, width)
+
+ img_vae_out, img_clip_out, text_out = self.unet(
+ img_vae_latents,
+ img_clip_latents,
+ prompt_embeds,
+ timestep_img=t,
+ timestep_text=max_timestep,
+ data_type=data_type,
+ )
+
+ img_out = self._combine(img_vae_out, img_clip_out)
+ return img_out
+
+ def check_latents_shape(self, latents_name, latents, expected_shape):
+ latents_shape = latents.shape
+ expected_num_dims = len(expected_shape) + 1 # expected dimensions plus the batch dimension
+ expected_shape_str = ", ".join(str(dim) for dim in expected_shape)
+ if len(latents_shape) != expected_num_dims:
+ raise ValueError(
+ f"`{latents_name}` should have shape (batch_size, {expected_shape_str}), but the current shape"
+ f" {latents_shape} has {len(latents_shape)} dimensions."
+ )
+ for i in range(1, expected_num_dims):
+ if latents_shape[i] != expected_shape[i - 1]:
+ raise ValueError(
+ f"`{latents_name}` should have shape (batch_size, {expected_shape_str}), but the current shape"
+ f" {latents_shape} has {latents_shape[i]} != {expected_shape[i - 1]} at dimension {i}."
+ )
+
+ def check_inputs(
+ self,
+ mode,
+ prompt,
+ image,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ latents=None,
+ prompt_latents=None,
+ vae_latents=None,
+ clip_latents=None,
+ ):
+ # Check inputs before running the generative process.
+ if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0:
+ raise ValueError(
+ f"`height` and `width` have to be divisible by {self.vae_scale_factor} but are {height} and {width}."
+ )
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if mode == "text2img":
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if mode == "img2text":
+ if image is None:
+ raise ValueError("`img2text` mode requires an image to be provided.")
+
+ # Check provided latents
+ latent_height = height // self.vae_scale_factor
+ latent_width = width // self.vae_scale_factor
+ full_latents_available = latents is not None
+ prompt_latents_available = prompt_latents is not None
+ vae_latents_available = vae_latents is not None
+ clip_latents_available = clip_latents is not None
+
+ if full_latents_available:
+ individual_latents_available = (
+ prompt_latents is not None or vae_latents is not None or clip_latents is not None
+ )
+ if individual_latents_available:
+ logger.warning(
+ "You have supplied both `latents` and at least one of `prompt_latents`, `vae_latents`, and"
+ " `clip_latents`. The value of `latents` will override the value of any individually supplied latents."
+ )
+ # Check shape of full latents
+ img_vae_dim = self.num_channels_latents * latent_height * latent_width
+ text_dim = self.text_encoder_seq_len * self.text_encoder_hidden_size
+ latents_dim = img_vae_dim + self.image_encoder_projection_dim + text_dim
+ latents_expected_shape = (latents_dim,)
+ self.check_latents_shape("latents", latents, latents_expected_shape)
+
+ # Check individual latent shapes, if present
+ if prompt_latents_available:
+ prompt_latents_expected_shape = (self.text_encoder_seq_len, self.text_encoder_hidden_size)
+ self.check_latents_shape("prompt_latents", prompt_latents, prompt_latents_expected_shape)
+
+ if vae_latents_available:
+ vae_latents_expected_shape = (self.num_channels_latents, latent_height, latent_width)
+ self.check_latents_shape("vae_latents", vae_latents, vae_latents_expected_shape)
+
+ if clip_latents_available:
+ clip_latents_expected_shape = (1, self.image_encoder_projection_dim)
+ self.check_latents_shape("clip_latents", clip_latents, clip_latents_expected_shape)
+
+ if mode in ["text2img", "img"] and vae_latents_available and clip_latents_available:
+ if vae_latents.shape[0] != clip_latents.shape[0]:
+ raise ValueError(
+ f"Both `vae_latents` and `clip_latents` are supplied, but their batch dimensions are not equal:"
+ f" {vae_latents.shape[0]} != {clip_latents.shape[0]}."
+ )
+
+ if mode == "joint" and prompt_latents_available and vae_latents_available and clip_latents_available:
+ if prompt_latents.shape[0] != vae_latents.shape[0] or prompt_latents.shape[0] != clip_latents.shape[0]:
+ raise ValueError(
+ f"All of `prompt_latents`, `vae_latents`, and `clip_latents` are supplied, but their batch"
+ f" dimensions are not equal: {prompt_latents.shape[0]} != {vae_latents.shape[0]}"
+ f" != {clip_latents.shape[0]}."
+ )
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ image: Optional[Union[torch.FloatTensor, PIL.Image.Image]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ data_type: Optional[int] = 1,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 8.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ num_prompts_per_image: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_latents: Optional[torch.FloatTensor] = None,
+ vae_latents: Optional[torch.FloatTensor] = None,
+ clip_latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ Required for text-conditioned image generation (`text2img`) mode.
+ image (`torch.FloatTensor` or `PIL.Image.Image`, *optional*):
+ `Image` or tensor representing an image batch. Required for image-conditioned text generation
+ (`img2text`) mode.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ data_type (`int`, *optional*, defaults to 1):
+ The data type (either 0 or 1). Only used if you are loading a checkpoint which supports a data type
+ embedding; this is added for compatibility with the
+ [UniDiffuser-v1](https://huggingface.co/thu-ml/unidiffuser-v1) checkpoint.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 8.0):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). Used in
+ text-conditioned image generation (`text2img`) mode.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt. Used in `text2img` (text-conditioned image generation) and
+ `img` mode. If the mode is joint and both `num_images_per_prompt` and `num_prompts_per_image` are
+ supplied, `min(num_images_per_prompt, num_prompts_per_image)` samples are generated.
+ num_prompts_per_image (`int`, *optional*, defaults to 1):
+ The number of prompts to generate per image. Used in `img2text` (image-conditioned text generation) and
+ `text` mode. If the mode is joint and both `num_images_per_prompt` and `num_prompts_per_image` are
+ supplied, `min(num_images_per_prompt, num_prompts_per_image)` samples are generated.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for joint
+ image-text generation. Can be used to tweak the same generation with different prompts. If not
+ provided, a latents tensor is generated by sampling using the supplied random `generator`. This assumes
+ a full set of VAE, CLIP, and text latents, if supplied, overrides the value of `prompt_latents`,
+ `vae_latents`, and `clip_latents`.
+ prompt_latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for text
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ vae_latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ clip_latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument. Used in text-conditioned
+ image generation (`text2img`) mode.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are be generated from the `negative_prompt` input argument. Used
+ in text-conditioned image generation (`text2img`) mode.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImageTextPipelineOutput`] instead of a plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+
+ Returns:
+ [`~pipelines.unidiffuser.ImageTextPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.unidiffuser.ImageTextPipelineOutput`] is returned, otherwise a
+ `tuple` is returned where the first element is a list with the generated images and the second element
+ is a list of generated texts.
+ """
+
+ # 0. Default height and width to unet
+ height = height or self.unet_resolution * self.vae_scale_factor
+ width = width or self.unet_resolution * self.vae_scale_factor
+
+ # 1. Check inputs
+ # Recalculate mode for each call to the pipeline.
+ mode = self._infer_mode(prompt, prompt_embeds, image, latents, prompt_latents, vae_latents, clip_latents)
+ self.check_inputs(
+ mode,
+ prompt,
+ image,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ latents,
+ prompt_latents,
+ vae_latents,
+ clip_latents,
+ )
+
+ # 2. Define call parameters
+ batch_size, multiplier = self._infer_batch_size(
+ mode,
+ prompt,
+ prompt_embeds,
+ image,
+ num_images_per_prompt,
+ num_prompts_per_image,
+ latents,
+ prompt_latents,
+ vae_latents,
+ clip_latents,
+ )
+ device = self._execution_device
+ reduce_text_emb_dim = self.text_intermediate_dim < self.text_encoder_hidden_size or self.mode != "text2img"
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ # Note that this differs from the formulation in the unidiffusers paper!
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # check if scheduler is in sigmas space
+ # scheduler_is_in_sigma_space = hasattr(self.scheduler, "sigmas")
+
+ # 3. Encode input prompt, if available; otherwise prepare text latents
+ if latents is not None:
+ # Overwrite individual latents
+ vae_latents, clip_latents, prompt_latents = self._split_joint(latents, height, width)
+
+ if mode in ["text2img"]:
+ # 3.1. Encode input prompt, if available
+ assert prompt is not None or prompt_embeds is not None
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=multiplier,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ # if do_classifier_free_guidance:
+ # prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+ else:
+ # 3.2. Prepare text latent variables, if input not available
+ prompt_embeds = self.prepare_text_latents(
+ batch_size=batch_size,
+ num_images_per_prompt=multiplier,
+ seq_len=self.text_encoder_seq_len,
+ hidden_size=self.text_encoder_hidden_size,
+ dtype=self.text_encoder.dtype, # Should work with both full precision and mixed precision
+ device=device,
+ generator=generator,
+ latents=prompt_latents,
+ )
+
+ if reduce_text_emb_dim:
+ prompt_embeds = self.text_decoder.encode(prompt_embeds)
+
+ # 4. Encode image, if available; otherwise prepare image latents
+ if mode in ["img2text"]:
+ # 4.1. Encode images, if available
+ assert image is not None, "`img2text` requires a conditioning image"
+ # Encode image using VAE
+ image_vae = self.image_processor.preprocess(image)
+ height, width = image_vae.shape[-2:]
+ image_vae_latents = self.encode_image_vae_latents(
+ image=image_vae,
+ batch_size=batch_size,
+ num_prompts_per_image=multiplier,
+ dtype=prompt_embeds.dtype,
+ device=device,
+ do_classifier_free_guidance=False, # Copied from InstructPix2Pix, don't use their version of CFG
+ generator=generator,
+ )
+
+ # Encode image using CLIP
+ image_clip_latents = self.encode_image_clip_latents(
+ image=image,
+ batch_size=batch_size,
+ num_prompts_per_image=multiplier,
+ dtype=prompt_embeds.dtype,
+ device=device,
+ generator=generator,
+ )
+ # (batch_size, clip_hidden_size) => (batch_size, 1, clip_hidden_size)
+ image_clip_latents = image_clip_latents.unsqueeze(1)
+ else:
+ # 4.2. Prepare image latent variables, if input not available
+ # Prepare image VAE latents in latent space
+ image_vae_latents = self.prepare_image_vae_latents(
+ batch_size=batch_size,
+ num_prompts_per_image=multiplier,
+ num_channels_latents=self.num_channels_latents,
+ height=height,
+ width=width,
+ dtype=prompt_embeds.dtype,
+ device=device,
+ generator=generator,
+ latents=vae_latents,
+ )
+
+ # Prepare image CLIP latents
+ image_clip_latents = self.prepare_image_clip_latents(
+ batch_size=batch_size,
+ num_prompts_per_image=multiplier,
+ clip_img_dim=self.image_encoder_projection_dim,
+ dtype=prompt_embeds.dtype,
+ device=device,
+ generator=generator,
+ latents=clip_latents,
+ )
+
+ # 5. Set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+ # max_timestep = timesteps[0]
+ max_timestep = self.scheduler.config.num_train_timesteps
+
+ # 6. Prepare latent variables
+ if mode == "joint":
+ latents = self._combine_joint(image_vae_latents, image_clip_latents, prompt_embeds)
+ elif mode in ["text2img", "img"]:
+ latents = self._combine(image_vae_latents, image_clip_latents)
+ elif mode in ["img2text", "text"]:
+ latents = prompt_embeds
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ logger.debug(f"Scheduler extra step kwargs: {extra_step_kwargs}")
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # predict the noise residual
+ # Also applies classifier-free guidance as described in the UniDiffuser paper
+ noise_pred = self._get_noise_pred(
+ mode,
+ latents,
+ t,
+ prompt_embeds,
+ image_vae_latents,
+ image_clip_latents,
+ max_timestep,
+ data_type,
+ guidance_scale,
+ generator,
+ device,
+ height,
+ width,
+ )
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 9. Post-processing
+ image = None
+ text = None
+ if mode == "joint":
+ image_vae_latents, image_clip_latents, text_latents = self._split_joint(latents, height, width)
+
+ if not output_type == "latent":
+ # Map latent VAE image back to pixel space
+ image = self.vae.decode(image_vae_latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = image_vae_latents
+
+ text = self.decode_text_latents(text_latents, device)
+ elif mode in ["text2img", "img"]:
+ image_vae_latents, image_clip_latents = self._split(latents, height, width)
+
+ if not output_type == "latent":
+ # Map latent VAE image back to pixel space
+ image = self.vae.decode(image_vae_latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ else:
+ image = image_vae_latents
+ elif mode in ["img2text", "text"]:
+ text_latents = latents
+ text = self.decode_text_latents(text_latents, device)
+
+ self.maybe_free_model_hooks()
+
+ # 10. Postprocess the image, if necessary
+ if image is not None:
+ do_denormalize = [True] * image.shape[0]
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload last model to CPU
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.final_offload_hook.offload()
+
+ if not return_dict:
+ return (image, text)
+
+ return ImageTextPipelineOutput(images=image, text=text)
diff --git a/model/diffusers_c/pipelines/wuerstchen/__init__.py b/model/diffusers_c/pipelines/wuerstchen/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..ddb852d1931558fe0948e81e16cf9a92fc2a114b
--- /dev/null
+++ b/model/diffusers_c/pipelines/wuerstchen/__init__.py
@@ -0,0 +1,56 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_torch_and_transformers_objects
+
+ _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+ _import_structure["modeling_paella_vq_model"] = ["PaellaVQModel"]
+ _import_structure["modeling_wuerstchen_diffnext"] = ["WuerstchenDiffNeXt"]
+ _import_structure["modeling_wuerstchen_prior"] = ["WuerstchenPrior"]
+ _import_structure["pipeline_wuerstchen"] = ["WuerstchenDecoderPipeline"]
+ _import_structure["pipeline_wuerstchen_combined"] = ["WuerstchenCombinedPipeline"]
+ _import_structure["pipeline_wuerstchen_prior"] = ["DEFAULT_STAGE_C_TIMESTEPS", "WuerstchenPriorPipeline"]
+
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ...utils.dummy_torch_and_transformers_objects import * # noqa F403
+ else:
+ from .modeling_paella_vq_model import PaellaVQModel
+ from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt
+ from .modeling_wuerstchen_prior import WuerstchenPrior
+ from .pipeline_wuerstchen import WuerstchenDecoderPipeline
+ from .pipeline_wuerstchen_combined import WuerstchenCombinedPipeline
+ from .pipeline_wuerstchen_prior import DEFAULT_STAGE_C_TIMESTEPS, WuerstchenPriorPipeline
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/pipelines/wuerstchen/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/pipelines/wuerstchen/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..84dbdf72c55bb09bc3e182dab82333abfeeb590c
Binary files /dev/null and b/model/diffusers_c/pipelines/wuerstchen/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_paella_vq_model.cpython-39.pyc b/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_paella_vq_model.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a4b16992dd615caf242b6c44d51680e852775a6b
Binary files /dev/null and b/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_paella_vq_model.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_common.cpython-39.pyc b/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_common.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..bd3f667326bbb160b731e284f49d334e27982683
Binary files /dev/null and b/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_common.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_diffnext.cpython-39.pyc b/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_diffnext.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..82fb51c0579e426393eafc9a4c5b4b3ce8b9b226
Binary files /dev/null and b/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_diffnext.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_prior.cpython-39.pyc b/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_prior.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d2eae3641c1b9c35cd2275f453fdde123a18e810
Binary files /dev/null and b/model/diffusers_c/pipelines/wuerstchen/__pycache__/modeling_wuerstchen_prior.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen.cpython-39.pyc b/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c8578c950be80f212607d3da648aa999bd8ac8cf
Binary files /dev/null and b/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_combined.cpython-39.pyc b/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_combined.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..05a5aaa7a03fb9911470e63ff60153453da70bad
Binary files /dev/null and b/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_combined.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_prior.cpython-39.pyc b/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_prior.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..af7dbe027803e51d6eecd02e3b48280f70d33706
Binary files /dev/null and b/model/diffusers_c/pipelines/wuerstchen/__pycache__/pipeline_wuerstchen_prior.cpython-39.pyc differ
diff --git a/model/diffusers_c/pipelines/wuerstchen/modeling_paella_vq_model.py b/model/diffusers_c/pipelines/wuerstchen/modeling_paella_vq_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..3b21dfb5f1f0ab1791f2237b38f059deb0897a0d
--- /dev/null
+++ b/model/diffusers_c/pipelines/wuerstchen/modeling_paella_vq_model.py
@@ -0,0 +1,172 @@
+# Copyright (c) 2022 Dominic Rampas MIT License
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models.autoencoders.vae import DecoderOutput, VectorQuantizer
+from ...models.modeling_utils import ModelMixin
+from ...models.vq_model import VQEncoderOutput
+from ...utils.accelerate_utils import apply_forward_hook
+
+
+class MixingResidualBlock(nn.Module):
+ """
+ Residual block with mixing used by Paella's VQ-VAE.
+ """
+
+ def __init__(self, inp_channels, embed_dim):
+ super().__init__()
+ # depthwise
+ self.norm1 = nn.LayerNorm(inp_channels, elementwise_affine=False, eps=1e-6)
+ self.depthwise = nn.Sequential(
+ nn.ReplicationPad2d(1), nn.Conv2d(inp_channels, inp_channels, kernel_size=3, groups=inp_channels)
+ )
+
+ # channelwise
+ self.norm2 = nn.LayerNorm(inp_channels, elementwise_affine=False, eps=1e-6)
+ self.channelwise = nn.Sequential(
+ nn.Linear(inp_channels, embed_dim), nn.GELU(), nn.Linear(embed_dim, inp_channels)
+ )
+
+ self.gammas = nn.Parameter(torch.zeros(6), requires_grad=True)
+
+ def forward(self, x):
+ mods = self.gammas
+ x_temp = self.norm1(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * (1 + mods[0]) + mods[1]
+ x = x + self.depthwise(x_temp) * mods[2]
+ x_temp = self.norm2(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * (1 + mods[3]) + mods[4]
+ x = x + self.channelwise(x_temp.permute(0, 2, 3, 1)).permute(0, 3, 1, 2) * mods[5]
+ return x
+
+
+class PaellaVQModel(ModelMixin, ConfigMixin):
+ r"""VQ-VAE model from Paella model.
+
+ This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library
+ implements for all the model (such as downloading or saving, etc.)
+
+ Parameters:
+ in_channels (int, *optional*, defaults to 3): Number of channels in the input image.
+ out_channels (int, *optional*, defaults to 3): Number of channels in the output.
+ up_down_scale_factor (int, *optional*, defaults to 2): Up and Downscale factor of the input image.
+ levels (int, *optional*, defaults to 2): Number of levels in the model.
+ bottleneck_blocks (int, *optional*, defaults to 12): Number of bottleneck blocks in the model.
+ embed_dim (int, *optional*, defaults to 384): Number of hidden channels in the model.
+ latent_channels (int, *optional*, defaults to 4): Number of latent channels in the VQ-VAE model.
+ num_vq_embeddings (int, *optional*, defaults to 8192): Number of codebook vectors in the VQ-VAE.
+ scale_factor (float, *optional*, defaults to 0.3764): Scaling factor of the latent space.
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 3,
+ out_channels: int = 3,
+ up_down_scale_factor: int = 2,
+ levels: int = 2,
+ bottleneck_blocks: int = 12,
+ embed_dim: int = 384,
+ latent_channels: int = 4,
+ num_vq_embeddings: int = 8192,
+ scale_factor: float = 0.3764,
+ ):
+ super().__init__()
+
+ c_levels = [embed_dim // (2**i) for i in reversed(range(levels))]
+ # Encoder blocks
+ self.in_block = nn.Sequential(
+ nn.PixelUnshuffle(up_down_scale_factor),
+ nn.Conv2d(in_channels * up_down_scale_factor**2, c_levels[0], kernel_size=1),
+ )
+ down_blocks = []
+ for i in range(levels):
+ if i > 0:
+ down_blocks.append(nn.Conv2d(c_levels[i - 1], c_levels[i], kernel_size=4, stride=2, padding=1))
+ block = MixingResidualBlock(c_levels[i], c_levels[i] * 4)
+ down_blocks.append(block)
+ down_blocks.append(
+ nn.Sequential(
+ nn.Conv2d(c_levels[-1], latent_channels, kernel_size=1, bias=False),
+ nn.BatchNorm2d(latent_channels), # then normalize them to have mean 0 and std 1
+ )
+ )
+ self.down_blocks = nn.Sequential(*down_blocks)
+
+ # Vector Quantizer
+ self.vquantizer = VectorQuantizer(num_vq_embeddings, vq_embed_dim=latent_channels, legacy=False, beta=0.25)
+
+ # Decoder blocks
+ up_blocks = [nn.Sequential(nn.Conv2d(latent_channels, c_levels[-1], kernel_size=1))]
+ for i in range(levels):
+ for j in range(bottleneck_blocks if i == 0 else 1):
+ block = MixingResidualBlock(c_levels[levels - 1 - i], c_levels[levels - 1 - i] * 4)
+ up_blocks.append(block)
+ if i < levels - 1:
+ up_blocks.append(
+ nn.ConvTranspose2d(
+ c_levels[levels - 1 - i], c_levels[levels - 2 - i], kernel_size=4, stride=2, padding=1
+ )
+ )
+ self.up_blocks = nn.Sequential(*up_blocks)
+ self.out_block = nn.Sequential(
+ nn.Conv2d(c_levels[0], out_channels * up_down_scale_factor**2, kernel_size=1),
+ nn.PixelShuffle(up_down_scale_factor),
+ )
+
+ @apply_forward_hook
+ def encode(self, x: torch.FloatTensor, return_dict: bool = True) -> VQEncoderOutput:
+ h = self.in_block(x)
+ h = self.down_blocks(h)
+
+ if not return_dict:
+ return (h,)
+
+ return VQEncoderOutput(latents=h)
+
+ @apply_forward_hook
+ def decode(
+ self, h: torch.FloatTensor, force_not_quantize: bool = True, return_dict: bool = True
+ ) -> Union[DecoderOutput, torch.FloatTensor]:
+ if not force_not_quantize:
+ quant, _, _ = self.vquantizer(h)
+ else:
+ quant = h
+
+ x = self.up_blocks(quant)
+ dec = self.out_block(x)
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
+
+ def forward(self, sample: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]:
+ r"""
+ Args:
+ sample (`torch.FloatTensor`): Input sample.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`DecoderOutput`] instead of a plain tuple.
+ """
+ x = sample
+ h = self.encode(x).latents
+ dec = self.decode(h).sample
+
+ if not return_dict:
+ return (dec,)
+
+ return DecoderOutput(sample=dec)
diff --git a/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_common.py b/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_common.py
new file mode 100644
index 0000000000000000000000000000000000000000..a4c6f9c6a8b97fa7592571b7c6519e6931afdbc2
--- /dev/null
+++ b/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_common.py
@@ -0,0 +1,83 @@
+import torch
+import torch.nn as nn
+
+from ...models.attention_processor import Attention
+from ...models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from ...utils import USE_PEFT_BACKEND
+
+
+class WuerstchenLayerNorm(nn.LayerNorm):
+ def __init__(self, *args, **kwargs):
+ super().__init__(*args, **kwargs)
+
+ def forward(self, x):
+ x = x.permute(0, 2, 3, 1)
+ x = super().forward(x)
+ return x.permute(0, 3, 1, 2)
+
+
+class TimestepBlock(nn.Module):
+ def __init__(self, c, c_timestep):
+ super().__init__()
+ linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+ self.mapper = linear_cls(c_timestep, c * 2)
+
+ def forward(self, x, t):
+ a, b = self.mapper(t)[:, :, None, None].chunk(2, dim=1)
+ return x * (1 + a) + b
+
+
+class ResBlock(nn.Module):
+ def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0):
+ super().__init__()
+
+ conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+ linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+ self.depthwise = conv_cls(c + c_skip, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c)
+ self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6)
+ self.channelwise = nn.Sequential(
+ linear_cls(c, c * 4), nn.GELU(), GlobalResponseNorm(c * 4), nn.Dropout(dropout), linear_cls(c * 4, c)
+ )
+
+ def forward(self, x, x_skip=None):
+ x_res = x
+ if x_skip is not None:
+ x = torch.cat([x, x_skip], dim=1)
+ x = self.norm(self.depthwise(x)).permute(0, 2, 3, 1)
+ x = self.channelwise(x).permute(0, 3, 1, 2)
+ return x + x_res
+
+
+# from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105
+class GlobalResponseNorm(nn.Module):
+ def __init__(self, dim):
+ super().__init__()
+ self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim))
+ self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim))
+
+ def forward(self, x):
+ agg_norm = torch.norm(x, p=2, dim=(1, 2), keepdim=True)
+ stand_div_norm = agg_norm / (agg_norm.mean(dim=-1, keepdim=True) + 1e-6)
+ return self.gamma * (x * stand_div_norm) + self.beta + x
+
+
+class AttnBlock(nn.Module):
+ def __init__(self, c, c_cond, nhead, self_attn=True, dropout=0.0):
+ super().__init__()
+
+ linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+ self.self_attn = self_attn
+ self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6)
+ self.attention = Attention(query_dim=c, heads=nhead, dim_head=c // nhead, dropout=dropout, bias=True)
+ self.kv_mapper = nn.Sequential(nn.SiLU(), linear_cls(c_cond, c))
+
+ def forward(self, x, kv):
+ kv = self.kv_mapper(kv)
+ norm_x = self.norm(x)
+ if self.self_attn:
+ batch_size, channel, _, _ = x.shape
+ kv = torch.cat([norm_x.view(batch_size, channel, -1).transpose(1, 2), kv], dim=1)
+ x = x + self.attention(norm_x, encoder_hidden_states=kv)
+ return x
diff --git a/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_diffnext.py b/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_diffnext.py
new file mode 100644
index 0000000000000000000000000000000000000000..6c06cc0e7303f2ec72c6c2958b9f03eaba5f8c67
--- /dev/null
+++ b/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_diffnext.py
@@ -0,0 +1,254 @@
+# Copyright (c) 2023 Dominic Rampas MIT License
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+
+import numpy as np
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...models.modeling_utils import ModelMixin
+from .modeling_wuerstchen_common import AttnBlock, GlobalResponseNorm, TimestepBlock, WuerstchenLayerNorm
+
+
+class WuerstchenDiffNeXt(ModelMixin, ConfigMixin):
+ @register_to_config
+ def __init__(
+ self,
+ c_in=4,
+ c_out=4,
+ c_r=64,
+ patch_size=2,
+ c_cond=1024,
+ c_hidden=[320, 640, 1280, 1280],
+ nhead=[-1, 10, 20, 20],
+ blocks=[4, 4, 14, 4],
+ level_config=["CT", "CTA", "CTA", "CTA"],
+ inject_effnet=[False, True, True, True],
+ effnet_embd=16,
+ clip_embd=1024,
+ kernel_size=3,
+ dropout=0.1,
+ ):
+ super().__init__()
+ self.c_r = c_r
+ self.c_cond = c_cond
+ if not isinstance(dropout, list):
+ dropout = [dropout] * len(c_hidden)
+
+ # CONDITIONING
+ self.clip_mapper = nn.Linear(clip_embd, c_cond)
+ self.effnet_mappers = nn.ModuleList(
+ [
+ nn.Conv2d(effnet_embd, c_cond, kernel_size=1) if inject else None
+ for inject in inject_effnet + list(reversed(inject_effnet))
+ ]
+ )
+ self.seq_norm = nn.LayerNorm(c_cond, elementwise_affine=False, eps=1e-6)
+
+ self.embedding = nn.Sequential(
+ nn.PixelUnshuffle(patch_size),
+ nn.Conv2d(c_in * (patch_size**2), c_hidden[0], kernel_size=1),
+ WuerstchenLayerNorm(c_hidden[0], elementwise_affine=False, eps=1e-6),
+ )
+
+ def get_block(block_type, c_hidden, nhead, c_skip=0, dropout=0):
+ if block_type == "C":
+ return ResBlockStageB(c_hidden, c_skip, kernel_size=kernel_size, dropout=dropout)
+ elif block_type == "A":
+ return AttnBlock(c_hidden, c_cond, nhead, self_attn=True, dropout=dropout)
+ elif block_type == "T":
+ return TimestepBlock(c_hidden, c_r)
+ else:
+ raise ValueError(f"Block type {block_type} not supported")
+
+ # BLOCKS
+ # -- down blocks
+ self.down_blocks = nn.ModuleList()
+ for i in range(len(c_hidden)):
+ down_block = nn.ModuleList()
+ if i > 0:
+ down_block.append(
+ nn.Sequential(
+ WuerstchenLayerNorm(c_hidden[i - 1], elementwise_affine=False, eps=1e-6),
+ nn.Conv2d(c_hidden[i - 1], c_hidden[i], kernel_size=2, stride=2),
+ )
+ )
+ for _ in range(blocks[i]):
+ for block_type in level_config[i]:
+ c_skip = c_cond if inject_effnet[i] else 0
+ down_block.append(get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i]))
+ self.down_blocks.append(down_block)
+
+ # -- up blocks
+ self.up_blocks = nn.ModuleList()
+ for i in reversed(range(len(c_hidden))):
+ up_block = nn.ModuleList()
+ for j in range(blocks[i]):
+ for k, block_type in enumerate(level_config[i]):
+ c_skip = c_hidden[i] if i < len(c_hidden) - 1 and j == k == 0 else 0
+ c_skip += c_cond if inject_effnet[i] else 0
+ up_block.append(get_block(block_type, c_hidden[i], nhead[i], c_skip=c_skip, dropout=dropout[i]))
+ if i > 0:
+ up_block.append(
+ nn.Sequential(
+ WuerstchenLayerNorm(c_hidden[i], elementwise_affine=False, eps=1e-6),
+ nn.ConvTranspose2d(c_hidden[i], c_hidden[i - 1], kernel_size=2, stride=2),
+ )
+ )
+ self.up_blocks.append(up_block)
+
+ # OUTPUT
+ self.clf = nn.Sequential(
+ WuerstchenLayerNorm(c_hidden[0], elementwise_affine=False, eps=1e-6),
+ nn.Conv2d(c_hidden[0], 2 * c_out * (patch_size**2), kernel_size=1),
+ nn.PixelShuffle(patch_size),
+ )
+
+ # --- WEIGHT INIT ---
+ self.apply(self._init_weights)
+
+ def _init_weights(self, m):
+ # General init
+ if isinstance(m, (nn.Conv2d, nn.Linear)):
+ nn.init.xavier_uniform_(m.weight)
+ if m.bias is not None:
+ nn.init.constant_(m.bias, 0)
+
+ for mapper in self.effnet_mappers:
+ if mapper is not None:
+ nn.init.normal_(mapper.weight, std=0.02) # conditionings
+ nn.init.normal_(self.clip_mapper.weight, std=0.02) # conditionings
+ nn.init.xavier_uniform_(self.embedding[1].weight, 0.02) # inputs
+ nn.init.constant_(self.clf[1].weight, 0) # outputs
+
+ # blocks
+ for level_block in self.down_blocks + self.up_blocks:
+ for block in level_block:
+ if isinstance(block, ResBlockStageB):
+ block.channelwise[-1].weight.data *= np.sqrt(1 / sum(self.config.blocks))
+ elif isinstance(block, TimestepBlock):
+ nn.init.constant_(block.mapper.weight, 0)
+
+ def gen_r_embedding(self, r, max_positions=10000):
+ r = r * max_positions
+ half_dim = self.c_r // 2
+ emb = math.log(max_positions) / (half_dim - 1)
+ emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp()
+ emb = r[:, None] * emb[None, :]
+ emb = torch.cat([emb.sin(), emb.cos()], dim=1)
+ if self.c_r % 2 == 1: # zero pad
+ emb = nn.functional.pad(emb, (0, 1), mode="constant")
+ return emb.to(dtype=r.dtype)
+
+ def gen_c_embeddings(self, clip):
+ clip = self.clip_mapper(clip)
+ clip = self.seq_norm(clip)
+ return clip
+
+ def _down_encode(self, x, r_embed, effnet, clip=None):
+ level_outputs = []
+ for i, down_block in enumerate(self.down_blocks):
+ effnet_c = None
+ for block in down_block:
+ if isinstance(block, ResBlockStageB):
+ if effnet_c is None and self.effnet_mappers[i] is not None:
+ dtype = effnet.dtype
+ effnet_c = self.effnet_mappers[i](
+ nn.functional.interpolate(
+ effnet.float(), size=x.shape[-2:], mode="bicubic", antialias=True, align_corners=True
+ ).to(dtype)
+ )
+ skip = effnet_c if self.effnet_mappers[i] is not None else None
+ x = block(x, skip)
+ elif isinstance(block, AttnBlock):
+ x = block(x, clip)
+ elif isinstance(block, TimestepBlock):
+ x = block(x, r_embed)
+ else:
+ x = block(x)
+ level_outputs.insert(0, x)
+ return level_outputs
+
+ def _up_decode(self, level_outputs, r_embed, effnet, clip=None):
+ x = level_outputs[0]
+ for i, up_block in enumerate(self.up_blocks):
+ effnet_c = None
+ for j, block in enumerate(up_block):
+ if isinstance(block, ResBlockStageB):
+ if effnet_c is None and self.effnet_mappers[len(self.down_blocks) + i] is not None:
+ dtype = effnet.dtype
+ effnet_c = self.effnet_mappers[len(self.down_blocks) + i](
+ nn.functional.interpolate(
+ effnet.float(), size=x.shape[-2:], mode="bicubic", antialias=True, align_corners=True
+ ).to(dtype)
+ )
+ skip = level_outputs[i] if j == 0 and i > 0 else None
+ if effnet_c is not None:
+ if skip is not None:
+ skip = torch.cat([skip, effnet_c], dim=1)
+ else:
+ skip = effnet_c
+ x = block(x, skip)
+ elif isinstance(block, AttnBlock):
+ x = block(x, clip)
+ elif isinstance(block, TimestepBlock):
+ x = block(x, r_embed)
+ else:
+ x = block(x)
+ return x
+
+ def forward(self, x, r, effnet, clip=None, x_cat=None, eps=1e-3, return_noise=True):
+ if x_cat is not None:
+ x = torch.cat([x, x_cat], dim=1)
+ # Process the conditioning embeddings
+ r_embed = self.gen_r_embedding(r)
+ if clip is not None:
+ clip = self.gen_c_embeddings(clip)
+
+ # Model Blocks
+ x_in = x
+ x = self.embedding(x)
+ level_outputs = self._down_encode(x, r_embed, effnet, clip)
+ x = self._up_decode(level_outputs, r_embed, effnet, clip)
+ a, b = self.clf(x).chunk(2, dim=1)
+ b = b.sigmoid() * (1 - eps * 2) + eps
+ if return_noise:
+ return (x_in - a) / b
+ else:
+ return a, b
+
+
+class ResBlockStageB(nn.Module):
+ def __init__(self, c, c_skip=0, kernel_size=3, dropout=0.0):
+ super().__init__()
+ self.depthwise = nn.Conv2d(c, c, kernel_size=kernel_size, padding=kernel_size // 2, groups=c)
+ self.norm = WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6)
+ self.channelwise = nn.Sequential(
+ nn.Linear(c + c_skip, c * 4),
+ nn.GELU(),
+ GlobalResponseNorm(c * 4),
+ nn.Dropout(dropout),
+ nn.Linear(c * 4, c),
+ )
+
+ def forward(self, x, x_skip=None):
+ x_res = x
+ x = self.norm(self.depthwise(x))
+ if x_skip is not None:
+ x = torch.cat([x, x_skip], dim=1)
+ x = self.channelwise(x.permute(0, 2, 3, 1)).permute(0, 3, 1, 2)
+ return x + x_res
diff --git a/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_prior.py b/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_prior.py
new file mode 100644
index 0000000000000000000000000000000000000000..c44c259ab0b473cdc07ff2be3114fef628321573
--- /dev/null
+++ b/model/diffusers_c/pipelines/wuerstchen/modeling_wuerstchen_prior.py
@@ -0,0 +1,201 @@
+# Copyright (c) 2023 Dominic Rampas MIT License
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import Dict, Union
+
+import torch
+import torch.nn as nn
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
+from ...models.attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from ...models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from ...models.modeling_utils import ModelMixin
+from ...utils import USE_PEFT_BACKEND, is_torch_version
+from .modeling_wuerstchen_common import AttnBlock, ResBlock, TimestepBlock, WuerstchenLayerNorm
+
+
+class WuerstchenPrior(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin):
+ unet_name = "prior"
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(self, c_in=16, c=1280, c_cond=1024, c_r=64, depth=16, nhead=16, dropout=0.1):
+ super().__init__()
+ conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+ linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+ self.c_r = c_r
+ self.projection = conv_cls(c_in, c, kernel_size=1)
+ self.cond_mapper = nn.Sequential(
+ linear_cls(c_cond, c),
+ nn.LeakyReLU(0.2),
+ linear_cls(c, c),
+ )
+
+ self.blocks = nn.ModuleList()
+ for _ in range(depth):
+ self.blocks.append(ResBlock(c, dropout=dropout))
+ self.blocks.append(TimestepBlock(c, c_r))
+ self.blocks.append(AttnBlock(c, c, nhead, self_attn=True, dropout=dropout))
+ self.out = nn.Sequential(
+ WuerstchenLayerNorm(c, elementwise_affine=False, eps=1e-6),
+ conv_cls(c, c_in * 2, kernel_size=1),
+ )
+
+ self.gradient_checkpointing = False
+ self.set_default_attn_processor()
+
+ @property
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+ def attn_processors(self) -> Dict[str, AttentionProcessor]:
+ r"""
+ Returns:
+ `dict` of attention processors: A dictionary containing all attention processors used in the model with
+ indexed by its weight name.
+ """
+ # set recursively
+ processors = {}
+
+ def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+ if hasattr(module, "get_processor"):
+ processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+ for sub_name, child in module.named_children():
+ fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+ return processors
+
+ for name, module in self.named_children():
+ fn_recursive_add_processors(name, module, processors)
+
+ return processors
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+ def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+ r"""
+ Sets the attention processor to use to compute attention.
+
+ Parameters:
+ processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+ The instantiated processor class or a dictionary of processor classes that will be set as the processor
+ for **all** `Attention` layers.
+
+ If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+ processor. This is strongly recommended when setting trainable attention processors.
+
+ """
+ count = len(self.attn_processors.keys())
+
+ if isinstance(processor, dict) and len(processor) != count:
+ raise ValueError(
+ f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+ f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+ )
+
+ def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+ if hasattr(module, "set_processor"):
+ if not isinstance(processor, dict):
+ module.set_processor(processor)
+ else:
+ module.set_processor(processor.pop(f"{name}.processor"))
+
+ for sub_name, child in module.named_children():
+ fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+ for name, module in self.named_children():
+ fn_recursive_attn_processor(name, module, processor)
+
+ # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+ def set_default_attn_processor(self):
+ """
+ Disables custom attention processors and sets the default attention implementation.
+ """
+ if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnAddedKVProcessor()
+ elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+ processor = AttnProcessor()
+ else:
+ raise ValueError(
+ f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+ )
+
+ self.set_attn_processor(processor)
+
+ def _set_gradient_checkpointing(self, module, value=False):
+ self.gradient_checkpointing = value
+
+ def gen_r_embedding(self, r, max_positions=10000):
+ r = r * max_positions
+ half_dim = self.c_r // 2
+ emb = math.log(max_positions) / (half_dim - 1)
+ emb = torch.arange(half_dim, device=r.device).float().mul(-emb).exp()
+ emb = r[:, None] * emb[None, :]
+ emb = torch.cat([emb.sin(), emb.cos()], dim=1)
+ if self.c_r % 2 == 1: # zero pad
+ emb = nn.functional.pad(emb, (0, 1), mode="constant")
+ return emb.to(dtype=r.dtype)
+
+ def forward(self, x, r, c):
+ x_in = x
+ x = self.projection(x)
+ c_embed = self.cond_mapper(c)
+ r_embed = self.gen_r_embedding(r)
+
+ if self.training and self.gradient_checkpointing:
+
+ def create_custom_forward(module):
+ def custom_forward(*inputs):
+ return module(*inputs)
+
+ return custom_forward
+
+ if is_torch_version(">=", "1.11.0"):
+ for block in self.blocks:
+ if isinstance(block, AttnBlock):
+ x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block), x, c_embed, use_reentrant=False
+ )
+ elif isinstance(block, TimestepBlock):
+ x = torch.utils.checkpoint.checkpoint(
+ create_custom_forward(block), x, r_embed, use_reentrant=False
+ )
+ else:
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(block), x, use_reentrant=False)
+ else:
+ for block in self.blocks:
+ if isinstance(block, AttnBlock):
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(block), x, c_embed)
+ elif isinstance(block, TimestepBlock):
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(block), x, r_embed)
+ else:
+ x = torch.utils.checkpoint.checkpoint(create_custom_forward(block), x)
+ else:
+ for block in self.blocks:
+ if isinstance(block, AttnBlock):
+ x = block(x, c_embed)
+ elif isinstance(block, TimestepBlock):
+ x = block(x, r_embed)
+ else:
+ x = block(x)
+ a, b = self.out(x).chunk(2, dim=1)
+ return (x_in - a) / ((1 - b).abs() + 1e-5)
diff --git a/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen.py b/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen.py
new file mode 100644
index 0000000000000000000000000000000000000000..e4277d58a006cfa9cdaf1084370d38b9038ec05d
--- /dev/null
+++ b/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen.py
@@ -0,0 +1,438 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ...schedulers import DDPMWuerstchenScheduler
+from ...utils import deprecate, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from .modeling_paella_vq_model import PaellaVQModel
+from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import WuerstchenPriorPipeline, WuerstchenDecoderPipeline
+
+ >>> prior_pipe = WuerstchenPriorPipeline.from_pretrained(
+ ... "warp-ai/wuerstchen-prior", torch_dtype=torch.float16
+ ... ).to("cuda")
+ >>> gen_pipe = WuerstchenDecoderPipeline.from_pretrain("warp-ai/wuerstchen", torch_dtype=torch.float16).to(
+ ... "cuda"
+ ... )
+
+ >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet"
+ >>> prior_output = pipe(prompt)
+ >>> images = gen_pipe(prior_output.image_embeddings, prompt=prompt)
+ ```
+"""
+
+
+class WuerstchenDecoderPipeline(DiffusionPipeline):
+ """
+ Pipeline for generating images from the Wuerstchen model.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ tokenizer (`CLIPTokenizer`):
+ The CLIP tokenizer.
+ text_encoder (`CLIPTextModel`):
+ The CLIP text encoder.
+ decoder ([`WuerstchenDiffNeXt`]):
+ The WuerstchenDiffNeXt unet decoder.
+ vqgan ([`PaellaVQModel`]):
+ The VQGAN model.
+ scheduler ([`DDPMWuerstchenScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ latent_dim_scale (float, `optional`, defaults to 10.67):
+ Multiplier to determine the VQ latent space size from the image embeddings. If the image embeddings are
+ height=24 and width=24, the VQ latent shape needs to be height=int(24*10.67)=256 and
+ width=int(24*10.67)=256 in order to match the training conditions.
+ """
+
+ model_cpu_offload_seq = "text_encoder->decoder->vqgan"
+ _callback_tensor_inputs = [
+ "latents",
+ "text_encoder_hidden_states",
+ "negative_prompt_embeds",
+ "image_embeddings",
+ ]
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModel,
+ decoder: WuerstchenDiffNeXt,
+ scheduler: DDPMWuerstchenScheduler,
+ vqgan: PaellaVQModel,
+ latent_dim_scale: float = 10.67,
+ ) -> None:
+ super().__init__()
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ decoder=decoder,
+ scheduler=scheduler,
+ vqgan=vqgan,
+ )
+ self.register_to_config(latent_dim_scale=latent_dim_scale)
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ ):
+ batch_size = len(prompt) if isinstance(prompt, list) else 1
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ attention_mask = text_inputs.attention_mask
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
+ removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1])
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+ attention_mask = attention_mask[:, : self.tokenizer.model_max_length]
+
+ text_encoder_output = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask.to(device))
+ text_encoder_hidden_states = text_encoder_output.last_hidden_state
+ text_encoder_hidden_states = text_encoder_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+
+ uncond_text_encoder_hidden_states = None
+ if do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(
+ uncond_input.input_ids.to(device), attention_mask=uncond_input.attention_mask.to(device)
+ )
+
+ uncond_text_encoder_hidden_states = negative_prompt_embeds_text_encoder_output.last_hidden_state
+
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = uncond_text_encoder_hidden_states.shape[1]
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.repeat(1, num_images_per_prompt, 1)
+ uncond_text_encoder_hidden_states = uncond_text_encoder_hidden_states.view(
+ batch_size * num_images_per_prompt, seq_len, -1
+ )
+ # done duplicates
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ return text_encoder_hidden_states, uncond_text_encoder_hidden_states
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ image_embeddings: Union[torch.FloatTensor, List[torch.FloatTensor]],
+ prompt: Union[str, List[str]] = None,
+ num_inference_steps: int = 12,
+ timesteps: Optional[List[float]] = None,
+ guidance_scale: float = 0.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ image_embedding (`torch.FloatTensor` or `List[torch.FloatTensor]`):
+ Image Embeddings either extracted from an image or generated by a Prior Model.
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ num_inference_steps (`int`, *optional*, defaults to 12):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 0.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `decoder_guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting
+ `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely
+ linked to the text `prompt`, usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `decoder_guidance_scale` is less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True,
+ otherwise a `tuple`. When returning a tuple, the first element is a list with the generated image
+ embeddings.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ # 0. Define commonly used variables
+ device = self._execution_device
+ dtype = self.decoder.dtype
+ self._guidance_scale = guidance_scale
+
+ # 1. Check inputs. Raise error if not correct
+ if not isinstance(prompt, list):
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ else:
+ raise TypeError(f"'prompt' must be of type 'list' or 'str', but got {type(prompt)}.")
+
+ if self.do_classifier_free_guidance:
+ if negative_prompt is not None and not isinstance(negative_prompt, list):
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+ else:
+ raise TypeError(
+ f"'negative_prompt' must be of type 'list' or 'str', but got {type(negative_prompt)}."
+ )
+
+ if isinstance(image_embeddings, list):
+ image_embeddings = torch.cat(image_embeddings, dim=0)
+ if isinstance(image_embeddings, np.ndarray):
+ image_embeddings = torch.Tensor(image_embeddings, device=device).to(dtype=dtype)
+ if not isinstance(image_embeddings, torch.Tensor):
+ raise TypeError(
+ f"'image_embeddings' must be of type 'torch.Tensor' or 'np.array', but got {type(image_embeddings)}."
+ )
+
+ if not isinstance(num_inference_steps, int):
+ raise TypeError(
+ f"'num_inference_steps' must be of type 'int', but got {type(num_inference_steps)}\
+ In Case you want to provide explicit timesteps, please use the 'timesteps' argument."
+ )
+
+ # 2. Encode caption
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ image_embeddings.size(0) * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ )
+ text_encoder_hidden_states = (
+ torch.cat([prompt_embeds, negative_prompt_embeds]) if negative_prompt_embeds is not None else prompt_embeds
+ )
+ effnet = (
+ torch.cat([image_embeddings, torch.zeros_like(image_embeddings)])
+ if self.do_classifier_free_guidance
+ else image_embeddings
+ )
+
+ # 3. Determine latent shape of latents
+ latent_height = int(image_embeddings.size(2) * self.config.latent_dim_scale)
+ latent_width = int(image_embeddings.size(3) * self.config.latent_dim_scale)
+ latent_features_shape = (image_embeddings.size(0) * num_images_per_prompt, 4, latent_height, latent_width)
+
+ # 4. Prepare and set timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latents
+ latents = self.prepare_latents(latent_features_shape, dtype, device, generator, latents, self.scheduler)
+
+ # 6. Run denoising loop
+ self._num_timesteps = len(timesteps[:-1])
+ for i, t in enumerate(self.progress_bar(timesteps[:-1])):
+ ratio = t.expand(latents.size(0)).to(dtype)
+ # 7. Denoise latents
+ predicted_latents = self.decoder(
+ torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents,
+ r=torch.cat([ratio] * 2) if self.do_classifier_free_guidance else ratio,
+ effnet=effnet,
+ clip=text_encoder_hidden_states,
+ )
+
+ # 8. Check for classifier free guidance and apply it
+ if self.do_classifier_free_guidance:
+ predicted_latents_text, predicted_latents_uncond = predicted_latents.chunk(2)
+ predicted_latents = torch.lerp(predicted_latents_uncond, predicted_latents_text, self.guidance_scale)
+
+ # 9. Renoise latents to next timestep
+ latents = self.scheduler.step(
+ model_output=predicted_latents,
+ timestep=ratio,
+ sample=latents,
+ generator=generator,
+ ).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ image_embeddings = callback_outputs.pop("image_embeddings", image_embeddings)
+ text_encoder_hidden_states = callback_outputs.pop(
+ "text_encoder_hidden_states", text_encoder_hidden_states
+ )
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ if output_type not in ["pt", "np", "pil", "latent"]:
+ raise ValueError(
+ f"Only the output types `pt`, `np`, `pil` and `latent` are supported not output_type={output_type}"
+ )
+
+ if not output_type == "latent":
+ # 10. Scale and decode the image latents with vq-vae
+ latents = self.vqgan.config.scale_factor * latents
+ images = self.vqgan.decode(latents).sample.clamp(0, 1)
+ if output_type == "np":
+ images = images.permute(0, 2, 3, 1).cpu().float().numpy()
+ elif output_type == "pil":
+ images = images.permute(0, 2, 3, 1).cpu().float().numpy()
+ images = self.numpy_to_pil(images)
+ else:
+ images = latents
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return images
+ return ImagePipelineOutput(images)
diff --git a/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen_combined.py b/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen_combined.py
new file mode 100644
index 0000000000000000000000000000000000000000..3a43ad5b9ee00a43171f32d13a712b521562008e
--- /dev/null
+++ b/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen_combined.py
@@ -0,0 +1,306 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Callable, Dict, List, Optional, Union
+
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ...schedulers import DDPMWuerstchenScheduler
+from ...utils import deprecate, replace_example_docstring
+from ..pipeline_utils import DiffusionPipeline
+from .modeling_paella_vq_model import PaellaVQModel
+from .modeling_wuerstchen_diffnext import WuerstchenDiffNeXt
+from .modeling_wuerstchen_prior import WuerstchenPrior
+from .pipeline_wuerstchen import WuerstchenDecoderPipeline
+from .pipeline_wuerstchen_prior import WuerstchenPriorPipeline
+
+
+TEXT2IMAGE_EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> from diffusions import WuerstchenCombinedPipeline
+
+ >>> pipe = WuerstchenCombinedPipeline.from_pretrained("warp-ai/Wuerstchen", torch_dtype=torch.float16).to(
+ ... "cuda"
+ ... )
+ >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet"
+ >>> images = pipe(prompt=prompt)
+ ```
+"""
+
+
+class WuerstchenCombinedPipeline(DiffusionPipeline):
+ """
+ Combined Pipeline for text-to-image generation using Wuerstchen
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ Args:
+ tokenizer (`CLIPTokenizer`):
+ The decoder tokenizer to be used for text inputs.
+ text_encoder (`CLIPTextModel`):
+ The decoder text encoder to be used for text inputs.
+ decoder (`WuerstchenDiffNeXt`):
+ The decoder model to be used for decoder image generation pipeline.
+ scheduler (`DDPMWuerstchenScheduler`):
+ The scheduler to be used for decoder image generation pipeline.
+ vqgan (`PaellaVQModel`):
+ The VQGAN model to be used for decoder image generation pipeline.
+ prior_tokenizer (`CLIPTokenizer`):
+ The prior tokenizer to be used for text inputs.
+ prior_text_encoder (`CLIPTextModel`):
+ The prior text encoder to be used for text inputs.
+ prior_prior (`WuerstchenPrior`):
+ The prior model to be used for prior pipeline.
+ prior_scheduler (`DDPMWuerstchenScheduler`):
+ The scheduler to be used for prior pipeline.
+ """
+
+ _load_connected_pipes = True
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModel,
+ decoder: WuerstchenDiffNeXt,
+ scheduler: DDPMWuerstchenScheduler,
+ vqgan: PaellaVQModel,
+ prior_tokenizer: CLIPTokenizer,
+ prior_text_encoder: CLIPTextModel,
+ prior_prior: WuerstchenPrior,
+ prior_scheduler: DDPMWuerstchenScheduler,
+ ):
+ super().__init__()
+
+ self.register_modules(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ decoder=decoder,
+ scheduler=scheduler,
+ vqgan=vqgan,
+ prior_prior=prior_prior,
+ prior_text_encoder=prior_text_encoder,
+ prior_tokenizer=prior_tokenizer,
+ prior_scheduler=prior_scheduler,
+ )
+ self.prior_pipe = WuerstchenPriorPipeline(
+ prior=prior_prior,
+ text_encoder=prior_text_encoder,
+ tokenizer=prior_tokenizer,
+ scheduler=prior_scheduler,
+ )
+ self.decoder_pipe = WuerstchenDecoderPipeline(
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ decoder=decoder,
+ scheduler=scheduler,
+ vqgan=vqgan,
+ )
+
+ def enable_xformers_memory_efficient_attention(self, attention_op: Optional[Callable] = None):
+ self.decoder_pipe.enable_xformers_memory_efficient_attention(attention_op)
+
+ def enable_model_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+ """
+ self.prior_pipe.enable_model_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_model_cpu_offload(gpu_id=gpu_id)
+
+ def enable_sequential_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models (`unet`, `text_encoder`, `vae`, and `safety checker` state dicts) to CPU using 🤗
+ Accelerate, significantly reducing memory usage. Models are moved to a `torch.device('meta')` and loaded on a
+ GPU only when their specific submodule's `forward` method is called. Offloading happens on a submodule basis.
+ Memory savings are higher than using `enable_model_cpu_offload`, but performance is lower.
+ """
+ self.prior_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+ self.decoder_pipe.enable_sequential_cpu_offload(gpu_id=gpu_id)
+
+ def progress_bar(self, iterable=None, total=None):
+ self.prior_pipe.progress_bar(iterable=iterable, total=total)
+ self.decoder_pipe.progress_bar(iterable=iterable, total=total)
+
+ def set_progress_bar_config(self, **kwargs):
+ self.prior_pipe.set_progress_bar_config(**kwargs)
+ self.decoder_pipe.set_progress_bar_config(**kwargs)
+
+ @torch.no_grad()
+ @replace_example_docstring(TEXT2IMAGE_EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ height: int = 512,
+ width: int = 512,
+ prior_num_inference_steps: int = 60,
+ prior_timesteps: Optional[List[float]] = None,
+ prior_guidance_scale: float = 4.0,
+ num_inference_steps: int = 12,
+ decoder_timesteps: Optional[List[float]] = None,
+ decoder_guidance_scale: float = 0.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ num_images_per_prompt: int = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ prior_callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ prior_callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation for the prior and decoder.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings for the prior. Can be used to easily tweak text inputs, *e.g.*
+ prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt`
+ input argument.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ height (`int`, *optional*, defaults to 512):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 512):
+ The width in pixels of the generated image.
+ prior_guidance_scale (`float`, *optional*, defaults to 4.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `prior_guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting
+ `prior_guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked
+ to the text `prompt`, usually at the expense of lower image quality.
+ prior_num_inference_steps (`Union[int, Dict[float, int]]`, *optional*, defaults to 60):
+ The number of prior denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. For more specific timestep spacing, you can pass customized
+ `prior_timesteps`
+ num_inference_steps (`int`, *optional*, defaults to 12):
+ The number of decoder denoising steps. More denoising steps usually lead to a higher quality image at
+ the expense of slower inference. For more specific timestep spacing, you can pass customized
+ `timesteps`
+ prior_timesteps (`List[float]`, *optional*):
+ Custom timesteps to use for the denoising process for the prior. If not defined, equal spaced
+ `prior_num_inference_steps` timesteps are used. Must be in descending order.
+ decoder_timesteps (`List[float]`, *optional*):
+ Custom timesteps to use for the denoising process for the decoder. If not defined, equal spaced
+ `num_inference_steps` timesteps are used. Must be in descending order.
+ decoder_guidance_scale (`float`, *optional*, defaults to 0.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ prior_callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `prior_callback_on_step_end(self: DiffusionPipeline, step: int, timestep:
+ int, callback_kwargs: Dict)`.
+ prior_callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `prior_callback_on_step_end` function. The tensors specified in the
+ list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in
+ the `._callback_tensor_inputs` attribute of your pipeline class.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.ImagePipelineOutput`] or `tuple` [`~pipelines.ImagePipelineOutput`] if `return_dict` is True,
+ otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images.
+ """
+ prior_kwargs = {}
+ if kwargs.get("prior_callback", None) is not None:
+ prior_kwargs["callback"] = kwargs.pop("prior_callback")
+ deprecate(
+ "prior_callback",
+ "1.0.0",
+ "Passing `prior_callback` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`",
+ )
+ if kwargs.get("prior_callback_steps", None) is not None:
+ deprecate(
+ "prior_callback_steps",
+ "1.0.0",
+ "Passing `prior_callback_steps` as an input argument to `__call__` is deprecated, consider use `prior_callback_on_step_end`",
+ )
+ prior_kwargs["callback_steps"] = kwargs.pop("prior_callback_steps")
+
+ prior_outputs = self.prior_pipe(
+ prompt=prompt if prompt_embeds is None else None,
+ height=height,
+ width=width,
+ num_inference_steps=prior_num_inference_steps,
+ timesteps=prior_timesteps,
+ guidance_scale=prior_guidance_scale,
+ negative_prompt=negative_prompt if negative_prompt_embeds is None else None,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ num_images_per_prompt=num_images_per_prompt,
+ generator=generator,
+ latents=latents,
+ output_type="pt",
+ return_dict=False,
+ callback_on_step_end=prior_callback_on_step_end,
+ callback_on_step_end_tensor_inputs=prior_callback_on_step_end_tensor_inputs,
+ **prior_kwargs,
+ )
+ image_embeddings = prior_outputs[0]
+
+ outputs = self.decoder_pipe(
+ image_embeddings=image_embeddings,
+ prompt=prompt if prompt is not None else "",
+ num_inference_steps=num_inference_steps,
+ timesteps=decoder_timesteps,
+ guidance_scale=decoder_guidance_scale,
+ negative_prompt=negative_prompt,
+ generator=generator,
+ output_type=output_type,
+ return_dict=return_dict,
+ callback_on_step_end=callback_on_step_end,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ **kwargs,
+ )
+
+ return outputs
diff --git a/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen_prior.py b/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen_prior.py
new file mode 100644
index 0000000000000000000000000000000000000000..4640f769676683066bb642df3d53bc8302cfe99b
--- /dev/null
+++ b/model/diffusers_c/pipelines/wuerstchen/pipeline_wuerstchen_prior.py
@@ -0,0 +1,516 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from math import ceil
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import torch
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from ...loaders import LoraLoaderMixin
+from ...schedulers import DDPMWuerstchenScheduler
+from ...utils import BaseOutput, deprecate, logging, replace_example_docstring
+from ...utils.torch_utils import randn_tensor
+from ..pipeline_utils import DiffusionPipeline
+from .modeling_wuerstchen_prior import WuerstchenPrior
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+DEFAULT_STAGE_C_TIMESTEPS = list(np.linspace(1.0, 2 / 3, 20)) + list(np.linspace(2 / 3, 0.0, 11))[1:]
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> import torch
+ >>> from diffusers import WuerstchenPriorPipeline
+
+ >>> prior_pipe = WuerstchenPriorPipeline.from_pretrained(
+ ... "warp-ai/wuerstchen-prior", torch_dtype=torch.float16
+ ... ).to("cuda")
+
+ >>> prompt = "an image of a shiba inu, donning a spacesuit and helmet"
+ >>> prior_output = pipe(prompt)
+ ```
+"""
+
+
+@dataclass
+class WuerstchenPriorPipelineOutput(BaseOutput):
+ """
+ Output class for WuerstchenPriorPipeline.
+
+ Args:
+ image_embeddings (`torch.FloatTensor` or `np.ndarray`)
+ Prior image embeddings for text prompt
+
+ """
+
+ image_embeddings: Union[torch.FloatTensor, np.ndarray]
+
+
+class WuerstchenPriorPipeline(DiffusionPipeline, LoraLoaderMixin):
+ """
+ Pipeline for generating image prior for Wuerstchen.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ prior ([`Prior`]):
+ The canonical unCLIP prior to approximate the image embedding from the text embedding.
+ text_encoder ([`CLIPTextModelWithProjection`]):
+ Frozen text-encoder.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ scheduler ([`DDPMWuerstchenScheduler`]):
+ A scheduler to be used in combination with `prior` to generate image embedding.
+ latent_mean ('float', *optional*, defaults to 42.0):
+ Mean value for latent diffusers.
+ latent_std ('float', *optional*, defaults to 1.0):
+ Standard value for latent diffusers.
+ resolution_multiple ('float', *optional*, defaults to 42.67):
+ Default resolution for multiple images generated.
+ """
+
+ unet_name = "prior"
+ text_encoder_name = "text_encoder"
+ model_cpu_offload_seq = "text_encoder->prior"
+ _callback_tensor_inputs = ["latents", "text_encoder_hidden_states", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ tokenizer: CLIPTokenizer,
+ text_encoder: CLIPTextModel,
+ prior: WuerstchenPrior,
+ scheduler: DDPMWuerstchenScheduler,
+ latent_mean: float = 42.0,
+ latent_std: float = 1.0,
+ resolution_multiple: float = 42.67,
+ ) -> None:
+ super().__init__()
+ self.register_modules(
+ tokenizer=tokenizer,
+ text_encoder=text_encoder,
+ prior=prior,
+ scheduler=scheduler,
+ )
+ self.register_to_config(
+ latent_mean=latent_mean, latent_std=latent_std, resolution_multiple=resolution_multiple
+ )
+
+ # Copied from diffusers.pipelines.unclip.pipeline_unclip.UnCLIPPipeline.prepare_latents
+ def prepare_latents(self, shape, dtype, device, generator, latents, scheduler):
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ if latents.shape != shape:
+ raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
+ latents = latents.to(device)
+
+ latents = latents * scheduler.init_noise_sigma
+ return latents
+
+ def encode_prompt(
+ self,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ prompt=None,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ):
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # get prompt text embeddings
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ attention_mask = text_inputs.attention_mask
+
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+ text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
+ attention_mask = attention_mask[:, : self.tokenizer.model_max_length]
+
+ text_encoder_output = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask.to(device)
+ )
+ prompt_embeds = text_encoder_output.last_hidden_state
+
+ prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+ prompt_embeds = prompt_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+
+ if negative_prompt_embeds is None and do_classifier_free_guidance:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ negative_prompt_embeds_text_encoder_output = self.text_encoder(
+ uncond_input.input_ids.to(device), attention_mask=uncond_input.attention_mask.to(device)
+ )
+
+ negative_prompt_embeds = negative_prompt_embeds_text_encoder_output.last_hidden_state
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+ # done duplicates
+
+ return prompt_embeds, negative_prompt_embeds
+
+ def check_inputs(
+ self,
+ prompt,
+ negative_prompt,
+ num_inference_steps,
+ do_classifier_free_guidance,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if not isinstance(num_inference_steps, int):
+ raise TypeError(
+ f"'num_inference_steps' must be of type 'int', but got {type(num_inference_steps)}\
+ In Case you want to provide explicit timesteps, please use the 'timesteps' argument."
+ )
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ height: int = 1024,
+ width: int = 1024,
+ num_inference_steps: int = 60,
+ timesteps: List[float] = None,
+ guidance_scale: float = 8.0,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pt",
+ return_dict: bool = True,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ """
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`):
+ The prompt or prompts to guide the image generation.
+ height (`int`, *optional*, defaults to 1024):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to 1024):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 60):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process. If not defined, equal spaced `num_inference_steps`
+ timesteps are used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 8.0):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `decoder_guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting
+ `decoder_guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely
+ linked to the text `prompt`, usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
+ if `decoder_guidance_scale` is less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between: `"pil"` (`PIL.Image.Image`), `"np"`
+ (`np.array`) or `"pt"` (`torch.Tensor`).
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.ImagePipelineOutput`] instead of a plain tuple.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.WuerstchenPriorPipelineOutput`] or `tuple` [`~pipelines.WuerstchenPriorPipelineOutput`] if
+ `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
+ generated image embeddings.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider use `callback_on_step_end`",
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ # 0. Define commonly used variables
+ device = self._execution_device
+ self._guidance_scale = guidance_scale
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ # 1. Check inputs. Raise error if not correct
+ if prompt is not None and not isinstance(prompt, list):
+ if isinstance(prompt, str):
+ prompt = [prompt]
+ else:
+ raise TypeError(f"'prompt' must be of type 'list' or 'str', but got {type(prompt)}.")
+
+ if self.do_classifier_free_guidance:
+ if negative_prompt is not None and not isinstance(negative_prompt, list):
+ if isinstance(negative_prompt, str):
+ negative_prompt = [negative_prompt]
+ else:
+ raise TypeError(
+ f"'negative_prompt' must be of type 'list' or 'str', but got {type(negative_prompt)}."
+ )
+
+ self.check_inputs(
+ prompt,
+ negative_prompt,
+ num_inference_steps,
+ self.do_classifier_free_guidance,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ # 2. Encode caption
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt=prompt,
+ device=device,
+ num_images_per_prompt=num_images_per_prompt,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ )
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ text_encoder_hidden_states = (
+ torch.cat([prompt_embeds, negative_prompt_embeds]) if negative_prompt_embeds is not None else prompt_embeds
+ )
+
+ # 3. Determine latent shape of image embeddings
+ dtype = text_encoder_hidden_states.dtype
+ latent_height = ceil(height / self.config.resolution_multiple)
+ latent_width = ceil(width / self.config.resolution_multiple)
+ num_channels = self.prior.config.c_in
+ effnet_features_shape = (num_images_per_prompt * batch_size, num_channels, latent_height, latent_width)
+
+ # 4. Prepare and set timesteps
+ if timesteps is not None:
+ self.scheduler.set_timesteps(timesteps=timesteps, device=device)
+ timesteps = self.scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 5. Prepare latents
+ latents = self.prepare_latents(effnet_features_shape, dtype, device, generator, latents, self.scheduler)
+
+ # 6. Run denoising loop
+ self._num_timesteps = len(timesteps[:-1])
+ for i, t in enumerate(self.progress_bar(timesteps[:-1])):
+ ratio = t.expand(latents.size(0)).to(dtype)
+
+ # 7. Denoise image embeddings
+ predicted_image_embedding = self.prior(
+ torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents,
+ r=torch.cat([ratio] * 2) if self.do_classifier_free_guidance else ratio,
+ c=text_encoder_hidden_states,
+ )
+
+ # 8. Check for classifier free guidance and apply it
+ if self.do_classifier_free_guidance:
+ predicted_image_embedding_text, predicted_image_embedding_uncond = predicted_image_embedding.chunk(2)
+ predicted_image_embedding = torch.lerp(
+ predicted_image_embedding_uncond, predicted_image_embedding_text, self.guidance_scale
+ )
+
+ # 9. Renoise latents to next timestep
+ latents = self.scheduler.step(
+ model_output=predicted_image_embedding,
+ timestep=ratio,
+ sample=latents,
+ generator=generator,
+ ).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ text_encoder_hidden_states = callback_outputs.pop(
+ "text_encoder_hidden_states", text_encoder_hidden_states
+ )
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # 10. Denormalize the latents
+ latents = latents * self.config.latent_mean - self.config.latent_std
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if output_type == "np":
+ latents = latents.cpu().float().numpy()
+
+ if not return_dict:
+ return (latents,)
+
+ return WuerstchenPriorPipelineOutput(latents)
diff --git a/model/diffusers_c/py.typed b/model/diffusers_c/py.typed
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/model/diffusers_c/schedulers/README.md b/model/diffusers_c/schedulers/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..cfff88fd287df6442a9d5c9fbf0ff8d49c600e06
--- /dev/null
+++ b/model/diffusers_c/schedulers/README.md
@@ -0,0 +1,3 @@
+# Schedulers
+
+For more information on the schedulers, please refer to the [docs](https://huggingface.co/docs/diffusers/api/schedulers/overview).
diff --git a/model/diffusers_c/schedulers/__init__.py b/model/diffusers_c/schedulers/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..720d8ea25e293e862488f72760dd5ff1b60ba15f
--- /dev/null
+++ b/model/diffusers_c/schedulers/__init__.py
@@ -0,0 +1,211 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ..utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_flax_available,
+ is_scipy_available,
+ is_torch_available,
+ is_torchsde_available,
+)
+
+
+_dummy_modules = {}
+_import_structure = {}
+
+try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_pt_objects # noqa F403
+
+ _dummy_modules.update(get_objects_from_module(dummy_pt_objects))
+
+else:
+ _import_structure["deprecated"] = ["KarrasVeScheduler", "ScoreSdeVpScheduler"]
+ _import_structure["scheduling_amused"] = ["AmusedScheduler"]
+ _import_structure["scheduling_consistency_decoder"] = ["ConsistencyDecoderScheduler"]
+ _import_structure["scheduling_consistency_models"] = ["CMStochasticIterativeScheduler"]
+ _import_structure["scheduling_ddim"] = ["DDIMScheduler"]
+ _import_structure["scheduling_ddim_inverse"] = ["DDIMInverseScheduler"]
+ _import_structure["scheduling_ddim_parallel"] = ["DDIMParallelScheduler"]
+ _import_structure["scheduling_ddpm"] = ["DDPMScheduler"]
+ _import_structure["scheduling_ddpm_parallel"] = ["DDPMParallelScheduler"]
+ _import_structure["scheduling_ddpm_wuerstchen"] = ["DDPMWuerstchenScheduler"]
+ _import_structure["scheduling_deis_multistep"] = ["DEISMultistepScheduler"]
+ _import_structure["scheduling_dpmsolver_multistep"] = ["DPMSolverMultistepScheduler"]
+ _import_structure["scheduling_dpmsolver_multistep_inverse"] = ["DPMSolverMultistepInverseScheduler"]
+ _import_structure["scheduling_dpmsolver_singlestep"] = ["DPMSolverSinglestepScheduler"]
+ _import_structure["scheduling_edm_dpmsolver_multistep"] = ["EDMDPMSolverMultistepScheduler"]
+ _import_structure["scheduling_edm_euler"] = ["EDMEulerScheduler"]
+ _import_structure["scheduling_euler_ancestral_discrete"] = ["EulerAncestralDiscreteScheduler"]
+ _import_structure["scheduling_euler_discrete"] = ["EulerDiscreteScheduler"]
+ _import_structure["scheduling_heun_discrete"] = ["HeunDiscreteScheduler"]
+ _import_structure["scheduling_ipndm"] = ["IPNDMScheduler"]
+ _import_structure["scheduling_k_dpm_2_ancestral_discrete"] = ["KDPM2AncestralDiscreteScheduler"]
+ _import_structure["scheduling_k_dpm_2_discrete"] = ["KDPM2DiscreteScheduler"]
+ _import_structure["scheduling_lcm"] = ["LCMScheduler"]
+ _import_structure["scheduling_pndm"] = ["PNDMScheduler"]
+ _import_structure["scheduling_repaint"] = ["RePaintScheduler"]
+ _import_structure["scheduling_sasolver"] = ["SASolverScheduler"]
+ _import_structure["scheduling_sde_ve"] = ["ScoreSdeVeScheduler"]
+ _import_structure["scheduling_tcd"] = ["TCDScheduler"]
+ _import_structure["scheduling_unclip"] = ["UnCLIPScheduler"]
+ _import_structure["scheduling_unipc_multistep"] = ["UniPCMultistepScheduler"]
+ _import_structure["scheduling_utils"] = ["KarrasDiffusionSchedulers", "SchedulerMixin"]
+ _import_structure["scheduling_vq_diffusion"] = ["VQDiffusionScheduler"]
+
+try:
+ if not is_flax_available():
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_flax_objects # noqa F403
+
+ _dummy_modules.update(get_objects_from_module(dummy_flax_objects))
+
+else:
+ _import_structure["scheduling_ddim_flax"] = ["FlaxDDIMScheduler"]
+ _import_structure["scheduling_ddpm_flax"] = ["FlaxDDPMScheduler"]
+ _import_structure["scheduling_dpmsolver_multistep_flax"] = ["FlaxDPMSolverMultistepScheduler"]
+ _import_structure["scheduling_euler_discrete_flax"] = ["FlaxEulerDiscreteScheduler"]
+ _import_structure["scheduling_karras_ve_flax"] = ["FlaxKarrasVeScheduler"]
+ _import_structure["scheduling_lms_discrete_flax"] = ["FlaxLMSDiscreteScheduler"]
+ _import_structure["scheduling_pndm_flax"] = ["FlaxPNDMScheduler"]
+ _import_structure["scheduling_sde_ve_flax"] = ["FlaxScoreSdeVeScheduler"]
+ _import_structure["scheduling_utils_flax"] = [
+ "FlaxKarrasDiffusionSchedulers",
+ "FlaxSchedulerMixin",
+ "FlaxSchedulerOutput",
+ "broadcast_to_shape_from_left",
+ ]
+
+
+try:
+ if not (is_torch_available() and is_scipy_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_torch_and_scipy_objects # noqa F403
+
+ _dummy_modules.update(get_objects_from_module(dummy_torch_and_scipy_objects))
+
+else:
+ _import_structure["scheduling_lms_discrete"] = ["LMSDiscreteScheduler"]
+
+try:
+ if not (is_torch_available() and is_torchsde_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ..utils import dummy_torch_and_torchsde_objects # noqa F403
+
+ _dummy_modules.update(get_objects_from_module(dummy_torch_and_torchsde_objects))
+
+else:
+ _import_structure["scheduling_dpmsolver_sde"] = ["DPMSolverSDEScheduler"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ from ..utils import (
+ OptionalDependencyNotAvailable,
+ is_flax_available,
+ is_scipy_available,
+ is_torch_available,
+ is_torchsde_available,
+ )
+
+ try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_pt_objects import * # noqa F403
+ else:
+ from .deprecated import KarrasVeScheduler, ScoreSdeVpScheduler
+ from .scheduling_amused import AmusedScheduler
+ from .scheduling_consistency_decoder import ConsistencyDecoderScheduler
+ from .scheduling_consistency_models import CMStochasticIterativeScheduler
+ from .scheduling_ddim import DDIMScheduler
+ from .scheduling_ddim_inverse import DDIMInverseScheduler
+ from .scheduling_ddim_parallel import DDIMParallelScheduler
+ from .scheduling_ddpm import DDPMScheduler
+ from .scheduling_ddpm_parallel import DDPMParallelScheduler
+ from .scheduling_ddpm_wuerstchen import DDPMWuerstchenScheduler
+ from .scheduling_deis_multistep import DEISMultistepScheduler
+ from .scheduling_dpmsolver_multistep import DPMSolverMultistepScheduler
+ from .scheduling_dpmsolver_multistep_inverse import DPMSolverMultistepInverseScheduler
+ from .scheduling_dpmsolver_singlestep import DPMSolverSinglestepScheduler
+ from .scheduling_edm_dpmsolver_multistep import EDMDPMSolverMultistepScheduler
+ from .scheduling_edm_euler import EDMEulerScheduler
+ from .scheduling_euler_ancestral_discrete import EulerAncestralDiscreteScheduler
+ from .scheduling_euler_discrete import EulerDiscreteScheduler
+ from .scheduling_heun_discrete import HeunDiscreteScheduler
+ from .scheduling_ipndm import IPNDMScheduler
+ from .scheduling_k_dpm_2_ancestral_discrete import KDPM2AncestralDiscreteScheduler
+ from .scheduling_k_dpm_2_discrete import KDPM2DiscreteScheduler
+ from .scheduling_lcm import LCMScheduler
+ from .scheduling_pndm import PNDMScheduler
+ from .scheduling_repaint import RePaintScheduler
+ from .scheduling_sasolver import SASolverScheduler
+ from .scheduling_sde_ve import ScoreSdeVeScheduler
+ from .scheduling_tcd import TCDScheduler
+ from .scheduling_unclip import UnCLIPScheduler
+ from .scheduling_unipc_multistep import UniPCMultistepScheduler
+ from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+ from .scheduling_vq_diffusion import VQDiffusionScheduler
+
+ try:
+ if not is_flax_available():
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_flax_objects import * # noqa F403
+ else:
+ from .scheduling_ddim_flax import FlaxDDIMScheduler
+ from .scheduling_ddpm_flax import FlaxDDPMScheduler
+ from .scheduling_dpmsolver_multistep_flax import FlaxDPMSolverMultistepScheduler
+ from .scheduling_euler_discrete_flax import FlaxEulerDiscreteScheduler
+ from .scheduling_karras_ve_flax import FlaxKarrasVeScheduler
+ from .scheduling_lms_discrete_flax import FlaxLMSDiscreteScheduler
+ from .scheduling_pndm_flax import FlaxPNDMScheduler
+ from .scheduling_sde_ve_flax import FlaxScoreSdeVeScheduler
+ from .scheduling_utils_flax import (
+ FlaxKarrasDiffusionSchedulers,
+ FlaxSchedulerMixin,
+ FlaxSchedulerOutput,
+ broadcast_to_shape_from_left,
+ )
+
+ try:
+ if not (is_torch_available() and is_scipy_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_torch_and_scipy_objects import * # noqa F403
+ else:
+ from .scheduling_lms_discrete import LMSDiscreteScheduler
+
+ try:
+ if not (is_torch_available() and is_torchsde_available()):
+ raise OptionalDependencyNotAvailable()
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_torch_and_torchsde_objects import * # noqa F403
+ else:
+ from .scheduling_dpmsolver_sde import DPMSolverSDEScheduler
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
+ for name, value in _dummy_modules.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/schedulers/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..01762ed728f80649689595571748b33d4056a05d
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f5ca8412c5cdf72fa4ab8e958e4eb0776486766c
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_amused.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_amused.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..16be7e0e72174b6b0c84bf82372424af1dec1461
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_amused.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_decoder.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_decoder.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5af1979e071099061a0e91a783698f4d4c61c3eb
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_decoder.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_decoder.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_decoder.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d0e9ccfe4ddc455c7b0d6e463ef487a94d689f60
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_decoder.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_models.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_models.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..87c0dd7dd2d81f87a1922822711ed43460cb323a
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_consistency_models.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_ddim.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_ddim.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e158fc0595691ea80edde3612198d8eb174eafc6
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_ddim.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_ddim.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_ddim.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e9b0e00c8a36d63dd6551151cf1ee2a6dea1ec42
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_ddim.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_ddim_inverse.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_ddim_inverse.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..67712a9125ce51fa002da19ac40d1b09fca210a8
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_ddim_inverse.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4c4dd6e85ba5068f18d0db32c0790a3706eeab79
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..768a982aed98b4d33959b9e908ae7ad28d9ec2c9
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm_wuerstchen.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm_wuerstchen.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..7294079d5beef4a7d5fea0cf892d85336da9713e
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_ddpm_wuerstchen.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_deis_multistep.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_deis_multistep.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..36319cd1b5cb267d51633b010546471934dd887b
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_deis_multistep.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..dba044ed29a90c2669ed885d98476b1cf24c79f5
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3d0d86d15b281b46e3f85f0d98c79e2ead8fc807
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_multistep.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_singlestep.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_singlestep.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6488092d9a243af9fda5a2f7d077205daa956c8d
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_dpmsolver_singlestep.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b23abf3d911e084c8d668a28a27a912c8b5ebc9f
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f80d8feecbcfec90dfe06b22fe2af8bdd34df90d
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_edm_dpmsolver_multistep.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_edm_euler.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_edm_euler.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a8b022310883ddb23544f420445ee4c5cc8f251f
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_edm_euler.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c367b182b5a502ce17729d1d660053787795ab09
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9c24f4b55980198b1f57e3b86425b9684233e8bd
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_euler_ancestral_discrete.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_euler_discrete.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_euler_discrete.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..554212a238da4a7e14a2b62671ca8f3239182f00
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_euler_discrete.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_euler_discrete.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_euler_discrete.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..123898f86441cd0cf0f51ad2002896f61cb47622
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_euler_discrete.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_heun_discrete.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_heun_discrete.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..673edac9bbf04b3997e75c42549050fde8ba86ef
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_heun_discrete.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_heun_discrete.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_heun_discrete.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f9c1f4e7b8fa29f0b20db7f9c3bea20ae95f2f75
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_heun_discrete.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_k_dpm_2_ancestral_discrete.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_k_dpm_2_ancestral_discrete.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..63940fd708d28c609ec647de296fade95f7c32a7
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_k_dpm_2_ancestral_discrete.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_k_dpm_2_discrete.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_k_dpm_2_discrete.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3d0b11a3da027e6d252b3631cedbf3232f5aba99
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_k_dpm_2_discrete.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_lcm.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_lcm.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..06699b85b2d09c6c9f6589feac1b37c761a65830
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_lcm.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_lms_discrete.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_lms_discrete.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..eeb9e199067c3cd40b1e6751bd4adc893dad4895
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_lms_discrete.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_lms_discrete.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_lms_discrete.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..06d69bd18f58b72000053bf69f0eb12c76637b3e
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_lms_discrete.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_pndm.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_pndm.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ba8fc78782889faafa36c1c8cb65fef2fcd4b9d7
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_pndm.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_pndm.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_pndm.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..ac9440078869a765a1124eb5e4e63e7e54b89559
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_pndm.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_repaint.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_repaint.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..13e6ace9a7a1373e1b55c44475568947c2e3f46d
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_repaint.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_sde_ve.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_sde_ve.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6a9f3a2fa47d9a8ef64181f24ccdcccd77110153
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_sde_ve.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_unclip.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_unclip.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..349e0f6ffe0e98190d13d99ae5bc47dbcb2d6a14
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_unclip.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_unipc_multistep.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_unipc_multistep.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9b387f35865a9a2c46053948c2cfa4f8e875012e
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_unipc_multistep.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_utils.cpython-310.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1e99901cd717d1bf71ad346d498bf913bb37f77d
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_utils.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5e8842d0f834c8b3e1a23f83ace4dae53ecb20d3
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/__pycache__/scheduling_vq_diffusion.cpython-39.pyc b/model/diffusers_c/schedulers/__pycache__/scheduling_vq_diffusion.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..888a1ebb63ab6d60ca2ed5be4037ddca4a49810e
Binary files /dev/null and b/model/diffusers_c/schedulers/__pycache__/scheduling_vq_diffusion.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/deprecated/__init__.py b/model/diffusers_c/schedulers/deprecated/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..786707f45206a46f761fa11cd4ced4e80033b2d2
--- /dev/null
+++ b/model/diffusers_c/schedulers/deprecated/__init__.py
@@ -0,0 +1,50 @@
+from typing import TYPE_CHECKING
+
+from ...utils import (
+ DIFFUSERS_SLOW_IMPORT,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_torch_available,
+ is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+try:
+ if not (is_transformers_available() and is_torch_available()):
+ raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+ from ...utils import dummy_pt_objects # noqa F403
+
+ _dummy_objects.update(get_objects_from_module(dummy_pt_objects))
+else:
+ _import_structure["scheduling_karras_ve"] = ["KarrasVeScheduler"]
+ _import_structure["scheduling_sde_vp"] = ["ScoreSdeVpScheduler"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+ try:
+ if not is_torch_available():
+ raise OptionalDependencyNotAvailable()
+
+ except OptionalDependencyNotAvailable:
+ from ..utils.dummy_pt_objects import * # noqa F403
+ else:
+ from .scheduling_karras_ve import KarrasVeScheduler
+ from .scheduling_sde_vp import ScoreSdeVpScheduler
+
+
+else:
+ import sys
+
+ sys.modules[__name__] = _LazyModule(
+ __name__,
+ globals()["__file__"],
+ _import_structure,
+ module_spec=__spec__,
+ )
+
+ for name, value in _dummy_objects.items():
+ setattr(sys.modules[__name__], name, value)
diff --git a/model/diffusers_c/schedulers/deprecated/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/schedulers/deprecated/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..9f07fbc5d8a1761353a9037da2a4e32eeb8a5df9
Binary files /dev/null and b/model/diffusers_c/schedulers/deprecated/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/deprecated/__pycache__/scheduling_karras_ve.cpython-39.pyc b/model/diffusers_c/schedulers/deprecated/__pycache__/scheduling_karras_ve.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..28ad73e74e199b192659c7c226459911a5acefa8
Binary files /dev/null and b/model/diffusers_c/schedulers/deprecated/__pycache__/scheduling_karras_ve.cpython-39.pyc differ
diff --git a/model/diffusers_c/schedulers/deprecated/scheduling_karras_ve.py b/model/diffusers_c/schedulers/deprecated/scheduling_karras_ve.py
new file mode 100644
index 0000000000000000000000000000000000000000..d776d989a143c40cdc8513f449be89520e44ea1b
--- /dev/null
+++ b/model/diffusers_c/schedulers/deprecated/scheduling_karras_ve.py
@@ -0,0 +1,243 @@
+# Copyright 2024 NVIDIA and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils import BaseOutput
+from ...utils.torch_utils import randn_tensor
+from ..scheduling_utils import SchedulerMixin
+
+
+@dataclass
+class KarrasVeOutput(BaseOutput):
+ """
+ Output class for the scheduler's step function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ derivative (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Derivative of predicted original image sample (x_0).
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample (x_{0}) based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ derivative: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+class KarrasVeScheduler(SchedulerMixin, ConfigMixin):
+ """
+ A stochastic scheduler tailored to variance-expanding models.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+
+
+ For more details on the parameters, see [Appendix E](https://arxiv.org/abs/2206.00364). The grid search values used
+ to find the optimal `{s_noise, s_churn, s_min, s_max}` for a specific model are described in Table 5 of the paper.
+
+
+
+ Args:
+ sigma_min (`float`, defaults to 0.02):
+ The minimum noise magnitude.
+ sigma_max (`float`, defaults to 100):
+ The maximum noise magnitude.
+ s_noise (`float`, defaults to 1.007):
+ The amount of additional noise to counteract loss of detail during sampling. A reasonable range is [1.000,
+ 1.011].
+ s_churn (`float`, defaults to 80):
+ The parameter controlling the overall amount of stochasticity. A reasonable range is [0, 100].
+ s_min (`float`, defaults to 0.05):
+ The start value of the sigma range to add noise (enable stochasticity). A reasonable range is [0, 10].
+ s_max (`float`, defaults to 50):
+ The end value of the sigma range to add noise. A reasonable range is [0.2, 80].
+ """
+
+ order = 2
+
+ @register_to_config
+ def __init__(
+ self,
+ sigma_min: float = 0.02,
+ sigma_max: float = 100,
+ s_noise: float = 1.007,
+ s_churn: float = 80,
+ s_min: float = 0.05,
+ s_max: float = 50,
+ ):
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = sigma_max
+
+ # setable values
+ self.num_inference_steps: int = None
+ self.timesteps: np.IntTensor = None
+ self.schedule: torch.FloatTensor = None # sigma(t_i)
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+ timesteps = np.arange(0, self.num_inference_steps)[::-1].copy()
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+ schedule = [
+ (
+ self.config.sigma_max**2
+ * (self.config.sigma_min**2 / self.config.sigma_max**2) ** (i / (num_inference_steps - 1))
+ )
+ for i in self.timesteps
+ ]
+ self.schedule = torch.tensor(schedule, dtype=torch.float32, device=device)
+
+ def add_noise_to_input(
+ self, sample: torch.FloatTensor, sigma: float, generator: Optional[torch.Generator] = None
+ ) -> Tuple[torch.FloatTensor, float]:
+ """
+ Explicit Langevin-like "churn" step of adding noise to the sample according to a `gamma_i ≥ 0` to reach a
+ higher noise level `sigma_hat = sigma_i + gamma_i*sigma_i`.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ sigma (`float`):
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ """
+ if self.config.s_min <= sigma <= self.config.s_max:
+ gamma = min(self.config.s_churn / self.num_inference_steps, 2**0.5 - 1)
+ else:
+ gamma = 0
+
+ # sample eps ~ N(0, S_noise^2 * I)
+ eps = self.config.s_noise * randn_tensor(sample.shape, generator=generator).to(sample.device)
+ sigma_hat = sigma + gamma * sigma
+ sample_hat = sample + ((sigma_hat**2 - sigma**2) ** 0.5 * eps)
+
+ return sample_hat, sigma_hat
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ sigma_hat: float,
+ sigma_prev: float,
+ sample_hat: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[KarrasVeOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ sigma_hat (`float`):
+ sigma_prev (`float`):
+ sample_hat (`torch.FloatTensor`):
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_karras_ve.KarrasVESchedulerOutput`] is returned,
+ otherwise a tuple is returned where the first element is the sample tensor.
+
+ """
+
+ pred_original_sample = sample_hat + sigma_hat * model_output
+ derivative = (sample_hat - pred_original_sample) / sigma_hat
+ sample_prev = sample_hat + (sigma_prev - sigma_hat) * derivative
+
+ if not return_dict:
+ return (sample_prev, derivative)
+
+ return KarrasVeOutput(
+ prev_sample=sample_prev, derivative=derivative, pred_original_sample=pred_original_sample
+ )
+
+ def step_correct(
+ self,
+ model_output: torch.FloatTensor,
+ sigma_hat: float,
+ sigma_prev: float,
+ sample_hat: torch.FloatTensor,
+ sample_prev: torch.FloatTensor,
+ derivative: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[KarrasVeOutput, Tuple]:
+ """
+ Corrects the predicted sample based on the `model_output` of the network.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ sigma_hat (`float`): TODO
+ sigma_prev (`float`): TODO
+ sample_hat (`torch.FloatTensor`): TODO
+ sample_prev (`torch.FloatTensor`): TODO
+ derivative (`torch.FloatTensor`): TODO
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`.
+
+ Returns:
+ prev_sample (TODO): updated sample in the diffusion chain. derivative (TODO): TODO
+
+ """
+ pred_original_sample = sample_prev + sigma_prev * model_output
+ derivative_corr = (sample_prev - pred_original_sample) / sigma_prev
+ sample_prev = sample_hat + (sigma_prev - sigma_hat) * (0.5 * derivative + 0.5 * derivative_corr)
+
+ if not return_dict:
+ return (sample_prev, derivative)
+
+ return KarrasVeOutput(
+ prev_sample=sample_prev, derivative=derivative, pred_original_sample=pred_original_sample
+ )
+
+ def add_noise(self, original_samples, noise, timesteps):
+ raise NotImplementedError()
diff --git a/model/diffusers_c/schedulers/deprecated/scheduling_sde_vp.py b/model/diffusers_c/schedulers/deprecated/scheduling_sde_vp.py
new file mode 100644
index 0000000000000000000000000000000000000000..09b02cadc4008b1d80184a355a06201a85f97e8f
--- /dev/null
+++ b/model/diffusers_c/schedulers/deprecated/scheduling_sde_vp.py
@@ -0,0 +1,109 @@
+# Copyright 2024 Google Brain and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/yang-song/score_sde_pytorch
+
+import math
+from typing import Union
+
+import torch
+
+from ...configuration_utils import ConfigMixin, register_to_config
+from ...utils.torch_utils import randn_tensor
+from ..scheduling_utils import SchedulerMixin
+
+
+class ScoreSdeVpScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `ScoreSdeVpScheduler` is a variance preserving stochastic differential equation (SDE) scheduler.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 2000):
+ The number of diffusion steps to train the model.
+ beta_min (`int`, defaults to 0.1):
+ beta_max (`int`, defaults to 20):
+ sampling_eps (`int`, defaults to 1e-3):
+ The end value of sampling where timesteps decrease progressively from 1 to epsilon.
+ """
+
+ order = 1
+
+ @register_to_config
+ def __init__(self, num_train_timesteps=2000, beta_min=0.1, beta_max=20, sampling_eps=1e-3):
+ self.sigmas = None
+ self.discrete_sigmas = None
+ self.timesteps = None
+
+ def set_timesteps(self, num_inference_steps, device: Union[str, torch.device] = None):
+ """
+ Sets the continuous timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.timesteps = torch.linspace(1, self.config.sampling_eps, num_inference_steps, device=device)
+
+ def step_pred(self, score, x, t, generator=None):
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ score ():
+ x ():
+ t ():
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ """
+ if self.timesteps is None:
+ raise ValueError(
+ "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ # TODO(Patrick) better comments + non-PyTorch
+ # postprocess model score
+ log_mean_coeff = -0.25 * t**2 * (self.config.beta_max - self.config.beta_min) - 0.5 * t * self.config.beta_min
+ std = torch.sqrt(1.0 - torch.exp(2.0 * log_mean_coeff))
+ std = std.flatten()
+ while len(std.shape) < len(score.shape):
+ std = std.unsqueeze(-1)
+ score = -score / std
+
+ # compute
+ dt = -1.0 / len(self.timesteps)
+
+ beta_t = self.config.beta_min + t * (self.config.beta_max - self.config.beta_min)
+ beta_t = beta_t.flatten()
+ while len(beta_t.shape) < len(x.shape):
+ beta_t = beta_t.unsqueeze(-1)
+ drift = -0.5 * beta_t * x
+
+ diffusion = torch.sqrt(beta_t)
+ drift = drift - diffusion**2 * score
+ x_mean = x + drift * dt
+
+ # add noise
+ noise = randn_tensor(x.shape, layout=x.layout, generator=generator, device=x.device, dtype=x.dtype)
+ x = x_mean + diffusion * math.sqrt(-dt) * noise
+
+ return x, x_mean
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_amused.py b/model/diffusers_c/schedulers/scheduling_amused.py
new file mode 100644
index 0000000000000000000000000000000000000000..51fbe6a4dc7da4282c9a264764c6b8e9c87d02e1
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_amused.py
@@ -0,0 +1,162 @@
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from .scheduling_utils import SchedulerMixin
+
+
+def gumbel_noise(t, generator=None):
+ device = generator.device if generator is not None else t.device
+ noise = torch.zeros_like(t, device=device).uniform_(0, 1, generator=generator).to(t.device)
+ return -torch.log((-torch.log(noise.clamp(1e-20))).clamp(1e-20))
+
+
+def mask_by_random_topk(mask_len, probs, temperature=1.0, generator=None):
+ confidence = torch.log(probs.clamp(1e-20)) + temperature * gumbel_noise(probs, generator=generator)
+ sorted_confidence = torch.sort(confidence, dim=-1).values
+ cut_off = torch.gather(sorted_confidence, 1, mask_len.long())
+ masking = confidence < cut_off
+ return masking
+
+
+@dataclass
+class AmusedSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: torch.FloatTensor = None
+
+
+class AmusedScheduler(SchedulerMixin, ConfigMixin):
+ order = 1
+
+ temperatures: torch.Tensor
+
+ @register_to_config
+ def __init__(
+ self,
+ mask_token_id: int,
+ masking_schedule: str = "cosine",
+ ):
+ self.temperatures = None
+ self.timesteps = None
+
+ def set_timesteps(
+ self,
+ num_inference_steps: int,
+ temperature: Union[int, Tuple[int, int], List[int]] = (2, 0),
+ device: Union[str, torch.device] = None,
+ ):
+ self.timesteps = torch.arange(num_inference_steps, device=device).flip(0)
+
+ if isinstance(temperature, (tuple, list)):
+ self.temperatures = torch.linspace(temperature[0], temperature[1], num_inference_steps, device=device)
+ else:
+ self.temperatures = torch.linspace(temperature, 0.01, num_inference_steps, device=device)
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: torch.long,
+ sample: torch.LongTensor,
+ starting_mask_ratio: int = 1,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[AmusedSchedulerOutput, Tuple]:
+ two_dim_input = sample.ndim == 3 and model_output.ndim == 4
+
+ if two_dim_input:
+ batch_size, codebook_size, height, width = model_output.shape
+ sample = sample.reshape(batch_size, height * width)
+ model_output = model_output.reshape(batch_size, codebook_size, height * width).permute(0, 2, 1)
+
+ unknown_map = sample == self.config.mask_token_id
+
+ probs = model_output.softmax(dim=-1)
+
+ device = probs.device
+ probs_ = probs.to(generator.device) if generator is not None else probs # handles when generator is on CPU
+ if probs_.device.type == "cpu" and probs_.dtype != torch.float32:
+ probs_ = probs_.float() # multinomial is not implemented for cpu half precision
+ probs_ = probs_.reshape(-1, probs.size(-1))
+ pred_original_sample = torch.multinomial(probs_, 1, generator=generator).to(device=device)
+ pred_original_sample = pred_original_sample[:, 0].view(*probs.shape[:-1])
+ pred_original_sample = torch.where(unknown_map, pred_original_sample, sample)
+
+ if timestep == 0:
+ prev_sample = pred_original_sample
+ else:
+ seq_len = sample.shape[1]
+ step_idx = (self.timesteps == timestep).nonzero()
+ ratio = (step_idx + 1) / len(self.timesteps)
+
+ if self.config.masking_schedule == "cosine":
+ mask_ratio = torch.cos(ratio * math.pi / 2)
+ elif self.config.masking_schedule == "linear":
+ mask_ratio = 1 - ratio
+ else:
+ raise ValueError(f"unknown masking schedule {self.config.masking_schedule}")
+
+ mask_ratio = starting_mask_ratio * mask_ratio
+
+ mask_len = (seq_len * mask_ratio).floor()
+ # do not mask more than amount previously masked
+ mask_len = torch.min(unknown_map.sum(dim=-1, keepdim=True) - 1, mask_len)
+ # mask at least one
+ mask_len = torch.max(torch.tensor([1], device=model_output.device), mask_len)
+
+ selected_probs = torch.gather(probs, -1, pred_original_sample[:, :, None])[:, :, 0]
+ # Ignores the tokens given in the input by overwriting their confidence.
+ selected_probs = torch.where(unknown_map, selected_probs, torch.finfo(selected_probs.dtype).max)
+
+ masking = mask_by_random_topk(mask_len, selected_probs, self.temperatures[step_idx], generator)
+
+ # Masks tokens with lower confidence.
+ prev_sample = torch.where(masking, self.config.mask_token_id, pred_original_sample)
+
+ if two_dim_input:
+ prev_sample = prev_sample.reshape(batch_size, height, width)
+ pred_original_sample = pred_original_sample.reshape(batch_size, height, width)
+
+ if not return_dict:
+ return (prev_sample, pred_original_sample)
+
+ return AmusedSchedulerOutput(prev_sample, pred_original_sample)
+
+ def add_noise(self, sample, timesteps, generator=None):
+ step_idx = (self.timesteps == timesteps).nonzero()
+ ratio = (step_idx + 1) / len(self.timesteps)
+
+ if self.config.masking_schedule == "cosine":
+ mask_ratio = torch.cos(ratio * math.pi / 2)
+ elif self.config.masking_schedule == "linear":
+ mask_ratio = 1 - ratio
+ else:
+ raise ValueError(f"unknown masking schedule {self.config.masking_schedule}")
+
+ mask_indices = (
+ torch.rand(
+ sample.shape, device=generator.device if generator is not None else sample.device, generator=generator
+ ).to(sample.device)
+ < mask_ratio
+ )
+
+ masked_sample = sample.clone()
+
+ masked_sample[mask_indices] = self.config.mask_token_id
+
+ return masked_sample
diff --git a/model/diffusers_c/schedulers/scheduling_consistency_decoder.py b/model/diffusers_c/schedulers/scheduling_consistency_decoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..69ca8a1737ecb4bc859952fba25c7bea0b889486
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_consistency_decoder.py
@@ -0,0 +1,180 @@
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+@dataclass
+class ConsistencyDecoderSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ """
+
+ prev_sample: torch.FloatTensor
+
+
+class ConsistencyDecoderScheduler(SchedulerMixin, ConfigMixin):
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1024,
+ sigma_data: float = 0.5,
+ ):
+ betas = betas_for_alpha_bar(num_train_timesteps)
+
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+
+ self.sqrt_alphas_cumprod = torch.sqrt(alphas_cumprod)
+ self.sqrt_one_minus_alphas_cumprod = torch.sqrt(1.0 - alphas_cumprod)
+
+ sigmas = torch.sqrt(1.0 / alphas_cumprod - 1)
+
+ sqrt_recip_alphas_cumprod = torch.sqrt(1.0 / alphas_cumprod)
+
+ self.c_skip = sqrt_recip_alphas_cumprod * sigma_data**2 / (sigmas**2 + sigma_data**2)
+ self.c_out = sigmas * sigma_data / (sigmas**2 + sigma_data**2) ** 0.5
+ self.c_in = sqrt_recip_alphas_cumprod / (sigmas**2 + sigma_data**2) ** 0.5
+
+ def set_timesteps(
+ self,
+ num_inference_steps: Optional[int] = None,
+ device: Union[str, torch.device] = None,
+ ):
+ if num_inference_steps != 2:
+ raise ValueError("Currently more than 2 inference steps are not supported.")
+
+ self.timesteps = torch.tensor([1008, 512], dtype=torch.long, device=device)
+ self.sqrt_alphas_cumprod = self.sqrt_alphas_cumprod.to(device)
+ self.sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod.to(device)
+ self.c_skip = self.c_skip.to(device)
+ self.c_out = self.c_out.to(device)
+ self.c_in = self.c_in.to(device)
+
+ @property
+ def init_noise_sigma(self):
+ return self.sqrt_one_minus_alphas_cumprod[self.timesteps[0]]
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample * self.c_in[timestep]
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ sample: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[ConsistencyDecoderSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ timestep (`float`):
+ The current timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a
+ [`~schedulers.scheduling_consistency_models.ConsistencyDecoderSchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_consistency_models.ConsistencyDecoderSchedulerOutput`] or `tuple`:
+ If return_dict is `True`,
+ [`~schedulers.scheduling_consistency_models.ConsistencyDecoderSchedulerOutput`] is returned, otherwise
+ a tuple is returned where the first element is the sample tensor.
+ """
+ x_0 = self.c_out[timestep] * model_output + self.c_skip[timestep] * sample
+
+ timestep_idx = torch.where(self.timesteps == timestep)[0]
+
+ if timestep_idx == len(self.timesteps) - 1:
+ prev_sample = x_0
+ else:
+ noise = randn_tensor(x_0.shape, generator=generator, dtype=x_0.dtype, device=x_0.device)
+ prev_sample = (
+ self.sqrt_alphas_cumprod[self.timesteps[timestep_idx + 1]].to(x_0.dtype) * x_0
+ + self.sqrt_one_minus_alphas_cumprod[self.timesteps[timestep_idx + 1]].to(x_0.dtype) * noise
+ )
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return ConsistencyDecoderSchedulerOutput(prev_sample=prev_sample)
diff --git a/model/diffusers_c/schedulers/scheduling_consistency_models.py b/model/diffusers_c/schedulers/scheduling_consistency_models.py
new file mode 100644
index 0000000000000000000000000000000000000000..14d37a390b5d522f5af01058c3c9cc16176273d2
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_consistency_models.py
@@ -0,0 +1,448 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class CMStochasticIterativeSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ """
+
+ prev_sample: torch.FloatTensor
+
+
+class CMStochasticIterativeScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Multistep and onestep sampling for consistency models.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 40):
+ The number of diffusion steps to train the model.
+ sigma_min (`float`, defaults to 0.002):
+ Minimum noise magnitude in the sigma schedule. Defaults to 0.002 from the original implementation.
+ sigma_max (`float`, defaults to 80.0):
+ Maximum noise magnitude in the sigma schedule. Defaults to 80.0 from the original implementation.
+ sigma_data (`float`, defaults to 0.5):
+ The standard deviation of the data distribution from the EDM
+ [paper](https://huggingface.co/papers/2206.00364). Defaults to 0.5 from the original implementation.
+ s_noise (`float`, defaults to 1.0):
+ The amount of additional noise to counteract loss of detail during sampling. A reasonable range is [1.000,
+ 1.011]. Defaults to 1.0 from the original implementation.
+ rho (`float`, defaults to 7.0):
+ The parameter for calculating the Karras sigma schedule from the EDM
+ [paper](https://huggingface.co/papers/2206.00364). Defaults to 7.0 from the original implementation.
+ clip_denoised (`bool`, defaults to `True`):
+ Whether to clip the denoised outputs to `(-1, 1)`.
+ timesteps (`List` or `np.ndarray` or `torch.Tensor`, *optional*):
+ An explicit timestep schedule that can be optionally specified. The timesteps are expected to be in
+ increasing order.
+ """
+
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 40,
+ sigma_min: float = 0.002,
+ sigma_max: float = 80.0,
+ sigma_data: float = 0.5,
+ s_noise: float = 1.0,
+ rho: float = 7.0,
+ clip_denoised: bool = True,
+ ):
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = sigma_max
+
+ ramp = np.linspace(0, 1, num_train_timesteps)
+ sigmas = self._convert_to_karras(ramp)
+ timesteps = self.sigma_to_t(sigmas)
+
+ # setable values
+ self.num_inference_steps = None
+ self.sigmas = torch.from_numpy(sigmas)
+ self.timesteps = torch.from_numpy(timesteps)
+ self.custom_timesteps = False
+ self.is_scale_input_called = False
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(
+ self, sample: torch.FloatTensor, timestep: Union[float, torch.FloatTensor]
+ ) -> torch.FloatTensor:
+ """
+ Scales the consistency model input by `(sigma**2 + sigma_data**2) ** 0.5`.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`float` or `torch.FloatTensor`):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ # Get sigma corresponding to timestep
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+
+ sample = sample / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+
+ self.is_scale_input_called = True
+ return sample
+
+ def sigma_to_t(self, sigmas: Union[float, np.ndarray]):
+ """
+ Gets scaled timesteps from the Karras sigmas for input to the consistency model.
+
+ Args:
+ sigmas (`float` or `np.ndarray`):
+ A single Karras sigma or an array of Karras sigmas.
+
+ Returns:
+ `float` or `np.ndarray`:
+ A scaled input timestep or scaled input timestep array.
+ """
+ if not isinstance(sigmas, np.ndarray):
+ sigmas = np.array(sigmas, dtype=np.float64)
+
+ timesteps = 1000 * 0.25 * np.log(sigmas + 1e-44)
+
+ return timesteps
+
+ def set_timesteps(
+ self,
+ num_inference_steps: Optional[int] = None,
+ device: Union[str, torch.device] = None,
+ timesteps: Optional[List[int]] = None,
+ ):
+ """
+ Sets the timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed,
+ `num_inference_steps` must be `None`.
+ """
+ if num_inference_steps is None and timesteps is None:
+ raise ValueError("Exactly one of `num_inference_steps` or `timesteps` must be supplied.")
+
+ if num_inference_steps is not None and timesteps is not None:
+ raise ValueError("Can only pass one of `num_inference_steps` or `timesteps`.")
+
+ # Follow DDPMScheduler custom timesteps logic
+ if timesteps is not None:
+ for i in range(1, len(timesteps)):
+ if timesteps[i] >= timesteps[i - 1]:
+ raise ValueError("`timesteps` must be in descending order.")
+
+ if timesteps[0] >= self.config.num_train_timesteps:
+ raise ValueError(
+ f"`timesteps` must start before `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps}."
+ )
+
+ timesteps = np.array(timesteps, dtype=np.int64)
+ self.custom_timesteps = True
+ else:
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ self.num_inference_steps = num_inference_steps
+
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
+ self.custom_timesteps = False
+
+ # Map timesteps to Karras sigmas directly for multistep sampling
+ # See https://github.com/openai/consistency_models/blob/main/cm/karras_diffusion.py#L675
+ num_train_timesteps = self.config.num_train_timesteps
+ ramp = timesteps[::-1].copy()
+ ramp = ramp / (num_train_timesteps - 1)
+ sigmas = self._convert_to_karras(ramp)
+ timesteps = self.sigma_to_t(sigmas)
+
+ sigmas = np.concatenate([sigmas, [self.sigma_min]]).astype(np.float32)
+ self.sigmas = torch.from_numpy(sigmas).to(device=device)
+
+ if str(device).startswith("mps"):
+ # mps does not support float64
+ self.timesteps = torch.from_numpy(timesteps).to(device, dtype=torch.float32)
+ else:
+ self.timesteps = torch.from_numpy(timesteps).to(device=device)
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Modified _convert_to_karras implementation that takes in ramp as argument
+ def _convert_to_karras(self, ramp):
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ sigma_min: float = self.config.sigma_min
+ sigma_max: float = self.config.sigma_max
+
+ rho = self.config.rho
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def get_scalings(self, sigma):
+ sigma_data = self.config.sigma_data
+
+ c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
+ c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+ return c_skip, c_out
+
+ def get_scalings_for_boundary_condition(self, sigma):
+ """
+ Gets the scalings used in the consistency model parameterization (from Appendix C of the
+ [paper](https://huggingface.co/papers/2303.01469)) to enforce boundary condition.
+
+
+
+ `epsilon` in the equations for `c_skip` and `c_out` is set to `sigma_min`.
+
+
+
+ Args:
+ sigma (`torch.FloatTensor`):
+ The current sigma in the Karras sigma schedule.
+
+ Returns:
+ `tuple`:
+ A two-element tuple where `c_skip` (which weights the current sample) is the first element and `c_out`
+ (which weights the consistency model output) is the second element.
+ """
+ sigma_min = self.config.sigma_min
+ sigma_data = self.config.sigma_data
+
+ c_skip = sigma_data**2 / ((sigma - sigma_min) ** 2 + sigma_data**2)
+ c_out = (sigma - sigma_min) * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+ return c_skip, c_out
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ sample: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[CMStochasticIterativeSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ timestep (`float`):
+ The current timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a
+ [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] or `tuple`:
+ If return_dict is `True`,
+ [`~schedulers.scheduling_consistency_models.CMStochasticIterativeSchedulerOutput`] is returned,
+ otherwise a tuple is returned where the first element is the sample tensor.
+ """
+
+ if (
+ isinstance(timestep, int)
+ or isinstance(timestep, torch.IntTensor)
+ or isinstance(timestep, torch.LongTensor)
+ ):
+ raise ValueError(
+ (
+ "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+ f" `{self.__class__}.step()` is not supported. Make sure to pass"
+ " one of the `scheduler.timesteps` as a timestep."
+ ),
+ )
+
+ if not self.is_scale_input_called:
+ logger.warning(
+ "The `scale_model_input` function should be called before `step` to ensure correct denoising. "
+ "See `StableDiffusionPipeline` for a usage example."
+ )
+
+ sigma_min = self.config.sigma_min
+ sigma_max = self.config.sigma_max
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ # sigma_next corresponds to next_t in original implementation
+ sigma = self.sigmas[self.step_index]
+ if self.step_index + 1 < self.config.num_train_timesteps:
+ sigma_next = self.sigmas[self.step_index + 1]
+ else:
+ # Set sigma_next to sigma_min
+ sigma_next = self.sigmas[-1]
+
+ # Get scalings for boundary conditions
+ c_skip, c_out = self.get_scalings_for_boundary_condition(sigma)
+
+ # 1. Denoise model output using boundary conditions
+ denoised = c_out * model_output + c_skip * sample
+ if self.config.clip_denoised:
+ denoised = denoised.clamp(-1, 1)
+
+ # 2. Sample z ~ N(0, s_noise^2 * I)
+ # Noise is not used for onestep sampling.
+ if len(self.timesteps) > 1:
+ noise = randn_tensor(
+ model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
+ )
+ else:
+ noise = torch.zeros_like(model_output)
+ z = noise * self.config.s_noise
+
+ sigma_hat = sigma_next.clamp(min=sigma_min, max=sigma_max)
+
+ # 3. Return noisy sample
+ # tau = sigma_hat, eps = sigma_min
+ prev_sample = denoised + z * (sigma_hat**2 - sigma_min**2) ** 0.5
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return CMStochasticIterativeSchedulerOutput(prev_sample=prev_sample)
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_ddim.py b/model/diffusers_c/schedulers/scheduling_ddim.py
new file mode 100644
index 0000000000000000000000000000000000000000..eac8c254a285dd7d9368c7a639f61621bae3c5d7
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ddim.py
@@ -0,0 +1,522 @@
+# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM
+class DDIMSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+def rescale_zero_terminal_snr(betas):
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class DDIMScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `DDIMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with
+ non-Markovian guidance.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ clip_sample (`bool`, defaults to `True`):
+ Clip the predicted sample for numerical stability.
+ clip_sample_range (`float`, defaults to 1.0):
+ The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ set_alpha_to_one (`bool`, defaults to `True`):
+ Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+ there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+ otherwise it uses the alpha value at step 0.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+ timestep_spacing (`str`, defaults to `"leading"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ clip_sample: bool = True,
+ set_alpha_to_one: bool = True,
+ steps_offset: int = 0,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ clip_sample_range: float = 1.0,
+ sample_max_value: float = 1.0,
+ timestep_spacing: str = "leading",
+ rescale_betas_zero_snr: bool = False,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ # Rescale for zero SNR
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # At every step in ddim, we are looking into the previous alphas_cumprod
+ # For the final step, there is no previous alphas_cumprod because we are already at 0
+ # `set_alpha_to_one` decides whether we set this parameter simply to one or
+ # whether we use the final alpha of the "non-previous" one.
+ self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def _get_variance(self, timestep, prev_timestep):
+ alpha_prod_t = self.alphas_cumprod[timestep]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+
+ return variance
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ """
+
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ self.num_inference_steps = num_inference_steps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
+ .round()[::-1]
+ .copy()
+ .astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'."
+ )
+
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ eta: float = 0.0,
+ use_clipped_model_output: bool = False,
+ generator=None,
+ variance_noise: Optional[torch.FloatTensor] = None,
+ return_dict: bool = True,
+ ) -> Union[DDIMSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ eta (`float`):
+ The weight of noise for added noise in diffusion step.
+ use_clipped_model_output (`bool`, defaults to `False`):
+ If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary
+ because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no
+ clipping has happened, "corrected" `model_output` would coincide with the one provided as input and
+ `use_clipped_model_output` has no effect.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ variance_noise (`torch.FloatTensor`):
+ Alternative to generating noise with `generator` by directly providing the noise for the variance
+ itself. Useful for methods such as [`CycleDiffusion`].
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.DDIMSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_ddim.DDIMSchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+ # Ideally, read DDIM paper in-detail understanding
+
+ # Notation ( ->
+ # - pred_noise_t -> e_theta(x_t, t)
+ # - pred_original_sample -> f_theta(x_t, t) or x_0
+ # - std_dev_t -> sigma_t
+ # - eta -> η
+ # - pred_sample_direction -> "direction pointing to x_t"
+ # - pred_prev_sample -> "x_t-1"
+
+ # 1. get previous step value (=t-1)
+ prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
+
+ # 2. compute alphas, betas
+ alpha_prod_t = self.alphas_cumprod[timestep]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+
+ beta_prod_t = 1 - alpha_prod_t
+
+ # 3. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ pred_epsilon = model_output
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction`"
+ )
+
+ # 4. Clip or threshold "predicted x_0"
+ if self.config.thresholding:
+ pred_original_sample = self._threshold_sample(pred_original_sample)
+ elif self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 5. compute variance: "sigma_t(η)" -> see formula (16)
+ # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
+ variance = self._get_variance(timestep, prev_timestep)
+ std_dev_t = eta * variance ** (0.5)
+
+ if use_clipped_model_output:
+ # the pred_epsilon is always re-derived from the clipped x_0 in Glide
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+
+ # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
+
+ # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+
+ if eta > 0:
+ if variance_noise is not None and generator is not None:
+ raise ValueError(
+ "Cannot pass both generator and variance_noise. Please make sure that either `generator` or"
+ " `variance_noise` stays `None`."
+ )
+
+ if variance_noise is None:
+ variance_noise = randn_tensor(
+ model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
+ )
+ variance = std_dev_t * variance_noise
+
+ prev_sample = prev_sample + variance
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+ # for the subsequent add_noise calls
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
+ def get_velocity(
+ self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as sample
+ self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype)
+ timesteps = timesteps.to(sample.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(sample.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+ return velocity
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_ddim_flax.py b/model/diffusers_c/schedulers/scheduling_ddim_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..0db2d61bf644615315423c5686b1b52ebf2e6c14
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ddim_flax.py
@@ -0,0 +1,315 @@
+# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import flax
+import jax.numpy as jnp
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils_flax import (
+ CommonSchedulerState,
+ FlaxKarrasDiffusionSchedulers,
+ FlaxSchedulerMixin,
+ FlaxSchedulerOutput,
+ add_noise_common,
+ get_velocity_common,
+)
+
+
+@flax.struct.dataclass
+class DDIMSchedulerState:
+ common: CommonSchedulerState
+ final_alpha_cumprod: jnp.ndarray
+
+ # setable values
+ init_noise_sigma: jnp.ndarray
+ timesteps: jnp.ndarray
+ num_inference_steps: Optional[int] = None
+
+ @classmethod
+ def create(
+ cls,
+ common: CommonSchedulerState,
+ final_alpha_cumprod: jnp.ndarray,
+ init_noise_sigma: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ):
+ return cls(
+ common=common,
+ final_alpha_cumprod=final_alpha_cumprod,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ )
+
+
+@dataclass
+class FlaxDDIMSchedulerOutput(FlaxSchedulerOutput):
+ state: DDIMSchedulerState
+
+
+class FlaxDDIMScheduler(FlaxSchedulerMixin, ConfigMixin):
+ """
+ Denoising diffusion implicit models is a scheduler that extends the denoising procedure introduced in denoising
+ diffusion probabilistic models (DDPMs) with non-Markovian guidance.
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ For more details, see the original paper: https://arxiv.org/abs/2010.02502
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ beta_start (`float`): the starting `beta` value of inference.
+ beta_end (`float`): the final `beta` value.
+ beta_schedule (`str`):
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`jnp.ndarray`, optional):
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+ clip_sample (`bool`, default `True`):
+ option to clip predicted sample between for numerical stability. The clip range is determined by `clip_sample_range`.
+ clip_sample_range (`float`, default `1.0`):
+ the maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ set_alpha_to_one (`bool`, default `True`):
+ each diffusion step uses the value of alphas product at that step and at the previous one. For the final
+ step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+ otherwise it uses the value of alpha at step 0.
+ steps_offset (`int`, default `0`):
+ an offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
+ stable diffusion.
+ prediction_type (`str`, default `epsilon`):
+ indicates whether the model predicts the noise (epsilon), or the samples. One of `epsilon`, `sample`.
+ `v-prediction` is not supported for this scheduler.
+ dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`):
+ the `dtype` used for params and computation.
+ """
+
+ _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers]
+
+ dtype: jnp.dtype
+
+ @property
+ def has_state(self):
+ return True
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[jnp.ndarray] = None,
+ clip_sample: bool = True,
+ clip_sample_range: float = 1.0,
+ set_alpha_to_one: bool = True,
+ steps_offset: int = 0,
+ prediction_type: str = "epsilon",
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ self.dtype = dtype
+
+ def create_state(self, common: Optional[CommonSchedulerState] = None) -> DDIMSchedulerState:
+ if common is None:
+ common = CommonSchedulerState.create(self)
+
+ # At every step in ddim, we are looking into the previous alphas_cumprod
+ # For the final step, there is no previous alphas_cumprod because we are already at 0
+ # `set_alpha_to_one` decides whether we set this parameter simply to one or
+ # whether we use the final alpha of the "non-previous" one.
+ final_alpha_cumprod = (
+ jnp.array(1.0, dtype=self.dtype) if self.config.set_alpha_to_one else common.alphas_cumprod[0]
+ )
+
+ # standard deviation of the initial noise distribution
+ init_noise_sigma = jnp.array(1.0, dtype=self.dtype)
+
+ timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1]
+
+ return DDIMSchedulerState.create(
+ common=common,
+ final_alpha_cumprod=final_alpha_cumprod,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ )
+
+ def scale_model_input(
+ self, state: DDIMSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None
+ ) -> jnp.ndarray:
+ """
+ Args:
+ state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance.
+ sample (`jnp.ndarray`): input sample
+ timestep (`int`, optional): current timestep
+
+ Returns:
+ `jnp.ndarray`: scaled input sample
+ """
+ return sample
+
+ def set_timesteps(
+ self, state: DDIMSchedulerState, num_inference_steps: int, shape: Tuple = ()
+ ) -> DDIMSchedulerState:
+ """
+ Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ state (`DDIMSchedulerState`):
+ the `FlaxDDIMScheduler` state data class instance.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ """
+ step_ratio = self.config.num_train_timesteps // num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # rounding to avoid issues when num_inference_step is power of 3
+ timesteps = (jnp.arange(0, num_inference_steps) * step_ratio).round()[::-1] + self.config.steps_offset
+
+ return state.replace(
+ num_inference_steps=num_inference_steps,
+ timesteps=timesteps,
+ )
+
+ def _get_variance(self, state: DDIMSchedulerState, timestep, prev_timestep):
+ alpha_prod_t = state.common.alphas_cumprod[timestep]
+ alpha_prod_t_prev = jnp.where(
+ prev_timestep >= 0, state.common.alphas_cumprod[prev_timestep], state.final_alpha_cumprod
+ )
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+
+ return variance
+
+ def step(
+ self,
+ state: DDIMSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ eta: float = 0.0,
+ return_dict: bool = True,
+ ) -> Union[FlaxDDIMSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ state (`DDIMSchedulerState`): the `FlaxDDIMScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ return_dict (`bool`): option for returning tuple rather than FlaxDDIMSchedulerOutput class
+
+ Returns:
+ [`FlaxDDIMSchedulerOutput`] or `tuple`: [`FlaxDDIMSchedulerOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+ if state.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+ # Ideally, read DDIM paper in-detail understanding
+
+ # Notation ( ->
+ # - pred_noise_t -> e_theta(x_t, t)
+ # - pred_original_sample -> f_theta(x_t, t) or x_0
+ # - std_dev_t -> sigma_t
+ # - eta -> η
+ # - pred_sample_direction -> "direction pointing to x_t"
+ # - pred_prev_sample -> "x_t-1"
+
+ # 1. get previous step value (=t-1)
+ prev_timestep = timestep - self.config.num_train_timesteps // state.num_inference_steps
+
+ alphas_cumprod = state.common.alphas_cumprod
+ final_alpha_cumprod = state.final_alpha_cumprod
+
+ # 2. compute alphas, betas
+ alpha_prod_t = alphas_cumprod[timestep]
+ alpha_prod_t_prev = jnp.where(prev_timestep >= 0, alphas_cumprod[prev_timestep], final_alpha_cumprod)
+
+ beta_prod_t = 1 - alpha_prod_t
+
+ # 3. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ pred_epsilon = model_output
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction`"
+ )
+
+ # 4. Clip or threshold "predicted x_0"
+ if self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clip(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 4. compute variance: "sigma_t(η)" -> see formula (16)
+ # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
+ variance = self._get_variance(state, timestep, prev_timestep)
+ std_dev_t = eta * variance ** (0.5)
+
+ # 5. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
+
+ # 6. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+
+ if not return_dict:
+ return (prev_sample, state)
+
+ return FlaxDDIMSchedulerOutput(prev_sample=prev_sample, state=state)
+
+ def add_noise(
+ self,
+ state: DDIMSchedulerState,
+ original_samples: jnp.ndarray,
+ noise: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ) -> jnp.ndarray:
+ return add_noise_common(state.common, original_samples, noise, timesteps)
+
+ def get_velocity(
+ self,
+ state: DDIMSchedulerState,
+ sample: jnp.ndarray,
+ noise: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ) -> jnp.ndarray:
+ return get_velocity_common(state.common, sample, noise, timesteps)
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_ddim_inverse.py b/model/diffusers_c/schedulers/scheduling_ddim_inverse.py
new file mode 100644
index 0000000000000000000000000000000000000000..7f34ef75b9f2b815d144db77e72c48666f33b14a
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ddim_inverse.py
@@ -0,0 +1,376 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+from diffusers.utils import BaseOutput, deprecate
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->DDIM
+class DDIMSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class DDIMInverseScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `DDIMInverseScheduler` is the reverse scheduler of [`DDIMScheduler`].
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ clip_sample (`bool`, defaults to `True`):
+ Clip the predicted sample for numerical stability.
+ clip_sample_range (`float`, defaults to 1.0):
+ The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ set_alpha_to_one (`bool`, defaults to `True`):
+ Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+ there is no previous alpha. When this option is `True` the previous alpha product is fixed to 0, otherwise
+ it uses the alpha value at step `num_train_timesteps - 1`.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use `num_train_timesteps - 1` for the previous alpha
+ product.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ timestep_spacing (`str`, defaults to `"leading"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ order = 1
+ ignore_for_config = ["kwargs"]
+ _deprecated_kwargs = ["set_alpha_to_zero"]
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ clip_sample: bool = True,
+ set_alpha_to_one: bool = True,
+ steps_offset: int = 0,
+ prediction_type: str = "epsilon",
+ clip_sample_range: float = 1.0,
+ timestep_spacing: str = "leading",
+ rescale_betas_zero_snr: bool = False,
+ **kwargs,
+ ):
+ if kwargs.get("set_alpha_to_zero", None) is not None:
+ deprecation_message = (
+ "The `set_alpha_to_zero` argument is deprecated. Please use `set_alpha_to_one` instead."
+ )
+ deprecate("set_alpha_to_zero", "1.0.0", deprecation_message, standard_warn=False)
+ set_alpha_to_one = kwargs["set_alpha_to_zero"]
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ # Rescale for zero SNR
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # At every step in inverted ddim, we are looking into the next alphas_cumprod
+ # For the initial step, there is no current alphas_cumprod, and the index is out of bounds
+ # `set_alpha_to_one` decides whether we set this parameter simply to one
+ # in this case, self.step() just output the predicted noise
+ # or whether we use the initial alpha used in training the diffusion model.
+ self.initial_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps).copy().astype(np.int64))
+
+ # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.scale_model_input
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ """
+
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ self.num_inference_steps = num_inference_steps
+
+ # "leading" and "trailing" corresponds to annotation of Table 1. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round().copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)[::-1]).astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'."
+ )
+
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[DDIMSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ eta (`float`):
+ The weight of noise for added noise in diffusion step.
+ use_clipped_model_output (`bool`, defaults to `False`):
+ If `True`, computes "corrected" `model_output` from the clipped predicted original sample. Necessary
+ because predicted original sample is clipped to [-1, 1] when `self.config.clip_sample` is `True`. If no
+ clipping has happened, "corrected" `model_output` would coincide with the one provided as input and
+ `use_clipped_model_output` has no effect.
+ variance_noise (`torch.FloatTensor`):
+ Alternative to generating noise with `generator` by directly providing the noise for the variance
+ itself. Useful for methods such as [`CycleDiffusion`].
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] or
+ `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_ddim_inverse.DDIMInverseSchedulerOutput`] is
+ returned, otherwise a tuple is returned where the first element is the sample tensor.
+
+ """
+ # 1. get previous step value (=t+1)
+ prev_timestep = timestep
+ timestep = min(
+ timestep - self.config.num_train_timesteps // self.num_inference_steps, self.config.num_train_timesteps - 1
+ )
+
+ # 2. compute alphas, betas
+ # change original implementation to exactly match noise levels for analogous forward process
+ alpha_prod_t = self.alphas_cumprod[timestep] if timestep >= 0 else self.initial_alpha_cumprod
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep]
+
+ beta_prod_t = 1 - alpha_prod_t
+
+ # 3. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ pred_epsilon = model_output
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction`"
+ )
+
+ # 4. Clip or threshold "predicted x_0"
+ if self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 5. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ pred_sample_direction = (1 - alpha_prod_t_prev) ** (0.5) * pred_epsilon
+
+ # 6. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+
+ if not return_dict:
+ return (prev_sample, pred_original_sample)
+ return DDIMSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_ddim_parallel.py b/model/diffusers_c/schedulers/scheduling_ddim_parallel.py
new file mode 100644
index 0000000000000000000000000000000000000000..3acb095747b2f3989c24b4deb48bdb2ebe23f8fb
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ddim_parallel.py
@@ -0,0 +1,647 @@
+# Copyright 2024 ParaDiGMS authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput
+class DDIMParallelSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class DDIMParallelScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Denoising diffusion implicit models is a scheduler that extends the denoising procedure introduced in denoising
+ diffusion probabilistic models (DDPMs) with non-Markovian guidance.
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ For more details, see the original paper: https://arxiv.org/abs/2010.02502
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ beta_start (`float`): the starting `beta` value of inference.
+ beta_end (`float`): the final `beta` value.
+ beta_schedule (`str`):
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, optional):
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+ clip_sample (`bool`, default `True`):
+ option to clip predicted sample for numerical stability.
+ clip_sample_range (`float`, default `1.0`):
+ the maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ set_alpha_to_one (`bool`, default `True`):
+ each diffusion step uses the value of alphas product at that step and at the previous one. For the final
+ step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+ otherwise it uses the value of alpha at step 0.
+ steps_offset (`int`, default `0`):
+ an offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
+ stable diffusion.
+ prediction_type (`str`, default `epsilon`, optional):
+ prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion
+ process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4
+ https://imagen.research.google/video/paper.pdf)
+ thresholding (`bool`, default `False`):
+ whether to use the "dynamic thresholding" method (introduced by Imagen, https://arxiv.org/abs/2205.11487).
+ Note that the thresholding method is unsuitable for latent-space diffusion models (such as
+ stable-diffusion).
+ dynamic_thresholding_ratio (`float`, default `0.995`):
+ the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen
+ (https://arxiv.org/abs/2205.11487). Valid only when `thresholding=True`.
+ sample_max_value (`float`, default `1.0`):
+ the threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+ timestep_spacing (`str`, default `"leading"`):
+ The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample
+ Steps are Flawed](https://arxiv.org/abs/2305.08891) for more information.
+ rescale_betas_zero_snr (`bool`, default `False`):
+ whether to rescale the betas to have zero terminal SNR (proposed by https://arxiv.org/pdf/2305.08891.pdf).
+ This can enable the model to generate very bright and dark samples instead of limiting it to samples with
+ medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+ _is_ode_scheduler = True
+
+ @register_to_config
+ # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.__init__
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ clip_sample: bool = True,
+ set_alpha_to_one: bool = True,
+ steps_offset: int = 0,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ clip_sample_range: float = 1.0,
+ sample_max_value: float = 1.0,
+ timestep_spacing: str = "leading",
+ rescale_betas_zero_snr: bool = False,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ # Rescale for zero SNR
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # At every step in ddim, we are looking into the previous alphas_cumprod
+ # For the final step, there is no previous alphas_cumprod because we are already at 0
+ # `set_alpha_to_one` decides whether we set this parameter simply to one or
+ # whether we use the final alpha of the "non-previous" one.
+ self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
+
+ # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.scale_model_input
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def _get_variance(self, timestep, prev_timestep=None):
+ if prev_timestep is None:
+ prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
+
+ alpha_prod_t = self.alphas_cumprod[timestep]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+
+ return variance
+
+ def _batch_get_variance(self, t, prev_t):
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)]
+ alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0)
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+
+ return variance
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler.set_timesteps
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ """
+
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ self.num_inference_steps = num_inference_steps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
+ .round()[::-1]
+ .copy()
+ .astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'leading' or 'trailing'."
+ )
+
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ eta: float = 0.0,
+ use_clipped_model_output: bool = False,
+ generator=None,
+ variance_noise: Optional[torch.FloatTensor] = None,
+ return_dict: bool = True,
+ ) -> Union[DDIMParallelSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ current instance of sample being created by diffusion process.
+ eta (`float`): weight of noise for added noise in diffusion step.
+ use_clipped_model_output (`bool`): if `True`, compute "corrected" `model_output` from the clipped
+ predicted original sample. Necessary because predicted original sample is clipped to [-1, 1] when
+ `self.config.clip_sample` is `True`. If no clipping has happened, "corrected" `model_output` would
+ coincide with the one provided as input and `use_clipped_model_output` will have not effect.
+ generator: random number generator.
+ variance_noise (`torch.FloatTensor`): instead of generating noise for the variance using `generator`, we
+ can directly provide the noise for the variance itself. This is useful for methods such as
+ CycleDiffusion. (https://arxiv.org/abs/2210.05559)
+ return_dict (`bool`): option for returning tuple rather than DDIMParallelSchedulerOutput class
+
+ Returns:
+ [`~schedulers.scheduling_utils.DDIMParallelSchedulerOutput`] or `tuple`:
+ [`~schedulers.scheduling_utils.DDIMParallelSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`.
+ When returning a tuple, the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+ # Ideally, read DDIM paper in-detail understanding
+
+ # Notation ( ->
+ # - pred_noise_t -> e_theta(x_t, t)
+ # - pred_original_sample -> f_theta(x_t, t) or x_0
+ # - std_dev_t -> sigma_t
+ # - eta -> η
+ # - pred_sample_direction -> "direction pointing to x_t"
+ # - pred_prev_sample -> "x_t-1"
+
+ # 1. get previous step value (=t-1)
+ prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
+
+ # 2. compute alphas, betas
+ alpha_prod_t = self.alphas_cumprod[timestep]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+
+ beta_prod_t = 1 - alpha_prod_t
+
+ # 3. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ pred_epsilon = model_output
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction`"
+ )
+
+ # 4. Clip or threshold "predicted x_0"
+ if self.config.thresholding:
+ pred_original_sample = self._threshold_sample(pred_original_sample)
+ elif self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 5. compute variance: "sigma_t(η)" -> see formula (16)
+ # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
+ variance = self._get_variance(timestep, prev_timestep)
+ std_dev_t = eta * variance ** (0.5)
+
+ if use_clipped_model_output:
+ # the pred_epsilon is always re-derived from the clipped x_0 in Glide
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+
+ # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
+
+ # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+
+ if eta > 0:
+ if variance_noise is not None and generator is not None:
+ raise ValueError(
+ "Cannot pass both generator and variance_noise. Please make sure that either `generator` or"
+ " `variance_noise` stays `None`."
+ )
+
+ if variance_noise is None:
+ variance_noise = randn_tensor(
+ model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
+ )
+ variance = std_dev_t * variance_noise
+
+ prev_sample = prev_sample + variance
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return DDIMParallelSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+
+ def batch_step_no_noise(
+ self,
+ model_output: torch.FloatTensor,
+ timesteps: List[int],
+ sample: torch.FloatTensor,
+ eta: float = 0.0,
+ use_clipped_model_output: bool = False,
+ ) -> torch.FloatTensor:
+ """
+ Batched version of the `step` function, to be able to reverse the SDE for multiple samples/timesteps at once.
+ Also, does not add any noise to the predicted sample, which is necessary for parallel sampling where the noise
+ is pre-sampled by the pipeline.
+
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+ timesteps (`List[int]`):
+ current discrete timesteps in the diffusion chain. This is now a list of integers.
+ sample (`torch.FloatTensor`):
+ current instance of sample being created by diffusion process.
+ eta (`float`): weight of noise for added noise in diffusion step.
+ use_clipped_model_output (`bool`): if `True`, compute "corrected" `model_output` from the clipped
+ predicted original sample. Necessary because predicted original sample is clipped to [-1, 1] when
+ `self.config.clip_sample` is `True`. If no clipping has happened, "corrected" `model_output` would
+ coincide with the one provided as input and `use_clipped_model_output` will have not effect.
+
+ Returns:
+ `torch.FloatTensor`: sample tensor at previous timestep.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ assert eta == 0.0
+
+ # See formulas (12) and (16) of DDIM paper https://arxiv.org/pdf/2010.02502.pdf
+ # Ideally, read DDIM paper in-detail understanding
+
+ # Notation ( ->
+ # - pred_noise_t -> e_theta(x_t, t)
+ # - pred_original_sample -> f_theta(x_t, t) or x_0
+ # - std_dev_t -> sigma_t
+ # - eta -> η
+ # - pred_sample_direction -> "direction pointing to x_t"
+ # - pred_prev_sample -> "x_t-1"
+
+ # 1. get previous step value (=t-1)
+ t = timesteps
+ prev_t = t - self.config.num_train_timesteps // self.num_inference_steps
+
+ t = t.view(-1, *([1] * (model_output.ndim - 1)))
+ prev_t = prev_t.view(-1, *([1] * (model_output.ndim - 1)))
+
+ # 1. compute alphas, betas
+ self.alphas_cumprod = self.alphas_cumprod.to(model_output.device)
+ self.final_alpha_cumprod = self.final_alpha_cumprod.to(model_output.device)
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)]
+ alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0)
+
+ beta_prod_t = 1 - alpha_prod_t
+
+ # 3. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ pred_epsilon = model_output
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction`"
+ )
+
+ # 4. Clip or threshold "predicted x_0"
+ if self.config.thresholding:
+ pred_original_sample = self._threshold_sample(pred_original_sample)
+ elif self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 5. compute variance: "sigma_t(η)" -> see formula (16)
+ # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
+ variance = self._batch_get_variance(t, prev_t).to(model_output.device).view(*alpha_prod_t_prev.shape)
+ std_dev_t = eta * variance ** (0.5)
+
+ if use_clipped_model_output:
+ # the pred_epsilon is always re-derived from the clipped x_0 in Glide
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+
+ # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * pred_epsilon
+
+ # 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
+
+ return prev_sample
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+ # for the subsequent add_noise calls
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
+ def get_velocity(
+ self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as sample
+ self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype)
+ timesteps = timesteps.to(sample.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(sample.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+ return velocity
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_ddpm.py b/model/diffusers_c/schedulers/scheduling_ddpm.py
new file mode 100644
index 0000000000000000000000000000000000000000..0fb3561c8ce23025b984aa95b11fc439a27e3fee
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ddpm.py
@@ -0,0 +1,564 @@
+# Copyright 2024 UC Berkeley Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+@dataclass
+class DDPMSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class DDPMScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `DDPMScheduler` explores the connections between denoising score matching and Langevin dynamics sampling.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ An array of betas to pass directly to the constructor without using `beta_start` and `beta_end`.
+ variance_type (`str`, defaults to `"fixed_small"`):
+ Clip the variance when adding noise to the denoised sample. Choose from `fixed_small`, `fixed_small_log`,
+ `fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
+ clip_sample (`bool`, defaults to `True`):
+ Clip the predicted sample for numerical stability.
+ clip_sample_range (`float`, defaults to 1.0):
+ The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+ timestep_spacing (`str`, defaults to `"leading"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ variance_type: str = "fixed_small",
+ clip_sample: bool = True,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ clip_sample_range: float = 1.0,
+ sample_max_value: float = 1.0,
+ timestep_spacing: str = "leading",
+ steps_offset: int = 0,
+ rescale_betas_zero_snr: int = False,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ elif beta_schedule == "sigmoid":
+ # GeoDiff sigmoid schedule
+ betas = torch.linspace(-6, 6, num_train_timesteps)
+ self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ # Rescale for zero SNR
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ self.one = torch.tensor(1.0)
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.custom_timesteps = False
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy())
+
+ self.variance_type = variance_type
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: Optional[int] = None,
+ device: Union[str, torch.device] = None,
+ timesteps: Optional[List[int]] = None,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed,
+ `num_inference_steps` must be `None`.
+
+ """
+ if num_inference_steps is not None and timesteps is not None:
+ raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
+
+ if timesteps is not None:
+ for i in range(1, len(timesteps)):
+ if timesteps[i] >= timesteps[i - 1]:
+ raise ValueError("`custom_timesteps` must be in descending order.")
+
+ if timesteps[0] >= self.config.num_train_timesteps:
+ raise ValueError(
+ f"`timesteps` must start before `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps}."
+ )
+
+ timesteps = np.array(timesteps, dtype=np.int64)
+ self.custom_timesteps = True
+ else:
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ self.num_inference_steps = num_inference_steps
+ self.custom_timesteps = False
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
+ .round()[::-1]
+ .copy()
+ .astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ def _get_variance(self, t, predicted_variance=None, variance_type=None):
+ prev_t = self.previous_timestep(t)
+
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
+ current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev
+
+ # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
+ # and sample from it to get previous sample
+ # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
+ variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t
+
+ # we always take the log of variance, so clamp it to ensure it's not 0
+ variance = torch.clamp(variance, min=1e-20)
+
+ if variance_type is None:
+ variance_type = self.config.variance_type
+
+ # hacks - were probably added for training stability
+ if variance_type == "fixed_small":
+ variance = variance
+ # for rl-diffuser https://arxiv.org/abs/2205.09991
+ elif variance_type == "fixed_small_log":
+ variance = torch.log(variance)
+ variance = torch.exp(0.5 * variance)
+ elif variance_type == "fixed_large":
+ variance = current_beta_t
+ elif variance_type == "fixed_large_log":
+ # Glide max_log
+ variance = torch.log(current_beta_t)
+ elif variance_type == "learned":
+ return predicted_variance
+ elif variance_type == "learned_range":
+ min_log = torch.log(variance)
+ max_log = torch.log(current_beta_t)
+ frac = (predicted_variance + 1) / 2
+ variance = frac * max_log + (1 - frac) * min_log
+
+ return variance
+
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator=None,
+ return_dict: bool = True,
+ ) -> Union[DDPMSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_ddpm.DDPMSchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ t = timestep
+
+ prev_t = self.previous_timestep(t)
+
+ if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]:
+ model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1)
+ else:
+ predicted_variance = None
+
+ # 1. compute alphas, betas
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+ current_alpha_t = alpha_prod_t / alpha_prod_t_prev
+ current_beta_t = 1 - current_alpha_t
+
+ # 2. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
+ " `v_prediction` for the DDPMScheduler."
+ )
+
+ # 3. Clip or threshold "predicted x_0"
+ if self.config.thresholding:
+ pred_original_sample = self._threshold_sample(pred_original_sample)
+ elif self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t
+ current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
+
+ # 5. Compute predicted previous sample µ_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
+
+ # 6. Add noise
+ variance = 0
+ if t > 0:
+ device = model_output.device
+ variance_noise = randn_tensor(
+ model_output.shape, generator=generator, device=device, dtype=model_output.dtype
+ )
+ if self.variance_type == "fixed_small_log":
+ variance = self._get_variance(t, predicted_variance=predicted_variance) * variance_noise
+ elif self.variance_type == "learned_range":
+ variance = self._get_variance(t, predicted_variance=predicted_variance)
+ variance = torch.exp(0.5 * variance) * variance_noise
+ else:
+ variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * variance_noise
+
+ pred_prev_sample = pred_prev_sample + variance
+
+ if not return_dict:
+ return (pred_prev_sample,)
+
+ return DDPMSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample)
+
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+ # for the subsequent add_noise calls
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+ def get_velocity(
+ self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as sample
+ self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype)
+ timesteps = timesteps.to(sample.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(sample.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+ return velocity
+
+ def __len__(self):
+ return self.config.num_train_timesteps
+
+ def previous_timestep(self, timestep):
+ if self.custom_timesteps:
+ index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0]
+ if index == self.timesteps.shape[0] - 1:
+ prev_t = torch.tensor(-1)
+ else:
+ prev_t = self.timesteps[index + 1]
+ else:
+ num_inference_steps = (
+ self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps
+ )
+ prev_t = timestep - self.config.num_train_timesteps // num_inference_steps
+
+ return prev_t
diff --git a/model/diffusers_c/schedulers/scheduling_ddpm_flax.py b/model/diffusers_c/schedulers/scheduling_ddpm_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..6bdfa5eb5e0d4eae5f0b172252f99c9f2f48caab
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ddpm_flax.py
@@ -0,0 +1,299 @@
+# Copyright 2024 UC Berkeley Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import flax
+import jax
+import jax.numpy as jnp
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils_flax import (
+ CommonSchedulerState,
+ FlaxKarrasDiffusionSchedulers,
+ FlaxSchedulerMixin,
+ FlaxSchedulerOutput,
+ add_noise_common,
+ get_velocity_common,
+)
+
+
+@flax.struct.dataclass
+class DDPMSchedulerState:
+ common: CommonSchedulerState
+
+ # setable values
+ init_noise_sigma: jnp.ndarray
+ timesteps: jnp.ndarray
+ num_inference_steps: Optional[int] = None
+
+ @classmethod
+ def create(cls, common: CommonSchedulerState, init_noise_sigma: jnp.ndarray, timesteps: jnp.ndarray):
+ return cls(common=common, init_noise_sigma=init_noise_sigma, timesteps=timesteps)
+
+
+@dataclass
+class FlaxDDPMSchedulerOutput(FlaxSchedulerOutput):
+ state: DDPMSchedulerState
+
+
+class FlaxDDPMScheduler(FlaxSchedulerMixin, ConfigMixin):
+ """
+ Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and
+ Langevin dynamics sampling.
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ For more details, see the original paper: https://arxiv.org/abs/2006.11239
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ beta_start (`float`): the starting `beta` value of inference.
+ beta_end (`float`): the final `beta` value.
+ beta_schedule (`str`):
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, optional):
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+ variance_type (`str`):
+ options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`,
+ `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
+ clip_sample (`bool`, default `True`):
+ option to clip predicted sample between -1 and 1 for numerical stability.
+ prediction_type (`str`, default `epsilon`):
+ indicates whether the model predicts the noise (epsilon), or the samples. One of `epsilon`, `sample`.
+ `v-prediction` is not supported for this scheduler.
+ dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`):
+ the `dtype` used for params and computation.
+ """
+
+ _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers]
+
+ dtype: jnp.dtype
+
+ @property
+ def has_state(self):
+ return True
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[jnp.ndarray] = None,
+ variance_type: str = "fixed_small",
+ clip_sample: bool = True,
+ prediction_type: str = "epsilon",
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ self.dtype = dtype
+
+ def create_state(self, common: Optional[CommonSchedulerState] = None) -> DDPMSchedulerState:
+ if common is None:
+ common = CommonSchedulerState.create(self)
+
+ # standard deviation of the initial noise distribution
+ init_noise_sigma = jnp.array(1.0, dtype=self.dtype)
+
+ timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1]
+
+ return DDPMSchedulerState.create(
+ common=common,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ )
+
+ def scale_model_input(
+ self, state: DDPMSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None
+ ) -> jnp.ndarray:
+ """
+ Args:
+ state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance.
+ sample (`jnp.ndarray`): input sample
+ timestep (`int`, optional): current timestep
+
+ Returns:
+ `jnp.ndarray`: scaled input sample
+ """
+ return sample
+
+ def set_timesteps(
+ self, state: DDPMSchedulerState, num_inference_steps: int, shape: Tuple = ()
+ ) -> DDPMSchedulerState:
+ """
+ Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ state (`DDIMSchedulerState`):
+ the `FlaxDDPMScheduler` state data class instance.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ """
+
+ step_ratio = self.config.num_train_timesteps // num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # rounding to avoid issues when num_inference_step is power of 3
+ timesteps = (jnp.arange(0, num_inference_steps) * step_ratio).round()[::-1]
+
+ return state.replace(
+ num_inference_steps=num_inference_steps,
+ timesteps=timesteps,
+ )
+
+ def _get_variance(self, state: DDPMSchedulerState, t, predicted_variance=None, variance_type=None):
+ alpha_prod_t = state.common.alphas_cumprod[t]
+ alpha_prod_t_prev = jnp.where(t > 0, state.common.alphas_cumprod[t - 1], jnp.array(1.0, dtype=self.dtype))
+
+ # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
+ # and sample from it to get previous sample
+ # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
+ variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * state.common.betas[t]
+
+ if variance_type is None:
+ variance_type = self.config.variance_type
+
+ # hacks - were probably added for training stability
+ if variance_type == "fixed_small":
+ variance = jnp.clip(variance, a_min=1e-20)
+ # for rl-diffuser https://arxiv.org/abs/2205.09991
+ elif variance_type == "fixed_small_log":
+ variance = jnp.log(jnp.clip(variance, a_min=1e-20))
+ elif variance_type == "fixed_large":
+ variance = state.common.betas[t]
+ elif variance_type == "fixed_large_log":
+ # Glide max_log
+ variance = jnp.log(state.common.betas[t])
+ elif variance_type == "learned":
+ return predicted_variance
+ elif variance_type == "learned_range":
+ min_log = variance
+ max_log = state.common.betas[t]
+ frac = (predicted_variance + 1) / 2
+ variance = frac * max_log + (1 - frac) * min_log
+
+ return variance
+
+ def step(
+ self,
+ state: DDPMSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ key: Optional[jax.Array] = None,
+ return_dict: bool = True,
+ ) -> Union[FlaxDDPMSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ state (`DDPMSchedulerState`): the `FlaxDDPMScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ key (`jax.Array`): a PRNG key.
+ return_dict (`bool`): option for returning tuple rather than FlaxDDPMSchedulerOutput class
+
+ Returns:
+ [`FlaxDDPMSchedulerOutput`] or `tuple`: [`FlaxDDPMSchedulerOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+ t = timestep
+
+ if key is None:
+ key = jax.random.PRNGKey(0)
+
+ if model_output.shape[1] == sample.shape[1] * 2 and self.config.variance_type in ["learned", "learned_range"]:
+ model_output, predicted_variance = jnp.split(model_output, sample.shape[1], axis=1)
+ else:
+ predicted_variance = None
+
+ # 1. compute alphas, betas
+ alpha_prod_t = state.common.alphas_cumprod[t]
+ alpha_prod_t_prev = jnp.where(t > 0, state.common.alphas_cumprod[t - 1], jnp.array(1.0, dtype=self.dtype))
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ # 2. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` "
+ " for the FlaxDDPMScheduler."
+ )
+
+ # 3. Clip "predicted x_0"
+ if self.config.clip_sample:
+ pred_original_sample = jnp.clip(pred_original_sample, -1, 1)
+
+ # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * state.common.betas[t]) / beta_prod_t
+ current_sample_coeff = state.common.alphas[t] ** (0.5) * beta_prod_t_prev / beta_prod_t
+
+ # 5. Compute predicted previous sample µ_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
+
+ # 6. Add noise
+ def random_variance():
+ split_key = jax.random.split(key, num=1)
+ noise = jax.random.normal(split_key, shape=model_output.shape, dtype=self.dtype)
+ return (self._get_variance(state, t, predicted_variance=predicted_variance) ** 0.5) * noise
+
+ variance = jnp.where(t > 0, random_variance(), jnp.zeros(model_output.shape, dtype=self.dtype))
+
+ pred_prev_sample = pred_prev_sample + variance
+
+ if not return_dict:
+ return (pred_prev_sample, state)
+
+ return FlaxDDPMSchedulerOutput(prev_sample=pred_prev_sample, state=state)
+
+ def add_noise(
+ self,
+ state: DDPMSchedulerState,
+ original_samples: jnp.ndarray,
+ noise: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ) -> jnp.ndarray:
+ return add_noise_common(state.common, original_samples, noise, timesteps)
+
+ def get_velocity(
+ self,
+ state: DDPMSchedulerState,
+ sample: jnp.ndarray,
+ noise: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ) -> jnp.ndarray:
+ return get_velocity_common(state.common, sample, noise, timesteps)
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_ddpm_parallel.py b/model/diffusers_c/schedulers/scheduling_ddpm_parallel.py
new file mode 100644
index 0000000000000000000000000000000000000000..0ddfec4f865b6a7de04e3eba410b05ef4b39d643
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ddpm_parallel.py
@@ -0,0 +1,655 @@
+# Copyright 2024 ParaDiGMS authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput
+class DDPMParallelSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class DDPMParallelScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and
+ Langevin dynamics sampling.
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ For more details, see the original paper: https://arxiv.org/abs/2006.11239
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ beta_start (`float`): the starting `beta` value of inference.
+ beta_end (`float`): the final `beta` value.
+ beta_schedule (`str`):
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, `squaredcos_cap_v2` or `sigmoid`.
+ trained_betas (`np.ndarray`, optional):
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+ variance_type (`str`):
+ options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small`,
+ `fixed_small_log`, `fixed_large`, `fixed_large_log`, `learned` or `learned_range`.
+ clip_sample (`bool`, default `True`):
+ option to clip predicted sample for numerical stability.
+ clip_sample_range (`float`, default `1.0`):
+ the maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ prediction_type (`str`, default `epsilon`, optional):
+ prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion
+ process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4
+ https://imagen.research.google/video/paper.pdf)
+ thresholding (`bool`, default `False`):
+ whether to use the "dynamic thresholding" method (introduced by Imagen, https://arxiv.org/abs/2205.11487).
+ Note that the thresholding method is unsuitable for latent-space diffusion models (such as
+ stable-diffusion).
+ dynamic_thresholding_ratio (`float`, default `0.995`):
+ the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen
+ (https://arxiv.org/abs/2205.11487). Valid only when `thresholding=True`.
+ sample_max_value (`float`, default `1.0`):
+ the threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+ timestep_spacing (`str`, default `"leading"`):
+ The way the timesteps should be scaled. Refer to Table 2. of [Common Diffusion Noise Schedules and Sample
+ Steps are Flawed](https://arxiv.org/abs/2305.08891) for more information.
+ steps_offset (`int`, default `0`):
+ an offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
+ stable diffusion.
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+ _is_ode_scheduler = False
+
+ @register_to_config
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.__init__
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ variance_type: str = "fixed_small",
+ clip_sample: bool = True,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ clip_sample_range: float = 1.0,
+ sample_max_value: float = 1.0,
+ timestep_spacing: str = "leading",
+ steps_offset: int = 0,
+ rescale_betas_zero_snr: int = False,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ elif beta_schedule == "sigmoid":
+ # GeoDiff sigmoid schedule
+ betas = torch.linspace(-6, 6, num_train_timesteps)
+ self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ # Rescale for zero SNR
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ self.one = torch.tensor(1.0)
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.custom_timesteps = False
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy())
+
+ self.variance_type = variance_type
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.scale_model_input
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.set_timesteps
+ def set_timesteps(
+ self,
+ num_inference_steps: Optional[int] = None,
+ device: Union[str, torch.device] = None,
+ timesteps: Optional[List[int]] = None,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of equal spacing between timesteps is used. If `timesteps` is passed,
+ `num_inference_steps` must be `None`.
+
+ """
+ if num_inference_steps is not None and timesteps is not None:
+ raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
+
+ if timesteps is not None:
+ for i in range(1, len(timesteps)):
+ if timesteps[i] >= timesteps[i - 1]:
+ raise ValueError("`custom_timesteps` must be in descending order.")
+
+ if timesteps[0] >= self.config.num_train_timesteps:
+ raise ValueError(
+ f"`timesteps` must start before `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps}."
+ )
+
+ timesteps = np.array(timesteps, dtype=np.int64)
+ self.custom_timesteps = True
+ else:
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ self.num_inference_steps = num_inference_steps
+ self.custom_timesteps = False
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps)
+ .round()[::-1]
+ .copy()
+ .astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio)).astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._get_variance
+ def _get_variance(self, t, predicted_variance=None, variance_type=None):
+ prev_t = self.previous_timestep(t)
+
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
+ current_beta_t = 1 - alpha_prod_t / alpha_prod_t_prev
+
+ # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
+ # and sample from it to get previous sample
+ # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
+ variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * current_beta_t
+
+ # we always take the log of variance, so clamp it to ensure it's not 0
+ variance = torch.clamp(variance, min=1e-20)
+
+ if variance_type is None:
+ variance_type = self.config.variance_type
+
+ # hacks - were probably added for training stability
+ if variance_type == "fixed_small":
+ variance = variance
+ # for rl-diffuser https://arxiv.org/abs/2205.09991
+ elif variance_type == "fixed_small_log":
+ variance = torch.log(variance)
+ variance = torch.exp(0.5 * variance)
+ elif variance_type == "fixed_large":
+ variance = current_beta_t
+ elif variance_type == "fixed_large_log":
+ # Glide max_log
+ variance = torch.log(current_beta_t)
+ elif variance_type == "learned":
+ return predicted_variance
+ elif variance_type == "learned_range":
+ min_log = torch.log(variance)
+ max_log = torch.log(current_beta_t)
+ frac = (predicted_variance + 1) / 2
+ variance = frac * max_log + (1 - frac) * min_log
+
+ return variance
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator=None,
+ return_dict: bool = True,
+ ) -> Union[DDPMParallelSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ current instance of sample being created by diffusion process.
+ generator: random number generator.
+ return_dict (`bool`): option for returning tuple rather than DDPMParallelSchedulerOutput class
+
+ Returns:
+ [`~schedulers.scheduling_utils.DDPMParallelSchedulerOutput`] or `tuple`:
+ [`~schedulers.scheduling_utils.DDPMParallelSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`.
+ When returning a tuple, the first element is the sample tensor.
+
+ """
+ t = timestep
+
+ prev_t = self.previous_timestep(t)
+
+ if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]:
+ model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1)
+ else:
+ predicted_variance = None
+
+ # 1. compute alphas, betas
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_t] if prev_t >= 0 else self.one
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+ current_alpha_t = alpha_prod_t / alpha_prod_t_prev
+ current_beta_t = 1 - current_alpha_t
+
+ # 2. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
+ " `v_prediction` for the DDPMScheduler."
+ )
+
+ # 3. Clip or threshold "predicted x_0"
+ if self.config.thresholding:
+ pred_original_sample = self._threshold_sample(pred_original_sample)
+ elif self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t
+ current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
+
+ # 5. Compute predicted previous sample µ_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
+
+ # 6. Add noise
+ variance = 0
+ if t > 0:
+ device = model_output.device
+ variance_noise = randn_tensor(
+ model_output.shape, generator=generator, device=device, dtype=model_output.dtype
+ )
+ if self.variance_type == "fixed_small_log":
+ variance = self._get_variance(t, predicted_variance=predicted_variance) * variance_noise
+ elif self.variance_type == "learned_range":
+ variance = self._get_variance(t, predicted_variance=predicted_variance)
+ variance = torch.exp(0.5 * variance) * variance_noise
+ else:
+ variance = (self._get_variance(t, predicted_variance=predicted_variance) ** 0.5) * variance_noise
+
+ pred_prev_sample = pred_prev_sample + variance
+
+ if not return_dict:
+ return (pred_prev_sample,)
+
+ return DDPMParallelSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample)
+
+ def batch_step_no_noise(
+ self,
+ model_output: torch.FloatTensor,
+ timesteps: List[int],
+ sample: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ """
+ Batched version of the `step` function, to be able to reverse the SDE for multiple samples/timesteps at once.
+ Also, does not add any noise to the predicted sample, which is necessary for parallel sampling where the noise
+ is pre-sampled by the pipeline.
+
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+ timesteps (`List[int]`):
+ current discrete timesteps in the diffusion chain. This is now a list of integers.
+ sample (`torch.FloatTensor`):
+ current instance of sample being created by diffusion process.
+
+ Returns:
+ `torch.FloatTensor`: sample tensor at previous timestep.
+ """
+ t = timesteps
+ num_inference_steps = self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps
+ prev_t = t - self.config.num_train_timesteps // num_inference_steps
+
+ t = t.view(-1, *([1] * (model_output.ndim - 1)))
+ prev_t = prev_t.view(-1, *([1] * (model_output.ndim - 1)))
+
+ if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type in ["learned", "learned_range"]:
+ model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1)
+ else:
+ pass
+
+ # 1. compute alphas, betas
+ self.alphas_cumprod = self.alphas_cumprod.to(model_output.device)
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[torch.clip(prev_t, min=0)]
+ alpha_prod_t_prev[prev_t < 0] = torch.tensor(1.0)
+
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+ current_alpha_t = alpha_prod_t / alpha_prod_t_prev
+ current_beta_t = 1 - current_alpha_t
+
+ # 2. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ elif self.config.prediction_type == "v_prediction":
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
+ " `v_prediction` for the DDPMParallelScheduler."
+ )
+
+ # 3. Clip or threshold "predicted x_0"
+ if self.config.thresholding:
+ pred_original_sample = self._threshold_sample(pred_original_sample)
+ elif self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * current_beta_t) / beta_prod_t
+ current_sample_coeff = current_alpha_t ** (0.5) * beta_prod_t_prev / beta_prod_t
+
+ # 5. Compute predicted previous sample µ_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
+
+ return pred_prev_sample
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+ # for the subsequent add_noise calls
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
+ def get_velocity(
+ self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as sample
+ self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype)
+ timesteps = timesteps.to(sample.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(sample.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+ return velocity
+
+ def __len__(self):
+ return self.config.num_train_timesteps
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep
+ def previous_timestep(self, timestep):
+ if self.custom_timesteps:
+ index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0]
+ if index == self.timesteps.shape[0] - 1:
+ prev_t = torch.tensor(-1)
+ else:
+ prev_t = self.timesteps[index + 1]
+ else:
+ num_inference_steps = (
+ self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps
+ )
+ prev_t = timestep - self.config.num_train_timesteps // num_inference_steps
+
+ return prev_t
diff --git a/model/diffusers_c/schedulers/scheduling_ddpm_wuerstchen.py b/model/diffusers_c/schedulers/scheduling_ddpm_wuerstchen.py
new file mode 100644
index 0000000000000000000000000000000000000000..ad4e4f414ef9db3d9a898ee4cebd1b60dfbcc5ca
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ddpm_wuerstchen.py
@@ -0,0 +1,230 @@
+# Copyright (c) 2022 Pablo Pernías MIT License
+# Copyright 2024 UC Berkeley Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin
+
+
+@dataclass
+class DDPMWuerstchenSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's step function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ """
+
+ prev_sample: torch.FloatTensor
+
+
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class DDPMWuerstchenScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Denoising diffusion probabilistic models (DDPMs) explores the connections between denoising score matching and
+ Langevin dynamics sampling.
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ For more details, see the original paper: https://arxiv.org/abs/2006.11239
+
+ Args:
+ scaler (`float`): ....
+ s (`float`): ....
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ scaler: float = 1.0,
+ s: float = 0.008,
+ ):
+ self.scaler = scaler
+ self.s = torch.tensor([s])
+ self._init_alpha_cumprod = torch.cos(self.s / (1 + self.s) * torch.pi * 0.5) ** 2
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ def _alpha_cumprod(self, t, device):
+ if self.scaler > 1:
+ t = 1 - (1 - t) ** self.scaler
+ elif self.scaler < 1:
+ t = t**self.scaler
+ alpha_cumprod = torch.cos(
+ (t + self.s.to(device)) / (1 + self.s.to(device)) * torch.pi * 0.5
+ ) ** 2 / self._init_alpha_cumprod.to(device)
+ return alpha_cumprod.clamp(0.0001, 0.9999)
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`): input sample
+ timestep (`int`, optional): current timestep
+
+ Returns:
+ `torch.FloatTensor`: scaled input sample
+ """
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: int = None,
+ timesteps: Optional[List[int]] = None,
+ device: Union[str, torch.device] = None,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ num_inference_steps (`Dict[float, int]`):
+ the number of diffusion steps used when generating samples with a pre-trained model. If passed, then
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, optional):
+ the device to which the timesteps are moved to. {2 / 3: 20, 0.0: 10}
+ """
+ if timesteps is None:
+ timesteps = torch.linspace(1.0, 0.0, num_inference_steps + 1, device=device)
+ if not isinstance(timesteps, torch.Tensor):
+ timesteps = torch.Tensor(timesteps).to(device)
+ self.timesteps = timesteps
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator=None,
+ return_dict: bool = True,
+ ) -> Union[DDPMWuerstchenSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ current instance of sample being created by diffusion process.
+ generator: random number generator.
+ return_dict (`bool`): option for returning tuple rather than DDPMWuerstchenSchedulerOutput class
+
+ Returns:
+ [`DDPMWuerstchenSchedulerOutput`] or `tuple`: [`DDPMWuerstchenSchedulerOutput`] if `return_dict` is True,
+ otherwise a `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+ dtype = model_output.dtype
+ device = model_output.device
+ t = timestep
+
+ prev_t = self.previous_timestep(t)
+
+ alpha_cumprod = self._alpha_cumprod(t, device).view(t.size(0), *[1 for _ in sample.shape[1:]])
+ alpha_cumprod_prev = self._alpha_cumprod(prev_t, device).view(prev_t.size(0), *[1 for _ in sample.shape[1:]])
+ alpha = alpha_cumprod / alpha_cumprod_prev
+
+ mu = (1.0 / alpha).sqrt() * (sample - (1 - alpha) * model_output / (1 - alpha_cumprod).sqrt())
+
+ std_noise = randn_tensor(mu.shape, generator=generator, device=model_output.device, dtype=model_output.dtype)
+ std = ((1 - alpha) * (1.0 - alpha_cumprod_prev) / (1.0 - alpha_cumprod)).sqrt() * std_noise
+ pred = mu + std * (prev_t != 0).float().view(prev_t.size(0), *[1 for _ in sample.shape[1:]])
+
+ if not return_dict:
+ return (pred.to(dtype),)
+
+ return DDPMWuerstchenSchedulerOutput(prev_sample=pred.to(dtype))
+
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ device = original_samples.device
+ dtype = original_samples.dtype
+ alpha_cumprod = self._alpha_cumprod(timesteps, device=device).view(
+ timesteps.size(0), *[1 for _ in original_samples.shape[1:]]
+ )
+ noisy_samples = alpha_cumprod.sqrt() * original_samples + (1 - alpha_cumprod).sqrt() * noise
+ return noisy_samples.to(dtype=dtype)
+
+ def __len__(self):
+ return self.config.num_train_timesteps
+
+ def previous_timestep(self, timestep):
+ index = (self.timesteps - timestep[0]).abs().argmin().item()
+ prev_t = self.timesteps[index + 1][None].expand(timestep.shape[0])
+ return prev_t
diff --git a/model/diffusers_c/schedulers/scheduling_deis_multistep.py b/model/diffusers_c/schedulers/scheduling_deis_multistep.py
new file mode 100644
index 0000000000000000000000000000000000000000..ee6db2c821a850044fd97dcdd91eaf372ced7278
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_deis_multistep.py
@@ -0,0 +1,788 @@
+# Copyright 2024 FLAIR Lab and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: check https://arxiv.org/abs/2204.13902 and https://github.com/qsh-zh/deis for more info
+# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import deprecate
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class DEISMultistepScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `DEISMultistepScheduler` is a fast high order solver for diffusion ordinary differential equations (ODEs).
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ solver_order (`int`, defaults to 2):
+ The DEIS order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
+ sampling, and `solver_order=3` for unconditional sampling.
+ prediction_type (`str`, defaults to `epsilon`):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+ algorithm_type (`str`, defaults to `deis`):
+ The algorithm type for the solver.
+ lower_order_final (`bool`, defaults to `True`):
+ Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[np.ndarray] = None,
+ solver_order: int = 2,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ algorithm_type: str = "deis",
+ solver_type: str = "logrho",
+ lower_order_final: bool = True,
+ use_karras_sigmas: Optional[bool] = False,
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ # Currently we only support VP-type noise schedule
+ self.alpha_t = torch.sqrt(self.alphas_cumprod)
+ self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
+ self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
+ self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # settings for DEIS
+ if algorithm_type not in ["deis"]:
+ if algorithm_type in ["dpmsolver", "dpmsolver++"]:
+ self.register_to_config(algorithm_type="deis")
+ else:
+ raise NotImplementedError(f"{algorithm_type} does is not implemented for {self.__class__}")
+
+ if solver_type not in ["logrho"]:
+ if solver_type in ["midpoint", "heun", "bh1", "bh2"]:
+ self.register_to_config(solver_type="logrho")
+ else:
+ raise NotImplementedError(f"solver type {solver_type} does is not implemented for {self.__class__}")
+
+ # setable values
+ self.num_inference_steps = None
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
+ self.timesteps = torch.from_numpy(timesteps)
+ self.model_outputs = [None] * solver_order
+ self.lower_order_nums = 0
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1)
+ .round()[::-1][:-1]
+ .copy()
+ .astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // (num_inference_steps + 1)
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.arange(self.config.num_train_timesteps, 0, -step_ratio).round().copy().astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ if self.config.use_karras_sigmas:
+ log_sigmas = np.log(sigmas)
+ sigmas = np.flip(sigmas).copy()
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+ sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
+ else:
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+ sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
+ sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
+
+ self.sigmas = torch.from_numpy(sigmas)
+ self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
+
+ self.num_inference_steps = len(timesteps)
+
+ self.model_outputs = [
+ None,
+ ] * self.config.solver_order
+ self.lower_order_nums = 0
+
+ # add an index counter for schedulers that allow duplicated timesteps
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
+ def _sigma_to_alpha_sigma_t(self, sigma):
+ alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+ sigma_t = sigma * alpha_t
+
+ return alpha_t, sigma_t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def convert_model_output(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ Convert the model output to the corresponding type the DEIS algorithm needs.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The converted model output.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ if sample is None:
+ if len(args) > 1:
+ sample = args[1]
+ else:
+ raise ValueError("missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ if self.config.prediction_type == "epsilon":
+ x0_pred = (sample - sigma_t * model_output) / alpha_t
+ elif self.config.prediction_type == "sample":
+ x0_pred = model_output
+ elif self.config.prediction_type == "v_prediction":
+ x0_pred = alpha_t * sample - sigma_t * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the DEISMultistepScheduler."
+ )
+
+ if self.config.thresholding:
+ x0_pred = self._threshold_sample(x0_pred)
+
+ if self.config.algorithm_type == "deis":
+ return (sample - alpha_t * x0_pred) / sigma_t
+ else:
+ raise NotImplementedError("only support log-rho multistep deis now")
+
+ def deis_first_order_update(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the first-order DEIS (equivalent to DDIM).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ prev_timestep (`int`):
+ The previous discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s = torch.log(alpha_s) - torch.log(sigma_s)
+
+ h = lambda_t - lambda_s
+ if self.config.algorithm_type == "deis":
+ x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
+ else:
+ raise NotImplementedError("only support log-rho multistep deis now")
+ return x_t
+
+ def multistep_deis_second_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the second-order multistep DEIS.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s0, sigma_s1 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+
+ m0, m1 = model_output_list[-1], model_output_list[-2]
+
+ rho_t, rho_s0, rho_s1 = sigma_t / alpha_t, sigma_s0 / alpha_s0, sigma_s1 / alpha_s1
+
+ if self.config.algorithm_type == "deis":
+
+ def ind_fn(t, b, c):
+ # Integrate[(log(t) - log(c)) / (log(b) - log(c)), {t}]
+ return t * (-np.log(c) + np.log(t) - 1) / (np.log(b) - np.log(c))
+
+ coef1 = ind_fn(rho_t, rho_s0, rho_s1) - ind_fn(rho_s0, rho_s0, rho_s1)
+ coef2 = ind_fn(rho_t, rho_s1, rho_s0) - ind_fn(rho_s0, rho_s1, rho_s0)
+
+ x_t = alpha_t * (sample / alpha_s0 + coef1 * m0 + coef2 * m1)
+ return x_t
+ else:
+ raise NotImplementedError("only support log-rho multistep deis now")
+
+ def multistep_deis_third_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the third-order multistep DEIS.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing`sample` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ self.sigmas[self.step_index - 2],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+ alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)
+
+ m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
+
+ rho_t, rho_s0, rho_s1, rho_s2 = (
+ sigma_t / alpha_t,
+ sigma_s0 / alpha_s0,
+ sigma_s1 / alpha_s1,
+ sigma_s2 / alpha_s2,
+ )
+
+ if self.config.algorithm_type == "deis":
+
+ def ind_fn(t, b, c, d):
+ # Integrate[(log(t) - log(c))(log(t) - log(d)) / (log(b) - log(c))(log(b) - log(d)), {t}]
+ numerator = t * (
+ np.log(c) * (np.log(d) - np.log(t) + 1)
+ - np.log(d) * np.log(t)
+ + np.log(d)
+ + np.log(t) ** 2
+ - 2 * np.log(t)
+ + 2
+ )
+ denominator = (np.log(b) - np.log(c)) * (np.log(b) - np.log(d))
+ return numerator / denominator
+
+ coef1 = ind_fn(rho_t, rho_s0, rho_s1, rho_s2) - ind_fn(rho_s0, rho_s0, rho_s1, rho_s2)
+ coef2 = ind_fn(rho_t, rho_s1, rho_s2, rho_s0) - ind_fn(rho_s0, rho_s1, rho_s2, rho_s0)
+ coef3 = ind_fn(rho_t, rho_s2, rho_s0, rho_s1) - ind_fn(rho_s0, rho_s2, rho_s0, rho_s1)
+
+ x_t = alpha_t * (sample / alpha_s0 + coef1 * m0 + coef2 * m1 + coef3 * m2)
+
+ return x_t
+ else:
+ raise NotImplementedError("only support log-rho multistep deis now")
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ index_candidates = (schedule_timesteps == timestep).nonzero()
+
+ if len(index_candidates) == 0:
+ step_index = len(self.timesteps) - 1
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ elif len(index_candidates) > 1:
+ step_index = index_candidates[1].item()
+ else:
+ step_index = index_candidates[0].item()
+
+ return step_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ """
+ Initialize the step_index counter for the scheduler.
+ """
+
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the multistep DEIS.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ lower_order_final = (
+ (self.step_index == len(self.timesteps) - 1) and self.config.lower_order_final and len(self.timesteps) < 15
+ )
+ lower_order_second = (
+ (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
+ )
+
+ model_output = self.convert_model_output(model_output, sample=sample)
+ for i in range(self.config.solver_order - 1):
+ self.model_outputs[i] = self.model_outputs[i + 1]
+ self.model_outputs[-1] = model_output
+
+ if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
+ prev_sample = self.deis_first_order_update(model_output, sample=sample)
+ elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
+ prev_sample = self.multistep_deis_second_order_update(self.model_outputs, sample=sample)
+ else:
+ prev_sample = self.multistep_deis_third_order_update(self.model_outputs, sample=sample)
+
+ if self.lower_order_nums < self.config.solver_order:
+ self.lower_order_nums += 1
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # begin_index is None when the scheduler is used for training
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ noisy_samples = alpha_t * original_samples + sigma_t * noise
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep.py b/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep.py
new file mode 100644
index 0000000000000000000000000000000000000000..5c6d03fa229ceaa98fc3ee19e1c022737987c82f
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep.py
@@ -0,0 +1,1026 @@
+# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import deprecate
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class DPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `DPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ solver_order (`int`, defaults to 2):
+ The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
+ sampling, and `solver_order=3` for unconditional sampling.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
+ `algorithm_type="dpmsolver++"`.
+ algorithm_type (`str`, defaults to `dpmsolver++`):
+ Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, `sde-dpmsolver` or `sde-dpmsolver++`. The
+ `dpmsolver` type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927)
+ paper, and the `dpmsolver++` type implements the algorithms in the
+ [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or
+ `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion.
+ solver_type (`str`, defaults to `midpoint`):
+ Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
+ sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
+ lower_order_final (`bool`, defaults to `True`):
+ Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
+ stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
+ euler_at_final (`bool`, defaults to `False`):
+ Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail
+ richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
+ steps, but sometimes may result in blurring.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ use_lu_lambdas (`bool`, *optional*, defaults to `False`):
+ Whether to use the uniform-logSNR for step sizes proposed by Lu's DPM-Solver in the noise schedule during
+ the sampling process. If `True`, the sigmas and time steps are determined according to a sequence of
+ `lambda(t)`.
+ final_sigmas_type (`str`, defaults to `"zero"`):
+ The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final sigma
+ is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
+ lambda_min_clipped (`float`, defaults to `-inf`):
+ Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
+ cosine (`squaredcos_cap_v2`) noise schedule.
+ variance_type (`str`, *optional*):
+ Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
+ contains the predicted Gaussian variance.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ solver_order: int = 2,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ algorithm_type: str = "dpmsolver++",
+ solver_type: str = "midpoint",
+ lower_order_final: bool = True,
+ euler_at_final: bool = False,
+ use_karras_sigmas: Optional[bool] = False,
+ use_lu_lambdas: Optional[bool] = False,
+ final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min"
+ lambda_min_clipped: float = -float("inf"),
+ variance_type: Optional[str] = None,
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ rescale_betas_zero_snr: bool = False,
+ ):
+ if algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
+ deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
+ deprecate("algorithm_types dpmsolver and sde-dpmsolver", "1.0.0", deprecation_message)
+
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ if rescale_betas_zero_snr:
+ # Close to 0 without being 0 so first sigma is not inf
+ # FP16 smallest positive subnormal works well here
+ self.alphas_cumprod[-1] = 2**-24
+
+ # Currently we only support VP-type noise schedule
+ self.alpha_t = torch.sqrt(self.alphas_cumprod)
+ self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
+ self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
+ self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # settings for DPM-Solver
+ if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"]:
+ if algorithm_type == "deis":
+ self.register_to_config(algorithm_type="dpmsolver++")
+ else:
+ raise NotImplementedError(f"{algorithm_type} does is not implemented for {self.__class__}")
+
+ if solver_type not in ["midpoint", "heun"]:
+ if solver_type in ["logrho", "bh1", "bh2"]:
+ self.register_to_config(solver_type="midpoint")
+ else:
+ raise NotImplementedError(f"{solver_type} does is not implemented for {self.__class__}")
+
+ if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero":
+ raise ValueError(
+ f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead."
+ )
+
+ # setable values
+ self.num_inference_steps = None
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
+ self.timesteps = torch.from_numpy(timesteps)
+ self.model_outputs = [None] * solver_order
+ self.lower_order_nums = 0
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ # Clipping the minimum of all lambda(t) for numerical stability.
+ # This is critical for cosine (squaredcos_cap_v2) noise schedule.
+ clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped)
+ last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item()
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, last_timestep - 1, num_inference_steps + 1).round()[::-1][:-1].copy().astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = last_timestep // (num_inference_steps + 1)
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.arange(last_timestep, 0, -step_ratio).round().copy().astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ log_sigmas = np.log(sigmas)
+
+ if self.config.use_karras_sigmas:
+ sigmas = np.flip(sigmas).copy()
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+ elif self.config.use_lu_lambdas:
+ lambdas = np.flip(log_sigmas.copy())
+ lambdas = self._convert_to_lu(in_lambdas=lambdas, num_inference_steps=num_inference_steps)
+ sigmas = np.exp(lambdas)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+ else:
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+
+ if self.config.final_sigmas_type == "sigma_min":
+ sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
+ elif self.config.final_sigmas_type == "zero":
+ sigma_last = 0
+ else:
+ raise ValueError(
+ f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}"
+ )
+
+ sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
+
+ self.sigmas = torch.from_numpy(sigmas)
+ self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
+
+ self.num_inference_steps = len(timesteps)
+
+ self.model_outputs = [
+ None,
+ ] * self.config.solver_order
+ self.lower_order_nums = 0
+
+ # add an index counter for schedulers that allow duplicated timesteps
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ def _sigma_to_alpha_sigma_t(self, sigma):
+ alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+ sigma_t = sigma * alpha_t
+
+ return alpha_t, sigma_t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def _convert_to_lu(self, in_lambdas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Lu et al. (2022)."""
+
+ lambda_min: float = in_lambdas[-1].item()
+ lambda_max: float = in_lambdas[0].item()
+
+ rho = 1.0 # 1.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = lambda_min ** (1 / rho)
+ max_inv_rho = lambda_max ** (1 / rho)
+ lambdas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return lambdas
+
+ def convert_model_output(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
+ designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
+ integral of the data prediction model.
+
+
+
+ The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise
+ prediction and data prediction models.
+
+
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The converted model output.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ if sample is None:
+ if len(args) > 1:
+ sample = args[1]
+ else:
+ raise ValueError("missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ # DPM-Solver++ needs to solve an integral of the data prediction model.
+ if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]:
+ if self.config.prediction_type == "epsilon":
+ # DPM-Solver and DPM-Solver++ only need the "mean" output.
+ if self.config.variance_type in ["learned", "learned_range"]:
+ model_output = model_output[:, :3]
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ x0_pred = (sample - sigma_t * model_output) / alpha_t
+ elif self.config.prediction_type == "sample":
+ x0_pred = model_output
+ elif self.config.prediction_type == "v_prediction":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ x0_pred = alpha_t * sample - sigma_t * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the DPMSolverMultistepScheduler."
+ )
+
+ if self.config.thresholding:
+ x0_pred = self._threshold_sample(x0_pred)
+
+ return x0_pred
+
+ # DPM-Solver needs to solve an integral of the noise prediction model.
+ elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
+ if self.config.prediction_type == "epsilon":
+ # DPM-Solver and DPM-Solver++ only need the "mean" output.
+ if self.config.variance_type in ["learned", "learned_range"]:
+ epsilon = model_output[:, :3]
+ else:
+ epsilon = model_output
+ elif self.config.prediction_type == "sample":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ epsilon = (sample - alpha_t * model_output) / sigma_t
+ elif self.config.prediction_type == "v_prediction":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ epsilon = alpha_t * model_output + sigma_t * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the DPMSolverMultistepScheduler."
+ )
+
+ if self.config.thresholding:
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ x0_pred = (sample - sigma_t * epsilon) / alpha_t
+ x0_pred = self._threshold_sample(x0_pred)
+ epsilon = (sample - alpha_t * x0_pred) / sigma_t
+
+ return epsilon
+
+ def dpm_solver_first_order_update(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ noise: Optional[torch.FloatTensor] = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the first-order DPMSolver (equivalent to DDIM).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s = torch.log(alpha_s) - torch.log(sigma_s)
+
+ h = lambda_t - lambda_s
+ if self.config.algorithm_type == "dpmsolver++":
+ x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
+ elif self.config.algorithm_type == "dpmsolver":
+ x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
+ elif self.config.algorithm_type == "sde-dpmsolver++":
+ assert noise is not None
+ x_t = (
+ (sigma_t / sigma_s * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+ elif self.config.algorithm_type == "sde-dpmsolver":
+ assert noise is not None
+ x_t = (
+ (alpha_t / alpha_s) * sample
+ - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output
+ + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+ )
+ return x_t
+
+ def multistep_dpm_solver_second_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ noise: Optional[torch.FloatTensor] = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the second-order multistep DPMSolver.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s0, sigma_s1 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+
+ m0, m1 = model_output_list[-1], model_output_list[-2]
+
+ h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
+ r0 = h_0 / h
+ D0, D1 = m0, (1.0 / r0) * (m0 - m1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2211.01095 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+ )
+ elif self.config.algorithm_type == "dpmsolver":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+ )
+ elif self.config.algorithm_type == "sde-dpmsolver++":
+ assert noise is not None
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
+ + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
+ + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+ elif self.config.algorithm_type == "sde-dpmsolver":
+ assert noise is not None
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * (torch.exp(h) - 1.0)) * D1
+ + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+ + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+ )
+ return x_t
+
+ def multistep_dpm_solver_third_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the third-order multistep DPMSolver.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing`sample` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ self.sigmas[self.step_index - 2],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+ alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+ lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)
+
+ m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
+
+ h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2
+ r0, r1 = h_0 / h, h_1 / h
+ D0 = m0
+ D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2)
+ D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
+ D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+ - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
+ )
+ elif self.config.algorithm_type == "dpmsolver":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+ - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
+ )
+ return x_t
+
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ index_candidates = (schedule_timesteps == timestep).nonzero()
+
+ if len(index_candidates) == 0:
+ step_index = len(self.timesteps) - 1
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ elif len(index_candidates) > 1:
+ step_index = index_candidates[1].item()
+ else:
+ step_index = index_candidates[0].item()
+
+ return step_index
+
+ def _init_step_index(self, timestep):
+ """
+ Initialize the step_index counter for the scheduler.
+ """
+
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator=None,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the multistep DPMSolver.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ # Improve numerical stability for small number of steps
+ lower_order_final = (self.step_index == len(self.timesteps) - 1) and (
+ self.config.euler_at_final
+ or (self.config.lower_order_final and len(self.timesteps) < 15)
+ or self.config.final_sigmas_type == "zero"
+ )
+ lower_order_second = (
+ (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
+ )
+
+ model_output = self.convert_model_output(model_output, sample=sample)
+ for i in range(self.config.solver_order - 1):
+ self.model_outputs[i] = self.model_outputs[i + 1]
+ self.model_outputs[-1] = model_output
+
+ # Upcast to avoid precision issues when computing prev_sample
+ sample = sample.to(torch.float32)
+
+ if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]:
+ noise = randn_tensor(
+ model_output.shape, generator=generator, device=model_output.device, dtype=torch.float32
+ )
+ else:
+ noise = None
+
+ if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
+ prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise)
+ elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
+ prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise)
+ else:
+ prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample)
+
+ if self.lower_order_nums < self.config.solver_order:
+ self.lower_order_nums += 1
+
+ # Cast sample back to expected dtype
+ prev_sample = prev_sample.to(model_output.dtype)
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # begin_index is None when the scheduler is used for training
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ noisy_samples = alpha_t * original_samples + sigma_t * noise
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep_flax.py b/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..0b48b499d36e92705341f5d8866238ee68d936a2
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep_flax.py
@@ -0,0 +1,643 @@
+# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver
+
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import flax
+import jax
+import jax.numpy as jnp
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils_flax import (
+ CommonSchedulerState,
+ FlaxKarrasDiffusionSchedulers,
+ FlaxSchedulerMixin,
+ FlaxSchedulerOutput,
+ add_noise_common,
+)
+
+
+@flax.struct.dataclass
+class DPMSolverMultistepSchedulerState:
+ common: CommonSchedulerState
+ alpha_t: jnp.ndarray
+ sigma_t: jnp.ndarray
+ lambda_t: jnp.ndarray
+
+ # setable values
+ init_noise_sigma: jnp.ndarray
+ timesteps: jnp.ndarray
+ num_inference_steps: Optional[int] = None
+
+ # running values
+ model_outputs: Optional[jnp.ndarray] = None
+ lower_order_nums: Optional[jnp.int32] = None
+ prev_timestep: Optional[jnp.int32] = None
+ cur_sample: Optional[jnp.ndarray] = None
+
+ @classmethod
+ def create(
+ cls,
+ common: CommonSchedulerState,
+ alpha_t: jnp.ndarray,
+ sigma_t: jnp.ndarray,
+ lambda_t: jnp.ndarray,
+ init_noise_sigma: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ):
+ return cls(
+ common=common,
+ alpha_t=alpha_t,
+ sigma_t=sigma_t,
+ lambda_t=lambda_t,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ )
+
+
+@dataclass
+class FlaxDPMSolverMultistepSchedulerOutput(FlaxSchedulerOutput):
+ state: DPMSolverMultistepSchedulerState
+
+
+class FlaxDPMSolverMultistepScheduler(FlaxSchedulerMixin, ConfigMixin):
+ """
+ DPM-Solver (and the improved version DPM-Solver++) is a fast dedicated high-order solver for diffusion ODEs with
+ the convergence order guarantee. Empirically, sampling by DPM-Solver with only 20 steps can generate high-quality
+ samples, and it can generate quite good samples even in only 10 steps.
+
+ For more details, see the original paper: https://arxiv.org/abs/2206.00927 and https://arxiv.org/abs/2211.01095
+
+ Currently, we support the multistep DPM-Solver for both noise prediction models and data prediction models. We
+ recommend to use `solver_order=2` for guided sampling, and `solver_order=3` for unconditional sampling.
+
+ We also support the "dynamic thresholding" method in Imagen (https://arxiv.org/abs/2205.11487). For pixel-space
+ diffusion models, you can set both `algorithm_type="dpmsolver++"` and `thresholding=True` to use the dynamic
+ thresholding. Note that the thresholding method is unsuitable for latent-space diffusion models (such as
+ stable-diffusion).
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ For more details, see the original paper: https://arxiv.org/abs/2206.00927 and https://arxiv.org/abs/2211.01095
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ beta_start (`float`): the starting `beta` value of inference.
+ beta_end (`float`): the final `beta` value.
+ beta_schedule (`str`):
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, optional):
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+ solver_order (`int`, default `2`):
+ the order of DPM-Solver; can be `1` or `2` or `3`. We recommend to use `solver_order=2` for guided
+ sampling, and `solver_order=3` for unconditional sampling.
+ prediction_type (`str`, default `epsilon`):
+ indicates whether the model predicts the noise (epsilon), or the data / `x0`. One of `epsilon`, `sample`,
+ or `v-prediction`.
+ thresholding (`bool`, default `False`):
+ whether to use the "dynamic thresholding" method (introduced by Imagen, https://arxiv.org/abs/2205.11487).
+ For pixel-space diffusion models, you can set both `algorithm_type=dpmsolver++` and `thresholding=True` to
+ use the dynamic thresholding. Note that the thresholding method is unsuitable for latent-space diffusion
+ models (such as stable-diffusion).
+ dynamic_thresholding_ratio (`float`, default `0.995`):
+ the ratio for the dynamic thresholding method. Default is `0.995`, the same as Imagen
+ (https://arxiv.org/abs/2205.11487).
+ sample_max_value (`float`, default `1.0`):
+ the threshold value for dynamic thresholding. Valid only when `thresholding=True` and
+ `algorithm_type="dpmsolver++`.
+ algorithm_type (`str`, default `dpmsolver++`):
+ the algorithm type for the solver. Either `dpmsolver` or `dpmsolver++`. The `dpmsolver` type implements the
+ algorithms in https://arxiv.org/abs/2206.00927, and the `dpmsolver++` type implements the algorithms in
+ https://arxiv.org/abs/2211.01095. We recommend to use `dpmsolver++` with `solver_order=2` for guided
+ sampling (e.g. stable-diffusion).
+ solver_type (`str`, default `midpoint`):
+ the solver type for the second-order solver. Either `midpoint` or `heun`. The solver type slightly affects
+ the sample quality, especially for small number of steps. We empirically find that `midpoint` solvers are
+ slightly better, so we recommend to use the `midpoint` type.
+ lower_order_final (`bool`, default `True`):
+ whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. We empirically
+ find this trick can stabilize the sampling of DPM-Solver for steps < 15, especially for steps <= 10.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`):
+ the `dtype` used for params and computation.
+ """
+
+ _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers]
+
+ dtype: jnp.dtype
+
+ @property
+ def has_state(self):
+ return True
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[jnp.ndarray] = None,
+ solver_order: int = 2,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ algorithm_type: str = "dpmsolver++",
+ solver_type: str = "midpoint",
+ lower_order_final: bool = True,
+ timestep_spacing: str = "linspace",
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ self.dtype = dtype
+
+ def create_state(self, common: Optional[CommonSchedulerState] = None) -> DPMSolverMultistepSchedulerState:
+ if common is None:
+ common = CommonSchedulerState.create(self)
+
+ # Currently we only support VP-type noise schedule
+ alpha_t = jnp.sqrt(common.alphas_cumprod)
+ sigma_t = jnp.sqrt(1 - common.alphas_cumprod)
+ lambda_t = jnp.log(alpha_t) - jnp.log(sigma_t)
+
+ # settings for DPM-Solver
+ if self.config.algorithm_type not in ["dpmsolver", "dpmsolver++"]:
+ raise NotImplementedError(f"{self.config.algorithm_type} does is not implemented for {self.__class__}")
+ if self.config.solver_type not in ["midpoint", "heun"]:
+ raise NotImplementedError(f"{self.config.solver_type} does is not implemented for {self.__class__}")
+
+ # standard deviation of the initial noise distribution
+ init_noise_sigma = jnp.array(1.0, dtype=self.dtype)
+
+ timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1]
+
+ return DPMSolverMultistepSchedulerState.create(
+ common=common,
+ alpha_t=alpha_t,
+ sigma_t=sigma_t,
+ lambda_t=lambda_t,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ )
+
+ def set_timesteps(
+ self, state: DPMSolverMultistepSchedulerState, num_inference_steps: int, shape: Tuple
+ ) -> DPMSolverMultistepSchedulerState:
+ """
+ Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ state (`DPMSolverMultistepSchedulerState`):
+ the `FlaxDPMSolverMultistepScheduler` state data class instance.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ shape (`Tuple`):
+ the shape of the samples to be generated.
+ """
+ last_timestep = self.config.num_train_timesteps
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ jnp.linspace(0, last_timestep - 1, num_inference_steps + 1).round()[::-1][:-1].astype(jnp.int32)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = last_timestep // (num_inference_steps + 1)
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (
+ (jnp.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(jnp.int32)
+ )
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = jnp.arange(last_timestep, 0, -step_ratio).round().copy().astype(jnp.int32)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ # initial running values
+
+ model_outputs = jnp.zeros((self.config.solver_order,) + shape, dtype=self.dtype)
+ lower_order_nums = jnp.int32(0)
+ prev_timestep = jnp.int32(-1)
+ cur_sample = jnp.zeros(shape, dtype=self.dtype)
+
+ return state.replace(
+ num_inference_steps=num_inference_steps,
+ timesteps=timesteps,
+ model_outputs=model_outputs,
+ lower_order_nums=lower_order_nums,
+ prev_timestep=prev_timestep,
+ cur_sample=cur_sample,
+ )
+
+ def convert_model_output(
+ self,
+ state: DPMSolverMultistepSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ ) -> jnp.ndarray:
+ """
+ Convert the model output to the corresponding type that the algorithm (DPM-Solver / DPM-Solver++) needs.
+
+ DPM-Solver is designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to
+ discretize an integral of the data prediction model. So we need to first convert the model output to the
+ corresponding type to match the algorithm.
+
+ Note that the algorithm type and the model type is decoupled. That is to say, we can use either DPM-Solver or
+ DPM-Solver++ for both noise prediction model and data prediction model.
+
+ Args:
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+
+ Returns:
+ `jnp.ndarray`: the converted model output.
+ """
+ # DPM-Solver++ needs to solve an integral of the data prediction model.
+ if self.config.algorithm_type == "dpmsolver++":
+ if self.config.prediction_type == "epsilon":
+ alpha_t, sigma_t = state.alpha_t[timestep], state.sigma_t[timestep]
+ x0_pred = (sample - sigma_t * model_output) / alpha_t
+ elif self.config.prediction_type == "sample":
+ x0_pred = model_output
+ elif self.config.prediction_type == "v_prediction":
+ alpha_t, sigma_t = state.alpha_t[timestep], state.sigma_t[timestep]
+ x0_pred = alpha_t * sample - sigma_t * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+ " or `v_prediction` for the FlaxDPMSolverMultistepScheduler."
+ )
+
+ if self.config.thresholding:
+ # Dynamic thresholding in https://arxiv.org/abs/2205.11487
+ dynamic_max_val = jnp.percentile(
+ jnp.abs(x0_pred), self.config.dynamic_thresholding_ratio, axis=tuple(range(1, x0_pred.ndim))
+ )
+ dynamic_max_val = jnp.maximum(
+ dynamic_max_val, self.config.sample_max_value * jnp.ones_like(dynamic_max_val)
+ )
+ x0_pred = jnp.clip(x0_pred, -dynamic_max_val, dynamic_max_val) / dynamic_max_val
+ return x0_pred
+ # DPM-Solver needs to solve an integral of the noise prediction model.
+ elif self.config.algorithm_type == "dpmsolver":
+ if self.config.prediction_type == "epsilon":
+ return model_output
+ elif self.config.prediction_type == "sample":
+ alpha_t, sigma_t = state.alpha_t[timestep], state.sigma_t[timestep]
+ epsilon = (sample - alpha_t * model_output) / sigma_t
+ return epsilon
+ elif self.config.prediction_type == "v_prediction":
+ alpha_t, sigma_t = state.alpha_t[timestep], state.sigma_t[timestep]
+ epsilon = alpha_t * model_output + sigma_t * sample
+ return epsilon
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, "
+ " or `v_prediction` for the FlaxDPMSolverMultistepScheduler."
+ )
+
+ def dpm_solver_first_order_update(
+ self,
+ state: DPMSolverMultistepSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ prev_timestep: int,
+ sample: jnp.ndarray,
+ ) -> jnp.ndarray:
+ """
+ One step for the first-order DPM-Solver (equivalent to DDIM).
+
+ See https://arxiv.org/abs/2206.00927 for the detailed derivation.
+
+ Args:
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ prev_timestep (`int`): previous discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+
+ Returns:
+ `jnp.ndarray`: the sample tensor at the previous timestep.
+ """
+ t, s0 = prev_timestep, timestep
+ m0 = model_output
+ lambda_t, lambda_s = state.lambda_t[t], state.lambda_t[s0]
+ alpha_t, alpha_s = state.alpha_t[t], state.alpha_t[s0]
+ sigma_t, sigma_s = state.sigma_t[t], state.sigma_t[s0]
+ h = lambda_t - lambda_s
+ if self.config.algorithm_type == "dpmsolver++":
+ x_t = (sigma_t / sigma_s) * sample - (alpha_t * (jnp.exp(-h) - 1.0)) * m0
+ elif self.config.algorithm_type == "dpmsolver":
+ x_t = (alpha_t / alpha_s) * sample - (sigma_t * (jnp.exp(h) - 1.0)) * m0
+ return x_t
+
+ def multistep_dpm_solver_second_order_update(
+ self,
+ state: DPMSolverMultistepSchedulerState,
+ model_output_list: jnp.ndarray,
+ timestep_list: List[int],
+ prev_timestep: int,
+ sample: jnp.ndarray,
+ ) -> jnp.ndarray:
+ """
+ One step for the second-order multistep DPM-Solver.
+
+ Args:
+ model_output_list (`List[jnp.ndarray]`):
+ direct outputs from learned diffusion model at current and latter timesteps.
+ timestep (`int`): current and latter discrete timestep in the diffusion chain.
+ prev_timestep (`int`): previous discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+
+ Returns:
+ `jnp.ndarray`: the sample tensor at the previous timestep.
+ """
+ t, s0, s1 = prev_timestep, timestep_list[-1], timestep_list[-2]
+ m0, m1 = model_output_list[-1], model_output_list[-2]
+ lambda_t, lambda_s0, lambda_s1 = state.lambda_t[t], state.lambda_t[s0], state.lambda_t[s1]
+ alpha_t, alpha_s0 = state.alpha_t[t], state.alpha_t[s0]
+ sigma_t, sigma_s0 = state.sigma_t[t], state.sigma_t[s0]
+ h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
+ r0 = h_0 / h
+ D0, D1 = m0, (1.0 / r0) * (m0 - m1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2211.01095 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (jnp.exp(-h) - 1.0)) * D0
+ - 0.5 * (alpha_t * (jnp.exp(-h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (jnp.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((jnp.exp(-h) - 1.0) / h + 1.0)) * D1
+ )
+ elif self.config.algorithm_type == "dpmsolver":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (jnp.exp(h) - 1.0)) * D0
+ - 0.5 * (sigma_t * (jnp.exp(h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (jnp.exp(h) - 1.0)) * D0
+ - (sigma_t * ((jnp.exp(h) - 1.0) / h - 1.0)) * D1
+ )
+ return x_t
+
+ def multistep_dpm_solver_third_order_update(
+ self,
+ state: DPMSolverMultistepSchedulerState,
+ model_output_list: jnp.ndarray,
+ timestep_list: List[int],
+ prev_timestep: int,
+ sample: jnp.ndarray,
+ ) -> jnp.ndarray:
+ """
+ One step for the third-order multistep DPM-Solver.
+
+ Args:
+ model_output_list (`List[jnp.ndarray]`):
+ direct outputs from learned diffusion model at current and latter timesteps.
+ timestep (`int`): current and latter discrete timestep in the diffusion chain.
+ prev_timestep (`int`): previous discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+
+ Returns:
+ `jnp.ndarray`: the sample tensor at the previous timestep.
+ """
+ t, s0, s1, s2 = prev_timestep, timestep_list[-1], timestep_list[-2], timestep_list[-3]
+ m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
+ lambda_t, lambda_s0, lambda_s1, lambda_s2 = (
+ state.lambda_t[t],
+ state.lambda_t[s0],
+ state.lambda_t[s1],
+ state.lambda_t[s2],
+ )
+ alpha_t, alpha_s0 = state.alpha_t[t], state.alpha_t[s0]
+ sigma_t, sigma_s0 = state.sigma_t[t], state.sigma_t[s0]
+ h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2
+ r0, r1 = h_0 / h, h_1 / h
+ D0 = m0
+ D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2)
+ D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
+ D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (jnp.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((jnp.exp(-h) - 1.0) / h + 1.0)) * D1
+ - (alpha_t * ((jnp.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
+ )
+ elif self.config.algorithm_type == "dpmsolver":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (jnp.exp(h) - 1.0)) * D0
+ - (sigma_t * ((jnp.exp(h) - 1.0) / h - 1.0)) * D1
+ - (sigma_t * ((jnp.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
+ )
+ return x_t
+
+ def step(
+ self,
+ state: DPMSolverMultistepSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ return_dict: bool = True,
+ ) -> Union[FlaxDPMSolverMultistepSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by DPM-Solver. Core function to propagate the diffusion process
+ from the learned model outputs (most often the predicted noise).
+
+ Args:
+ state (`DPMSolverMultistepSchedulerState`):
+ the `FlaxDPMSolverMultistepScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ return_dict (`bool`): option for returning tuple rather than FlaxDPMSolverMultistepSchedulerOutput class
+
+ Returns:
+ [`FlaxDPMSolverMultistepSchedulerOutput`] or `tuple`: [`FlaxDPMSolverMultistepSchedulerOutput`] if
+ `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+ if state.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ (step_index,) = jnp.where(state.timesteps == timestep, size=1)
+ step_index = step_index[0]
+
+ prev_timestep = jax.lax.select(step_index == len(state.timesteps) - 1, 0, state.timesteps[step_index + 1])
+
+ model_output = self.convert_model_output(state, model_output, timestep, sample)
+
+ model_outputs_new = jnp.roll(state.model_outputs, -1, axis=0)
+ model_outputs_new = model_outputs_new.at[-1].set(model_output)
+ state = state.replace(
+ model_outputs=model_outputs_new,
+ prev_timestep=prev_timestep,
+ cur_sample=sample,
+ )
+
+ def step_1(state: DPMSolverMultistepSchedulerState) -> jnp.ndarray:
+ return self.dpm_solver_first_order_update(
+ state,
+ state.model_outputs[-1],
+ state.timesteps[step_index],
+ state.prev_timestep,
+ state.cur_sample,
+ )
+
+ def step_23(state: DPMSolverMultistepSchedulerState) -> jnp.ndarray:
+ def step_2(state: DPMSolverMultistepSchedulerState) -> jnp.ndarray:
+ timestep_list = jnp.array([state.timesteps[step_index - 1], state.timesteps[step_index]])
+ return self.multistep_dpm_solver_second_order_update(
+ state,
+ state.model_outputs,
+ timestep_list,
+ state.prev_timestep,
+ state.cur_sample,
+ )
+
+ def step_3(state: DPMSolverMultistepSchedulerState) -> jnp.ndarray:
+ timestep_list = jnp.array(
+ [
+ state.timesteps[step_index - 2],
+ state.timesteps[step_index - 1],
+ state.timesteps[step_index],
+ ]
+ )
+ return self.multistep_dpm_solver_third_order_update(
+ state,
+ state.model_outputs,
+ timestep_list,
+ state.prev_timestep,
+ state.cur_sample,
+ )
+
+ step_2_output = step_2(state)
+ step_3_output = step_3(state)
+
+ if self.config.solver_order == 2:
+ return step_2_output
+ elif self.config.lower_order_final and len(state.timesteps) < 15:
+ return jax.lax.select(
+ state.lower_order_nums < 2,
+ step_2_output,
+ jax.lax.select(
+ step_index == len(state.timesteps) - 2,
+ step_2_output,
+ step_3_output,
+ ),
+ )
+ else:
+ return jax.lax.select(
+ state.lower_order_nums < 2,
+ step_2_output,
+ step_3_output,
+ )
+
+ step_1_output = step_1(state)
+ step_23_output = step_23(state)
+
+ if self.config.solver_order == 1:
+ prev_sample = step_1_output
+
+ elif self.config.lower_order_final and len(state.timesteps) < 15:
+ prev_sample = jax.lax.select(
+ state.lower_order_nums < 1,
+ step_1_output,
+ jax.lax.select(
+ step_index == len(state.timesteps) - 1,
+ step_1_output,
+ step_23_output,
+ ),
+ )
+
+ else:
+ prev_sample = jax.lax.select(
+ state.lower_order_nums < 1,
+ step_1_output,
+ step_23_output,
+ )
+
+ state = state.replace(
+ lower_order_nums=jnp.minimum(state.lower_order_nums + 1, self.config.solver_order),
+ )
+
+ if not return_dict:
+ return (prev_sample, state)
+
+ return FlaxDPMSolverMultistepSchedulerOutput(prev_sample=prev_sample, state=state)
+
+ def scale_model_input(
+ self, state: DPMSolverMultistepSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None
+ ) -> jnp.ndarray:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ state (`DPMSolverMultistepSchedulerState`):
+ the `FlaxDPMSolverMultistepScheduler` state data class instance.
+ sample (`jnp.ndarray`): input sample
+ timestep (`int`, optional): current timestep
+
+ Returns:
+ `jnp.ndarray`: scaled input sample
+ """
+ return sample
+
+ def add_noise(
+ self,
+ state: DPMSolverMultistepSchedulerState,
+ original_samples: jnp.ndarray,
+ noise: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ) -> jnp.ndarray:
+ return add_noise_common(state.common, original_samples, noise, timesteps)
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep_inverse.py b/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep_inverse.py
new file mode 100644
index 0000000000000000000000000000000000000000..f775c35c4f105e6518f9cb43ca124afb9d7f2356
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_dpmsolver_multistep_inverse.py
@@ -0,0 +1,917 @@
+# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import deprecate
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class DPMSolverMultistepInverseScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `DPMSolverMultistepInverseScheduler` is the reverse scheduler of [`DPMSolverMultistepScheduler`].
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ solver_order (`int`, defaults to 2):
+ The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
+ sampling, and `solver_order=3` for unconditional sampling.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
+ `algorithm_type="dpmsolver++"`.
+ algorithm_type (`str`, defaults to `dpmsolver++`):
+ Algorithm type for the solver; can be `dpmsolver`, `dpmsolver++`, `sde-dpmsolver` or `sde-dpmsolver++`. The
+ `dpmsolver` type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927)
+ paper, and the `dpmsolver++` type implements the algorithms in the
+ [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or
+ `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion.
+ solver_type (`str`, defaults to `midpoint`):
+ Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
+ sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
+ lower_order_final (`bool`, defaults to `True`):
+ Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
+ stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
+ euler_at_final (`bool`, defaults to `False`):
+ Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail
+ richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
+ steps, but sometimes may result in blurring.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ lambda_min_clipped (`float`, defaults to `-inf`):
+ Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
+ cosine (`squaredcos_cap_v2`) noise schedule.
+ variance_type (`str`, *optional*):
+ Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
+ contains the predicted Gaussian variance.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ solver_order: int = 2,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ algorithm_type: str = "dpmsolver++",
+ solver_type: str = "midpoint",
+ lower_order_final: bool = True,
+ euler_at_final: bool = False,
+ use_karras_sigmas: Optional[bool] = False,
+ lambda_min_clipped: float = -float("inf"),
+ variance_type: Optional[str] = None,
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
+ deprecation_message = f"algorithm_type {algorithm_type} is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
+ deprecate("algorithm_types dpmsolver and sde-dpmsolver", "1.0.0", deprecation_message)
+
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ # Currently we only support VP-type noise schedule
+ self.alpha_t = torch.sqrt(self.alphas_cumprod)
+ self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
+ self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
+ self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # settings for DPM-Solver
+ if algorithm_type not in ["dpmsolver", "dpmsolver++", "sde-dpmsolver", "sde-dpmsolver++"]:
+ if algorithm_type == "deis":
+ self.register_to_config(algorithm_type="dpmsolver++")
+ else:
+ raise NotImplementedError(f"{algorithm_type} does is not implemented for {self.__class__}")
+
+ if solver_type not in ["midpoint", "heun"]:
+ if solver_type in ["logrho", "bh1", "bh2"]:
+ self.register_to_config(solver_type="midpoint")
+ else:
+ raise NotImplementedError(f"{solver_type} does is not implemented for {self.__class__}")
+
+ # setable values
+ self.num_inference_steps = None
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32).copy()
+ self.timesteps = torch.from_numpy(timesteps)
+ self.model_outputs = [None] * solver_order
+ self.lower_order_nums = 0
+ self._step_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+ self.use_karras_sigmas = use_karras_sigmas
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ # Clipping the minimum of all lambda(t) for numerical stability.
+ # This is critical for cosine (squaredcos_cap_v2) noise schedule.
+ clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.lambda_min_clipped).item()
+ self.noisiest_timestep = self.config.num_train_timesteps - 1 - clipped_idx
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, self.noisiest_timestep, num_inference_steps + 1).round()[:-1].copy().astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = (self.noisiest_timestep + 1) // (num_inference_steps + 1)
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[:-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.arange(self.noisiest_timestep + 1, 0, -step_ratio).round()[::-1].copy().astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', "
+ "'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ log_sigmas = np.log(sigmas)
+
+ if self.config.use_karras_sigmas:
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+ timesteps = timesteps.copy().astype(np.int64)
+ sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
+ else:
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+ sigma_max = (
+ (1 - self.alphas_cumprod[self.noisiest_timestep]) / self.alphas_cumprod[self.noisiest_timestep]
+ ) ** 0.5
+ sigmas = np.concatenate([sigmas, [sigma_max]]).astype(np.float32)
+
+ self.sigmas = torch.from_numpy(sigmas)
+
+ # when num_inference_steps == num_train_timesteps, we can end up with
+ # duplicates in timesteps.
+ _, unique_indices = np.unique(timesteps, return_index=True)
+ timesteps = timesteps[np.sort(unique_indices)]
+
+ self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
+
+ self.num_inference_steps = len(timesteps)
+
+ self.model_outputs = [
+ None,
+ ] * self.config.solver_order
+ self.lower_order_nums = 0
+
+ # add an index counter for schedulers that allow duplicated timesteps
+ self._step_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
+ def _sigma_to_alpha_sigma_t(self, sigma):
+ alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+ sigma_t = sigma * alpha_t
+
+ return alpha_t, sigma_t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.convert_model_output
+ def convert_model_output(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
+ designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
+ integral of the data prediction model.
+
+
+
+ The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise
+ prediction and data prediction models.
+
+
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The converted model output.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ if sample is None:
+ if len(args) > 1:
+ sample = args[1]
+ else:
+ raise ValueError("missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ # DPM-Solver++ needs to solve an integral of the data prediction model.
+ if self.config.algorithm_type in ["dpmsolver++", "sde-dpmsolver++"]:
+ if self.config.prediction_type == "epsilon":
+ # DPM-Solver and DPM-Solver++ only need the "mean" output.
+ if self.config.variance_type in ["learned", "learned_range"]:
+ model_output = model_output[:, :3]
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ x0_pred = (sample - sigma_t * model_output) / alpha_t
+ elif self.config.prediction_type == "sample":
+ x0_pred = model_output
+ elif self.config.prediction_type == "v_prediction":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ x0_pred = alpha_t * sample - sigma_t * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the DPMSolverMultistepScheduler."
+ )
+
+ if self.config.thresholding:
+ x0_pred = self._threshold_sample(x0_pred)
+
+ return x0_pred
+
+ # DPM-Solver needs to solve an integral of the noise prediction model.
+ elif self.config.algorithm_type in ["dpmsolver", "sde-dpmsolver"]:
+ if self.config.prediction_type == "epsilon":
+ # DPM-Solver and DPM-Solver++ only need the "mean" output.
+ if self.config.variance_type in ["learned", "learned_range"]:
+ epsilon = model_output[:, :3]
+ else:
+ epsilon = model_output
+ elif self.config.prediction_type == "sample":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ epsilon = (sample - alpha_t * model_output) / sigma_t
+ elif self.config.prediction_type == "v_prediction":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ epsilon = alpha_t * model_output + sigma_t * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the DPMSolverMultistepScheduler."
+ )
+
+ if self.config.thresholding:
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ x0_pred = (sample - sigma_t * epsilon) / alpha_t
+ x0_pred = self._threshold_sample(x0_pred)
+ epsilon = (sample - alpha_t * x0_pred) / sigma_t
+
+ return epsilon
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.dpm_solver_first_order_update
+ def dpm_solver_first_order_update(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ noise: Optional[torch.FloatTensor] = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the first-order DPMSolver (equivalent to DDIM).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s = torch.log(alpha_s) - torch.log(sigma_s)
+
+ h = lambda_t - lambda_s
+ if self.config.algorithm_type == "dpmsolver++":
+ x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
+ elif self.config.algorithm_type == "dpmsolver":
+ x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
+ elif self.config.algorithm_type == "sde-dpmsolver++":
+ assert noise is not None
+ x_t = (
+ (sigma_t / sigma_s * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+ elif self.config.algorithm_type == "sde-dpmsolver":
+ assert noise is not None
+ x_t = (
+ (alpha_t / alpha_s) * sample
+ - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * model_output
+ + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+ )
+ return x_t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.multistep_dpm_solver_second_order_update
+ def multistep_dpm_solver_second_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ noise: Optional[torch.FloatTensor] = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the second-order multistep DPMSolver.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s0, sigma_s1 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+
+ m0, m1 = model_output_list[-1], model_output_list[-2]
+
+ h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
+ r0 = h_0 / h
+ D0, D1 = m0, (1.0 / r0) * (m0 - m1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2211.01095 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+ )
+ elif self.config.algorithm_type == "dpmsolver":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+ )
+ elif self.config.algorithm_type == "sde-dpmsolver++":
+ assert noise is not None
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
+ + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
+ + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+ elif self.config.algorithm_type == "sde-dpmsolver":
+ assert noise is not None
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * (torch.exp(h) - 1.0)) * D1
+ + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - 2.0 * (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - 2.0 * (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+ + sigma_t * torch.sqrt(torch.exp(2 * h) - 1.0) * noise
+ )
+ return x_t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.multistep_dpm_solver_third_order_update
+ def multistep_dpm_solver_third_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the third-order multistep DPMSolver.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing`sample` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ self.sigmas[self.step_index - 2],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+ alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+ lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)
+
+ m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
+
+ h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2
+ r0, r1 = h_0 / h, h_1 / h
+ D0 = m0
+ D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2)
+ D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
+ D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+ - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
+ )
+ elif self.config.algorithm_type == "dpmsolver":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ x_t = (
+ (alpha_t / alpha_s0) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+ - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
+ )
+ return x_t
+
+ def _init_step_index(self, timestep):
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+
+ index_candidates = (self.timesteps == timestep).nonzero()
+
+ if len(index_candidates) == 0:
+ step_index = len(self.timesteps) - 1
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ elif len(index_candidates) > 1:
+ step_index = index_candidates[1].item()
+ else:
+ step_index = index_candidates[0].item()
+
+ self._step_index = step_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator=None,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the multistep DPMSolver.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ # Improve numerical stability for small number of steps
+ lower_order_final = (self.step_index == len(self.timesteps) - 1) and (
+ self.config.euler_at_final or (self.config.lower_order_final and len(self.timesteps) < 15)
+ )
+ lower_order_second = (
+ (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
+ )
+
+ model_output = self.convert_model_output(model_output, sample=sample)
+ for i in range(self.config.solver_order - 1):
+ self.model_outputs[i] = self.model_outputs[i + 1]
+ self.model_outputs[-1] = model_output
+
+ if self.config.algorithm_type in ["sde-dpmsolver", "sde-dpmsolver++"]:
+ noise = randn_tensor(
+ model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
+ )
+ else:
+ noise = None
+
+ if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
+ prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise)
+ elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
+ prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise)
+ else:
+ prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample)
+
+ if self.lower_order_nums < self.config.solver_order:
+ self.lower_order_nums += 1
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.scale_model_input
+ def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ step_indices = []
+ for timestep in timesteps:
+ index_candidates = (schedule_timesteps == timestep).nonzero()
+ if len(index_candidates) == 0:
+ step_index = len(schedule_timesteps) - 1
+ elif len(index_candidates) > 1:
+ step_index = index_candidates[1].item()
+ else:
+ step_index = index_candidates[0].item()
+ step_indices.append(step_index)
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ noisy_samples = alpha_t * original_samples + sigma_t * noise
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_dpmsolver_sde.py b/model/diffusers_c/schedulers/scheduling_dpmsolver_sde.py
new file mode 100644
index 0000000000000000000000000000000000000000..81a42a6e37e6d1a7c6115090b5745a63033a7595
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_dpmsolver_sde.py
@@ -0,0 +1,559 @@
+# Copyright 2024 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torchsde
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+class BatchedBrownianTree:
+ """A wrapper around torchsde.BrownianTree that enables batches of entropy."""
+
+ def __init__(self, x, t0, t1, seed=None, **kwargs):
+ t0, t1, self.sign = self.sort(t0, t1)
+ w0 = kwargs.get("w0", torch.zeros_like(x))
+ if seed is None:
+ seed = torch.randint(0, 2**63 - 1, []).item()
+ self.batched = True
+ try:
+ assert len(seed) == x.shape[0]
+ w0 = w0[0]
+ except TypeError:
+ seed = [seed]
+ self.batched = False
+ self.trees = [torchsde.BrownianTree(t0, w0, t1, entropy=s, **kwargs) for s in seed]
+
+ @staticmethod
+ def sort(a, b):
+ return (a, b, 1) if a < b else (b, a, -1)
+
+ def __call__(self, t0, t1):
+ t0, t1, sign = self.sort(t0, t1)
+ w = torch.stack([tree(t0, t1) for tree in self.trees]) * (self.sign * sign)
+ return w if self.batched else w[0]
+
+
+class BrownianTreeNoiseSampler:
+ """A noise sampler backed by a torchsde.BrownianTree.
+
+ Args:
+ x (Tensor): The tensor whose shape, device and dtype to use to generate
+ random samples.
+ sigma_min (float): The low end of the valid interval.
+ sigma_max (float): The high end of the valid interval.
+ seed (int or List[int]): The random seed. If a list of seeds is
+ supplied instead of a single integer, then the noise sampler will use one BrownianTree per batch item, each
+ with its own seed.
+ transform (callable): A function that maps sigma to the sampler's
+ internal timestep.
+ """
+
+ def __init__(self, x, sigma_min, sigma_max, seed=None, transform=lambda x: x):
+ self.transform = transform
+ t0, t1 = self.transform(torch.as_tensor(sigma_min)), self.transform(torch.as_tensor(sigma_max))
+ self.tree = BatchedBrownianTree(x, t0, t1, seed)
+
+ def __call__(self, sigma, sigma_next):
+ t0, t1 = self.transform(torch.as_tensor(sigma)), self.transform(torch.as_tensor(sigma_next))
+ return self.tree(t0, t1) / (t1 - t0).abs().sqrt()
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class DPMSolverSDEScheduler(SchedulerMixin, ConfigMixin):
+ """
+ DPMSolverSDEScheduler implements the stochastic sampler from the [Elucidating the Design Space of Diffusion-Based
+ Generative Models](https://huggingface.co/papers/2206.00364) paper.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.00085):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.012):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ noise_sampler_seed (`int`, *optional*, defaults to `None`):
+ The random seed to use for the noise sampler. If `None`, a random seed is generated.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 2
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.00085, # sensible defaults
+ beta_end: float = 0.012,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ prediction_type: str = "epsilon",
+ use_karras_sigmas: Optional[bool] = False,
+ noise_sampler_seed: Optional[int] = None,
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # set all values
+ self.set_timesteps(num_train_timesteps, None, num_train_timesteps)
+ self.use_karras_sigmas = use_karras_sigmas
+ self.noise_sampler = None
+ self.noise_sampler_seed = noise_sampler_seed
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ return self.sigmas.max()
+
+ return (self.sigmas.max() ** 2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+ sigma_input = sigma if self.state_in_first_order else self.mid_point_sigma
+ sample = sample / ((sigma_input**2 + 1) ** 0.5)
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: int,
+ device: Union[str, torch.device] = None,
+ num_train_timesteps: Optional[int] = None,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+
+ num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=float)[::-1].copy()
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(float)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(float)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ log_sigmas = np.log(sigmas)
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+
+ if self.use_karras_sigmas:
+ sigmas = self._convert_to_karras(in_sigmas=sigmas)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+
+ second_order_timesteps = self._second_order_timesteps(sigmas, log_sigmas)
+
+ sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+ sigmas = torch.from_numpy(sigmas).to(device=device)
+ self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]])
+
+ timesteps = torch.from_numpy(timesteps)
+ second_order_timesteps = torch.from_numpy(second_order_timesteps)
+ timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)])
+ timesteps[1::2] = second_order_timesteps
+
+ if str(device).startswith("mps"):
+ # mps does not support float64
+ self.timesteps = timesteps.to(device, dtype=torch.float32)
+ else:
+ self.timesteps = timesteps.to(device=device)
+
+ # empty first order variables
+ self.sample = None
+ self.mid_point_sigma = None
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+ self.noise_sampler = None
+
+ def _second_order_timesteps(self, sigmas, log_sigmas):
+ def sigma_fn(_t):
+ return np.exp(-_t)
+
+ def t_fn(_sigma):
+ return -np.log(_sigma)
+
+ midpoint_ratio = 0.5
+ t = t_fn(sigmas)
+ delta_time = np.diff(t)
+ t_proposed = t[:-1] + delta_time * midpoint_ratio
+ sig_proposed = sigma_fn(t_proposed)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sig_proposed])
+ return timesteps
+
+ # copied from diffusers.schedulers.scheduling_euler_discrete._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # copied from diffusers.schedulers.scheduling_euler_discrete._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ sigma_min: float = in_sigmas[-1].item()
+ sigma_max: float = in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, self.num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ @property
+ def state_in_first_order(self):
+ return self.sample is None
+
+ def step(
+ self,
+ model_output: Union[torch.FloatTensor, np.ndarray],
+ timestep: Union[float, torch.FloatTensor],
+ sample: Union[torch.FloatTensor, np.ndarray],
+ return_dict: bool = True,
+ s_noise: float = 1.0,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor` or `np.ndarray`):
+ The direct output from learned diffusion model.
+ timestep (`float` or `torch.FloatTensor`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor` or `np.ndarray`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+ s_noise (`float`, *optional*, defaults to 1.0):
+ Scaling factor for noise added to the sample.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ # Create a noise sampler if it hasn't been created yet
+ if self.noise_sampler is None:
+ min_sigma, max_sigma = self.sigmas[self.sigmas > 0].min(), self.sigmas.max()
+ self.noise_sampler = BrownianTreeNoiseSampler(sample, min_sigma, max_sigma, self.noise_sampler_seed)
+
+ # Define functions to compute sigma and t from each other
+ def sigma_fn(_t: torch.FloatTensor) -> torch.FloatTensor:
+ return _t.neg().exp()
+
+ def t_fn(_sigma: torch.FloatTensor) -> torch.FloatTensor:
+ return _sigma.log().neg()
+
+ if self.state_in_first_order:
+ sigma = self.sigmas[self.step_index]
+ sigma_next = self.sigmas[self.step_index + 1]
+ else:
+ # 2nd order
+ sigma = self.sigmas[self.step_index - 1]
+ sigma_next = self.sigmas[self.step_index]
+
+ # Set the midpoint and step size for the current step
+ midpoint_ratio = 0.5
+ t, t_next = t_fn(sigma), t_fn(sigma_next)
+ delta_time = t_next - t
+ t_proposed = t + delta_time * midpoint_ratio
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ if self.config.prediction_type == "epsilon":
+ sigma_input = sigma if self.state_in_first_order else sigma_fn(t_proposed)
+ pred_original_sample = sample - sigma_input * model_output
+ elif self.config.prediction_type == "v_prediction":
+ sigma_input = sigma if self.state_in_first_order else sigma_fn(t_proposed)
+ pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + (
+ sample / (sigma_input**2 + 1)
+ )
+ elif self.config.prediction_type == "sample":
+ raise NotImplementedError("prediction_type not implemented yet: sample")
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ if sigma_next == 0:
+ derivative = (sample - pred_original_sample) / sigma
+ dt = sigma_next - sigma
+ prev_sample = sample + derivative * dt
+ else:
+ if self.state_in_first_order:
+ t_next = t_proposed
+ else:
+ sample = self.sample
+
+ sigma_from = sigma_fn(t)
+ sigma_to = sigma_fn(t_next)
+ sigma_up = min(sigma_to, (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5)
+ sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5
+ ancestral_t = t_fn(sigma_down)
+ prev_sample = (sigma_fn(ancestral_t) / sigma_fn(t)) * sample - (
+ t - ancestral_t
+ ).expm1() * pred_original_sample
+ prev_sample = prev_sample + self.noise_sampler(sigma_fn(t), sigma_fn(t_next)) * s_noise * sigma_up
+
+ if self.state_in_first_order:
+ # store for 2nd order step
+ self.sample = sample
+ self.mid_point_sigma = sigma_fn(t_next)
+ else:
+ # free for "first order mode"
+ self.sample = None
+ self.mid_point_sigma = None
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_dpmsolver_singlestep.py b/model/diffusers_c/schedulers/scheduling_dpmsolver_singlestep.py
new file mode 100644
index 0000000000000000000000000000000000000000..9806df845243b2dee176aaa598938fce99319726
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_dpmsolver_singlestep.py
@@ -0,0 +1,977 @@
+# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import deprecate, logging
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class DPMSolverSinglestepScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `DPMSolverSinglestepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ solver_order (`int`, defaults to 2):
+ The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
+ sampling, and `solver_order=3` for unconditional sampling.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
+ `algorithm_type="dpmsolver++"`.
+ algorithm_type (`str`, defaults to `dpmsolver++`):
+ Algorithm type for the solver; can be `dpmsolver` or `dpmsolver++`. The
+ `dpmsolver` type implements the algorithms in the [DPMSolver](https://huggingface.co/papers/2206.00927)
+ paper, and the `dpmsolver++` type implements the algorithms in the
+ [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or
+ `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion.
+ solver_type (`str`, defaults to `midpoint`):
+ Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
+ sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
+ lower_order_final (`bool`, defaults to `True`):
+ Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
+ stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ final_sigmas_type (`str`, *optional*, defaults to `"zero"`):
+ The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final sigma
+ is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
+ lambda_min_clipped (`float`, defaults to `-inf`):
+ Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
+ cosine (`squaredcos_cap_v2`) noise schedule.
+ variance_type (`str`, *optional*):
+ Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
+ contains the predicted Gaussian variance.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[np.ndarray] = None,
+ solver_order: int = 2,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ algorithm_type: str = "dpmsolver++",
+ solver_type: str = "midpoint",
+ lower_order_final: bool = False,
+ use_karras_sigmas: Optional[bool] = False,
+ final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min"
+ lambda_min_clipped: float = -float("inf"),
+ variance_type: Optional[str] = None,
+ ):
+ if algorithm_type == "dpmsolver":
+ deprecation_message = "algorithm_type `dpmsolver` is deprecated and will be removed in a future version. Choose from `dpmsolver++` or `sde-dpmsolver++` instead"
+ deprecate("algorithm_types=dpmsolver", "1.0.0", deprecation_message)
+
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ # Currently we only support VP-type noise schedule
+ self.alpha_t = torch.sqrt(self.alphas_cumprod)
+ self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
+ self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
+ self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # settings for DPM-Solver
+ if algorithm_type not in ["dpmsolver", "dpmsolver++"]:
+ if algorithm_type == "deis":
+ self.register_to_config(algorithm_type="dpmsolver++")
+ else:
+ raise NotImplementedError(f"{algorithm_type} does is not implemented for {self.__class__}")
+ if solver_type not in ["midpoint", "heun"]:
+ if solver_type in ["logrho", "bh1", "bh2"]:
+ self.register_to_config(solver_type="midpoint")
+ else:
+ raise NotImplementedError(f"{solver_type} does is not implemented for {self.__class__}")
+
+ if algorithm_type != "dpmsolver++" and final_sigmas_type == "zero":
+ raise ValueError(
+ f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please chooose `sigma_min` instead."
+ )
+
+ # setable values
+ self.num_inference_steps = None
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
+ self.timesteps = torch.from_numpy(timesteps)
+ self.model_outputs = [None] * solver_order
+ self.sample = None
+ self.order_list = self.get_order_list(num_train_timesteps)
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ def get_order_list(self, num_inference_steps: int) -> List[int]:
+ """
+ Computes the solver order at each time step.
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ """
+ steps = num_inference_steps
+ order = self.config.solver_order
+ if self.config.lower_order_final:
+ if order == 3:
+ if steps % 3 == 0:
+ orders = [1, 2, 3] * (steps // 3 - 1) + [1, 2] + [1]
+ elif steps % 3 == 1:
+ orders = [1, 2, 3] * (steps // 3) + [1]
+ else:
+ orders = [1, 2, 3] * (steps // 3) + [1, 2]
+ elif order == 2:
+ if steps % 2 == 0:
+ orders = [1, 2] * (steps // 2 - 1) + [1, 1]
+ else:
+ orders = [1, 2] * (steps // 2) + [1]
+ elif order == 1:
+ orders = [1] * steps
+ else:
+ if order == 3:
+ orders = [1, 2, 3] * (steps // 3)
+ elif order == 2:
+ orders = [1, 2] * (steps // 2)
+ elif order == 1:
+ orders = [1] * steps
+ return orders
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+ # Clipping the minimum of all lambda(t) for numerical stability.
+ # This is critical for cosine (squaredcos_cap_v2) noise schedule.
+ clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped)
+ timesteps = (
+ np.linspace(0, self.config.num_train_timesteps - 1 - clipped_idx, num_inference_steps + 1)
+ .round()[::-1][:-1]
+ .copy()
+ .astype(np.int64)
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ if self.config.use_karras_sigmas:
+ log_sigmas = np.log(sigmas)
+ sigmas = np.flip(sigmas).copy()
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+ else:
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+
+ if self.config.final_sigmas_type == "sigma_min":
+ sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
+ elif self.config.final_sigmas_type == "zero":
+ sigma_last = 0
+ else:
+ raise ValueError(
+ f" `final_sigmas_type` must be one of `sigma_min` or `zero`, but got {self.config.final_sigmas_type}"
+ )
+ sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
+
+ self.sigmas = torch.from_numpy(sigmas).to(device=device)
+
+ self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
+ self.model_outputs = [None] * self.config.solver_order
+ self.sample = None
+
+ if not self.config.lower_order_final and num_inference_steps % self.config.solver_order != 0:
+ logger.warn(
+ "Changing scheduler {self.config} to have `lower_order_final` set to True to handle uneven amount of inference steps. Please make sure to always use an even number of `num_inference steps when using `lower_order_final=False`."
+ )
+ self.register_to_config(lower_order_final=True)
+
+ if not self.config.lower_order_final and self.config.final_sigmas_type == "zero":
+ logger.warn(
+ " `last_sigmas_type='zero'` is not supported for `lower_order_final=False`. Changing scheduler {self.config} to have `lower_order_final` set to True."
+ )
+ self.register_to_config(lower_order_final=True)
+
+ self.order_list = self.get_order_list(num_inference_steps)
+
+ # add an index counter for schedulers that allow duplicated timesteps
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
+ def _sigma_to_alpha_sigma_t(self, sigma):
+ alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+ sigma_t = sigma * alpha_t
+
+ return alpha_t, sigma_t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def convert_model_output(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
+ designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
+ integral of the data prediction model.
+
+
+
+ The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise
+ prediction and data prediction models.
+
+
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The converted model output.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ if sample is None:
+ if len(args) > 1:
+ sample = args[1]
+ else:
+ raise ValueError("missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+ # DPM-Solver++ needs to solve an integral of the data prediction model.
+ if self.config.algorithm_type == "dpmsolver++":
+ if self.config.prediction_type == "epsilon":
+ # DPM-Solver and DPM-Solver++ only need the "mean" output.
+ if self.config.variance_type in ["learned_range"]:
+ model_output = model_output[:, :3]
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ x0_pred = (sample - sigma_t * model_output) / alpha_t
+ elif self.config.prediction_type == "sample":
+ x0_pred = model_output
+ elif self.config.prediction_type == "v_prediction":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ x0_pred = alpha_t * sample - sigma_t * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the DPMSolverSinglestepScheduler."
+ )
+
+ if self.config.thresholding:
+ x0_pred = self._threshold_sample(x0_pred)
+
+ return x0_pred
+ # DPM-Solver needs to solve an integral of the noise prediction model.
+ elif self.config.algorithm_type == "dpmsolver":
+ if self.config.prediction_type == "epsilon":
+ # DPM-Solver and DPM-Solver++ only need the "mean" output.
+ if self.config.variance_type in ["learned_range"]:
+ model_output = model_output[:, :3]
+ return model_output
+ elif self.config.prediction_type == "sample":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ epsilon = (sample - alpha_t * model_output) / sigma_t
+ return epsilon
+ elif self.config.prediction_type == "v_prediction":
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ epsilon = alpha_t * model_output + sigma_t * sample
+ return epsilon
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the DPMSolverSinglestepScheduler."
+ )
+
+ def dpm_solver_first_order_update(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the first-order DPMSolver (equivalent to DDIM).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ prev_timestep (`int`):
+ The previous discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+ sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s = torch.log(alpha_s) - torch.log(sigma_s)
+ h = lambda_t - lambda_s
+ if self.config.algorithm_type == "dpmsolver++":
+ x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
+ elif self.config.algorithm_type == "dpmsolver":
+ x_t = (alpha_t / alpha_s) * sample - (sigma_t * (torch.exp(h) - 1.0)) * model_output
+ return x_t
+
+ def singlestep_dpm_solver_second_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the second-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the
+ time `timestep_list[-2]`.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ timestep (`int`):
+ The current and latter discrete timestep in the diffusion chain.
+ prev_timestep (`int`):
+ The previous discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+ sigma_t, sigma_s0, sigma_s1 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+
+ m0, m1 = model_output_list[-1], model_output_list[-2]
+
+ h, h_0 = lambda_t - lambda_s1, lambda_s0 - lambda_s1
+ r0 = h_0 / h
+ D0, D1 = m1, (1.0 / r0) * (m0 - m1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2211.01095 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s1) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s1) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+ )
+ elif self.config.algorithm_type == "dpmsolver":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (alpha_t / alpha_s1) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - 0.5 * (sigma_t * (torch.exp(h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (alpha_t / alpha_s1) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+ )
+ return x_t
+
+ def singlestep_dpm_solver_third_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the third-order singlestep DPMSolver that computes the solution at time `prev_timestep` from the
+ time `timestep_list[-3]`.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ timestep (`int`):
+ The current and latter discrete timestep in the diffusion chain.
+ prev_timestep (`int`):
+ The previous discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing`sample` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ self.sigmas[self.step_index - 2],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+ alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+ lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)
+
+ m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
+
+ h, h_0, h_1 = lambda_t - lambda_s2, lambda_s0 - lambda_s2, lambda_s1 - lambda_s2
+ r0, r1 = h_0 / h, h_1 / h
+ D0 = m2
+ D1_0, D1_1 = (1.0 / r1) * (m1 - m2), (1.0 / r0) * (m0 - m2)
+ D1 = (r0 * D1_0 - r1 * D1_1) / (r0 - r1)
+ D2 = 2.0 * (D1_1 - D1_0) / (r0 - r1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s2) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1_1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s2) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+ - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
+ )
+ elif self.config.algorithm_type == "dpmsolver":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (alpha_t / alpha_s2) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1_1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (alpha_t / alpha_s2) * sample
+ - (sigma_t * (torch.exp(h) - 1.0)) * D0
+ - (sigma_t * ((torch.exp(h) - 1.0) / h - 1.0)) * D1
+ - (sigma_t * ((torch.exp(h) - 1.0 - h) / h**2 - 0.5)) * D2
+ )
+ return x_t
+
+ def singlestep_dpm_solver_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ *args,
+ sample: torch.FloatTensor = None,
+ order: int = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the singlestep DPMSolver.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ timestep (`int`):
+ The current and latter discrete timestep in the diffusion chain.
+ prev_timestep (`int`):
+ The previous discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by diffusion process.
+ order (`int`):
+ The solver order at this step.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ timestep_list = args[0] if len(args) > 0 else kwargs.pop("timestep_list", None)
+ prev_timestep = args[1] if len(args) > 1 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 2:
+ sample = args[2]
+ else:
+ raise ValueError(" missing`sample` as a required keyward argument")
+ if order is None:
+ if len(args) > 3:
+ order = args[3]
+ else:
+ raise ValueError(" missing `order` as a required keyward argument")
+ if timestep_list is not None:
+ deprecate(
+ "timestep_list",
+ "1.0.0",
+ "Passing `timestep_list` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ if order == 1:
+ return self.dpm_solver_first_order_update(model_output_list[-1], sample=sample)
+ elif order == 2:
+ return self.singlestep_dpm_solver_second_order_update(model_output_list, sample=sample)
+ elif order == 3:
+ return self.singlestep_dpm_solver_third_order_update(model_output_list, sample=sample)
+ else:
+ raise ValueError(f"Order must be 1, 2, 3, got {order}")
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ index_candidates = (schedule_timesteps == timestep).nonzero()
+
+ if len(index_candidates) == 0:
+ step_index = len(self.timesteps) - 1
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ elif len(index_candidates) > 1:
+ step_index = index_candidates[1].item()
+ else:
+ step_index = index_candidates[0].item()
+
+ return step_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ """
+ Initialize the step_index counter for the scheduler.
+ """
+
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the singlestep DPMSolver.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ model_output = self.convert_model_output(model_output, sample=sample)
+ for i in range(self.config.solver_order - 1):
+ self.model_outputs[i] = self.model_outputs[i + 1]
+ self.model_outputs[-1] = model_output
+
+ order = self.order_list[self.step_index]
+
+ # For img2img denoising might start with order>1 which is not possible
+ # In this case make sure that the first two steps are both order=1
+ while self.model_outputs[-order] is None:
+ order -= 1
+
+ # For single-step solvers, we use the initial value at each time with order = 1.
+ if order == 1:
+ self.sample = sample
+
+ prev_sample = self.singlestep_dpm_solver_update(self.model_outputs, sample=self.sample, order=order)
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # begin_index is None when the scheduler is used for training
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ noisy_samples = alpha_t * original_samples + sigma_t * noise
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_edm_dpmsolver_multistep.py b/model/diffusers_c/schedulers/scheduling_edm_dpmsolver_multistep.py
new file mode 100644
index 0000000000000000000000000000000000000000..5fea89bb887b2e5551e8ba3230d3de0cb7a75e23
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_edm_dpmsolver_multistep.py
@@ -0,0 +1,683 @@
+# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/LuChengTHU/dpm-solver and https://github.com/NVlabs/edm
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin, SchedulerOutput
+
+
+class EDMDPMSolverMultistepScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Implements DPMSolverMultistepScheduler in EDM formulation as presented in Karras et al. 2022 [1].
+ `EDMDPMSolverMultistepScheduler` is a fast dedicated high-order solver for diffusion ODEs.
+
+ [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models."
+ https://arxiv.org/abs/2206.00364
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ sigma_min (`float`, *optional*, defaults to 0.002):
+ Minimum noise magnitude in the sigma schedule. This was set to 0.002 in the EDM paper [1]; a reasonable
+ range is [0, 10].
+ sigma_max (`float`, *optional*, defaults to 80.0):
+ Maximum noise magnitude in the sigma schedule. This was set to 80.0 in the EDM paper [1]; a reasonable
+ range is [0.2, 80.0].
+ sigma_data (`float`, *optional*, defaults to 0.5):
+ The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1].
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ solver_order (`int`, defaults to 2):
+ The DPMSolver order which can be `1` or `2` or `3`. It is recommended to use `solver_order=2` for guided
+ sampling, and `solver_order=3` for unconditional sampling.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
+ `algorithm_type="dpmsolver++"`.
+ algorithm_type (`str`, defaults to `dpmsolver++`):
+ Algorithm type for the solver; can be `dpmsolver++` or `sde-dpmsolver++`. The
+ `dpmsolver++` type implements the algorithms in the
+ [DPMSolver++](https://huggingface.co/papers/2211.01095) paper. It is recommended to use `dpmsolver++` or
+ `sde-dpmsolver++` with `solver_order=2` for guided sampling like in Stable Diffusion.
+ solver_type (`str`, defaults to `midpoint`):
+ Solver type for the second-order solver; can be `midpoint` or `heun`. The solver type slightly affects the
+ sample quality, especially for a small number of steps. It is recommended to use `midpoint` solvers.
+ lower_order_final (`bool`, defaults to `True`):
+ Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
+ stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
+ euler_at_final (`bool`, defaults to `False`):
+ Whether to use Euler's method in the final step. It is a trade-off between numerical stability and detail
+ richness. This can stabilize the sampling of the SDE variant of DPMSolver for small number of inference
+ steps, but sometimes may result in blurring.
+ final_sigmas_type (`str`, defaults to `"zero"`):
+ The final `sigma` value for the noise schedule during the sampling process. If `"sigma_min"`, the final sigma
+ is the same as the last sigma in the training schedule. If `zero`, the final sigma is set to 0.
+ """
+
+ _compatibles = []
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ sigma_min: float = 0.002,
+ sigma_max: float = 80.0,
+ sigma_data: float = 0.5,
+ num_train_timesteps: int = 1000,
+ prediction_type: str = "epsilon",
+ rho: float = 7.0,
+ solver_order: int = 2,
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ algorithm_type: str = "dpmsolver++",
+ solver_type: str = "midpoint",
+ lower_order_final: bool = True,
+ euler_at_final: bool = False,
+ final_sigmas_type: Optional[str] = "zero", # "zero", "sigma_min"
+ ):
+ # settings for DPM-Solver
+ if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"]:
+ if algorithm_type == "deis":
+ self.register_to_config(algorithm_type="dpmsolver++")
+ else:
+ raise NotImplementedError(f"{algorithm_type} is not implemented for {self.__class__}")
+
+ if solver_type not in ["midpoint", "heun"]:
+ if solver_type in ["logrho", "bh1", "bh2"]:
+ self.register_to_config(solver_type="midpoint")
+ else:
+ raise NotImplementedError(f"{solver_type} does is not implemented for {self.__class__}")
+
+ if algorithm_type not in ["dpmsolver++", "sde-dpmsolver++"] and final_sigmas_type == "zero":
+ raise ValueError(
+ f"`final_sigmas_type` {final_sigmas_type} is not supported for `algorithm_type` {algorithm_type}. Please choose `sigma_min` instead."
+ )
+
+ ramp = torch.linspace(0, 1, num_train_timesteps)
+ sigmas = self._compute_sigmas(ramp)
+ self.timesteps = self.precondition_noise(sigmas)
+
+ self.sigmas = self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+
+ # setable values
+ self.num_inference_steps = None
+ self.model_outputs = [None] * solver_order
+ self.lower_order_nums = 0
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ return (self.config.sigma_max**2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_inputs
+ def precondition_inputs(self, sample, sigma):
+ c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+ scaled_sample = sample * c_in
+ return scaled_sample
+
+ # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_noise
+ def precondition_noise(self, sigma):
+ if not isinstance(sigma, torch.Tensor):
+ sigma = torch.tensor([sigma])
+
+ c_noise = 0.25 * torch.log(sigma)
+
+ return c_noise
+
+ # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.precondition_outputs
+ def precondition_outputs(self, sample, model_output, sigma):
+ sigma_data = self.config.sigma_data
+ c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
+
+ if self.config.prediction_type == "epsilon":
+ c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+ elif self.config.prediction_type == "v_prediction":
+ c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+ else:
+ raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
+
+ denoised = c_skip * sample + c_out * model_output
+
+ return denoised
+
+ # Copied from diffusers.schedulers.scheduling_edm_euler.EDMEulerScheduler.scale_model_input
+ def scale_model_input(
+ self, sample: torch.FloatTensor, timestep: Union[float, torch.FloatTensor]
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+ sample = self.precondition_inputs(sample, sigma)
+
+ self.is_scale_input_called = True
+ return sample
+
+ def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+
+ self.num_inference_steps = num_inference_steps
+
+ ramp = np.linspace(0, 1, self.num_inference_steps)
+ sigmas = self._compute_sigmas(ramp)
+
+ sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
+ self.timesteps = self.precondition_noise(sigmas)
+
+ if self.config.final_sigmas_type == "sigma_min":
+ sigma_last = self.config.sigma_min
+ elif self.config.final_sigmas_type == "zero":
+ sigma_last = 0
+ else:
+ raise ValueError(
+ f"`final_sigmas_type` must be one of 'zero', or 'sigma_min', but got {self.config.final_sigmas_type}"
+ )
+
+ self.sigmas = torch.cat([sigmas, torch.tensor([sigma_last], dtype=torch.float32, device=device)])
+
+ self.model_outputs = [
+ None,
+ ] * self.config.solver_order
+ self.lower_order_nums = 0
+
+ # add an index counter for schedulers that allow duplicated timesteps
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Taken from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L17
+ def _compute_sigmas(self, ramp, sigma_min=None, sigma_max=None) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ sigma_min = sigma_min or self.config.sigma_min
+ sigma_max = sigma_max or self.config.sigma_max
+
+ rho = self.config.rho
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ def _sigma_to_alpha_sigma_t(self, sigma):
+ alpha_t = torch.tensor(1) # Inputs are pre-scaled before going into unet, so alpha_t = 1
+ sigma_t = sigma
+
+ return alpha_t, sigma_t
+
+ def convert_model_output(
+ self,
+ model_output: torch.FloatTensor,
+ sample: torch.FloatTensor = None,
+ ) -> torch.FloatTensor:
+ """
+ Convert the model output to the corresponding type the DPMSolver/DPMSolver++ algorithm needs. DPM-Solver is
+ designed to discretize an integral of the noise prediction model, and DPM-Solver++ is designed to discretize an
+ integral of the data prediction model.
+
+
+
+ The algorithm and model type are decoupled. You can use either DPMSolver or DPMSolver++ for both noise
+ prediction and data prediction models.
+
+
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The converted model output.
+ """
+ sigma = self.sigmas[self.step_index]
+ x0_pred = self.precondition_outputs(sample, model_output, sigma)
+
+ if self.config.thresholding:
+ x0_pred = self._threshold_sample(x0_pred)
+
+ return x0_pred
+
+ def dpm_solver_first_order_update(
+ self,
+ model_output: torch.FloatTensor,
+ sample: torch.FloatTensor = None,
+ noise: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ """
+ One step for the first-order DPMSolver (equivalent to DDIM).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ sigma_t, sigma_s = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s, sigma_s = self._sigma_to_alpha_sigma_t(sigma_s)
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s = torch.log(alpha_s) - torch.log(sigma_s)
+
+ h = lambda_t - lambda_s
+ if self.config.algorithm_type == "dpmsolver++":
+ x_t = (sigma_t / sigma_s) * sample - (alpha_t * (torch.exp(-h) - 1.0)) * model_output
+ elif self.config.algorithm_type == "sde-dpmsolver++":
+ assert noise is not None
+ x_t = (
+ (sigma_t / sigma_s * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+
+ return x_t
+
+ def multistep_dpm_solver_second_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ sample: torch.FloatTensor = None,
+ noise: Optional[torch.FloatTensor] = None,
+ ) -> torch.FloatTensor:
+ """
+ One step for the second-order multistep DPMSolver.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ sigma_t, sigma_s0, sigma_s1 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+
+ m0, m1 = model_output_list[-1], model_output_list[-2]
+
+ h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1
+ r0 = h_0 / h
+ D0, D1 = m0, (1.0 / r0) * (m0 - m1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2211.01095 for detailed derivations
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ - 0.5 * (alpha_t * (torch.exp(-h) - 1.0)) * D1
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+ )
+ elif self.config.algorithm_type == "sde-dpmsolver++":
+ assert noise is not None
+ if self.config.solver_type == "midpoint":
+ x_t = (
+ (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
+ + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+ elif self.config.solver_type == "heun":
+ x_t = (
+ (sigma_t / sigma_s0 * torch.exp(-h)) * sample
+ + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0
+ + (alpha_t * ((1.0 - torch.exp(-2.0 * h)) / (-2.0 * h) + 1.0)) * D1
+ + sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) * noise
+ )
+
+ return x_t
+
+ def multistep_dpm_solver_third_order_update(
+ self,
+ model_output_list: List[torch.FloatTensor],
+ sample: torch.FloatTensor = None,
+ ) -> torch.FloatTensor:
+ """
+ One step for the third-order multistep DPMSolver.
+
+ Args:
+ model_output_list (`List[torch.FloatTensor]`):
+ The direct outputs from learned diffusion model at current and latter timesteps.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ sigma_t, sigma_s0, sigma_s1, sigma_s2 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ self.sigmas[self.step_index - 2],
+ )
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ alpha_s1, sigma_s1 = self._sigma_to_alpha_sigma_t(sigma_s1)
+ alpha_s2, sigma_s2 = self._sigma_to_alpha_sigma_t(sigma_s2)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ lambda_s1 = torch.log(alpha_s1) - torch.log(sigma_s1)
+ lambda_s2 = torch.log(alpha_s2) - torch.log(sigma_s2)
+
+ m0, m1, m2 = model_output_list[-1], model_output_list[-2], model_output_list[-3]
+
+ h, h_0, h_1 = lambda_t - lambda_s0, lambda_s0 - lambda_s1, lambda_s1 - lambda_s2
+ r0, r1 = h_0 / h, h_1 / h
+ D0 = m0
+ D1_0, D1_1 = (1.0 / r0) * (m0 - m1), (1.0 / r1) * (m1 - m2)
+ D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
+ D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
+ if self.config.algorithm_type == "dpmsolver++":
+ # See https://arxiv.org/abs/2206.00927 for detailed derivations
+ x_t = (
+ (sigma_t / sigma_s0) * sample
+ - (alpha_t * (torch.exp(-h) - 1.0)) * D0
+ + (alpha_t * ((torch.exp(-h) - 1.0) / h + 1.0)) * D1
+ - (alpha_t * ((torch.exp(-h) - 1.0 + h) / h**2 - 0.5)) * D2
+ )
+
+ return x_t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ index_candidates = (schedule_timesteps == timestep).nonzero()
+
+ if len(index_candidates) == 0:
+ step_index = len(self.timesteps) - 1
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ elif len(index_candidates) > 1:
+ step_index = index_candidates[1].item()
+ else:
+ step_index = index_candidates[0].item()
+
+ return step_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ """
+ Initialize the step_index counter for the scheduler.
+ """
+
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator=None,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the multistep DPMSolver.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ # Improve numerical stability for small number of steps
+ lower_order_final = (self.step_index == len(self.timesteps) - 1) and (
+ self.config.euler_at_final
+ or (self.config.lower_order_final and len(self.timesteps) < 15)
+ or self.config.final_sigmas_type == "zero"
+ )
+ lower_order_second = (
+ (self.step_index == len(self.timesteps) - 2) and self.config.lower_order_final and len(self.timesteps) < 15
+ )
+
+ model_output = self.convert_model_output(model_output, sample=sample)
+ for i in range(self.config.solver_order - 1):
+ self.model_outputs[i] = self.model_outputs[i + 1]
+ self.model_outputs[-1] = model_output
+
+ if self.config.algorithm_type == "sde-dpmsolver++":
+ noise = randn_tensor(
+ model_output.shape, generator=generator, device=model_output.device, dtype=model_output.dtype
+ )
+ else:
+ noise = None
+
+ if self.config.solver_order == 1 or self.lower_order_nums < 1 or lower_order_final:
+ prev_sample = self.dpm_solver_first_order_update(model_output, sample=sample, noise=noise)
+ elif self.config.solver_order == 2 or self.lower_order_nums < 2 or lower_order_second:
+ prev_sample = self.multistep_dpm_solver_second_order_update(self.model_outputs, sample=sample, noise=noise)
+ else:
+ prev_sample = self.multistep_dpm_solver_third_order_update(self.model_outputs, sample=sample)
+
+ if self.lower_order_nums < self.config.solver_order:
+ self.lower_order_nums += 1
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_edm_euler.py b/model/diffusers_c/schedulers/scheduling_edm_euler.py
new file mode 100644
index 0000000000000000000000000000000000000000..e62a486cc214fc549aedf5c3b4bc77924e0ffd32
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_edm_euler.py
@@ -0,0 +1,381 @@
+# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerDiscrete
+class EDMEulerSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+class EDMEulerScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Implements the Euler scheduler in EDM formulation as presented in Karras et al. 2022 [1].
+
+ [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models."
+ https://arxiv.org/abs/2206.00364
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ sigma_min (`float`, *optional*, defaults to 0.002):
+ Minimum noise magnitude in the sigma schedule. This was set to 0.002 in the EDM paper [1]; a reasonable
+ range is [0, 10].
+ sigma_max (`float`, *optional*, defaults to 80.0):
+ Maximum noise magnitude in the sigma schedule. This was set to 80.0 in the EDM paper [1]; a reasonable
+ range is [0.2, 80.0].
+ sigma_data (`float`, *optional*, defaults to 0.5):
+ The standard deviation of the data distribution. This is set to 0.5 in the EDM paper [1].
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ rho (`float`, *optional*, defaults to 7.0):
+ The rho parameter used for calculating the Karras sigma schedule, which is set to 7.0 in the EDM paper [1].
+ """
+
+ _compatibles = []
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ sigma_min: float = 0.002,
+ sigma_max: float = 80.0,
+ sigma_data: float = 0.5,
+ num_train_timesteps: int = 1000,
+ prediction_type: str = "epsilon",
+ rho: float = 7.0,
+ ):
+ # setable values
+ self.num_inference_steps = None
+
+ ramp = torch.linspace(0, 1, num_train_timesteps)
+ sigmas = self._compute_sigmas(ramp)
+ self.timesteps = self.precondition_noise(sigmas)
+
+ self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+
+ self.is_scale_input_called = False
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ return (self.config.sigma_max**2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def precondition_inputs(self, sample, sigma):
+ c_in = 1 / ((sigma**2 + self.config.sigma_data**2) ** 0.5)
+ scaled_sample = sample * c_in
+ return scaled_sample
+
+ def precondition_noise(self, sigma):
+ if not isinstance(sigma, torch.Tensor):
+ sigma = torch.tensor([sigma])
+
+ c_noise = 0.25 * torch.log(sigma)
+
+ return c_noise
+
+ def precondition_outputs(self, sample, model_output, sigma):
+ sigma_data = self.config.sigma_data
+ c_skip = sigma_data**2 / (sigma**2 + sigma_data**2)
+
+ if self.config.prediction_type == "epsilon":
+ c_out = sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+ elif self.config.prediction_type == "v_prediction":
+ c_out = -sigma * sigma_data / (sigma**2 + sigma_data**2) ** 0.5
+ else:
+ raise ValueError(f"Prediction type {self.config.prediction_type} is not supported.")
+
+ denoised = c_skip * sample + c_out * model_output
+
+ return denoised
+
+ def scale_model_input(
+ self, sample: torch.FloatTensor, timestep: Union[float, torch.FloatTensor]
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+ sample = self.precondition_inputs(sample, sigma)
+
+ self.is_scale_input_called = True
+ return sample
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+
+ ramp = np.linspace(0, 1, self.num_inference_steps)
+ sigmas = self._compute_sigmas(ramp)
+
+ sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
+ self.timesteps = self.precondition_noise(sigmas)
+
+ self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Taken from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L17
+ def _compute_sigmas(self, ramp, sigma_min=None, sigma_max=None) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ sigma_min = sigma_min or self.config.sigma_min
+ sigma_max = sigma_max or self.config.sigma_max
+
+ rho = self.config.rho
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ sample: torch.FloatTensor,
+ s_churn: float = 0.0,
+ s_tmin: float = 0.0,
+ s_tmax: float = float("inf"),
+ s_noise: float = 1.0,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[EDMEulerSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ s_churn (`float`):
+ s_tmin (`float`):
+ s_tmax (`float`):
+ s_noise (`float`, defaults to 1.0):
+ Scaling factor for noise added to the sample.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_euler_discrete.EDMEulerSchedulerOutput`] or
+ tuple.
+
+ Returns:
+ [`~schedulers.scheduling_euler_discrete.EDMEulerSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EDMEulerSchedulerOutput`] is
+ returned, otherwise a tuple is returned where the first element is the sample tensor.
+ """
+
+ if (
+ isinstance(timestep, int)
+ or isinstance(timestep, torch.IntTensor)
+ or isinstance(timestep, torch.LongTensor)
+ ):
+ raise ValueError(
+ (
+ "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+ " `EDMEulerScheduler.step()` is not supported. Make sure to pass"
+ " one of the `scheduler.timesteps` as a timestep."
+ ),
+ )
+
+ if not self.is_scale_input_called:
+ logger.warning(
+ "The `scale_model_input` function should be called before `step` to ensure correct denoising. "
+ "See `StableDiffusionPipeline` for a usage example."
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ # Upcast to avoid precision issues when computing prev_sample
+ sample = sample.to(torch.float32)
+
+ sigma = self.sigmas[self.step_index]
+
+ gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
+
+ noise = randn_tensor(
+ model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
+ )
+
+ eps = noise * s_noise
+ sigma_hat = sigma * (gamma + 1)
+
+ if gamma > 0:
+ sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ pred_original_sample = self.precondition_outputs(sample, model_output, sigma_hat)
+
+ # 2. Convert to an ODE derivative
+ derivative = (sample - pred_original_sample) / sigma_hat
+
+ dt = self.sigmas[self.step_index + 1] - sigma_hat
+
+ prev_sample = sample + derivative * dt
+
+ # Cast sample back to model compatible dtype
+ prev_sample = prev_sample.to(model_output.dtype)
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return EDMEulerSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_euler_ancestral_discrete.py b/model/diffusers_c/schedulers/scheduling_euler_ancestral_discrete.py
new file mode 100644
index 0000000000000000000000000000000000000000..a4f3f418ebc67a79a7dc045eabb230fbc2c60a9d
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_euler_ancestral_discrete.py
@@ -0,0 +1,483 @@
+# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerAncestralDiscrete
+class EulerAncestralDiscreteSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class EulerAncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Ancestral sampling with Euler method steps.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ prediction_type: str = "epsilon",
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ rescale_betas_zero_snr: bool = False,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ if rescale_betas_zero_snr:
+ # Close to 0 without being 0 so first sigma is not inf
+ # FP16 smallest positive subnormal works well here
+ self.alphas_cumprod[-1] = 2**-24
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32)
+ self.sigmas = torch.from_numpy(sigmas)
+
+ # setable values
+ self.num_inference_steps = None
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy()
+ self.timesteps = torch.from_numpy(timesteps)
+ self.is_scale_input_called = False
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ return self.sigmas.max()
+
+ return (self.sigmas.max() ** 2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(
+ self, sample: torch.FloatTensor, timestep: Union[float, torch.FloatTensor]
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+ sample = sample / ((sigma**2 + 1) ** 0.5)
+ self.is_scale_input_called = True
+ return sample
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[
+ ::-1
+ ].copy()
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+ sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+ self.sigmas = torch.from_numpy(sigmas).to(device=device)
+
+ self.timesteps = torch.from_numpy(timesteps).to(device=device)
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ sample: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[EulerAncestralDiscreteSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`):
+ Whether or not to return a
+ [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or tuple.
+
+ Returns:
+ [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] or `tuple`:
+ If return_dict is `True`,
+ [`~schedulers.scheduling_euler_ancestral_discrete.EulerAncestralDiscreteSchedulerOutput`] is returned,
+ otherwise a tuple is returned where the first element is the sample tensor.
+
+ """
+
+ if (
+ isinstance(timestep, int)
+ or isinstance(timestep, torch.IntTensor)
+ or isinstance(timestep, torch.LongTensor)
+ ):
+ raise ValueError(
+ (
+ "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+ " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
+ " one of the `scheduler.timesteps` as a timestep."
+ ),
+ )
+
+ if not self.is_scale_input_called:
+ logger.warning(
+ "The `scale_model_input` function should be called before `step` to ensure correct denoising. "
+ "See `StableDiffusionPipeline` for a usage example."
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+
+ # Upcast to avoid precision issues when computing prev_sample
+ sample = sample.to(torch.float32)
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = sample - sigma * model_output
+ elif self.config.prediction_type == "v_prediction":
+ # * c_out + input * c_skip
+ pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
+ elif self.config.prediction_type == "sample":
+ raise NotImplementedError("prediction_type not implemented yet: sample")
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ sigma_from = self.sigmas[self.step_index]
+ sigma_to = self.sigmas[self.step_index + 1]
+ sigma_up = (sigma_to**2 * (sigma_from**2 - sigma_to**2) / sigma_from**2) ** 0.5
+ sigma_down = (sigma_to**2 - sigma_up**2) ** 0.5
+
+ # 2. Convert to an ODE derivative
+ derivative = (sample - pred_original_sample) / sigma
+
+ dt = sigma_down - sigma
+
+ prev_sample = sample + derivative * dt
+
+ device = model_output.device
+ noise = randn_tensor(model_output.shape, dtype=model_output.dtype, device=device, generator=generator)
+
+ prev_sample = prev_sample + noise * sigma_up
+
+ # Cast sample back to model compatible dtype
+ prev_sample = prev_sample.to(model_output.dtype)
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return EulerAncestralDiscreteSchedulerOutput(
+ prev_sample=prev_sample, pred_original_sample=pred_original_sample
+ )
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_euler_discrete.py b/model/diffusers_c/schedulers/scheduling_euler_discrete.py
new file mode 100644
index 0000000000000000000000000000000000000000..05ae94eff0fc989c641a58a5fb20aa89bde464cd
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_euler_discrete.py
@@ -0,0 +1,578 @@
+# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->EulerDiscrete
+class EulerDiscreteSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas):
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class EulerDiscreteScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Euler scheduler.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ interpolation_type(`str`, defaults to `"linear"`, *optional*):
+ The interpolation type to compute intermediate sigmas for the scheduler denoising steps. Should be on of
+ `"linear"` or `"log_linear"`.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ prediction_type: str = "epsilon",
+ interpolation_type: str = "linear",
+ use_karras_sigmas: Optional[bool] = False,
+ sigma_min: Optional[float] = None,
+ sigma_max: Optional[float] = None,
+ timestep_spacing: str = "linspace",
+ timestep_type: str = "discrete", # can be "discrete" or "continuous"
+ steps_offset: int = 0,
+ rescale_betas_zero_snr: bool = False,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ if rescale_betas_zero_snr:
+ # Close to 0 without being 0 so first sigma is not inf
+ # FP16 smallest positive subnormal works well here
+ self.alphas_cumprod[-1] = 2**-24
+
+ sigmas = (((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5).flip(0)
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=float)[::-1].copy()
+ timesteps = torch.from_numpy(timesteps).to(dtype=torch.float32)
+
+ # setable values
+ self.num_inference_steps = None
+
+ # TODO: Support the full EDM scalings for all prediction types and timestep types
+ if timestep_type == "continuous" and prediction_type == "v_prediction":
+ self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas])
+ else:
+ self.timesteps = timesteps
+
+ self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+
+ self.is_scale_input_called = False
+ self.use_karras_sigmas = use_karras_sigmas
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ max_sigma = max(self.sigmas) if isinstance(self.sigmas, list) else self.sigmas.max()
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ return max_sigma
+
+ return (max_sigma**2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(
+ self, sample: torch.FloatTensor, timestep: Union[float, torch.FloatTensor]
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep. Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+ sample = sample / ((sigma**2 + 1) ** 0.5)
+
+ self.is_scale_input_called = True
+ return sample
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[
+ ::-1
+ ].copy()
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ log_sigmas = np.log(sigmas)
+
+ if self.config.interpolation_type == "linear":
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+ elif self.config.interpolation_type == "log_linear":
+ sigmas = torch.linspace(np.log(sigmas[-1]), np.log(sigmas[0]), num_inference_steps + 1).exp().numpy()
+ else:
+ raise ValueError(
+ f"{self.config.interpolation_type} is not implemented. Please specify interpolation_type to either"
+ " 'linear' or 'log_linear'"
+ )
+
+ if self.use_karras_sigmas:
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+
+ sigmas = torch.from_numpy(sigmas).to(dtype=torch.float32, device=device)
+
+ # TODO: Support the full EDM scalings for all prediction types and timestep types
+ if self.config.timestep_type == "continuous" and self.config.prediction_type == "v_prediction":
+ self.timesteps = torch.Tensor([0.25 * sigma.log() for sigma in sigmas]).to(device=device)
+ else:
+ self.timesteps = torch.from_numpy(timesteps.astype(np.float32)).to(device=device)
+
+ self.sigmas = torch.cat([sigmas, torch.zeros(1, device=sigmas.device)])
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from https://github.com/crowsonkb/k-diffusion/blob/686dbad0f39640ea25c8a8c6a6e56bb40eacefa2/k_diffusion/sampling.py#L17
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ sample: torch.FloatTensor,
+ s_churn: float = 0.0,
+ s_tmin: float = 0.0,
+ s_tmax: float = float("inf"),
+ s_noise: float = 1.0,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[EulerDiscreteSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ s_churn (`float`):
+ s_tmin (`float`):
+ s_tmax (`float`):
+ s_noise (`float`, defaults to 1.0):
+ Scaling factor for noise added to the sample.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or
+ tuple.
+
+ Returns:
+ [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_euler_discrete.EulerDiscreteSchedulerOutput`] is
+ returned, otherwise a tuple is returned where the first element is the sample tensor.
+ """
+
+ if (
+ isinstance(timestep, int)
+ or isinstance(timestep, torch.IntTensor)
+ or isinstance(timestep, torch.LongTensor)
+ ):
+ raise ValueError(
+ (
+ "Passing integer indices (e.g. from `enumerate(timesteps)`) as timesteps to"
+ " `EulerDiscreteScheduler.step()` is not supported. Make sure to pass"
+ " one of the `scheduler.timesteps` as a timestep."
+ ),
+ )
+
+ if not self.is_scale_input_called:
+ logger.warning(
+ "The `scale_model_input` function should be called before `step` to ensure correct denoising. "
+ "See `StableDiffusionPipeline` for a usage example."
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ # Upcast to avoid precision issues when computing prev_sample
+ sample = sample.to(torch.float32)
+
+ sigma = self.sigmas[self.step_index]
+
+ gamma = min(s_churn / (len(self.sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigma <= s_tmax else 0.0
+
+ noise = randn_tensor(
+ model_output.shape, dtype=model_output.dtype, device=model_output.device, generator=generator
+ )
+
+ eps = noise * s_noise
+ sigma_hat = sigma * (gamma + 1)
+
+ if gamma > 0:
+ sample = sample + eps * (sigma_hat**2 - sigma**2) ** 0.5
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ # NOTE: "original_sample" should not be an expected prediction_type but is left in for
+ # backwards compatibility
+ if self.config.prediction_type == "original_sample" or self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ elif self.config.prediction_type == "epsilon":
+ pred_original_sample = sample - sigma_hat * model_output
+ elif self.config.prediction_type == "v_prediction":
+ # denoised = model_output * c_out + input * c_skip
+ pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ # 2. Convert to an ODE derivative
+ derivative = (sample - pred_original_sample) / sigma_hat
+
+ dt = self.sigmas[self.step_index + 1] - sigma_hat
+
+ prev_sample = sample + derivative * dt
+
+ # Cast sample back to model compatible dtype
+ prev_sample = prev_sample.to(model_output.dtype)
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return EulerDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_euler_discrete_flax.py b/model/diffusers_c/schedulers/scheduling_euler_discrete_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..55b0c2460a81d1d9ca01fa93a0f0061066a7fe63
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_euler_discrete_flax.py
@@ -0,0 +1,265 @@
+# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import flax
+import jax.numpy as jnp
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils_flax import (
+ CommonSchedulerState,
+ FlaxKarrasDiffusionSchedulers,
+ FlaxSchedulerMixin,
+ FlaxSchedulerOutput,
+ broadcast_to_shape_from_left,
+)
+
+
+@flax.struct.dataclass
+class EulerDiscreteSchedulerState:
+ common: CommonSchedulerState
+
+ # setable values
+ init_noise_sigma: jnp.ndarray
+ timesteps: jnp.ndarray
+ sigmas: jnp.ndarray
+ num_inference_steps: Optional[int] = None
+
+ @classmethod
+ def create(
+ cls, common: CommonSchedulerState, init_noise_sigma: jnp.ndarray, timesteps: jnp.ndarray, sigmas: jnp.ndarray
+ ):
+ return cls(common=common, init_noise_sigma=init_noise_sigma, timesteps=timesteps, sigmas=sigmas)
+
+
+@dataclass
+class FlaxEulerDiscreteSchedulerOutput(FlaxSchedulerOutput):
+ state: EulerDiscreteSchedulerState
+
+
+class FlaxEulerDiscreteScheduler(FlaxSchedulerMixin, ConfigMixin):
+ """
+ Euler scheduler (Algorithm 2) from Karras et al. (2022) https://arxiv.org/abs/2206.00364. . Based on the original
+ k-diffusion implementation by Katherine Crowson:
+ https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L51
+
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ beta_start (`float`): the starting `beta` value of inference.
+ beta_end (`float`): the final `beta` value.
+ beta_schedule (`str`):
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`jnp.ndarray`, optional):
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+ prediction_type (`str`, default `epsilon`, optional):
+ prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion
+ process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4
+ https://imagen.research.google/video/paper.pdf)
+ dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`):
+ the `dtype` used for params and computation.
+ """
+
+ _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers]
+
+ dtype: jnp.dtype
+
+ @property
+ def has_state(self):
+ return True
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[jnp.ndarray] = None,
+ prediction_type: str = "epsilon",
+ timestep_spacing: str = "linspace",
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ self.dtype = dtype
+
+ def create_state(self, common: Optional[CommonSchedulerState] = None) -> EulerDiscreteSchedulerState:
+ if common is None:
+ common = CommonSchedulerState.create(self)
+
+ timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1]
+ sigmas = ((1 - common.alphas_cumprod) / common.alphas_cumprod) ** 0.5
+ sigmas = jnp.interp(timesteps, jnp.arange(0, len(sigmas)), sigmas)
+ sigmas = jnp.concatenate([sigmas, jnp.array([0.0], dtype=self.dtype)])
+
+ # standard deviation of the initial noise distribution
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ init_noise_sigma = sigmas.max()
+ else:
+ init_noise_sigma = (sigmas.max() ** 2 + 1) ** 0.5
+
+ return EulerDiscreteSchedulerState.create(
+ common=common,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ sigmas=sigmas,
+ )
+
+ def scale_model_input(self, state: EulerDiscreteSchedulerState, sample: jnp.ndarray, timestep: int) -> jnp.ndarray:
+ """
+ Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the Euler algorithm.
+
+ Args:
+ state (`EulerDiscreteSchedulerState`):
+ the `FlaxEulerDiscreteScheduler` state data class instance.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ timestep (`int`):
+ current discrete timestep in the diffusion chain.
+
+ Returns:
+ `jnp.ndarray`: scaled input sample
+ """
+ (step_index,) = jnp.where(state.timesteps == timestep, size=1)
+ step_index = step_index[0]
+
+ sigma = state.sigmas[step_index]
+ sample = sample / ((sigma**2 + 1) ** 0.5)
+ return sample
+
+ def set_timesteps(
+ self, state: EulerDiscreteSchedulerState, num_inference_steps: int, shape: Tuple = ()
+ ) -> EulerDiscreteSchedulerState:
+ """
+ Sets the timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ state (`EulerDiscreteSchedulerState`):
+ the `FlaxEulerDiscreteScheduler` state data class instance.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ """
+
+ if self.config.timestep_spacing == "linspace":
+ timesteps = jnp.linspace(self.config.num_train_timesteps - 1, 0, num_inference_steps, dtype=self.dtype)
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // num_inference_steps
+ timesteps = (jnp.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(float)
+ timesteps += 1
+ else:
+ raise ValueError(
+ f"timestep_spacing must be one of ['linspace', 'leading'], got {self.config.timestep_spacing}"
+ )
+
+ sigmas = ((1 - state.common.alphas_cumprod) / state.common.alphas_cumprod) ** 0.5
+ sigmas = jnp.interp(timesteps, jnp.arange(0, len(sigmas)), sigmas)
+ sigmas = jnp.concatenate([sigmas, jnp.array([0.0], dtype=self.dtype)])
+
+ # standard deviation of the initial noise distribution
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ init_noise_sigma = sigmas.max()
+ else:
+ init_noise_sigma = (sigmas.max() ** 2 + 1) ** 0.5
+
+ return state.replace(
+ timesteps=timesteps,
+ sigmas=sigmas,
+ num_inference_steps=num_inference_steps,
+ init_noise_sigma=init_noise_sigma,
+ )
+
+ def step(
+ self,
+ state: EulerDiscreteSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ return_dict: bool = True,
+ ) -> Union[FlaxEulerDiscreteSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ state (`EulerDiscreteSchedulerState`):
+ the `FlaxEulerDiscreteScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ order: coefficient for multi-step inference.
+ return_dict (`bool`): option for returning tuple rather than FlaxEulerDiscreteScheduler class
+
+ Returns:
+ [`FlaxEulerDiscreteScheduler`] or `tuple`: [`FlaxEulerDiscreteScheduler`] if `return_dict` is True,
+ otherwise a `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+ if state.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ (step_index,) = jnp.where(state.timesteps == timestep, size=1)
+ step_index = step_index[0]
+
+ sigma = state.sigmas[step_index]
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = sample - sigma * model_output
+ elif self.config.prediction_type == "v_prediction":
+ # * c_out + input * c_skip
+ pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ # 2. Convert to an ODE derivative
+ derivative = (sample - pred_original_sample) / sigma
+
+ # dt = sigma_down - sigma
+ dt = state.sigmas[step_index + 1] - sigma
+
+ prev_sample = sample + derivative * dt
+
+ if not return_dict:
+ return (prev_sample, state)
+
+ return FlaxEulerDiscreteSchedulerOutput(prev_sample=prev_sample, state=state)
+
+ def add_noise(
+ self,
+ state: EulerDiscreteSchedulerState,
+ original_samples: jnp.ndarray,
+ noise: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ) -> jnp.ndarray:
+ sigma = state.sigmas[timesteps].flatten()
+ sigma = broadcast_to_shape_from_left(sigma, noise.shape)
+
+ noisy_samples = original_samples + noise * sigma
+
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_heun_discrete.py b/model/diffusers_c/schedulers/scheduling_heun_discrete.py
new file mode 100644
index 0000000000000000000000000000000000000000..aea512afc257036e9e318d802900181d978a0d23
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_heun_discrete.py
@@ -0,0 +1,484 @@
+# Copyright 2024 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class HeunDiscreteScheduler(SchedulerMixin, ConfigMixin):
+ """
+ Scheduler with Heun steps for discrete beta schedules.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ clip_sample (`bool`, defaults to `True`):
+ Clip the predicted sample for numerical stability.
+ clip_sample_range (`float`, defaults to 1.0):
+ The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 2
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.00085, # sensible defaults
+ beta_end: float = 0.012,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ prediction_type: str = "epsilon",
+ use_karras_sigmas: Optional[bool] = False,
+ clip_sample: Optional[bool] = False,
+ clip_sample_range: float = 1.0,
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps, alpha_transform_type="cosine")
+ elif beta_schedule == "exp":
+ self.betas = betas_for_alpha_bar(num_train_timesteps, alpha_transform_type="exp")
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # set all values
+ self.set_timesteps(num_train_timesteps, None, num_train_timesteps)
+ self.use_karras_sigmas = use_karras_sigmas
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ return self.sigmas.max()
+
+ return (self.sigmas.max() ** 2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+ sample = sample / ((sigma**2 + 1) ** 0.5)
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: int,
+ device: Union[str, torch.device] = None,
+ num_train_timesteps: Optional[int] = None,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+
+ num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy()
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ log_sigmas = np.log(sigmas)
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+
+ if self.config.use_karras_sigmas:
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=self.num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+
+ sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+ sigmas = torch.from_numpy(sigmas).to(device=device)
+ self.sigmas = torch.cat([sigmas[:1], sigmas[1:-1].repeat_interleave(2), sigmas[-1:]])
+
+ timesteps = torch.from_numpy(timesteps)
+ timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)])
+
+ self.timesteps = timesteps.to(device=device)
+
+ # empty dt and derivative
+ self.prev_derivative = None
+ self.dt = None
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ @property
+ def state_in_first_order(self):
+ return self.dt is None
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: Union[torch.FloatTensor, np.ndarray],
+ timestep: Union[float, torch.FloatTensor],
+ sample: Union[torch.FloatTensor, np.ndarray],
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ if self.state_in_first_order:
+ sigma = self.sigmas[self.step_index]
+ sigma_next = self.sigmas[self.step_index + 1]
+ else:
+ # 2nd order / Heun's method
+ sigma = self.sigmas[self.step_index - 1]
+ sigma_next = self.sigmas[self.step_index]
+
+ # currently only gamma=0 is supported. This usually works best anyways.
+ # We can support gamma in the future but then need to scale the timestep before
+ # passing it to the model which requires a change in API
+ gamma = 0
+ sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ if self.config.prediction_type == "epsilon":
+ sigma_input = sigma_hat if self.state_in_first_order else sigma_next
+ pred_original_sample = sample - sigma_input * model_output
+ elif self.config.prediction_type == "v_prediction":
+ sigma_input = sigma_hat if self.state_in_first_order else sigma_next
+ pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + (
+ sample / (sigma_input**2 + 1)
+ )
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ if self.config.clip_sample:
+ pred_original_sample = pred_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ if self.state_in_first_order:
+ # 2. Convert to an ODE derivative for 1st order
+ derivative = (sample - pred_original_sample) / sigma_hat
+ # 3. delta timestep
+ dt = sigma_next - sigma_hat
+
+ # store for 2nd order step
+ self.prev_derivative = derivative
+ self.dt = dt
+ self.sample = sample
+ else:
+ # 2. 2nd order / Heun's method
+ derivative = (sample - pred_original_sample) / sigma_next
+ derivative = (self.prev_derivative + derivative) / 2
+
+ # 3. take prev timestep & sample
+ dt = self.dt
+ sample = self.sample
+
+ # free dt and derivative
+ # Note, this puts the scheduler in "first order mode"
+ self.prev_derivative = None
+ self.dt = None
+ self.sample = None
+
+ prev_sample = sample + derivative * dt
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_ipndm.py b/model/diffusers_c/schedulers/scheduling_ipndm.py
new file mode 100644
index 0000000000000000000000000000000000000000..583afa4d256dc90ad2e835b6215f6b21cfa35756
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_ipndm.py
@@ -0,0 +1,224 @@
+# Copyright 2024 Zhejiang University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils import SchedulerMixin, SchedulerOutput
+
+
+class IPNDMScheduler(SchedulerMixin, ConfigMixin):
+ """
+ A fourth-order Improved Pseudo Linear Multistep scheduler.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ """
+
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self, num_train_timesteps: int = 1000, trained_betas: Optional[Union[np.ndarray, List[float]]] = None
+ ):
+ # set `betas`, `alphas`, `timesteps`
+ self.set_timesteps(num_train_timesteps)
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # For now we only support F-PNDM, i.e. the runge-kutta method
+ # For more information on the algorithm please take a look at the paper: https://arxiv.org/pdf/2202.09778.pdf
+ # mainly at formula (9), (12), (13) and the Algorithm 2.
+ self.pndm_order = 4
+
+ # running values
+ self.ets = []
+ self._step_index = None
+ self._begin_index = None
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+ steps = torch.linspace(1, 0, num_inference_steps + 1)[:-1]
+ steps = torch.cat([steps, torch.tensor([0.0])])
+
+ if self.config.trained_betas is not None:
+ self.betas = torch.tensor(self.config.trained_betas, dtype=torch.float32)
+ else:
+ self.betas = torch.sin(steps * math.pi / 2) ** 2
+
+ self.alphas = (1.0 - self.betas**2) ** 0.5
+
+ timesteps = (torch.atan2(self.betas, self.alphas) / math.pi * 2)[:-1]
+ self.timesteps = timesteps.to(device)
+
+ self.ets = []
+ self._step_index = None
+ self._begin_index = None
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the linear multistep method. It performs one forward pass multiple times to approximate the solution.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ timestep_index = self.step_index
+ prev_timestep_index = self.step_index + 1
+
+ ets = sample * self.betas[timestep_index] + model_output * self.alphas[timestep_index]
+ self.ets.append(ets)
+
+ if len(self.ets) == 1:
+ ets = self.ets[-1]
+ elif len(self.ets) == 2:
+ ets = (3 * self.ets[-1] - self.ets[-2]) / 2
+ elif len(self.ets) == 3:
+ ets = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12
+ else:
+ ets = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4])
+
+ prev_sample = self._get_prev_sample(sample, timestep_index, prev_timestep_index, ets)
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def _get_prev_sample(self, sample, timestep_index, prev_timestep_index, ets):
+ alpha = self.alphas[timestep_index]
+ sigma = self.betas[timestep_index]
+
+ next_alpha = self.alphas[prev_timestep_index]
+ next_sigma = self.betas[prev_timestep_index]
+
+ pred = (sample - sigma * ets) / max(alpha, 1e-8)
+ prev_sample = next_alpha * pred + ets * next_sigma
+
+ return prev_sample
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_k_dpm_2_ancestral_discrete.py b/model/diffusers_c/schedulers/scheduling_k_dpm_2_ancestral_discrete.py
new file mode 100644
index 0000000000000000000000000000000000000000..8ba3d2336db24c1b2a878e448cd9a15f76cda5ca
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_k_dpm_2_ancestral_discrete.py
@@ -0,0 +1,510 @@
+# Copyright 2024 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class KDPM2AncestralDiscreteScheduler(SchedulerMixin, ConfigMixin):
+ """
+ KDPM2DiscreteScheduler with ancestral sampling is inspired by the DPMSolver2 and Algorithm 2 from the [Elucidating
+ the Design Space of Diffusion-Based Generative Models](https://huggingface.co/papers/2206.00364) paper.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.00085):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.012):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 2
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.00085, # sensible defaults
+ beta_end: float = 0.012,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ use_karras_sigmas: Optional[bool] = False,
+ prediction_type: str = "epsilon",
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # set all values
+ self.set_timesteps(num_train_timesteps, None, num_train_timesteps)
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ return self.sigmas.max()
+
+ return (self.sigmas.max() ** 2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ if self.state_in_first_order:
+ sigma = self.sigmas[self.step_index]
+ else:
+ sigma = self.sigmas_interpol[self.step_index - 1]
+
+ sample = sample / ((sigma**2 + 1) ** 0.5)
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: int,
+ device: Union[str, torch.device] = None,
+ num_train_timesteps: Optional[int] = None,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+
+ num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy()
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ log_sigmas = np.log(sigmas)
+
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+
+ if self.config.use_karras_sigmas:
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+
+ self.log_sigmas = torch.from_numpy(log_sigmas).to(device)
+ sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+ sigmas = torch.from_numpy(sigmas).to(device=device)
+
+ # compute up and down sigmas
+ sigmas_next = sigmas.roll(-1)
+ sigmas_next[-1] = 0.0
+ sigmas_up = (sigmas_next**2 * (sigmas**2 - sigmas_next**2) / sigmas**2) ** 0.5
+ sigmas_down = (sigmas_next**2 - sigmas_up**2) ** 0.5
+ sigmas_down[-1] = 0.0
+
+ # compute interpolated sigmas
+ sigmas_interpol = sigmas.log().lerp(sigmas_down.log(), 0.5).exp()
+ sigmas_interpol[-2:] = 0.0
+
+ # set sigmas
+ self.sigmas = torch.cat([sigmas[:1], sigmas[1:].repeat_interleave(2), sigmas[-1:]])
+ self.sigmas_interpol = torch.cat(
+ [sigmas_interpol[:1], sigmas_interpol[1:].repeat_interleave(2), sigmas_interpol[-1:]]
+ )
+ self.sigmas_up = torch.cat([sigmas_up[:1], sigmas_up[1:].repeat_interleave(2), sigmas_up[-1:]])
+ self.sigmas_down = torch.cat([sigmas_down[:1], sigmas_down[1:].repeat_interleave(2), sigmas_down[-1:]])
+
+ if str(device).startswith("mps"):
+ timesteps = torch.from_numpy(timesteps).to(device, dtype=torch.float32)
+ else:
+ timesteps = torch.from_numpy(timesteps).to(device)
+
+ sigmas_interpol = sigmas_interpol.cpu()
+ log_sigmas = self.log_sigmas.cpu()
+ timesteps_interpol = np.array(
+ [self._sigma_to_t(sigma_interpol, log_sigmas) for sigma_interpol in sigmas_interpol]
+ )
+
+ timesteps_interpol = torch.from_numpy(timesteps_interpol).to(device, dtype=timesteps.dtype)
+ interleaved_timesteps = torch.stack((timesteps_interpol[:-2, None], timesteps[1:, None]), dim=-1).flatten()
+
+ self.timesteps = torch.cat([timesteps[:1], interleaved_timesteps])
+
+ self.sample = None
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ @property
+ def state_in_first_order(self):
+ return self.sample is None
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: Union[torch.FloatTensor, np.ndarray],
+ timestep: Union[float, torch.FloatTensor],
+ sample: Union[torch.FloatTensor, np.ndarray],
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_ddim.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ if self.state_in_first_order:
+ sigma = self.sigmas[self.step_index]
+ sigma_interpol = self.sigmas_interpol[self.step_index]
+ sigma_up = self.sigmas_up[self.step_index]
+ sigma_down = self.sigmas_down[self.step_index - 1]
+ else:
+ # 2nd order / KPDM2's method
+ sigma = self.sigmas[self.step_index - 1]
+ sigma_interpol = self.sigmas_interpol[self.step_index - 1]
+ sigma_up = self.sigmas_up[self.step_index - 1]
+ sigma_down = self.sigmas_down[self.step_index - 1]
+
+ # currently only gamma=0 is supported. This usually works best anyways.
+ # We can support gamma in the future but then need to scale the timestep before
+ # passing it to the model which requires a change in API
+ gamma = 0
+ sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now
+
+ device = model_output.device
+ noise = randn_tensor(model_output.shape, dtype=model_output.dtype, device=device, generator=generator)
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ if self.config.prediction_type == "epsilon":
+ sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol
+ pred_original_sample = sample - sigma_input * model_output
+ elif self.config.prediction_type == "v_prediction":
+ sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol
+ pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + (
+ sample / (sigma_input**2 + 1)
+ )
+ elif self.config.prediction_type == "sample":
+ raise NotImplementedError("prediction_type not implemented yet: sample")
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ if self.state_in_first_order:
+ # 2. Convert to an ODE derivative for 1st order
+ derivative = (sample - pred_original_sample) / sigma_hat
+ # 3. delta timestep
+ dt = sigma_interpol - sigma_hat
+
+ # store for 2nd order step
+ self.sample = sample
+ self.dt = dt
+ prev_sample = sample + derivative * dt
+ else:
+ # DPM-Solver-2
+ # 2. Convert to an ODE derivative for 2nd order
+ derivative = (sample - pred_original_sample) / sigma_interpol
+ # 3. delta timestep
+ dt = sigma_down - sigma_hat
+
+ sample = self.sample
+ self.sample = None
+
+ prev_sample = sample + derivative * dt
+ prev_sample = prev_sample + noise * sigma_up
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_k_dpm_2_discrete.py b/model/diffusers_c/schedulers/scheduling_k_dpm_2_discrete.py
new file mode 100644
index 0000000000000000000000000000000000000000..e4fa881dd5d90ee5379110e52272e1e17260693c
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_k_dpm_2_discrete.py
@@ -0,0 +1,485 @@
+# Copyright 2024 Katherine Crowson, The HuggingFace Team and hlky. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class KDPM2DiscreteScheduler(SchedulerMixin, ConfigMixin):
+ """
+ KDPM2DiscreteScheduler is inspired by the DPMSolver2 and Algorithm 2 from the [Elucidating the Design Space of
+ Diffusion-Based Generative Models](https://huggingface.co/papers/2206.00364) paper.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.00085):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.012):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 2
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.00085, # sensible defaults
+ beta_end: float = 0.012,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ use_karras_sigmas: Optional[bool] = False,
+ prediction_type: str = "epsilon",
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # set all values
+ self.set_timesteps(num_train_timesteps, None, num_train_timesteps)
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ return self.sigmas.max()
+
+ return (self.sigmas.max() ** 2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(
+ self,
+ sample: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ if self.state_in_first_order:
+ sigma = self.sigmas[self.step_index]
+ else:
+ sigma = self.sigmas_interpol[self.step_index]
+
+ sample = sample / ((sigma**2 + 1) ** 0.5)
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: int,
+ device: Union[str, torch.device] = None,
+ num_train_timesteps: Optional[int] = None,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+
+ num_train_timesteps = num_train_timesteps or self.config.num_train_timesteps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[::-1].copy()
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ log_sigmas = np.log(sigmas)
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+
+ if self.config.use_karras_sigmas:
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+
+ self.log_sigmas = torch.from_numpy(log_sigmas).to(device=device)
+ sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+ sigmas = torch.from_numpy(sigmas).to(device=device)
+
+ # interpolate sigmas
+ sigmas_interpol = sigmas.log().lerp(sigmas.roll(1).log(), 0.5).exp()
+
+ self.sigmas = torch.cat([sigmas[:1], sigmas[1:].repeat_interleave(2), sigmas[-1:]])
+ self.sigmas_interpol = torch.cat(
+ [sigmas_interpol[:1], sigmas_interpol[1:].repeat_interleave(2), sigmas_interpol[-1:]]
+ )
+
+ timesteps = torch.from_numpy(timesteps).to(device)
+
+ # interpolate timesteps
+ sigmas_interpol = sigmas_interpol.cpu()
+ log_sigmas = self.log_sigmas.cpu()
+ timesteps_interpol = np.array(
+ [self._sigma_to_t(sigma_interpol, log_sigmas) for sigma_interpol in sigmas_interpol]
+ )
+ timesteps_interpol = torch.from_numpy(timesteps_interpol).to(device, dtype=timesteps.dtype)
+ interleaved_timesteps = torch.stack((timesteps_interpol[1:-1, None], timesteps[1:, None]), dim=-1).flatten()
+
+ self.timesteps = torch.cat([timesteps[:1], interleaved_timesteps])
+
+ self.sample = None
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def state_in_first_order(self):
+ return self.sample is None
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def step(
+ self,
+ model_output: Union[torch.FloatTensor, np.ndarray],
+ timestep: Union[float, torch.FloatTensor],
+ sample: Union[torch.FloatTensor, np.ndarray],
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+ """
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ if self.state_in_first_order:
+ sigma = self.sigmas[self.step_index]
+ sigma_interpol = self.sigmas_interpol[self.step_index + 1]
+ sigma_next = self.sigmas[self.step_index + 1]
+ else:
+ # 2nd order / KDPM2's method
+ sigma = self.sigmas[self.step_index - 1]
+ sigma_interpol = self.sigmas_interpol[self.step_index]
+ sigma_next = self.sigmas[self.step_index]
+
+ # currently only gamma=0 is supported. This usually works best anyways.
+ # We can support gamma in the future but then need to scale the timestep before
+ # passing it to the model which requires a change in API
+ gamma = 0
+ sigma_hat = sigma * (gamma + 1) # Note: sigma_hat == sigma for now
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ if self.config.prediction_type == "epsilon":
+ sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol
+ pred_original_sample = sample - sigma_input * model_output
+ elif self.config.prediction_type == "v_prediction":
+ sigma_input = sigma_hat if self.state_in_first_order else sigma_interpol
+ pred_original_sample = model_output * (-sigma_input / (sigma_input**2 + 1) ** 0.5) + (
+ sample / (sigma_input**2 + 1)
+ )
+ elif self.config.prediction_type == "sample":
+ raise NotImplementedError("prediction_type not implemented yet: sample")
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ if self.state_in_first_order:
+ # 2. Convert to an ODE derivative for 1st order
+ derivative = (sample - pred_original_sample) / sigma_hat
+ # 3. delta timestep
+ dt = sigma_interpol - sigma_hat
+
+ # store for 2nd order step
+ self.sample = sample
+ else:
+ # DPM-Solver-2
+ # 2. Convert to an ODE derivative for 2nd order
+ derivative = (sample - pred_original_sample) / sigma_interpol
+
+ # 3. delta timestep
+ dt = sigma_next - sigma_hat
+
+ sample = self.sample
+ self.sample = None
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ prev_sample = sample + derivative * dt
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_karras_ve_flax.py b/model/diffusers_c/schedulers/scheduling_karras_ve_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..4d099604a9e5c6782eed66029f84af66ea7efb36
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_karras_ve_flax.py
@@ -0,0 +1,238 @@
+# Copyright 2024 NVIDIA and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import flax
+import jax
+import jax.numpy as jnp
+from jax import random
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from .scheduling_utils_flax import FlaxSchedulerMixin
+
+
+@flax.struct.dataclass
+class KarrasVeSchedulerState:
+ # setable values
+ num_inference_steps: Optional[int] = None
+ timesteps: Optional[jnp.ndarray] = None
+ schedule: Optional[jnp.ndarray] = None # sigma(t_i)
+
+ @classmethod
+ def create(cls):
+ return cls()
+
+
+@dataclass
+class FlaxKarrasVeOutput(BaseOutput):
+ """
+ Output class for the scheduler's step function output.
+
+ Args:
+ prev_sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ derivative (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images):
+ Derivative of predicted original image sample (x_0).
+ state (`KarrasVeSchedulerState`): the `FlaxKarrasVeScheduler` state data class.
+ """
+
+ prev_sample: jnp.ndarray
+ derivative: jnp.ndarray
+ state: KarrasVeSchedulerState
+
+
+class FlaxKarrasVeScheduler(FlaxSchedulerMixin, ConfigMixin):
+ """
+ Stochastic sampling from Karras et al. [1] tailored to the Variance-Expanding (VE) models [2]. Use Algorithm 2 and
+ the VE column of Table 1 from [1] for reference.
+
+ [1] Karras, Tero, et al. "Elucidating the Design Space of Diffusion-Based Generative Models."
+ https://arxiv.org/abs/2206.00364 [2] Song, Yang, et al. "Score-based generative modeling through stochastic
+ differential equations." https://arxiv.org/abs/2011.13456
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ For more details on the parameters, see the original paper's Appendix E.: "Elucidating the Design Space of
+ Diffusion-Based Generative Models." https://arxiv.org/abs/2206.00364. The grid search values used to find the
+ optimal {s_noise, s_churn, s_min, s_max} for a specific model are described in Table 5 of the paper.
+
+ Args:
+ sigma_min (`float`): minimum noise magnitude
+ sigma_max (`float`): maximum noise magnitude
+ s_noise (`float`): the amount of additional noise to counteract loss of detail during sampling.
+ A reasonable range is [1.000, 1.011].
+ s_churn (`float`): the parameter controlling the overall amount of stochasticity.
+ A reasonable range is [0, 100].
+ s_min (`float`): the start value of the sigma range where we add noise (enable stochasticity).
+ A reasonable range is [0, 10].
+ s_max (`float`): the end value of the sigma range where we add noise.
+ A reasonable range is [0.2, 80].
+ """
+
+ @property
+ def has_state(self):
+ return True
+
+ @register_to_config
+ def __init__(
+ self,
+ sigma_min: float = 0.02,
+ sigma_max: float = 100,
+ s_noise: float = 1.007,
+ s_churn: float = 80,
+ s_min: float = 0.05,
+ s_max: float = 50,
+ ):
+ pass
+
+ def create_state(self):
+ return KarrasVeSchedulerState.create()
+
+ def set_timesteps(
+ self, state: KarrasVeSchedulerState, num_inference_steps: int, shape: Tuple = ()
+ ) -> KarrasVeSchedulerState:
+ """
+ Sets the continuous timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ state (`KarrasVeSchedulerState`):
+ the `FlaxKarrasVeScheduler` state data class.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+
+ """
+ timesteps = jnp.arange(0, num_inference_steps)[::-1].copy()
+ schedule = [
+ (
+ self.config.sigma_max**2
+ * (self.config.sigma_min**2 / self.config.sigma_max**2) ** (i / (num_inference_steps - 1))
+ )
+ for i in timesteps
+ ]
+
+ return state.replace(
+ num_inference_steps=num_inference_steps,
+ schedule=jnp.array(schedule, dtype=jnp.float32),
+ timesteps=timesteps,
+ )
+
+ def add_noise_to_input(
+ self,
+ state: KarrasVeSchedulerState,
+ sample: jnp.ndarray,
+ sigma: float,
+ key: jax.Array,
+ ) -> Tuple[jnp.ndarray, float]:
+ """
+ Explicit Langevin-like "churn" step of adding noise to the sample according to a factor gamma_i ≥ 0 to reach a
+ higher noise level sigma_hat = sigma_i + gamma_i*sigma_i.
+
+ TODO Args:
+ """
+ if self.config.s_min <= sigma <= self.config.s_max:
+ gamma = min(self.config.s_churn / state.num_inference_steps, 2**0.5 - 1)
+ else:
+ gamma = 0
+
+ # sample eps ~ N(0, S_noise^2 * I)
+ key = random.split(key, num=1)
+ eps = self.config.s_noise * random.normal(key=key, shape=sample.shape)
+ sigma_hat = sigma + gamma * sigma
+ sample_hat = sample + ((sigma_hat**2 - sigma**2) ** 0.5 * eps)
+
+ return sample_hat, sigma_hat
+
+ def step(
+ self,
+ state: KarrasVeSchedulerState,
+ model_output: jnp.ndarray,
+ sigma_hat: float,
+ sigma_prev: float,
+ sample_hat: jnp.ndarray,
+ return_dict: bool = True,
+ ) -> Union[FlaxKarrasVeOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ state (`KarrasVeSchedulerState`): the `FlaxKarrasVeScheduler` state data class.
+ model_output (`torch.FloatTensor` or `np.ndarray`): direct output from learned diffusion model.
+ sigma_hat (`float`): TODO
+ sigma_prev (`float`): TODO
+ sample_hat (`torch.FloatTensor` or `np.ndarray`): TODO
+ return_dict (`bool`): option for returning tuple rather than FlaxKarrasVeOutput class
+
+ Returns:
+ [`~schedulers.scheduling_karras_ve_flax.FlaxKarrasVeOutput`] or `tuple`: Updated sample in the diffusion
+ chain and derivative. [`~schedulers.scheduling_karras_ve_flax.FlaxKarrasVeOutput`] if `return_dict` is
+ True, otherwise a `tuple`. When returning a tuple, the first element is the sample tensor.
+ """
+
+ pred_original_sample = sample_hat + sigma_hat * model_output
+ derivative = (sample_hat - pred_original_sample) / sigma_hat
+ sample_prev = sample_hat + (sigma_prev - sigma_hat) * derivative
+
+ if not return_dict:
+ return (sample_prev, derivative, state)
+
+ return FlaxKarrasVeOutput(prev_sample=sample_prev, derivative=derivative, state=state)
+
+ def step_correct(
+ self,
+ state: KarrasVeSchedulerState,
+ model_output: jnp.ndarray,
+ sigma_hat: float,
+ sigma_prev: float,
+ sample_hat: jnp.ndarray,
+ sample_prev: jnp.ndarray,
+ derivative: jnp.ndarray,
+ return_dict: bool = True,
+ ) -> Union[FlaxKarrasVeOutput, Tuple]:
+ """
+ Correct the predicted sample based on the output model_output of the network. TODO complete description
+
+ Args:
+ state (`KarrasVeSchedulerState`): the `FlaxKarrasVeScheduler` state data class.
+ model_output (`torch.FloatTensor` or `np.ndarray`): direct output from learned diffusion model.
+ sigma_hat (`float`): TODO
+ sigma_prev (`float`): TODO
+ sample_hat (`torch.FloatTensor` or `np.ndarray`): TODO
+ sample_prev (`torch.FloatTensor` or `np.ndarray`): TODO
+ derivative (`torch.FloatTensor` or `np.ndarray`): TODO
+ return_dict (`bool`): option for returning tuple rather than FlaxKarrasVeOutput class
+
+ Returns:
+ prev_sample (TODO): updated sample in the diffusion chain. derivative (TODO): TODO
+
+ """
+ pred_original_sample = sample_prev + sigma_prev * model_output
+ derivative_corr = (sample_prev - pred_original_sample) / sigma_prev
+ sample_prev = sample_hat + (sigma_prev - sigma_hat) * (0.5 * derivative + 0.5 * derivative_corr)
+
+ if not return_dict:
+ return (sample_prev, derivative, state)
+
+ return FlaxKarrasVeOutput(prev_sample=sample_prev, derivative=derivative, state=state)
+
+ def add_noise(self, state: KarrasVeSchedulerState, original_samples, noise, timesteps):
+ raise NotImplementedError()
diff --git a/model/diffusers_c/schedulers/scheduling_lcm.py b/model/diffusers_c/schedulers/scheduling_lcm.py
new file mode 100644
index 0000000000000000000000000000000000000000..5dc79321fc387097f1cd2704bd3349c228ac3153
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_lcm.py
@@ -0,0 +1,662 @@
+# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class LCMSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ denoised: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class LCMScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `LCMScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with
+ non-Markovian guidance.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config
+ attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be
+ accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving
+ functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ original_inference_steps (`int`, *optional*, defaults to 50):
+ The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we
+ will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule.
+ clip_sample (`bool`, defaults to `True`):
+ Clip the predicted sample for numerical stability.
+ clip_sample_range (`float`, defaults to 1.0):
+ The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ set_alpha_to_one (`bool`, defaults to `True`):
+ Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+ there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+ otherwise it uses the alpha value at step 0.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+ timestep_spacing (`str`, defaults to `"leading"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ timestep_scaling (`float`, defaults to 10.0):
+ The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions
+ `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation
+ error at the default of `10.0` is already pretty small).
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.00085,
+ beta_end: float = 0.012,
+ beta_schedule: str = "scaled_linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ original_inference_steps: int = 50,
+ clip_sample: bool = False,
+ clip_sample_range: float = 1.0,
+ set_alpha_to_one: bool = True,
+ steps_offset: int = 0,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ timestep_spacing: str = "leading",
+ timestep_scaling: float = 10.0,
+ rescale_betas_zero_snr: bool = False,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ # Rescale for zero SNR
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # At every step in ddim, we are looking into the previous alphas_cumprod
+ # For the final step, there is no previous alphas_cumprod because we are already at 0
+ # `set_alpha_to_one` decides whether we set this parameter simply to one or
+ # whether we use the final alpha of the "non-previous" one.
+ self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
+ self.custom_timesteps = False
+
+ self._step_index = None
+ self._begin_index = None
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ @property
+ def step_index(self):
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: Optional[int] = None,
+ device: Union[str, torch.device] = None,
+ original_inference_steps: Optional[int] = None,
+ timesteps: Optional[List[int]] = None,
+ strength: int = 1.0,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`, *optional*):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ original_inference_steps (`int`, *optional*):
+ The original number of inference steps, which will be used to generate a linearly-spaced timestep
+ schedule (which is different from the standard `diffusers` implementation). We will then take
+ `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as
+ our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of equal spacing between timesteps on the training/distillation timestep
+ schedule is used. If `timesteps` is passed, `num_inference_steps` must be `None`.
+ """
+ # 0. Check inputs
+ if num_inference_steps is None and timesteps is None:
+ raise ValueError("Must pass exactly one of `num_inference_steps` or `custom_timesteps`.")
+
+ if num_inference_steps is not None and timesteps is not None:
+ raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
+
+ # 1. Calculate the LCM original training/distillation timestep schedule.
+ original_steps = (
+ original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps
+ )
+
+ if original_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ # LCM Timesteps Setting
+ # The skipping step parameter k from the paper.
+ k = self.config.num_train_timesteps // original_steps
+ # LCM Training/Distillation Steps Schedule
+ # Currently, only a linearly-spaced schedule is supported (same as in the LCM distillation scripts).
+ lcm_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * k - 1
+
+ # 2. Calculate the LCM inference timestep schedule.
+ if timesteps is not None:
+ # 2.1 Handle custom timestep schedules.
+ train_timesteps = set(lcm_origin_timesteps)
+ non_train_timesteps = []
+ for i in range(1, len(timesteps)):
+ if timesteps[i] >= timesteps[i - 1]:
+ raise ValueError("`custom_timesteps` must be in descending order.")
+
+ if timesteps[i] not in train_timesteps:
+ non_train_timesteps.append(timesteps[i])
+
+ if timesteps[0] >= self.config.num_train_timesteps:
+ raise ValueError(
+ f"`timesteps` must start before `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps}."
+ )
+
+ # Raise warning if timestep schedule does not start with self.config.num_train_timesteps - 1
+ if strength == 1.0 and timesteps[0] != self.config.num_train_timesteps - 1:
+ logger.warning(
+ f"The first timestep on the custom timestep schedule is {timesteps[0]}, not"
+ f" `self.config.num_train_timesteps - 1`: {self.config.num_train_timesteps - 1}. You may get"
+ f" unexpected results when using this timestep schedule."
+ )
+
+ # Raise warning if custom timestep schedule contains timesteps not on original timestep schedule
+ if non_train_timesteps:
+ logger.warning(
+ f"The custom timestep schedule contains the following timesteps which are not on the original"
+ f" training/distillation timestep schedule: {non_train_timesteps}. You may get unexpected results"
+ f" when using this timestep schedule."
+ )
+
+ # Raise warning if custom timestep schedule is longer than original_steps
+ if len(timesteps) > original_steps:
+ logger.warning(
+ f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the"
+ f" the length of the timestep schedule used for training: {original_steps}. You may get some"
+ f" unexpected results when using this timestep schedule."
+ )
+
+ timesteps = np.array(timesteps, dtype=np.int64)
+ self.num_inference_steps = len(timesteps)
+ self.custom_timesteps = True
+
+ # Apply strength (e.g. for img2img pipelines) (see StableDiffusionImg2ImgPipeline.get_timesteps)
+ init_timestep = min(int(self.num_inference_steps * strength), self.num_inference_steps)
+ t_start = max(self.num_inference_steps - init_timestep, 0)
+ timesteps = timesteps[t_start * self.order :]
+ # TODO: also reset self.num_inference_steps?
+ else:
+ # 2.2 Create the "standard" LCM inference timestep schedule.
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ skipping_step = len(lcm_origin_timesteps) // num_inference_steps
+
+ if skipping_step < 1:
+ raise ValueError(
+ f"The combination of `original_steps x strength`: {original_steps} x {strength} is smaller than `num_inference_steps`: {num_inference_steps}. Make sure to either reduce `num_inference_steps` to a value smaller than {int(original_steps * strength)} or increase `strength` to a value higher than {float(num_inference_steps / original_steps)}."
+ )
+
+ self.num_inference_steps = num_inference_steps
+
+ if num_inference_steps > original_steps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:"
+ f" {original_steps} because the final timestep schedule will be a subset of the"
+ f" `original_inference_steps`-sized initial timestep schedule."
+ )
+
+ # LCM Inference Steps Schedule
+ lcm_origin_timesteps = lcm_origin_timesteps[::-1].copy()
+ # Select (approximately) evenly spaced indices from lcm_origin_timesteps.
+ inference_indices = np.linspace(0, len(lcm_origin_timesteps), num=num_inference_steps, endpoint=False)
+ inference_indices = np.floor(inference_indices).astype(np.int64)
+ timesteps = lcm_origin_timesteps[inference_indices]
+
+ self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.long)
+
+ self._step_index = None
+ self._begin_index = None
+
+ def get_scalings_for_boundary_condition_discrete(self, timestep):
+ self.sigma_data = 0.5 # Default: 0.5
+ scaled_timestep = timestep * self.config.timestep_scaling
+
+ c_skip = self.sigma_data**2 / (scaled_timestep**2 + self.sigma_data**2)
+ c_out = scaled_timestep / (scaled_timestep**2 + self.sigma_data**2) ** 0.5
+ return c_skip, c_out
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[LCMSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] or `tuple`.
+ Returns:
+ [`~schedulers.scheduling_utils.LCMSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ # 1. get previous step value
+ prev_step_index = self.step_index + 1
+ if prev_step_index < len(self.timesteps):
+ prev_timestep = self.timesteps[prev_step_index]
+ else:
+ prev_timestep = timestep
+
+ # 2. compute alphas, betas
+ alpha_prod_t = self.alphas_cumprod[timestep]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ # 3. Get scalings for boundary conditions
+ c_skip, c_out = self.get_scalings_for_boundary_condition_discrete(timestep)
+
+ # 4. Compute the predicted original sample x_0 based on the model parameterization
+ if self.config.prediction_type == "epsilon": # noise-prediction
+ predicted_original_sample = (sample - beta_prod_t.sqrt() * model_output) / alpha_prod_t.sqrt()
+ elif self.config.prediction_type == "sample": # x-prediction
+ predicted_original_sample = model_output
+ elif self.config.prediction_type == "v_prediction": # v-prediction
+ predicted_original_sample = alpha_prod_t.sqrt() * sample - beta_prod_t.sqrt() * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
+ " `v_prediction` for `LCMScheduler`."
+ )
+
+ # 5. Clip or threshold "predicted x_0"
+ if self.config.thresholding:
+ predicted_original_sample = self._threshold_sample(predicted_original_sample)
+ elif self.config.clip_sample:
+ predicted_original_sample = predicted_original_sample.clamp(
+ -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 6. Denoise model output using boundary conditions
+ denoised = c_out * predicted_original_sample + c_skip * sample
+
+ # 7. Sample and inject noise z ~ N(0, I) for MultiStep Inference
+ # Noise is not used on the final timestep of the timestep schedule.
+ # This also means that noise is not used for one-step sampling.
+ if self.step_index != self.num_inference_steps - 1:
+ noise = randn_tensor(
+ model_output.shape, generator=generator, device=model_output.device, dtype=denoised.dtype
+ )
+ prev_sample = alpha_prod_t_prev.sqrt() * denoised + beta_prod_t_prev.sqrt() * noise
+ else:
+ prev_sample = denoised
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample, denoised)
+
+ return LCMSchedulerOutput(prev_sample=prev_sample, denoised=denoised)
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+ # for the subsequent add_noise calls
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
+ def get_velocity(
+ self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as sample
+ self.alphas_cumprod = self.alphas_cumprod.to(device=sample.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=sample.dtype)
+ timesteps = timesteps.to(sample.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(sample.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+ return velocity
+
+ def __len__(self):
+ return self.config.num_train_timesteps
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.previous_timestep
+ def previous_timestep(self, timestep):
+ if self.custom_timesteps:
+ index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0]
+ if index == self.timesteps.shape[0] - 1:
+ prev_t = torch.tensor(-1)
+ else:
+ prev_t = self.timesteps[index + 1]
+ else:
+ num_inference_steps = (
+ self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps
+ )
+ prev_t = timestep - self.config.num_train_timesteps // num_inference_steps
+
+ return prev_t
diff --git a/model/diffusers_c/schedulers/scheduling_lms_discrete.py b/model/diffusers_c/schedulers/scheduling_lms_discrete.py
new file mode 100644
index 0000000000000000000000000000000000000000..65ea6e3b7c9efa74ee9b40d56f5539991a59cb13
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_lms_discrete.py
@@ -0,0 +1,477 @@
+# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+import warnings
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from scipy import integrate
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->LMSDiscrete
+class LMSDiscreteSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class LMSDiscreteScheduler(SchedulerMixin, ConfigMixin):
+ """
+ A linear multistep scheduler for discrete beta schedules.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ use_karras_sigmas: Optional[bool] = False,
+ prediction_type: str = "epsilon",
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ sigmas = np.concatenate([sigmas[::-1], [0.0]]).astype(np.float32)
+ self.sigmas = torch.from_numpy(sigmas)
+
+ # setable values
+ self.num_inference_steps = None
+ self.use_karras_sigmas = use_karras_sigmas
+ self.set_timesteps(num_train_timesteps, None)
+ self.derivatives = []
+ self.is_scale_input_called = False
+
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def init_noise_sigma(self):
+ # standard deviation of the initial noise distribution
+ if self.config.timestep_spacing in ["linspace", "trailing"]:
+ return self.sigmas.max()
+
+ return (self.sigmas.max() ** 2 + 1) ** 0.5
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(
+ self, sample: torch.FloatTensor, timestep: Union[float, torch.FloatTensor]
+ ) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`float` or `torch.FloatTensor`):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+ sample = sample / ((sigma**2 + 1) ** 0.5)
+ self.is_scale_input_called = True
+ return sample
+
+ def get_lms_coefficient(self, order, t, current_order):
+ """
+ Compute the linear multistep coefficient.
+
+ Args:
+ order ():
+ t ():
+ current_order ():
+ """
+
+ def lms_derivative(tau):
+ prod = 1.0
+ for k in range(order):
+ if current_order == k:
+ continue
+ prod *= (tau - self.sigmas[t - k]) / (self.sigmas[t - current_order] - self.sigmas[t - k])
+ return prod
+
+ integrated_coeff = integrate.quad(lms_derivative, self.sigmas[t], self.sigmas[t + 1], epsrel=1e-4)[0]
+
+ return integrated_coeff
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ self.num_inference_steps = num_inference_steps
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps, dtype=np.float32)[
+ ::-1
+ ].copy()
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.float32)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(self.config.num_train_timesteps, 0, -step_ratio)).round().copy().astype(np.float32)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ log_sigmas = np.log(sigmas)
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+
+ if self.use_karras_sigmas:
+ sigmas = self._convert_to_karras(in_sigmas=sigmas)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas])
+
+ sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
+
+ self.sigmas = torch.from_numpy(sigmas).to(device=device)
+ self.timesteps = torch.from_numpy(timesteps).to(device=device)
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ self.derivatives = []
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ # copied from diffusers.schedulers.scheduling_euler_discrete._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # copied from diffusers.schedulers.scheduling_euler_discrete._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ sigma_min: float = in_sigmas[-1].item()
+ sigma_max: float = in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, self.num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: Union[float, torch.FloatTensor],
+ sample: torch.FloatTensor,
+ order: int = 4,
+ return_dict: bool = True,
+ ) -> Union[LMSDiscreteSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`float` or `torch.FloatTensor`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ order (`int`, defaults to 4):
+ The order of the linear multistep method.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if not self.is_scale_input_called:
+ warnings.warn(
+ "The `scale_model_input` function should be called before `step` to ensure correct denoising. "
+ "See `StableDiffusionPipeline` for a usage example."
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ sigma = self.sigmas[self.step_index]
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = sample - sigma * model_output
+ elif self.config.prediction_type == "v_prediction":
+ # * c_out + input * c_skip
+ pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ # 2. Convert to an ODE derivative
+ derivative = (sample - pred_original_sample) / sigma
+ self.derivatives.append(derivative)
+ if len(self.derivatives) > order:
+ self.derivatives.pop(0)
+
+ # 3. Compute linear multistep coefficients
+ order = min(self.step_index + 1, order)
+ lms_coeffs = [self.get_lms_coefficient(order, self.step_index, curr_order) for curr_order in range(order)]
+
+ # 4. Compute previous sample based on the derivatives path
+ prev_sample = sample + sum(
+ coeff * derivative for coeff, derivative in zip(lms_coeffs, reversed(self.derivatives))
+ )
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return LMSDiscreteSchedulerOutput(prev_sample=prev_sample, pred_original_sample=pred_original_sample)
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # self.begin_index is None when scheduler is used for training, or pipeline does not implement set_begin_index
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ noisy_samples = original_samples + noise * sigma
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_lms_discrete_flax.py b/model/diffusers_c/schedulers/scheduling_lms_discrete_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..f1169cc90a7bd53d2ede80584e09647e5f800ccf
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_lms_discrete_flax.py
@@ -0,0 +1,283 @@
+# Copyright 2024 Katherine Crowson and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import flax
+import jax.numpy as jnp
+from scipy import integrate
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils_flax import (
+ CommonSchedulerState,
+ FlaxKarrasDiffusionSchedulers,
+ FlaxSchedulerMixin,
+ FlaxSchedulerOutput,
+ broadcast_to_shape_from_left,
+)
+
+
+@flax.struct.dataclass
+class LMSDiscreteSchedulerState:
+ common: CommonSchedulerState
+
+ # setable values
+ init_noise_sigma: jnp.ndarray
+ timesteps: jnp.ndarray
+ sigmas: jnp.ndarray
+ num_inference_steps: Optional[int] = None
+
+ # running values
+ derivatives: Optional[jnp.ndarray] = None
+
+ @classmethod
+ def create(
+ cls, common: CommonSchedulerState, init_noise_sigma: jnp.ndarray, timesteps: jnp.ndarray, sigmas: jnp.ndarray
+ ):
+ return cls(common=common, init_noise_sigma=init_noise_sigma, timesteps=timesteps, sigmas=sigmas)
+
+
+@dataclass
+class FlaxLMSSchedulerOutput(FlaxSchedulerOutput):
+ state: LMSDiscreteSchedulerState
+
+
+class FlaxLMSDiscreteScheduler(FlaxSchedulerMixin, ConfigMixin):
+ """
+ Linear Multistep Scheduler for discrete beta schedules. Based on the original k-diffusion implementation by
+ Katherine Crowson:
+ https://github.com/crowsonkb/k-diffusion/blob/481677d114f6ea445aa009cf5bd7a9cdee909e47/k_diffusion/sampling.py#L181
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ beta_start (`float`): the starting `beta` value of inference.
+ beta_end (`float`): the final `beta` value.
+ beta_schedule (`str`):
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear` or `scaled_linear`.
+ trained_betas (`jnp.ndarray`, optional):
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+ prediction_type (`str`, default `epsilon`, optional):
+ prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion
+ process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4
+ https://imagen.research.google/video/paper.pdf)
+ dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`):
+ the `dtype` used for params and computation.
+ """
+
+ _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers]
+
+ dtype: jnp.dtype
+
+ @property
+ def has_state(self):
+ return True
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[jnp.ndarray] = None,
+ prediction_type: str = "epsilon",
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ self.dtype = dtype
+
+ def create_state(self, common: Optional[CommonSchedulerState] = None) -> LMSDiscreteSchedulerState:
+ if common is None:
+ common = CommonSchedulerState.create(self)
+
+ timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1]
+ sigmas = ((1 - common.alphas_cumprod) / common.alphas_cumprod) ** 0.5
+
+ # standard deviation of the initial noise distribution
+ init_noise_sigma = sigmas.max()
+
+ return LMSDiscreteSchedulerState.create(
+ common=common,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ sigmas=sigmas,
+ )
+
+ def scale_model_input(self, state: LMSDiscreteSchedulerState, sample: jnp.ndarray, timestep: int) -> jnp.ndarray:
+ """
+ Scales the denoising model input by `(sigma**2 + 1) ** 0.5` to match the K-LMS algorithm.
+
+ Args:
+ state (`LMSDiscreteSchedulerState`):
+ the `FlaxLMSDiscreteScheduler` state data class instance.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ timestep (`int`):
+ current discrete timestep in the diffusion chain.
+
+ Returns:
+ `jnp.ndarray`: scaled input sample
+ """
+ (step_index,) = jnp.where(state.timesteps == timestep, size=1)
+ step_index = step_index[0]
+
+ sigma = state.sigmas[step_index]
+ sample = sample / ((sigma**2 + 1) ** 0.5)
+ return sample
+
+ def get_lms_coefficient(self, state: LMSDiscreteSchedulerState, order, t, current_order):
+ """
+ Compute a linear multistep coefficient.
+
+ Args:
+ order (TODO):
+ t (TODO):
+ current_order (TODO):
+ """
+
+ def lms_derivative(tau):
+ prod = 1.0
+ for k in range(order):
+ if current_order == k:
+ continue
+ prod *= (tau - state.sigmas[t - k]) / (state.sigmas[t - current_order] - state.sigmas[t - k])
+ return prod
+
+ integrated_coeff = integrate.quad(lms_derivative, state.sigmas[t], state.sigmas[t + 1], epsrel=1e-4)[0]
+
+ return integrated_coeff
+
+ def set_timesteps(
+ self, state: LMSDiscreteSchedulerState, num_inference_steps: int, shape: Tuple = ()
+ ) -> LMSDiscreteSchedulerState:
+ """
+ Sets the timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ state (`LMSDiscreteSchedulerState`):
+ the `FlaxLMSDiscreteScheduler` state data class instance.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ """
+
+ timesteps = jnp.linspace(self.config.num_train_timesteps - 1, 0, num_inference_steps, dtype=self.dtype)
+
+ low_idx = jnp.floor(timesteps).astype(jnp.int32)
+ high_idx = jnp.ceil(timesteps).astype(jnp.int32)
+
+ frac = jnp.mod(timesteps, 1.0)
+
+ sigmas = ((1 - state.common.alphas_cumprod) / state.common.alphas_cumprod) ** 0.5
+ sigmas = (1 - frac) * sigmas[low_idx] + frac * sigmas[high_idx]
+ sigmas = jnp.concatenate([sigmas, jnp.array([0.0], dtype=self.dtype)])
+
+ timesteps = timesteps.astype(jnp.int32)
+
+ # initial running values
+ derivatives = jnp.zeros((0,) + shape, dtype=self.dtype)
+
+ return state.replace(
+ timesteps=timesteps,
+ sigmas=sigmas,
+ num_inference_steps=num_inference_steps,
+ derivatives=derivatives,
+ )
+
+ def step(
+ self,
+ state: LMSDiscreteSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ order: int = 4,
+ return_dict: bool = True,
+ ) -> Union[FlaxLMSSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ state (`LMSDiscreteSchedulerState`): the `FlaxLMSDiscreteScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ order: coefficient for multi-step inference.
+ return_dict (`bool`): option for returning tuple rather than FlaxLMSSchedulerOutput class
+
+ Returns:
+ [`FlaxLMSSchedulerOutput`] or `tuple`: [`FlaxLMSSchedulerOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+ if state.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ sigma = state.sigmas[timestep]
+
+ # 1. compute predicted original sample (x_0) from sigma-scaled predicted noise
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = sample - sigma * model_output
+ elif self.config.prediction_type == "v_prediction":
+ # * c_out + input * c_skip
+ pred_original_sample = model_output * (-sigma / (sigma**2 + 1) ** 0.5) + (sample / (sigma**2 + 1))
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, or `v_prediction`"
+ )
+
+ # 2. Convert to an ODE derivative
+ derivative = (sample - pred_original_sample) / sigma
+ state = state.replace(derivatives=jnp.append(state.derivatives, derivative))
+ if len(state.derivatives) > order:
+ state = state.replace(derivatives=jnp.delete(state.derivatives, 0))
+
+ # 3. Compute linear multistep coefficients
+ order = min(timestep + 1, order)
+ lms_coeffs = [self.get_lms_coefficient(state, order, timestep, curr_order) for curr_order in range(order)]
+
+ # 4. Compute previous sample based on the derivatives path
+ prev_sample = sample + sum(
+ coeff * derivative for coeff, derivative in zip(lms_coeffs, reversed(state.derivatives))
+ )
+
+ if not return_dict:
+ return (prev_sample, state)
+
+ return FlaxLMSSchedulerOutput(prev_sample=prev_sample, state=state)
+
+ def add_noise(
+ self,
+ state: LMSDiscreteSchedulerState,
+ original_samples: jnp.ndarray,
+ noise: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ) -> jnp.ndarray:
+ sigma = state.sigmas[timesteps].flatten()
+ sigma = broadcast_to_shape_from_left(sigma, noise.shape)
+
+ noisy_samples = original_samples + noise * sigma
+
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_pndm.py b/model/diffusers_c/schedulers/scheduling_pndm.py
new file mode 100644
index 0000000000000000000000000000000000000000..4bfd9d6c3a7f8c562d465572fa14c514919d8b51
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_pndm.py
@@ -0,0 +1,478 @@
+# Copyright 2024 Zhejiang University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class PNDMScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `PNDMScheduler` uses pseudo numerical methods for diffusion models such as the Runge-Kutta and linear multi-step
+ method.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ skip_prk_steps (`bool`, defaults to `False`):
+ Allows the scheduler to skip the Runge-Kutta steps defined in the original paper as being required before
+ PLMS steps.
+ set_alpha_to_one (`bool`, defaults to `False`):
+ Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+ there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+ otherwise it uses the alpha value at step 0.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process)
+ or `v_prediction` (see section 2.4 of [Imagen Video](https://imagen.research.google/video/paper.pdf)
+ paper).
+ timestep_spacing (`str`, defaults to `"leading"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ skip_prk_steps: bool = False,
+ set_alpha_to_one: bool = False,
+ prediction_type: str = "epsilon",
+ timestep_spacing: str = "leading",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # For now we only support F-PNDM, i.e. the runge-kutta method
+ # For more information on the algorithm please take a look at the paper: https://arxiv.org/pdf/2202.09778.pdf
+ # mainly at formula (9), (12), (13) and the Algorithm 2.
+ self.pndm_order = 4
+
+ # running values
+ self.cur_model_output = 0
+ self.counter = 0
+ self.cur_sample = None
+ self.ets = []
+
+ # setable values
+ self.num_inference_steps = None
+ self._timesteps = np.arange(0, num_train_timesteps)[::-1].copy()
+ self.prk_timesteps = None
+ self.plms_timesteps = None
+ self.timesteps = None
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+
+ self.num_inference_steps = num_inference_steps
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ self._timesteps = (
+ np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps).round().astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ self._timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()
+ self._timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / self.num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ self._timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio))[::-1].astype(
+ np.int64
+ )
+ self._timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ if self.config.skip_prk_steps:
+ # for some models like stable diffusion the prk steps can/should be skipped to
+ # produce better results. When using PNDM with `self.config.skip_prk_steps` the implementation
+ # is based on crowsonkb's PLMS sampler implementation: https://github.com/CompVis/latent-diffusion/pull/51
+ self.prk_timesteps = np.array([])
+ self.plms_timesteps = np.concatenate([self._timesteps[:-1], self._timesteps[-2:-1], self._timesteps[-1:]])[
+ ::-1
+ ].copy()
+ else:
+ prk_timesteps = np.array(self._timesteps[-self.pndm_order :]).repeat(2) + np.tile(
+ np.array([0, self.config.num_train_timesteps // num_inference_steps // 2]), self.pndm_order
+ )
+ self.prk_timesteps = (prk_timesteps[:-1].repeat(2)[1:-1])[::-1].copy()
+ self.plms_timesteps = self._timesteps[:-3][
+ ::-1
+ ].copy() # we copy to avoid having negative strides which are not supported by torch.from_numpy
+
+ timesteps = np.concatenate([self.prk_timesteps, self.plms_timesteps]).astype(np.int64)
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ self.ets = []
+ self.counter = 0
+ self.cur_model_output = 0
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise), and calls [`~PNDMScheduler.step_prk`]
+ or [`~PNDMScheduler.step_plms`] depending on the internal variable `counter`.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.counter < len(self.prk_timesteps) and not self.config.skip_prk_steps:
+ return self.step_prk(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict)
+ else:
+ return self.step_plms(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict)
+
+ def step_prk(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the Runge-Kutta method. It performs four forward passes to approximate the solution to the differential
+ equation.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ diff_to_prev = 0 if self.counter % 2 else self.config.num_train_timesteps // self.num_inference_steps // 2
+ prev_timestep = timestep - diff_to_prev
+ timestep = self.prk_timesteps[self.counter // 4 * 4]
+
+ if self.counter % 4 == 0:
+ self.cur_model_output += 1 / 6 * model_output
+ self.ets.append(model_output)
+ self.cur_sample = sample
+ elif (self.counter - 1) % 4 == 0:
+ self.cur_model_output += 1 / 3 * model_output
+ elif (self.counter - 2) % 4 == 0:
+ self.cur_model_output += 1 / 3 * model_output
+ elif (self.counter - 3) % 4 == 0:
+ model_output = self.cur_model_output + 1 / 6 * model_output
+ self.cur_model_output = 0
+
+ # cur_sample should not be `None`
+ cur_sample = self.cur_sample if self.cur_sample is not None else sample
+
+ prev_sample = self._get_prev_sample(cur_sample, timestep, prev_timestep, model_output)
+ self.counter += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def step_plms(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the linear multistep method. It performs one forward pass multiple times to approximate the solution.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if not self.config.skip_prk_steps and len(self.ets) < 3:
+ raise ValueError(
+ f"{self.__class__} can only be run AFTER scheduler has been run "
+ "in 'prk' mode for at least 12 iterations "
+ "See: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pipeline_pndm.py "
+ "for more information."
+ )
+
+ prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
+
+ if self.counter != 1:
+ self.ets = self.ets[-3:]
+ self.ets.append(model_output)
+ else:
+ prev_timestep = timestep
+ timestep = timestep + self.config.num_train_timesteps // self.num_inference_steps
+
+ if len(self.ets) == 1 and self.counter == 0:
+ model_output = model_output
+ self.cur_sample = sample
+ elif len(self.ets) == 1 and self.counter == 1:
+ model_output = (model_output + self.ets[-1]) / 2
+ sample = self.cur_sample
+ self.cur_sample = None
+ elif len(self.ets) == 2:
+ model_output = (3 * self.ets[-1] - self.ets[-2]) / 2
+ elif len(self.ets) == 3:
+ model_output = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12
+ else:
+ model_output = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4])
+
+ prev_sample = self._get_prev_sample(sample, timestep, prev_timestep, model_output)
+ self.counter += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def _get_prev_sample(self, sample, timestep, prev_timestep, model_output):
+ # See formula (9) of PNDM paper https://arxiv.org/pdf/2202.09778.pdf
+ # this function computes x_(t−δ) using the formula of (9)
+ # Note that x_t needs to be added to both sides of the equation
+
+ # Notation ( ->
+ # alpha_prod_t -> α_t
+ # alpha_prod_t_prev -> α_(t−δ)
+ # beta_prod_t -> (1 - α_t)
+ # beta_prod_t_prev -> (1 - α_(t−δ))
+ # sample -> x_t
+ # model_output -> e_θ(x_t, t)
+ # prev_sample -> x_(t−δ)
+ alpha_prod_t = self.alphas_cumprod[timestep]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ if self.config.prediction_type == "v_prediction":
+ model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ elif self.config.prediction_type != "epsilon":
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `v_prediction`"
+ )
+
+ # corresponds to (α_(t−δ) - α_t) divided by
+ # denominator of x_t in formula (9) and plus 1
+ # Note: (α_(t−δ) - α_t) / (sqrt(α_t) * (sqrt(α_(t−δ)) + sqr(α_t))) =
+ # sqrt(α_(t−δ)) / sqrt(α_t))
+ sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5)
+
+ # corresponds to denominator of e_θ(x_t, t) in formula (9)
+ model_output_denom_coeff = alpha_prod_t * beta_prod_t_prev ** (0.5) + (
+ alpha_prod_t * beta_prod_t * alpha_prod_t_prev
+ ) ** (0.5)
+
+ # full formula (9)
+ prev_sample = (
+ sample_coeff * sample - (alpha_prod_t_prev - alpha_prod_t) * model_output / model_output_denom_coeff
+ )
+
+ return prev_sample
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+ # for the subsequent add_noise calls
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_pndm_flax.py b/model/diffusers_c/schedulers/scheduling_pndm_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..e2aec3f8db157ae5fa1d35800344af9ac2ba32df
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_pndm_flax.py
@@ -0,0 +1,511 @@
+# Copyright 2024 Zhejiang University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import flax
+import jax
+import jax.numpy as jnp
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils_flax import (
+ CommonSchedulerState,
+ FlaxKarrasDiffusionSchedulers,
+ FlaxSchedulerMixin,
+ FlaxSchedulerOutput,
+ add_noise_common,
+)
+
+
+@flax.struct.dataclass
+class PNDMSchedulerState:
+ common: CommonSchedulerState
+ final_alpha_cumprod: jnp.ndarray
+
+ # setable values
+ init_noise_sigma: jnp.ndarray
+ timesteps: jnp.ndarray
+ num_inference_steps: Optional[int] = None
+ prk_timesteps: Optional[jnp.ndarray] = None
+ plms_timesteps: Optional[jnp.ndarray] = None
+
+ # running values
+ cur_model_output: Optional[jnp.ndarray] = None
+ counter: Optional[jnp.int32] = None
+ cur_sample: Optional[jnp.ndarray] = None
+ ets: Optional[jnp.ndarray] = None
+
+ @classmethod
+ def create(
+ cls,
+ common: CommonSchedulerState,
+ final_alpha_cumprod: jnp.ndarray,
+ init_noise_sigma: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ):
+ return cls(
+ common=common,
+ final_alpha_cumprod=final_alpha_cumprod,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ )
+
+
+@dataclass
+class FlaxPNDMSchedulerOutput(FlaxSchedulerOutput):
+ state: PNDMSchedulerState
+
+
+class FlaxPNDMScheduler(FlaxSchedulerMixin, ConfigMixin):
+ """
+ Pseudo numerical methods for diffusion models (PNDM) proposes using more advanced ODE integration techniques,
+ namely Runge-Kutta method and a linear multi-step method.
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ For more details, see the original paper: https://arxiv.org/abs/2202.09778
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ beta_start (`float`): the starting `beta` value of inference.
+ beta_end (`float`): the final `beta` value.
+ beta_schedule (`str`):
+ the beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`jnp.ndarray`, optional):
+ option to pass an array of betas directly to the constructor to bypass `beta_start`, `beta_end` etc.
+ skip_prk_steps (`bool`):
+ allows the scheduler to skip the Runge-Kutta steps that are defined in the original paper as being required
+ before plms steps; defaults to `False`.
+ set_alpha_to_one (`bool`, default `False`):
+ each diffusion step uses the value of alphas product at that step and at the previous one. For the final
+ step there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+ otherwise it uses the value of alpha at step 0.
+ steps_offset (`int`, default `0`):
+ an offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False`, to make the last step use step 0 for the previous alpha product, as done in
+ stable diffusion.
+ prediction_type (`str`, default `epsilon`, optional):
+ prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion
+ process), `sample` (directly predicting the noisy sample`) or `v_prediction` (see section 2.4
+ https://imagen.research.google/video/paper.pdf)
+ dtype (`jnp.dtype`, *optional*, defaults to `jnp.float32`):
+ the `dtype` used for params and computation.
+ """
+
+ _compatibles = [e.name for e in FlaxKarrasDiffusionSchedulers]
+
+ dtype: jnp.dtype
+ pndm_order: int
+
+ @property
+ def has_state(self):
+ return True
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[jnp.ndarray] = None,
+ skip_prk_steps: bool = False,
+ set_alpha_to_one: bool = False,
+ steps_offset: int = 0,
+ prediction_type: str = "epsilon",
+ dtype: jnp.dtype = jnp.float32,
+ ):
+ self.dtype = dtype
+
+ # For now we only support F-PNDM, i.e. the runge-kutta method
+ # For more information on the algorithm please take a look at the paper: https://arxiv.org/pdf/2202.09778.pdf
+ # mainly at formula (9), (12), (13) and the Algorithm 2.
+ self.pndm_order = 4
+
+ def create_state(self, common: Optional[CommonSchedulerState] = None) -> PNDMSchedulerState:
+ if common is None:
+ common = CommonSchedulerState.create(self)
+
+ # At every step in ddim, we are looking into the previous alphas_cumprod
+ # For the final step, there is no previous alphas_cumprod because we are already at 0
+ # `set_alpha_to_one` decides whether we set this parameter simply to one or
+ # whether we use the final alpha of the "non-previous" one.
+ final_alpha_cumprod = (
+ jnp.array(1.0, dtype=self.dtype) if self.config.set_alpha_to_one else common.alphas_cumprod[0]
+ )
+
+ # standard deviation of the initial noise distribution
+ init_noise_sigma = jnp.array(1.0, dtype=self.dtype)
+
+ timesteps = jnp.arange(0, self.config.num_train_timesteps).round()[::-1]
+
+ return PNDMSchedulerState.create(
+ common=common,
+ final_alpha_cumprod=final_alpha_cumprod,
+ init_noise_sigma=init_noise_sigma,
+ timesteps=timesteps,
+ )
+
+ def set_timesteps(self, state: PNDMSchedulerState, num_inference_steps: int, shape: Tuple) -> PNDMSchedulerState:
+ """
+ Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ state (`PNDMSchedulerState`):
+ the `FlaxPNDMScheduler` state data class instance.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ shape (`Tuple`):
+ the shape of the samples to be generated.
+ """
+
+ step_ratio = self.config.num_train_timesteps // num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # rounding to avoid issues when num_inference_step is power of 3
+ _timesteps = (jnp.arange(0, num_inference_steps) * step_ratio).round() + self.config.steps_offset
+
+ if self.config.skip_prk_steps:
+ # for some models like stable diffusion the prk steps can/should be skipped to
+ # produce better results. When using PNDM with `self.config.skip_prk_steps` the implementation
+ # is based on crowsonkb's PLMS sampler implementation: https://github.com/CompVis/latent-diffusion/pull/51
+
+ prk_timesteps = jnp.array([], dtype=jnp.int32)
+ plms_timesteps = jnp.concatenate([_timesteps[:-1], _timesteps[-2:-1], _timesteps[-1:]])[::-1]
+
+ else:
+ prk_timesteps = _timesteps[-self.pndm_order :].repeat(2) + jnp.tile(
+ jnp.array([0, self.config.num_train_timesteps // num_inference_steps // 2], dtype=jnp.int32),
+ self.pndm_order,
+ )
+
+ prk_timesteps = (prk_timesteps[:-1].repeat(2)[1:-1])[::-1]
+ plms_timesteps = _timesteps[:-3][::-1]
+
+ timesteps = jnp.concatenate([prk_timesteps, plms_timesteps])
+
+ # initial running values
+
+ cur_model_output = jnp.zeros(shape, dtype=self.dtype)
+ counter = jnp.int32(0)
+ cur_sample = jnp.zeros(shape, dtype=self.dtype)
+ ets = jnp.zeros((4,) + shape, dtype=self.dtype)
+
+ return state.replace(
+ timesteps=timesteps,
+ num_inference_steps=num_inference_steps,
+ prk_timesteps=prk_timesteps,
+ plms_timesteps=plms_timesteps,
+ cur_model_output=cur_model_output,
+ counter=counter,
+ cur_sample=cur_sample,
+ ets=ets,
+ )
+
+ def scale_model_input(
+ self, state: PNDMSchedulerState, sample: jnp.ndarray, timestep: Optional[int] = None
+ ) -> jnp.ndarray:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance.
+ sample (`jnp.ndarray`): input sample
+ timestep (`int`, optional): current timestep
+
+ Returns:
+ `jnp.ndarray`: scaled input sample
+ """
+ return sample
+
+ def step(
+ self,
+ state: PNDMSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ return_dict: bool = True,
+ ) -> Union[FlaxPNDMSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ This function calls `step_prk()` or `step_plms()` depending on the internal variable `counter`.
+
+ Args:
+ state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ return_dict (`bool`): option for returning tuple rather than FlaxPNDMSchedulerOutput class
+
+ Returns:
+ [`FlaxPNDMSchedulerOutput`] or `tuple`: [`FlaxPNDMSchedulerOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+
+ if state.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.config.skip_prk_steps:
+ prev_sample, state = self.step_plms(state, model_output, timestep, sample)
+ else:
+ prk_prev_sample, prk_state = self.step_prk(state, model_output, timestep, sample)
+ plms_prev_sample, plms_state = self.step_plms(state, model_output, timestep, sample)
+
+ cond = state.counter < len(state.prk_timesteps)
+
+ prev_sample = jax.lax.select(cond, prk_prev_sample, plms_prev_sample)
+
+ state = state.replace(
+ cur_model_output=jax.lax.select(cond, prk_state.cur_model_output, plms_state.cur_model_output),
+ ets=jax.lax.select(cond, prk_state.ets, plms_state.ets),
+ cur_sample=jax.lax.select(cond, prk_state.cur_sample, plms_state.cur_sample),
+ counter=jax.lax.select(cond, prk_state.counter, plms_state.counter),
+ )
+
+ if not return_dict:
+ return (prev_sample, state)
+
+ return FlaxPNDMSchedulerOutput(prev_sample=prev_sample, state=state)
+
+ def step_prk(
+ self,
+ state: PNDMSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ ) -> Union[FlaxPNDMSchedulerOutput, Tuple]:
+ """
+ Step function propagating the sample with the Runge-Kutta method. RK takes 4 forward passes to approximate the
+ solution to the differential equation.
+
+ Args:
+ state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ return_dict (`bool`): option for returning tuple rather than FlaxPNDMSchedulerOutput class
+
+ Returns:
+ [`FlaxPNDMSchedulerOutput`] or `tuple`: [`FlaxPNDMSchedulerOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+
+ if state.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ diff_to_prev = jnp.where(
+ state.counter % 2, 0, self.config.num_train_timesteps // state.num_inference_steps // 2
+ )
+ prev_timestep = timestep - diff_to_prev
+ timestep = state.prk_timesteps[state.counter // 4 * 4]
+
+ model_output = jax.lax.select(
+ (state.counter % 4) != 3,
+ model_output, # remainder 0, 1, 2
+ state.cur_model_output + 1 / 6 * model_output, # remainder 3
+ )
+
+ state = state.replace(
+ cur_model_output=jax.lax.select_n(
+ state.counter % 4,
+ state.cur_model_output + 1 / 6 * model_output, # remainder 0
+ state.cur_model_output + 1 / 3 * model_output, # remainder 1
+ state.cur_model_output + 1 / 3 * model_output, # remainder 2
+ jnp.zeros_like(state.cur_model_output), # remainder 3
+ ),
+ ets=jax.lax.select(
+ (state.counter % 4) == 0,
+ state.ets.at[0:3].set(state.ets[1:4]).at[3].set(model_output), # remainder 0
+ state.ets, # remainder 1, 2, 3
+ ),
+ cur_sample=jax.lax.select(
+ (state.counter % 4) == 0,
+ sample, # remainder 0
+ state.cur_sample, # remainder 1, 2, 3
+ ),
+ )
+
+ cur_sample = state.cur_sample
+ prev_sample = self._get_prev_sample(state, cur_sample, timestep, prev_timestep, model_output)
+ state = state.replace(counter=state.counter + 1)
+
+ return (prev_sample, state)
+
+ def step_plms(
+ self,
+ state: PNDMSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ ) -> Union[FlaxPNDMSchedulerOutput, Tuple]:
+ """
+ Step function propagating the sample with the linear multi-step method. This has one forward pass with multiple
+ times to approximate the solution.
+
+ Args:
+ state (`PNDMSchedulerState`): the `FlaxPNDMScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ return_dict (`bool`): option for returning tuple rather than FlaxPNDMSchedulerOutput class
+
+ Returns:
+ [`FlaxPNDMSchedulerOutput`] or `tuple`: [`FlaxPNDMSchedulerOutput`] if `return_dict` is True, otherwise a
+ `tuple`. When returning a tuple, the first element is the sample tensor.
+
+ """
+
+ if state.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ # NOTE: There is no way to check in the jitted runtime if the prk mode was ran before
+
+ prev_timestep = timestep - self.config.num_train_timesteps // state.num_inference_steps
+ prev_timestep = jnp.where(prev_timestep > 0, prev_timestep, 0)
+
+ # Reference:
+ # if state.counter != 1:
+ # state.ets.append(model_output)
+ # else:
+ # prev_timestep = timestep
+ # timestep = timestep + self.config.num_train_timesteps // state.num_inference_steps
+
+ prev_timestep = jnp.where(state.counter == 1, timestep, prev_timestep)
+ timestep = jnp.where(
+ state.counter == 1, timestep + self.config.num_train_timesteps // state.num_inference_steps, timestep
+ )
+
+ # Reference:
+ # if len(state.ets) == 1 and state.counter == 0:
+ # model_output = model_output
+ # state.cur_sample = sample
+ # elif len(state.ets) == 1 and state.counter == 1:
+ # model_output = (model_output + state.ets[-1]) / 2
+ # sample = state.cur_sample
+ # state.cur_sample = None
+ # elif len(state.ets) == 2:
+ # model_output = (3 * state.ets[-1] - state.ets[-2]) / 2
+ # elif len(state.ets) == 3:
+ # model_output = (23 * state.ets[-1] - 16 * state.ets[-2] + 5 * state.ets[-3]) / 12
+ # else:
+ # model_output = (1 / 24) * (55 * state.ets[-1] - 59 * state.ets[-2] + 37 * state.ets[-3] - 9 * state.ets[-4])
+
+ state = state.replace(
+ ets=jax.lax.select(
+ state.counter != 1,
+ state.ets.at[0:3].set(state.ets[1:4]).at[3].set(model_output), # counter != 1
+ state.ets, # counter 1
+ ),
+ cur_sample=jax.lax.select(
+ state.counter != 1,
+ sample, # counter != 1
+ state.cur_sample, # counter 1
+ ),
+ )
+
+ state = state.replace(
+ cur_model_output=jax.lax.select_n(
+ jnp.clip(state.counter, 0, 4),
+ model_output, # counter 0
+ (model_output + state.ets[-1]) / 2, # counter 1
+ (3 * state.ets[-1] - state.ets[-2]) / 2, # counter 2
+ (23 * state.ets[-1] - 16 * state.ets[-2] + 5 * state.ets[-3]) / 12, # counter 3
+ (1 / 24)
+ * (55 * state.ets[-1] - 59 * state.ets[-2] + 37 * state.ets[-3] - 9 * state.ets[-4]), # counter >= 4
+ ),
+ )
+
+ sample = state.cur_sample
+ model_output = state.cur_model_output
+ prev_sample = self._get_prev_sample(state, sample, timestep, prev_timestep, model_output)
+ state = state.replace(counter=state.counter + 1)
+
+ return (prev_sample, state)
+
+ def _get_prev_sample(self, state: PNDMSchedulerState, sample, timestep, prev_timestep, model_output):
+ # See formula (9) of PNDM paper https://arxiv.org/pdf/2202.09778.pdf
+ # this function computes x_(t−δ) using the formula of (9)
+ # Note that x_t needs to be added to both sides of the equation
+
+ # Notation ( ->
+ # alpha_prod_t -> α_t
+ # alpha_prod_t_prev -> α_(t−δ)
+ # beta_prod_t -> (1 - α_t)
+ # beta_prod_t_prev -> (1 - α_(t−δ))
+ # sample -> x_t
+ # model_output -> e_θ(x_t, t)
+ # prev_sample -> x_(t−δ)
+ alpha_prod_t = state.common.alphas_cumprod[timestep]
+ alpha_prod_t_prev = jnp.where(
+ prev_timestep >= 0, state.common.alphas_cumprod[prev_timestep], state.final_alpha_cumprod
+ )
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ if self.config.prediction_type == "v_prediction":
+ model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ elif self.config.prediction_type != "epsilon":
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `v_prediction`"
+ )
+
+ # corresponds to (α_(t−δ) - α_t) divided by
+ # denominator of x_t in formula (9) and plus 1
+ # Note: (α_(t−δ) - α_t) / (sqrt(α_t) * (sqrt(α_(t−δ)) + sqr(α_t))) =
+ # sqrt(α_(t−δ)) / sqrt(α_t))
+ sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5)
+
+ # corresponds to denominator of e_θ(x_t, t) in formula (9)
+ model_output_denom_coeff = alpha_prod_t * beta_prod_t_prev ** (0.5) + (
+ alpha_prod_t * beta_prod_t * alpha_prod_t_prev
+ ) ** (0.5)
+
+ # full formula (9)
+ prev_sample = (
+ sample_coeff * sample - (alpha_prod_t_prev - alpha_prod_t) * model_output / model_output_denom_coeff
+ )
+
+ return prev_sample
+
+ def add_noise(
+ self,
+ state: PNDMSchedulerState,
+ original_samples: jnp.ndarray,
+ noise: jnp.ndarray,
+ timesteps: jnp.ndarray,
+ ) -> jnp.ndarray:
+ return add_noise_common(state.common, original_samples, noise, timesteps)
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_repaint.py b/model/diffusers_c/schedulers/scheduling_repaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..0bff07c4fdf348f154638eed0805ff943aa8f99f
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_repaint.py
@@ -0,0 +1,361 @@
+# Copyright 2024 ETH Zurich Computer Vision Lab and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin
+
+
+@dataclass
+class RePaintSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's step function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample (x_{0}) based on the model output from
+ the current timestep. `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: torch.FloatTensor
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class RePaintScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `RePaintScheduler` is a scheduler for DDPM inpainting inside a given mask.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, `squaredcos_cap_v2`, or `sigmoid`.
+ eta (`float`):
+ The weight of noise for added noise in diffusion step. If its value is between 0.0 and 1.0 it corresponds
+ to the DDIM scheduler, and if its value is between -0.0 and 1.0 it corresponds to the DDPM scheduler.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ clip_sample (`bool`, defaults to `True`):
+ Clip the predicted sample between -1 and 1 for numerical stability.
+
+ """
+
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ eta: float = 0.0,
+ trained_betas: Optional[np.ndarray] = None,
+ clip_sample: bool = True,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.from_numpy(trained_betas)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ elif beta_schedule == "sigmoid":
+ # GeoDiff sigmoid schedule
+ betas = torch.linspace(-6, 6, num_train_timesteps)
+ self.betas = torch.sigmoid(betas) * (beta_end - beta_start) + beta_start
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ self.one = torch.tensor(1.0)
+
+ self.final_alpha_cumprod = torch.tensor(1.0)
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy())
+
+ self.eta = eta
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: int,
+ jump_length: int = 10,
+ jump_n_sample: int = 10,
+ device: Union[str, torch.device] = None,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ jump_length (`int`, defaults to 10):
+ The number of steps taken forward in time before going backward in time for a single jump (“j” in
+ RePaint paper). Take a look at Figure 9 and 10 in the paper.
+ jump_n_sample (`int`, defaults to 10):
+ The number of times to make a forward time jump for a given chosen time sample. Take a look at Figure 9
+ and 10 in the paper.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+
+ """
+ num_inference_steps = min(self.config.num_train_timesteps, num_inference_steps)
+ self.num_inference_steps = num_inference_steps
+
+ timesteps = []
+
+ jumps = {}
+ for j in range(0, num_inference_steps - jump_length, jump_length):
+ jumps[j] = jump_n_sample - 1
+
+ t = num_inference_steps
+ while t >= 1:
+ t = t - 1
+ timesteps.append(t)
+
+ if jumps.get(t, 0) > 0:
+ jumps[t] = jumps[t] - 1
+ for _ in range(jump_length):
+ t = t + 1
+ timesteps.append(t)
+
+ timesteps = np.array(timesteps) * (self.config.num_train_timesteps // self.num_inference_steps)
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ def _get_variance(self, t):
+ prev_timestep = t - self.config.num_train_timesteps // self.num_inference_steps
+
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ # For t > 0, compute predicted variance βt (see formula (6) and (7) from
+ # https://arxiv.org/pdf/2006.11239.pdf) and sample from it to get
+ # previous sample x_{t-1} ~ N(pred_prev_sample, variance) == add
+ # variance to pred_sample
+ # Is equivalent to formula (16) in https://arxiv.org/pdf/2010.02502.pdf
+ # without eta.
+ # variance = (1 - alpha_prod_t_prev) / (1 - alpha_prod_t) * self.betas[t]
+ variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+
+ return variance
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ original_image: torch.FloatTensor,
+ mask: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[RePaintSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ original_image (`torch.FloatTensor`):
+ The original image to inpaint on.
+ mask (`torch.FloatTensor`):
+ The mask where a value of 0.0 indicates which part of the original image to inpaint.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_repaint.RePaintSchedulerOutput`] is returned,
+ otherwise a tuple is returned where the first element is the sample tensor.
+
+ """
+ t = timestep
+ prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps
+
+ # 1. compute alphas, betas
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+ beta_prod_t = 1 - alpha_prod_t
+
+ # 2. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_original_sample = (sample - beta_prod_t**0.5 * model_output) / alpha_prod_t**0.5
+
+ # 3. Clip "predicted x_0"
+ if self.config.clip_sample:
+ pred_original_sample = torch.clamp(pred_original_sample, -1, 1)
+
+ # We choose to follow RePaint Algorithm 1 to get x_{t-1}, however we
+ # substitute formula (7) in the algorithm coming from DDPM paper
+ # (formula (4) Algorithm 2 - Sampling) with formula (12) from DDIM paper.
+ # DDIM schedule gives the same results as DDPM with eta = 1.0
+ # Noise is being reused in 7. and 8., but no impact on quality has
+ # been observed.
+
+ # 5. Add noise
+ device = model_output.device
+ noise = randn_tensor(model_output.shape, generator=generator, device=device, dtype=model_output.dtype)
+ std_dev_t = self.eta * self._get_variance(timestep) ** 0.5
+
+ variance = 0
+ if t > 0 and self.eta > 0:
+ variance = std_dev_t * noise
+
+ # 6. compute "direction pointing to x_t" of formula (12)
+ # from https://arxiv.org/pdf/2010.02502.pdf
+ pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** 0.5 * model_output
+
+ # 7. compute x_{t-1} of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
+ prev_unknown_part = alpha_prod_t_prev**0.5 * pred_original_sample + pred_sample_direction + variance
+
+ # 8. Algorithm 1 Line 5 https://arxiv.org/pdf/2201.09865.pdf
+ prev_known_part = (alpha_prod_t_prev**0.5) * original_image + ((1 - alpha_prod_t_prev) ** 0.5) * noise
+
+ # 9. Algorithm 1 Line 8 https://arxiv.org/pdf/2201.09865.pdf
+ pred_prev_sample = mask * prev_known_part + (1.0 - mask) * prev_unknown_part
+
+ if not return_dict:
+ return (
+ pred_prev_sample,
+ pred_original_sample,
+ )
+
+ return RePaintSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample)
+
+ def undo_step(self, sample, timestep, generator=None):
+ n = self.config.num_train_timesteps // self.num_inference_steps
+
+ for i in range(n):
+ beta = self.betas[timestep + i]
+ if sample.device.type == "mps":
+ # randn does not work reproducibly on mps
+ noise = randn_tensor(sample.shape, dtype=sample.dtype, generator=generator)
+ noise = noise.to(sample.device)
+ else:
+ noise = randn_tensor(sample.shape, generator=generator, device=sample.device, dtype=sample.dtype)
+
+ # 10. Algorithm 1 Line 10 https://arxiv.org/pdf/2201.09865.pdf
+ sample = (1 - beta) ** 0.5 * sample + beta**0.5 * noise
+
+ return sample
+
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ raise NotImplementedError("Use `DDPMScheduler.add_noise()` to train for sampling with RePaint.")
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_sasolver.py b/model/diffusers_c/schedulers/scheduling_sasolver.py
new file mode 100644
index 0000000000000000000000000000000000000000..7657b43aecf35e5b1b50746ca45cac6fd62899c7
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_sasolver.py
@@ -0,0 +1,1126 @@
+# Copyright 2024 Shuchen Xue, etc. in University of Chinese Academy of Sciences Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: check https://arxiv.org/abs/2309.05019
+# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py
+
+import math
+from typing import Callable, List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import deprecate
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class SASolverScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `SASolverScheduler` is a fast dedicated high-order solver for diffusion SDEs.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ predictor_order (`int`, defaults to 2):
+ The predictor order which can be `1` or `2` or `3` or '4'. It is recommended to use `predictor_order=2` for guided
+ sampling, and `predictor_order=3` for unconditional sampling.
+ corrector_order (`int`, defaults to 2):
+ The corrector order which can be `1` or `2` or `3` or '4'. It is recommended to use `corrector_order=2` for guided
+ sampling, and `corrector_order=3` for unconditional sampling.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ tau_func (`Callable`, *optional*):
+ Stochasticity during the sampling. Default in init is `lambda t: 1 if t >= 200 and t <= 800 else 0`. SA-Solver
+ will sample from vanilla diffusion ODE if tau_func is set to `lambda t: 0`. SA-Solver will sample from vanilla
+ diffusion SDE if tau_func is set to `lambda t: 1`. For more details, please check https://arxiv.org/abs/2309.05019
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True` and
+ `algorithm_type="dpmsolver++"`.
+ algorithm_type (`str`, defaults to `data_prediction`):
+ Algorithm type for the solver; can be `data_prediction` or `noise_prediction`. It is recommended to use `data_prediction`
+ with `solver_order=2` for guided sampling like in Stable Diffusion.
+ lower_order_final (`bool`, defaults to `True`):
+ Whether to use lower-order solvers in the final steps. Default = True.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ lambda_min_clipped (`float`, defaults to `-inf`):
+ Clipping threshold for the minimum value of `lambda(t)` for numerical stability. This is critical for the
+ cosine (`squaredcos_cap_v2`) noise schedule.
+ variance_type (`str`, *optional*):
+ Set to "learned" or "learned_range" for diffusion models that predict variance. If set, the model's output
+ contains the predicted Gaussian variance.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ predictor_order: int = 2,
+ corrector_order: int = 2,
+ prediction_type: str = "epsilon",
+ tau_func: Optional[Callable] = None,
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ algorithm_type: str = "data_prediction",
+ lower_order_final: bool = True,
+ use_karras_sigmas: Optional[bool] = False,
+ lambda_min_clipped: float = -float("inf"),
+ variance_type: Optional[str] = None,
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = (
+ torch.linspace(
+ beta_start**0.5,
+ beta_end**0.5,
+ num_train_timesteps,
+ dtype=torch.float32,
+ )
+ ** 2
+ )
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ # Currently we only support VP-type noise schedule
+ self.alpha_t = torch.sqrt(self.alphas_cumprod)
+ self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
+ self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
+ self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ if algorithm_type not in ["data_prediction", "noise_prediction"]:
+ raise NotImplementedError(f"{algorithm_type} does is not implemented for {self.__class__}")
+
+ # setable values
+ self.num_inference_steps = None
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
+ self.timesteps = torch.from_numpy(timesteps)
+ self.timestep_list = [None] * max(predictor_order, corrector_order - 1)
+ self.model_outputs = [None] * max(predictor_order, corrector_order - 1)
+
+ if tau_func is None:
+ self.tau_func = lambda t: 1 if t >= 200 and t <= 800 else 0
+ else:
+ self.tau_func = tau_func
+ self.predict_x0 = algorithm_type == "data_prediction"
+ self.lower_order_nums = 0
+ self.last_sample = None
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def set_timesteps(self, num_inference_steps: int = None, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ # Clipping the minimum of all lambda(t) for numerical stability.
+ # This is critical for cosine (squaredcos_cap_v2) noise schedule.
+ clipped_idx = torch.searchsorted(torch.flip(self.lambda_t, [0]), self.config.lambda_min_clipped)
+ last_timestep = ((self.config.num_train_timesteps - clipped_idx).numpy()).item()
+
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, last_timestep - 1, num_inference_steps + 1).round()[::-1][:-1].copy().astype(np.int64)
+ )
+
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = last_timestep // (num_inference_steps + 1)
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.arange(last_timestep, 0, -step_ratio).round().copy().astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ if self.config.use_karras_sigmas:
+ log_sigmas = np.log(sigmas)
+ sigmas = np.flip(sigmas).copy()
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+ sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
+ else:
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+ sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
+ sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
+
+ self.sigmas = torch.from_numpy(sigmas)
+ self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
+
+ self.num_inference_steps = len(timesteps)
+ self.model_outputs = [
+ None,
+ ] * max(self.config.predictor_order, self.config.corrector_order - 1)
+ self.lower_order_nums = 0
+ self.last_sample = None
+
+ # add an index counter for schedulers that allow duplicated timesteps
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
+ def _sigma_to_alpha_sigma_t(self, sigma):
+ alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+ sigma_t = sigma * alpha_t
+
+ return alpha_t, sigma_t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def convert_model_output(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ Convert the model output to the corresponding type the data_prediction/noise_prediction algorithm needs. Noise_prediction is
+ designed to discretize an integral of the noise prediction model, and data_prediction is designed to discretize an
+ integral of the data prediction model.
+
+
+
+ The algorithm and model type are decoupled. You can use either data_prediction or noise_prediction for both noise
+ prediction and data prediction models.
+
+
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The converted model output.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ if sample is None:
+ if len(args) > 1:
+ sample = args[1]
+ else:
+ raise ValueError("missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ # SA-Solver_data_prediction needs to solve an integral of the data prediction model.
+ if self.config.algorithm_type in ["data_prediction"]:
+ if self.config.prediction_type == "epsilon":
+ # SA-Solver only needs the "mean" output.
+ if self.config.variance_type in ["learned", "learned_range"]:
+ model_output = model_output[:, :3]
+ x0_pred = (sample - sigma_t * model_output) / alpha_t
+ elif self.config.prediction_type == "sample":
+ x0_pred = model_output
+ elif self.config.prediction_type == "v_prediction":
+ x0_pred = alpha_t * sample - sigma_t * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the SASolverScheduler."
+ )
+
+ if self.config.thresholding:
+ x0_pred = self._threshold_sample(x0_pred)
+
+ return x0_pred
+
+ # SA-Solver_noise_prediction needs to solve an integral of the noise prediction model.
+ elif self.config.algorithm_type in ["noise_prediction"]:
+ if self.config.prediction_type == "epsilon":
+ # SA-Solver only needs the "mean" output.
+ if self.config.variance_type in ["learned", "learned_range"]:
+ epsilon = model_output[:, :3]
+ else:
+ epsilon = model_output
+ elif self.config.prediction_type == "sample":
+ epsilon = (sample - alpha_t * model_output) / sigma_t
+ elif self.config.prediction_type == "v_prediction":
+ epsilon = alpha_t * model_output + sigma_t * sample
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the SASolverScheduler."
+ )
+
+ if self.config.thresholding:
+ alpha_t, sigma_t = self.alpha_t[timestep], self.sigma_t[timestep]
+ x0_pred = (sample - sigma_t * epsilon) / alpha_t
+ x0_pred = self._threshold_sample(x0_pred)
+ epsilon = (sample - alpha_t * x0_pred) / sigma_t
+
+ return epsilon
+
+ def get_coefficients_exponential_negative(self, order, interval_start, interval_end):
+ """
+ Calculate the integral of exp(-x) * x^order dx from interval_start to interval_end
+ """
+ assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
+
+ if order == 0:
+ return torch.exp(-interval_end) * (torch.exp(interval_end - interval_start) - 1)
+ elif order == 1:
+ return torch.exp(-interval_end) * (
+ (interval_start + 1) * torch.exp(interval_end - interval_start) - (interval_end + 1)
+ )
+ elif order == 2:
+ return torch.exp(-interval_end) * (
+ (interval_start**2 + 2 * interval_start + 2) * torch.exp(interval_end - interval_start)
+ - (interval_end**2 + 2 * interval_end + 2)
+ )
+ elif order == 3:
+ return torch.exp(-interval_end) * (
+ (interval_start**3 + 3 * interval_start**2 + 6 * interval_start + 6)
+ * torch.exp(interval_end - interval_start)
+ - (interval_end**3 + 3 * interval_end**2 + 6 * interval_end + 6)
+ )
+
+ def get_coefficients_exponential_positive(self, order, interval_start, interval_end, tau):
+ """
+ Calculate the integral of exp(x(1+tau^2)) * x^order dx from interval_start to interval_end
+ """
+ assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
+
+ # after change of variable(cov)
+ interval_end_cov = (1 + tau**2) * interval_end
+ interval_start_cov = (1 + tau**2) * interval_start
+
+ if order == 0:
+ return (
+ torch.exp(interval_end_cov) * (1 - torch.exp(-(interval_end_cov - interval_start_cov))) / (1 + tau**2)
+ )
+ elif order == 1:
+ return (
+ torch.exp(interval_end_cov)
+ * (
+ (interval_end_cov - 1)
+ - (interval_start_cov - 1) * torch.exp(-(interval_end_cov - interval_start_cov))
+ )
+ / ((1 + tau**2) ** 2)
+ )
+ elif order == 2:
+ return (
+ torch.exp(interval_end_cov)
+ * (
+ (interval_end_cov**2 - 2 * interval_end_cov + 2)
+ - (interval_start_cov**2 - 2 * interval_start_cov + 2)
+ * torch.exp(-(interval_end_cov - interval_start_cov))
+ )
+ / ((1 + tau**2) ** 3)
+ )
+ elif order == 3:
+ return (
+ torch.exp(interval_end_cov)
+ * (
+ (interval_end_cov**3 - 3 * interval_end_cov**2 + 6 * interval_end_cov - 6)
+ - (interval_start_cov**3 - 3 * interval_start_cov**2 + 6 * interval_start_cov - 6)
+ * torch.exp(-(interval_end_cov - interval_start_cov))
+ )
+ / ((1 + tau**2) ** 4)
+ )
+
+ def lagrange_polynomial_coefficient(self, order, lambda_list):
+ """
+ Calculate the coefficient of lagrange polynomial
+ """
+
+ assert order in [0, 1, 2, 3]
+ assert order == len(lambda_list) - 1
+ if order == 0:
+ return [[1]]
+ elif order == 1:
+ return [
+ [
+ 1 / (lambda_list[0] - lambda_list[1]),
+ -lambda_list[1] / (lambda_list[0] - lambda_list[1]),
+ ],
+ [
+ 1 / (lambda_list[1] - lambda_list[0]),
+ -lambda_list[0] / (lambda_list[1] - lambda_list[0]),
+ ],
+ ]
+ elif order == 2:
+ denominator1 = (lambda_list[0] - lambda_list[1]) * (lambda_list[0] - lambda_list[2])
+ denominator2 = (lambda_list[1] - lambda_list[0]) * (lambda_list[1] - lambda_list[2])
+ denominator3 = (lambda_list[2] - lambda_list[0]) * (lambda_list[2] - lambda_list[1])
+ return [
+ [
+ 1 / denominator1,
+ (-lambda_list[1] - lambda_list[2]) / denominator1,
+ lambda_list[1] * lambda_list[2] / denominator1,
+ ],
+ [
+ 1 / denominator2,
+ (-lambda_list[0] - lambda_list[2]) / denominator2,
+ lambda_list[0] * lambda_list[2] / denominator2,
+ ],
+ [
+ 1 / denominator3,
+ (-lambda_list[0] - lambda_list[1]) / denominator3,
+ lambda_list[0] * lambda_list[1] / denominator3,
+ ],
+ ]
+ elif order == 3:
+ denominator1 = (
+ (lambda_list[0] - lambda_list[1])
+ * (lambda_list[0] - lambda_list[2])
+ * (lambda_list[0] - lambda_list[3])
+ )
+ denominator2 = (
+ (lambda_list[1] - lambda_list[0])
+ * (lambda_list[1] - lambda_list[2])
+ * (lambda_list[1] - lambda_list[3])
+ )
+ denominator3 = (
+ (lambda_list[2] - lambda_list[0])
+ * (lambda_list[2] - lambda_list[1])
+ * (lambda_list[2] - lambda_list[3])
+ )
+ denominator4 = (
+ (lambda_list[3] - lambda_list[0])
+ * (lambda_list[3] - lambda_list[1])
+ * (lambda_list[3] - lambda_list[2])
+ )
+ return [
+ [
+ 1 / denominator1,
+ (-lambda_list[1] - lambda_list[2] - lambda_list[3]) / denominator1,
+ (
+ lambda_list[1] * lambda_list[2]
+ + lambda_list[1] * lambda_list[3]
+ + lambda_list[2] * lambda_list[3]
+ )
+ / denominator1,
+ (-lambda_list[1] * lambda_list[2] * lambda_list[3]) / denominator1,
+ ],
+ [
+ 1 / denominator2,
+ (-lambda_list[0] - lambda_list[2] - lambda_list[3]) / denominator2,
+ (
+ lambda_list[0] * lambda_list[2]
+ + lambda_list[0] * lambda_list[3]
+ + lambda_list[2] * lambda_list[3]
+ )
+ / denominator2,
+ (-lambda_list[0] * lambda_list[2] * lambda_list[3]) / denominator2,
+ ],
+ [
+ 1 / denominator3,
+ (-lambda_list[0] - lambda_list[1] - lambda_list[3]) / denominator3,
+ (
+ lambda_list[0] * lambda_list[1]
+ + lambda_list[0] * lambda_list[3]
+ + lambda_list[1] * lambda_list[3]
+ )
+ / denominator3,
+ (-lambda_list[0] * lambda_list[1] * lambda_list[3]) / denominator3,
+ ],
+ [
+ 1 / denominator4,
+ (-lambda_list[0] - lambda_list[1] - lambda_list[2]) / denominator4,
+ (
+ lambda_list[0] * lambda_list[1]
+ + lambda_list[0] * lambda_list[2]
+ + lambda_list[1] * lambda_list[2]
+ )
+ / denominator4,
+ (-lambda_list[0] * lambda_list[1] * lambda_list[2]) / denominator4,
+ ],
+ ]
+
+ def get_coefficients_fn(self, order, interval_start, interval_end, lambda_list, tau):
+ assert order in [1, 2, 3, 4]
+ assert order == len(lambda_list), "the length of lambda list must be equal to the order"
+ coefficients = []
+ lagrange_coefficient = self.lagrange_polynomial_coefficient(order - 1, lambda_list)
+ for i in range(order):
+ coefficient = 0
+ for j in range(order):
+ if self.predict_x0:
+ coefficient += lagrange_coefficient[i][j] * self.get_coefficients_exponential_positive(
+ order - 1 - j, interval_start, interval_end, tau
+ )
+ else:
+ coefficient += lagrange_coefficient[i][j] * self.get_coefficients_exponential_negative(
+ order - 1 - j, interval_start, interval_end
+ )
+ coefficients.append(coefficient)
+ assert len(coefficients) == order, "the length of coefficients does not match the order"
+ return coefficients
+
+ def stochastic_adams_bashforth_update(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ order: int,
+ tau: torch.FloatTensor,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the SA-Predictor.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model at the current timestep.
+ prev_timestep (`int`):
+ The previous discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ order (`int`):
+ The order of SA-Predictor at this timestep.
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ prev_timestep = args[0] if len(args) > 0 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 1:
+ sample = args[1]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if noise is None:
+ if len(args) > 2:
+ noise = args[2]
+ else:
+ raise ValueError(" missing `noise` as a required keyward argument")
+ if order is None:
+ if len(args) > 3:
+ order = args[3]
+ else:
+ raise ValueError(" missing `order` as a required keyward argument")
+ if tau is None:
+ if len(args) > 4:
+ tau = args[4]
+ else:
+ raise ValueError(" missing `tau` as a required keyward argument")
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+ model_output_list = self.model_outputs
+ sigma_t, sigma_s0 = (
+ self.sigmas[self.step_index + 1],
+ self.sigmas[self.step_index],
+ )
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+
+ gradient_part = torch.zeros_like(sample)
+ h = lambda_t - lambda_s0
+ lambda_list = []
+
+ for i in range(order):
+ si = self.step_index - i
+ alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si])
+ lambda_si = torch.log(alpha_si) - torch.log(sigma_si)
+ lambda_list.append(lambda_si)
+
+ gradient_coefficients = self.get_coefficients_fn(order, lambda_s0, lambda_t, lambda_list, tau)
+
+ x = sample
+
+ if self.predict_x0:
+ if (
+ order == 2
+ ): ## if order = 2 we do a modification that does not influence the convergence order similar to unipc. Note: This is used only for few steps sampling.
+ # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
+ # ODE case
+ # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
+ # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
+ temp_sigma = self.sigmas[self.step_index - 1]
+ temp_alpha_s, temp_sigma_s = self._sigma_to_alpha_sigma_t(temp_sigma)
+ temp_lambda_s = torch.log(temp_alpha_s) - torch.log(temp_sigma_s)
+ gradient_coefficients[0] += (
+ 1.0
+ * torch.exp((1 + tau**2) * lambda_t)
+ * (h**2 / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2))
+ / (lambda_s0 - temp_lambda_s)
+ )
+ gradient_coefficients[1] -= (
+ 1.0
+ * torch.exp((1 + tau**2) * lambda_t)
+ * (h**2 / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2))
+ / (lambda_s0 - temp_lambda_s)
+ )
+
+ for i in range(order):
+ if self.predict_x0:
+ gradient_part += (
+ (1 + tau**2)
+ * sigma_t
+ * torch.exp(-(tau**2) * lambda_t)
+ * gradient_coefficients[i]
+ * model_output_list[-(i + 1)]
+ )
+ else:
+ gradient_part += -(1 + tau**2) * alpha_t * gradient_coefficients[i] * model_output_list[-(i + 1)]
+
+ if self.predict_x0:
+ noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+ else:
+ noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+ if self.predict_x0:
+ x_t = torch.exp(-(tau**2) * h) * (sigma_t / sigma_s0) * x + gradient_part + noise_part
+ else:
+ x_t = (alpha_t / alpha_s0) * x + gradient_part + noise_part
+
+ x_t = x_t.to(x.dtype)
+ return x_t
+
+ def stochastic_adams_moulton_update(
+ self,
+ this_model_output: torch.FloatTensor,
+ *args,
+ last_sample: torch.FloatTensor,
+ last_noise: torch.FloatTensor,
+ this_sample: torch.FloatTensor,
+ order: int,
+ tau: torch.FloatTensor,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the SA-Corrector.
+
+ Args:
+ this_model_output (`torch.FloatTensor`):
+ The model outputs at `x_t`.
+ this_timestep (`int`):
+ The current timestep `t`.
+ last_sample (`torch.FloatTensor`):
+ The generated sample before the last predictor `x_{t-1}`.
+ this_sample (`torch.FloatTensor`):
+ The generated sample after the last predictor `x_{t}`.
+ order (`int`):
+ The order of SA-Corrector at this step.
+
+ Returns:
+ `torch.FloatTensor`:
+ The corrected sample tensor at the current timestep.
+ """
+
+ this_timestep = args[0] if len(args) > 0 else kwargs.pop("this_timestep", None)
+ if last_sample is None:
+ if len(args) > 1:
+ last_sample = args[1]
+ else:
+ raise ValueError(" missing`last_sample` as a required keyward argument")
+ if last_noise is None:
+ if len(args) > 2:
+ last_noise = args[2]
+ else:
+ raise ValueError(" missing`last_noise` as a required keyward argument")
+ if this_sample is None:
+ if len(args) > 3:
+ this_sample = args[3]
+ else:
+ raise ValueError(" missing`this_sample` as a required keyward argument")
+ if order is None:
+ if len(args) > 4:
+ order = args[4]
+ else:
+ raise ValueError(" missing`order` as a required keyward argument")
+ if tau is None:
+ if len(args) > 5:
+ tau = args[5]
+ else:
+ raise ValueError(" missing`tau` as a required keyward argument")
+ if this_timestep is not None:
+ deprecate(
+ "this_timestep",
+ "1.0.0",
+ "Passing `this_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ model_output_list = self.model_outputs
+ sigma_t, sigma_s0 = (
+ self.sigmas[self.step_index],
+ self.sigmas[self.step_index - 1],
+ )
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+ gradient_part = torch.zeros_like(this_sample)
+ h = lambda_t - lambda_s0
+ lambda_list = []
+ for i in range(order):
+ si = self.step_index - i
+ alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si])
+ lambda_si = torch.log(alpha_si) - torch.log(sigma_si)
+ lambda_list.append(lambda_si)
+
+ model_prev_list = model_output_list + [this_model_output]
+
+ gradient_coefficients = self.get_coefficients_fn(order, lambda_s0, lambda_t, lambda_list, tau)
+
+ x = last_sample
+
+ if self.predict_x0:
+ if (
+ order == 2
+ ): ## if order = 2 we do a modification that does not influence the convergence order similar to UniPC. Note: This is used only for few steps sampling.
+ # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
+ # ODE case
+ # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
+ # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
+ gradient_coefficients[0] += (
+ 1.0
+ * torch.exp((1 + tau**2) * lambda_t)
+ * (h / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2 * h))
+ )
+ gradient_coefficients[1] -= (
+ 1.0
+ * torch.exp((1 + tau**2) * lambda_t)
+ * (h / 2 - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h))) / ((1 + tau**2) ** 2 * h))
+ )
+
+ for i in range(order):
+ if self.predict_x0:
+ gradient_part += (
+ (1 + tau**2)
+ * sigma_t
+ * torch.exp(-(tau**2) * lambda_t)
+ * gradient_coefficients[i]
+ * model_prev_list[-(i + 1)]
+ )
+ else:
+ gradient_part += -(1 + tau**2) * alpha_t * gradient_coefficients[i] * model_prev_list[-(i + 1)]
+
+ if self.predict_x0:
+ noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * last_noise
+ else:
+ noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * last_noise
+
+ if self.predict_x0:
+ x_t = torch.exp(-(tau**2) * h) * (sigma_t / sigma_s0) * x + gradient_part + noise_part
+ else:
+ x_t = (alpha_t / alpha_s0) * x + gradient_part + noise_part
+
+ x_t = x_t.to(x.dtype)
+ return x_t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ index_candidates = (schedule_timesteps == timestep).nonzero()
+
+ if len(index_candidates) == 0:
+ step_index = len(self.timesteps) - 1
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ elif len(index_candidates) > 1:
+ step_index = index_candidates[1].item()
+ else:
+ step_index = index_candidates[0].item()
+
+ return step_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ """
+ Initialize the step_index counter for the scheduler.
+ """
+
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator=None,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the SA-Solver.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ use_corrector = self.step_index > 0 and self.last_sample is not None
+
+ model_output_convert = self.convert_model_output(model_output, sample=sample)
+
+ if use_corrector:
+ current_tau = self.tau_func(self.timestep_list[-1])
+ sample = self.stochastic_adams_moulton_update(
+ this_model_output=model_output_convert,
+ last_sample=self.last_sample,
+ last_noise=self.last_noise,
+ this_sample=sample,
+ order=self.this_corrector_order,
+ tau=current_tau,
+ )
+
+ for i in range(max(self.config.predictor_order, self.config.corrector_order - 1) - 1):
+ self.model_outputs[i] = self.model_outputs[i + 1]
+ self.timestep_list[i] = self.timestep_list[i + 1]
+
+ self.model_outputs[-1] = model_output_convert
+ self.timestep_list[-1] = timestep
+
+ noise = randn_tensor(
+ model_output.shape,
+ generator=generator,
+ device=model_output.device,
+ dtype=model_output.dtype,
+ )
+
+ if self.config.lower_order_final:
+ this_predictor_order = min(self.config.predictor_order, len(self.timesteps) - self.step_index)
+ this_corrector_order = min(self.config.corrector_order, len(self.timesteps) - self.step_index + 1)
+ else:
+ this_predictor_order = self.config.predictor_order
+ this_corrector_order = self.config.corrector_order
+
+ self.this_predictor_order = min(this_predictor_order, self.lower_order_nums + 1) # warmup for multistep
+ self.this_corrector_order = min(this_corrector_order, self.lower_order_nums + 2) # warmup for multistep
+ assert self.this_predictor_order > 0
+ assert self.this_corrector_order > 0
+
+ self.last_sample = sample
+ self.last_noise = noise
+
+ current_tau = self.tau_func(self.timestep_list[-1])
+ prev_sample = self.stochastic_adams_bashforth_update(
+ model_output=model_output_convert,
+ sample=sample,
+ noise=noise,
+ order=self.this_predictor_order,
+ tau=current_tau,
+ )
+
+ if self.lower_order_nums < max(self.config.predictor_order, self.config.corrector_order - 1):
+ self.lower_order_nums += 1
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+ # for the subsequent add_noise calls
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_sde_ve.py b/model/diffusers_c/schedulers/scheduling_sde_ve.py
new file mode 100644
index 0000000000000000000000000000000000000000..8f8dd1877331f6bf2fbb9908092b3a1174f1228f
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_sde_ve.py
@@ -0,0 +1,301 @@
+# Copyright 2024 Google Brain and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/yang-song/score_sde_pytorch
+
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin, SchedulerOutput
+
+
+@dataclass
+class SdeVeOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ prev_sample_mean (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Mean averaged `prev_sample` over previous timesteps.
+ """
+
+ prev_sample: torch.FloatTensor
+ prev_sample_mean: torch.FloatTensor
+
+
+class ScoreSdeVeScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `ScoreSdeVeScheduler` is a variance exploding stochastic differential equation (SDE) scheduler.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ snr (`float`, defaults to 0.15):
+ A coefficient weighting the step from the `model_output` sample (from the network) to the random noise.
+ sigma_min (`float`, defaults to 0.01):
+ The initial noise scale for the sigma sequence in the sampling procedure. The minimum sigma should mirror
+ the distribution of the data.
+ sigma_max (`float`, defaults to 1348.0):
+ The maximum value used for the range of continuous timesteps passed into the model.
+ sampling_eps (`float`, defaults to 1e-5):
+ The end value of sampling where timesteps decrease progressively from 1 to epsilon.
+ correct_steps (`int`, defaults to 1):
+ The number of correction steps performed on a produced sample.
+ """
+
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 2000,
+ snr: float = 0.15,
+ sigma_min: float = 0.01,
+ sigma_max: float = 1348.0,
+ sampling_eps: float = 1e-5,
+ correct_steps: int = 1,
+ ):
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = sigma_max
+
+ # setable values
+ self.timesteps = None
+
+ self.set_sigmas(num_train_timesteps, sigma_min, sigma_max, sampling_eps)
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ def set_timesteps(
+ self, num_inference_steps: int, sampling_eps: float = None, device: Union[str, torch.device] = None
+ ):
+ """
+ Sets the continuous timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ sampling_eps (`float`, *optional*):
+ The final timestep value (overrides value given during scheduler instantiation).
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+
+ """
+ sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps
+
+ self.timesteps = torch.linspace(1, sampling_eps, num_inference_steps, device=device)
+
+ def set_sigmas(
+ self, num_inference_steps: int, sigma_min: float = None, sigma_max: float = None, sampling_eps: float = None
+ ):
+ """
+ Sets the noise scales used for the diffusion chain (to be run before inference). The sigmas control the weight
+ of the `drift` and `diffusion` components of the sample update.
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ sigma_min (`float`, optional):
+ The initial noise scale value (overrides value given during scheduler instantiation).
+ sigma_max (`float`, optional):
+ The final noise scale value (overrides value given during scheduler instantiation).
+ sampling_eps (`float`, optional):
+ The final timestep value (overrides value given during scheduler instantiation).
+
+ """
+ sigma_min = sigma_min if sigma_min is not None else self.config.sigma_min
+ sigma_max = sigma_max if sigma_max is not None else self.config.sigma_max
+ sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps
+ if self.timesteps is None:
+ self.set_timesteps(num_inference_steps, sampling_eps)
+
+ self.sigmas = sigma_min * (sigma_max / sigma_min) ** (self.timesteps / sampling_eps)
+ self.discrete_sigmas = torch.exp(torch.linspace(math.log(sigma_min), math.log(sigma_max), num_inference_steps))
+ self.sigmas = torch.tensor([sigma_min * (sigma_max / sigma_min) ** t for t in self.timesteps])
+
+ def get_adjacent_sigma(self, timesteps, t):
+ return torch.where(
+ timesteps == 0,
+ torch.zeros_like(t.to(timesteps.device)),
+ self.discrete_sigmas[timesteps - 1].to(timesteps.device),
+ )
+
+ def step_pred(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[SdeVeOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_sde_ve.SdeVeOutput`] is returned, otherwise a tuple
+ is returned where the first element is the sample tensor.
+
+ """
+ if self.timesteps is None:
+ raise ValueError(
+ "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ timestep = timestep * torch.ones(
+ sample.shape[0], device=sample.device
+ ) # torch.repeat_interleave(timestep, sample.shape[0])
+ timesteps = (timestep * (len(self.timesteps) - 1)).long()
+
+ # mps requires indices to be in the same device, so we use cpu as is the default with cuda
+ timesteps = timesteps.to(self.discrete_sigmas.device)
+
+ sigma = self.discrete_sigmas[timesteps].to(sample.device)
+ adjacent_sigma = self.get_adjacent_sigma(timesteps, timestep).to(sample.device)
+ drift = torch.zeros_like(sample)
+ diffusion = (sigma**2 - adjacent_sigma**2) ** 0.5
+
+ # equation 6 in the paper: the model_output modeled by the network is grad_x log pt(x)
+ # also equation 47 shows the analog from SDE models to ancestral sampling methods
+ diffusion = diffusion.flatten()
+ while len(diffusion.shape) < len(sample.shape):
+ diffusion = diffusion.unsqueeze(-1)
+ drift = drift - diffusion**2 * model_output
+
+ # equation 6: sample noise for the diffusion term of
+ noise = randn_tensor(
+ sample.shape, layout=sample.layout, generator=generator, device=sample.device, dtype=sample.dtype
+ )
+ prev_sample_mean = sample - drift # subtract because `dt` is a small negative timestep
+ # TODO is the variable diffusion the correct scaling term for the noise?
+ prev_sample = prev_sample_mean + diffusion * noise # add impact of diffusion field g
+
+ if not return_dict:
+ return (prev_sample, prev_sample_mean)
+
+ return SdeVeOutput(prev_sample=prev_sample, prev_sample_mean=prev_sample_mean)
+
+ def step_correct(
+ self,
+ model_output: torch.FloatTensor,
+ sample: torch.FloatTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Correct the predicted sample based on the `model_output` of the network. This is often run repeatedly after
+ making the prediction for the previous timestep.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_sde_ve.SdeVeOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_sde_ve.SdeVeOutput`] is returned, otherwise a tuple
+ is returned where the first element is the sample tensor.
+
+ """
+ if self.timesteps is None:
+ raise ValueError(
+ "`self.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ # For small batch sizes, the paper "suggest replacing norm(z) with sqrt(d), where d is the dim. of z"
+ # sample noise for correction
+ noise = randn_tensor(sample.shape, layout=sample.layout, generator=generator).to(sample.device)
+
+ # compute step size from the model_output, the noise, and the snr
+ grad_norm = torch.norm(model_output.reshape(model_output.shape[0], -1), dim=-1).mean()
+ noise_norm = torch.norm(noise.reshape(noise.shape[0], -1), dim=-1).mean()
+ step_size = (self.config.snr * noise_norm / grad_norm) ** 2 * 2
+ step_size = step_size * torch.ones(sample.shape[0]).to(sample.device)
+ # self.repeat_scalar(step_size, sample.shape[0])
+
+ # compute corrected sample: model_output term and noise term
+ step_size = step_size.flatten()
+ while len(step_size.shape) < len(sample.shape):
+ step_size = step_size.unsqueeze(-1)
+ prev_sample_mean = sample + step_size * model_output
+ prev_sample = prev_sample_mean + ((step_size * 2) ** 0.5) * noise
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.FloatTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ timesteps = timesteps.to(original_samples.device)
+ sigmas = self.discrete_sigmas.to(original_samples.device)[timesteps]
+ noise = (
+ noise * sigmas[:, None, None, None]
+ if noise is not None
+ else torch.randn_like(original_samples) * sigmas[:, None, None, None]
+ )
+ noisy_samples = noise + original_samples
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_sde_ve_flax.py b/model/diffusers_c/schedulers/scheduling_sde_ve_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..0a8d45d4acbcfb6dd922aa86058aee8bb30893b5
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_sde_ve_flax.py
@@ -0,0 +1,280 @@
+# Copyright 2024 Google Brain and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This file is strongly influenced by https://github.com/yang-song/score_sde_pytorch
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import flax
+import jax
+import jax.numpy as jnp
+from jax import random
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from .scheduling_utils_flax import FlaxSchedulerMixin, FlaxSchedulerOutput, broadcast_to_shape_from_left
+
+
+@flax.struct.dataclass
+class ScoreSdeVeSchedulerState:
+ # setable values
+ timesteps: Optional[jnp.ndarray] = None
+ discrete_sigmas: Optional[jnp.ndarray] = None
+ sigmas: Optional[jnp.ndarray] = None
+
+ @classmethod
+ def create(cls):
+ return cls()
+
+
+@dataclass
+class FlaxSdeVeOutput(FlaxSchedulerOutput):
+ """
+ Output class for the ScoreSdeVeScheduler's step function output.
+
+ Args:
+ state (`ScoreSdeVeSchedulerState`):
+ prev_sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ prev_sample_mean (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images):
+ Mean averaged `prev_sample`. Same as `prev_sample`, only mean-averaged over previous timesteps.
+ """
+
+ state: ScoreSdeVeSchedulerState
+ prev_sample: jnp.ndarray
+ prev_sample_mean: Optional[jnp.ndarray] = None
+
+
+class FlaxScoreSdeVeScheduler(FlaxSchedulerMixin, ConfigMixin):
+ """
+ The variance exploding stochastic differential equation (SDE) scheduler.
+
+ For more information, see the original paper: https://arxiv.org/abs/2011.13456
+
+ [`~ConfigMixin`] takes care of storing all config attributes that are passed in the scheduler's `__init__`
+ function, such as `num_train_timesteps`. They can be accessed via `scheduler.config.num_train_timesteps`.
+ [`SchedulerMixin`] provides general loading and saving functionality via the [`SchedulerMixin.save_pretrained`] and
+ [`~SchedulerMixin.from_pretrained`] functions.
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ snr (`float`):
+ coefficient weighting the step from the model_output sample (from the network) to the random noise.
+ sigma_min (`float`):
+ initial noise scale for sigma sequence in sampling procedure. The minimum sigma should mirror the
+ distribution of the data.
+ sigma_max (`float`): maximum value used for the range of continuous timesteps passed into the model.
+ sampling_eps (`float`): the end value of sampling, where timesteps decrease progressively from 1 to
+ epsilon.
+ correct_steps (`int`): number of correction steps performed on a produced sample.
+ """
+
+ @property
+ def has_state(self):
+ return True
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 2000,
+ snr: float = 0.15,
+ sigma_min: float = 0.01,
+ sigma_max: float = 1348.0,
+ sampling_eps: float = 1e-5,
+ correct_steps: int = 1,
+ ):
+ pass
+
+ def create_state(self):
+ state = ScoreSdeVeSchedulerState.create()
+ return self.set_sigmas(
+ state,
+ self.config.num_train_timesteps,
+ self.config.sigma_min,
+ self.config.sigma_max,
+ self.config.sampling_eps,
+ )
+
+ def set_timesteps(
+ self, state: ScoreSdeVeSchedulerState, num_inference_steps: int, shape: Tuple = (), sampling_eps: float = None
+ ) -> ScoreSdeVeSchedulerState:
+ """
+ Sets the continuous timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Args:
+ state (`ScoreSdeVeSchedulerState`): the `FlaxScoreSdeVeScheduler` state data class instance.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ sampling_eps (`float`, optional):
+ final timestep value (overrides value given at Scheduler instantiation).
+
+ """
+ sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps
+
+ timesteps = jnp.linspace(1, sampling_eps, num_inference_steps)
+ return state.replace(timesteps=timesteps)
+
+ def set_sigmas(
+ self,
+ state: ScoreSdeVeSchedulerState,
+ num_inference_steps: int,
+ sigma_min: float = None,
+ sigma_max: float = None,
+ sampling_eps: float = None,
+ ) -> ScoreSdeVeSchedulerState:
+ """
+ Sets the noise scales used for the diffusion chain. Supporting function to be run before inference.
+
+ The sigmas control the weight of the `drift` and `diffusion` components of sample update.
+
+ Args:
+ state (`ScoreSdeVeSchedulerState`): the `FlaxScoreSdeVeScheduler` state data class instance.
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ sigma_min (`float`, optional):
+ initial noise scale value (overrides value given at Scheduler instantiation).
+ sigma_max (`float`, optional):
+ final noise scale value (overrides value given at Scheduler instantiation).
+ sampling_eps (`float`, optional):
+ final timestep value (overrides value given at Scheduler instantiation).
+ """
+ sigma_min = sigma_min if sigma_min is not None else self.config.sigma_min
+ sigma_max = sigma_max if sigma_max is not None else self.config.sigma_max
+ sampling_eps = sampling_eps if sampling_eps is not None else self.config.sampling_eps
+ if state.timesteps is None:
+ state = self.set_timesteps(state, num_inference_steps, sampling_eps)
+
+ discrete_sigmas = jnp.exp(jnp.linspace(jnp.log(sigma_min), jnp.log(sigma_max), num_inference_steps))
+ sigmas = jnp.array([sigma_min * (sigma_max / sigma_min) ** t for t in state.timesteps])
+
+ return state.replace(discrete_sigmas=discrete_sigmas, sigmas=sigmas)
+
+ def get_adjacent_sigma(self, state, timesteps, t):
+ return jnp.where(timesteps == 0, jnp.zeros_like(t), state.discrete_sigmas[timesteps - 1])
+
+ def step_pred(
+ self,
+ state: ScoreSdeVeSchedulerState,
+ model_output: jnp.ndarray,
+ timestep: int,
+ sample: jnp.ndarray,
+ key: jax.Array,
+ return_dict: bool = True,
+ ) -> Union[FlaxSdeVeOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ state (`ScoreSdeVeSchedulerState`): the `FlaxScoreSdeVeScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ generator: random number generator.
+ return_dict (`bool`): option for returning tuple rather than FlaxSdeVeOutput class
+
+ Returns:
+ [`FlaxSdeVeOutput`] or `tuple`: [`FlaxSdeVeOutput`] if `return_dict` is True, otherwise a `tuple`. When
+ returning a tuple, the first element is the sample tensor.
+
+ """
+ if state.timesteps is None:
+ raise ValueError(
+ "`state.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ timestep = timestep * jnp.ones(
+ sample.shape[0],
+ )
+ timesteps = (timestep * (len(state.timesteps) - 1)).long()
+
+ sigma = state.discrete_sigmas[timesteps]
+ adjacent_sigma = self.get_adjacent_sigma(state, timesteps, timestep)
+ drift = jnp.zeros_like(sample)
+ diffusion = (sigma**2 - adjacent_sigma**2) ** 0.5
+
+ # equation 6 in the paper: the model_output modeled by the network is grad_x log pt(x)
+ # also equation 47 shows the analog from SDE models to ancestral sampling methods
+ diffusion = diffusion.flatten()
+ diffusion = broadcast_to_shape_from_left(diffusion, sample.shape)
+ drift = drift - diffusion**2 * model_output
+
+ # equation 6: sample noise for the diffusion term of
+ key = random.split(key, num=1)
+ noise = random.normal(key=key, shape=sample.shape)
+ prev_sample_mean = sample - drift # subtract because `dt` is a small negative timestep
+ # TODO is the variable diffusion the correct scaling term for the noise?
+ prev_sample = prev_sample_mean + diffusion * noise # add impact of diffusion field g
+
+ if not return_dict:
+ return (prev_sample, prev_sample_mean, state)
+
+ return FlaxSdeVeOutput(prev_sample=prev_sample, prev_sample_mean=prev_sample_mean, state=state)
+
+ def step_correct(
+ self,
+ state: ScoreSdeVeSchedulerState,
+ model_output: jnp.ndarray,
+ sample: jnp.ndarray,
+ key: jax.Array,
+ return_dict: bool = True,
+ ) -> Union[FlaxSdeVeOutput, Tuple]:
+ """
+ Correct the predicted sample based on the output model_output of the network. This is often run repeatedly
+ after making the prediction for the previous timestep.
+
+ Args:
+ state (`ScoreSdeVeSchedulerState`): the `FlaxScoreSdeVeScheduler` state data class instance.
+ model_output (`jnp.ndarray`): direct output from learned diffusion model.
+ sample (`jnp.ndarray`):
+ current instance of sample being created by diffusion process.
+ generator: random number generator.
+ return_dict (`bool`): option for returning tuple rather than FlaxSdeVeOutput class
+
+ Returns:
+ [`FlaxSdeVeOutput`] or `tuple`: [`FlaxSdeVeOutput`] if `return_dict` is True, otherwise a `tuple`. When
+ returning a tuple, the first element is the sample tensor.
+
+ """
+ if state.timesteps is None:
+ raise ValueError(
+ "`state.timesteps` is not set, you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ # For small batch sizes, the paper "suggest replacing norm(z) with sqrt(d), where d is the dim. of z"
+ # sample noise for correction
+ key = random.split(key, num=1)
+ noise = random.normal(key=key, shape=sample.shape)
+
+ # compute step size from the model_output, the noise, and the snr
+ grad_norm = jnp.linalg.norm(model_output)
+ noise_norm = jnp.linalg.norm(noise)
+ step_size = (self.config.snr * noise_norm / grad_norm) ** 2 * 2
+ step_size = step_size * jnp.ones(sample.shape[0])
+
+ # compute corrected sample: model_output term and noise term
+ step_size = step_size.flatten()
+ step_size = broadcast_to_shape_from_left(step_size, sample.shape)
+ prev_sample_mean = sample + step_size * model_output
+ prev_sample = prev_sample_mean + ((step_size * 2) ** 0.5) * noise
+
+ if not return_dict:
+ return (prev_sample, state)
+
+ return FlaxSdeVeOutput(prev_sample=prev_sample, state=state)
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_tcd.py b/model/diffusers_c/schedulers/scheduling_tcd.py
new file mode 100644
index 0000000000000000000000000000000000000000..2f1234256ca506bfc1f3cfd0b4e7b6fe1901248c
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_tcd.py
@@ -0,0 +1,688 @@
+# Copyright 2024 Stanford University Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..schedulers.scheduling_utils import SchedulerMixin
+from ..utils import BaseOutput, logging
+from ..utils.torch_utils import randn_tensor
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class TCDSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_noised_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted noised sample `(x_{s})` based on the model output from the current timestep.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_noised_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+ Args:
+ betas (`torch.FloatTensor`):
+ the betas that the scheduler is being initialized with.
+
+ Returns:
+ `torch.FloatTensor`: rescaled betas with zero terminal SNR
+ """
+ # Convert betas to alphas_bar_sqrt
+ alphas = 1.0 - betas
+ alphas_cumprod = torch.cumprod(alphas, dim=0)
+ alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+ # Store old values.
+ alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+ alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+ # Shift so the last timestep is zero.
+ alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+ # Scale so the first timestep is back to the old value.
+ alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+ # Convert alphas_bar_sqrt to betas
+ alphas_bar = alphas_bar_sqrt**2 # Revert sqrt
+ alphas = alphas_bar[1:] / alphas_bar[:-1] # Revert cumprod
+ alphas = torch.cat([alphas_bar[0:1], alphas])
+ betas = 1 - alphas
+
+ return betas
+
+
+class TCDScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `TCDScheduler` incorporates the `Strategic Stochastic Sampling` introduced by the paper `Trajectory Consistency Distillation`,
+ extending the original Multistep Consistency Sampling to enable unrestricted trajectory traversal.
+
+ This code is based on the official repo of TCD(https://github.com/jabir-zheng/TCD).
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config
+ attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be
+ accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving
+ functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ original_inference_steps (`int`, *optional*, defaults to 50):
+ The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we
+ will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule.
+ clip_sample (`bool`, defaults to `True`):
+ Clip the predicted sample for numerical stability.
+ clip_sample_range (`float`, defaults to 1.0):
+ The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+ set_alpha_to_one (`bool`, defaults to `True`):
+ Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+ there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+ otherwise it uses the alpha value at step 0.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+ timestep_spacing (`str`, defaults to `"leading"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ timestep_scaling (`float`, defaults to 10.0):
+ The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions
+ `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation
+ error at the default of `10.0` is already pretty small).
+ rescale_betas_zero_snr (`bool`, defaults to `False`):
+ Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+ dark samples instead of limiting it to samples with medium brightness. Loosely related to
+ [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+ """
+
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.00085,
+ beta_end: float = 0.012,
+ beta_schedule: str = "scaled_linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ original_inference_steps: int = 50,
+ clip_sample: bool = False,
+ clip_sample_range: float = 1.0,
+ set_alpha_to_one: bool = True,
+ steps_offset: int = 0,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ timestep_spacing: str = "leading",
+ timestep_scaling: float = 10.0,
+ rescale_betas_zero_snr: bool = False,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ # Rescale for zero SNR
+ if rescale_betas_zero_snr:
+ self.betas = rescale_zero_terminal_snr(self.betas)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+ # At every step in ddim, we are looking into the previous alphas_cumprod
+ # For the final step, there is no previous alphas_cumprod because we are already at 0
+ # `set_alpha_to_one` decides whether we set this parameter simply to one or
+ # whether we use the final alpha of the "non-previous" one.
+ self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
+ self.custom_timesteps = False
+
+ self._step_index = None
+ self._begin_index = None
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ indices = (schedule_timesteps == timestep).nonzero()
+
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ pos = 1 if len(indices) > 1 else 0
+
+ return indices[pos].item()
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ @property
+ def step_index(self):
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+ timestep (`int`, *optional*):
+ The current timestep in the diffusion chain.
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_ddim.DDIMScheduler._get_variance
+ def _get_variance(self, timestep, prev_timestep):
+ alpha_prod_t = self.alphas_cumprod[timestep]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
+
+ return variance
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ def set_timesteps(
+ self,
+ num_inference_steps: Optional[int] = None,
+ device: Union[str, torch.device] = None,
+ original_inference_steps: Optional[int] = None,
+ timesteps: Optional[List[int]] = None,
+ strength: int = 1.0,
+ ):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`, *optional*):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ original_inference_steps (`int`, *optional*):
+ The original number of inference steps, which will be used to generate a linearly-spaced timestep
+ schedule (which is different from the standard `diffusers` implementation). We will then take
+ `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as
+ our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of equal spacing between timesteps on the training/distillation timestep
+ schedule is used. If `timesteps` is passed, `num_inference_steps` must be `None`.
+ """
+ # 0. Check inputs
+ if num_inference_steps is None and timesteps is None:
+ raise ValueError("Must pass exactly one of `num_inference_steps` or `custom_timesteps`.")
+
+ if num_inference_steps is not None and timesteps is not None:
+ raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
+
+ # 1. Calculate the TCD original training/distillation timestep schedule.
+ original_steps = (
+ original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps
+ )
+
+ if original_inference_steps is None:
+ # default option, timesteps align with discrete inference steps
+ if original_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+ # TCD Timesteps Setting
+ # The skipping step parameter k from the paper.
+ k = self.config.num_train_timesteps // original_steps
+ # TCD Training/Distillation Steps Schedule
+ tcd_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * k - 1
+ else:
+ # customised option, sampled timesteps can be any arbitrary value
+ tcd_origin_timesteps = np.asarray(list(range(0, int(self.config.num_train_timesteps * strength))))
+
+ # 2. Calculate the TCD inference timestep schedule.
+ if timesteps is not None:
+ # 2.1 Handle custom timestep schedules.
+ train_timesteps = set(tcd_origin_timesteps)
+ non_train_timesteps = []
+ for i in range(1, len(timesteps)):
+ if timesteps[i] >= timesteps[i - 1]:
+ raise ValueError("`custom_timesteps` must be in descending order.")
+
+ if timesteps[i] not in train_timesteps:
+ non_train_timesteps.append(timesteps[i])
+
+ if timesteps[0] >= self.config.num_train_timesteps:
+ raise ValueError(
+ f"`timesteps` must start before `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps}."
+ )
+
+ # Raise warning if timestep schedule does not start with self.config.num_train_timesteps - 1
+ if strength == 1.0 and timesteps[0] != self.config.num_train_timesteps - 1:
+ logger.warning(
+ f"The first timestep on the custom timestep schedule is {timesteps[0]}, not"
+ f" `self.config.num_train_timesteps - 1`: {self.config.num_train_timesteps - 1}. You may get"
+ f" unexpected results when using this timestep schedule."
+ )
+
+ # Raise warning if custom timestep schedule contains timesteps not on original timestep schedule
+ if non_train_timesteps:
+ logger.warning(
+ f"The custom timestep schedule contains the following timesteps which are not on the original"
+ f" training/distillation timestep schedule: {non_train_timesteps}. You may get unexpected results"
+ f" when using this timestep schedule."
+ )
+
+ # Raise warning if custom timestep schedule is longer than original_steps
+ if original_steps is not None:
+ if len(timesteps) > original_steps:
+ logger.warning(
+ f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the"
+ f" the length of the timestep schedule used for training: {original_steps}. You may get some"
+ f" unexpected results when using this timestep schedule."
+ )
+ else:
+ if len(timesteps) > self.config.num_train_timesteps:
+ logger.warning(
+ f"The number of timesteps in the custom timestep schedule is {len(timesteps)}, which exceeds the"
+ f" the length of the timestep schedule used for training: {self.config.num_train_timesteps}. You may get some"
+ f" unexpected results when using this timestep schedule."
+ )
+
+ timesteps = np.array(timesteps, dtype=np.int64)
+ self.num_inference_steps = len(timesteps)
+ self.custom_timesteps = True
+
+ # Apply strength (e.g. for img2img pipelines) (see StableDiffusionImg2ImgPipeline.get_timesteps)
+ init_timestep = min(int(self.num_inference_steps * strength), self.num_inference_steps)
+ t_start = max(self.num_inference_steps - init_timestep, 0)
+ timesteps = timesteps[t_start * self.order :]
+ # TODO: also reset self.num_inference_steps?
+ else:
+ # 2.2 Create the "standard" TCD inference timestep schedule.
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+ f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+ f" maximal {self.config.num_train_timesteps} timesteps."
+ )
+
+ if original_steps is not None:
+ skipping_step = len(tcd_origin_timesteps) // num_inference_steps
+
+ if skipping_step < 1:
+ raise ValueError(
+ f"The combination of `original_steps x strength`: {original_steps} x {strength} is smaller than `num_inference_steps`: {num_inference_steps}. Make sure to either reduce `num_inference_steps` to a value smaller than {int(original_steps * strength)} or increase `strength` to a value higher than {float(num_inference_steps / original_steps)}."
+ )
+
+ self.num_inference_steps = num_inference_steps
+
+ if original_steps is not None:
+ if num_inference_steps > original_steps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:"
+ f" {original_steps} because the final timestep schedule will be a subset of the"
+ f" `original_inference_steps`-sized initial timestep schedule."
+ )
+ else:
+ if num_inference_steps > self.config.num_train_timesteps:
+ raise ValueError(
+ f"`num_inference_steps`: {num_inference_steps} cannot be larger than `num_train_timesteps`:"
+ f" {self.config.num_train_timesteps} because the final timestep schedule will be a subset of the"
+ f" `num_train_timesteps`-sized initial timestep schedule."
+ )
+
+ # TCD Inference Steps Schedule
+ tcd_origin_timesteps = tcd_origin_timesteps[::-1].copy()
+ # Select (approximately) evenly spaced indices from tcd_origin_timesteps.
+ inference_indices = np.linspace(0, len(tcd_origin_timesteps), num=num_inference_steps, endpoint=False)
+ inference_indices = np.floor(inference_indices).astype(np.int64)
+ timesteps = tcd_origin_timesteps[inference_indices]
+
+ self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.long)
+
+ self._step_index = None
+ self._begin_index = None
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ eta: float = 0.3,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[TCDSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ eta (`float`):
+ A stochastic parameter (referred to as `gamma` in the paper) used to control the stochasticity in every step.
+ When eta = 0, it represents deterministic sampling, whereas eta = 1 indicates full stochastic sampling.
+ generator (`torch.Generator`, *optional*):
+ A random number generator.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] or `tuple`.
+ Returns:
+ [`~schedulers.scheduling_utils.TCDSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_tcd.TCDSchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ assert 0 <= eta <= 1.0, "gamma must be less than or equal to 1.0"
+
+ # 1. get previous step value
+ prev_step_index = self.step_index + 1
+ if prev_step_index < len(self.timesteps):
+ prev_timestep = self.timesteps[prev_step_index]
+ else:
+ prev_timestep = torch.tensor(0)
+
+ timestep_s = torch.floor((1 - eta) * prev_timestep).to(dtype=torch.long)
+
+ # 2. compute alphas, betas
+ alpha_prod_t = self.alphas_cumprod[timestep]
+ beta_prod_t = 1 - alpha_prod_t
+
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod
+
+ alpha_prod_s = self.alphas_cumprod[timestep_s]
+ beta_prod_s = 1 - alpha_prod_s
+
+ # 3. Compute the predicted noised sample x_s based on the model parameterization
+ if self.config.prediction_type == "epsilon": # noise-prediction
+ pred_original_sample = (sample - beta_prod_t.sqrt() * model_output) / alpha_prod_t.sqrt()
+ pred_epsilon = model_output
+ pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon
+ elif self.config.prediction_type == "sample": # x-prediction
+ pred_original_sample = model_output
+ pred_epsilon = (sample - alpha_prod_t ** (0.5) * pred_original_sample) / beta_prod_t ** (0.5)
+ pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon
+ elif self.config.prediction_type == "v_prediction": # v-prediction
+ pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output
+ pred_epsilon = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample
+ pred_noised_sample = alpha_prod_s.sqrt() * pred_original_sample + beta_prod_s.sqrt() * pred_epsilon
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
+ " `v_prediction` for `TCDScheduler`."
+ )
+
+ # 4. Sample and inject noise z ~ N(0, I) for MultiStep Inference
+ # Noise is not used on the final timestep of the timestep schedule.
+ # This also means that noise is not used for one-step sampling.
+ # Eta (referred to as "gamma" in the paper) was introduced to control the stochasticity in every step.
+ # When eta = 0, it represents deterministic sampling, whereas eta = 1 indicates full stochastic sampling.
+ if eta > 0:
+ if self.step_index != self.num_inference_steps - 1:
+ noise = randn_tensor(
+ model_output.shape, generator=generator, device=model_output.device, dtype=pred_noised_sample.dtype
+ )
+ prev_sample = (alpha_prod_t_prev / alpha_prod_s).sqrt() * pred_noised_sample + (
+ 1 - alpha_prod_t_prev / alpha_prod_s
+ ).sqrt() * noise
+ else:
+ prev_sample = pred_noised_sample
+ else:
+ prev_sample = pred_noised_sample
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample, pred_noised_sample)
+
+ return TCDSchedulerOutput(prev_sample=prev_sample, pred_noised_sample=pred_noised_sample)
+
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+ def get_velocity(
+ self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as sample
+ alphas_cumprod = self.alphas_cumprod.to(device=sample.device, dtype=sample.dtype)
+ timesteps = timesteps.to(sample.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(sample.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+ return velocity
+
+ def __len__(self):
+ return self.config.num_train_timesteps
+
+ def previous_timestep(self, timestep):
+ if self.custom_timesteps:
+ index = (self.timesteps == timestep).nonzero(as_tuple=True)[0][0]
+ if index == self.timesteps.shape[0] - 1:
+ prev_t = torch.tensor(-1)
+ else:
+ prev_t = self.timesteps[index + 1]
+ else:
+ num_inference_steps = (
+ self.num_inference_steps if self.num_inference_steps else self.config.num_train_timesteps
+ )
+ prev_t = timestep - self.config.num_train_timesteps // num_inference_steps
+
+ return prev_t
diff --git a/model/diffusers_c/schedulers/scheduling_unclip.py b/model/diffusers_c/schedulers/scheduling_unclip.py
new file mode 100644
index 0000000000000000000000000000000000000000..7bc0a0f9b85b09e9082b42f2ae6d18081c2a4413
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_unclip.py
@@ -0,0 +1,352 @@
+# Copyright 2024 Kakao Brain and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from ..utils.torch_utils import randn_tensor
+from .scheduling_utils import SchedulerMixin
+
+
+@dataclass
+# Copied from diffusers.schedulers.scheduling_ddpm.DDPMSchedulerOutput with DDPM->UnCLIP
+class UnCLIPSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's `step` function output.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+ `pred_original_sample` can be used to preview progress or for guidance.
+ """
+
+ prev_sample: torch.FloatTensor
+ pred_original_sample: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class UnCLIPScheduler(SchedulerMixin, ConfigMixin):
+ """
+ NOTE: do not use this scheduler. The DDPM scheduler has been updated to support the changes made here. This
+ scheduler will be removed and replaced with DDPM.
+
+ This is a modified DDPM Scheduler specifically for the karlo unCLIP model.
+
+ This scheduler has some minor variations in how it calculates the learned range variance and dynamically
+ re-calculates betas based off the timesteps it is skipping.
+
+ The scheduler also uses a slightly different step ratio when computing timesteps to use for inference.
+
+ See [`~DDPMScheduler`] for more information on DDPM scheduling
+
+ Args:
+ num_train_timesteps (`int`): number of diffusion steps used to train the model.
+ variance_type (`str`):
+ options to clip the variance used when adding noise to the denoised sample. Choose from `fixed_small_log`
+ or `learned_range`.
+ clip_sample (`bool`, default `True`):
+ option to clip predicted sample between `-clip_sample_range` and `clip_sample_range` for numerical
+ stability.
+ clip_sample_range (`float`, default `1.0`):
+ The range to clip the sample between. See `clip_sample`.
+ prediction_type (`str`, default `epsilon`, optional):
+ prediction type of the scheduler function, one of `epsilon` (predicting the noise of the diffusion process)
+ or `sample` (directly predicting the noisy sample`)
+ """
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ variance_type: str = "fixed_small_log",
+ clip_sample: bool = True,
+ clip_sample_range: Optional[float] = 1.0,
+ prediction_type: str = "epsilon",
+ beta_schedule: str = "squaredcos_cap_v2",
+ ):
+ if beta_schedule != "squaredcos_cap_v2":
+ raise ValueError("UnCLIPScheduler only supports `beta_schedule`: 'squaredcos_cap_v2'")
+
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ self.one = torch.tensor(1.0)
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ # setable values
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy())
+
+ self.variance_type = variance_type
+
+ def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`): input sample
+ timestep (`int`, optional): current timestep
+
+ Returns:
+ `torch.FloatTensor`: scaled input sample
+ """
+ return sample
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain. Supporting function to be run before inference.
+
+ Note that this scheduler uses a slightly different step ratio than the other diffusers schedulers. The
+ different step ratio is to mimic the original karlo implementation and does not affect the quality or accuracy
+ of the results.
+
+ Args:
+ num_inference_steps (`int`):
+ the number of diffusion steps used when generating samples with a pre-trained model.
+ """
+ self.num_inference_steps = num_inference_steps
+ step_ratio = (self.config.num_train_timesteps - 1) / (self.num_inference_steps - 1)
+ timesteps = (np.arange(0, num_inference_steps) * step_ratio).round()[::-1].copy().astype(np.int64)
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ def _get_variance(self, t, prev_timestep=None, predicted_variance=None, variance_type=None):
+ if prev_timestep is None:
+ prev_timestep = t - 1
+
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.one
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ if prev_timestep == t - 1:
+ beta = self.betas[t]
+ else:
+ beta = 1 - alpha_prod_t / alpha_prod_t_prev
+
+ # For t > 0, compute predicted variance βt (see formula (6) and (7) from https://arxiv.org/pdf/2006.11239.pdf)
+ # and sample from it to get previous sample
+ # x_{t-1} ~ N(pred_prev_sample, variance) == add variance to pred_sample
+ variance = beta_prod_t_prev / beta_prod_t * beta
+
+ if variance_type is None:
+ variance_type = self.config.variance_type
+
+ # hacks - were probably added for training stability
+ if variance_type == "fixed_small_log":
+ variance = torch.log(torch.clamp(variance, min=1e-20))
+ variance = torch.exp(0.5 * variance)
+ elif variance_type == "learned_range":
+ # NOTE difference with DDPM scheduler
+ min_log = variance.log()
+ max_log = beta.log()
+
+ frac = (predicted_variance + 1) / 2
+ variance = frac * max_log + (1 - frac) * min_log
+
+ return variance
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ prev_timestep: Optional[int] = None,
+ generator=None,
+ return_dict: bool = True,
+ ) -> Union[UnCLIPSchedulerOutput, Tuple]:
+ """
+ Predict the sample at the previous timestep by reversing the SDE. Core function to propagate the diffusion
+ process from the learned model outputs (most often the predicted noise).
+
+ Args:
+ model_output (`torch.FloatTensor`): direct output from learned diffusion model.
+ timestep (`int`): current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ current instance of sample being created by diffusion process.
+ prev_timestep (`int`, *optional*): The previous timestep to predict the previous sample at.
+ Used to dynamically compute beta. If not given, `t-1` is used and the pre-computed beta is used.
+ generator: random number generator.
+ return_dict (`bool`): option for returning tuple rather than UnCLIPSchedulerOutput class
+
+ Returns:
+ [`~schedulers.scheduling_utils.UnCLIPSchedulerOutput`] or `tuple`:
+ [`~schedulers.scheduling_utils.UnCLIPSchedulerOutput`] if `return_dict` is True, otherwise a `tuple`. When
+ returning a tuple, the first element is the sample tensor.
+
+ """
+ t = timestep
+
+ if model_output.shape[1] == sample.shape[1] * 2 and self.variance_type == "learned_range":
+ model_output, predicted_variance = torch.split(model_output, sample.shape[1], dim=1)
+ else:
+ predicted_variance = None
+
+ # 1. compute alphas, betas
+ if prev_timestep is None:
+ prev_timestep = t - 1
+
+ alpha_prod_t = self.alphas_cumprod[t]
+ alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.one
+ beta_prod_t = 1 - alpha_prod_t
+ beta_prod_t_prev = 1 - alpha_prod_t_prev
+
+ if prev_timestep == t - 1:
+ beta = self.betas[t]
+ alpha = self.alphas[t]
+ else:
+ beta = 1 - alpha_prod_t / alpha_prod_t_prev
+ alpha = 1 - beta
+
+ # 2. compute predicted original sample from predicted noise also called
+ # "predicted x_0" of formula (15) from https://arxiv.org/pdf/2006.11239.pdf
+ if self.config.prediction_type == "epsilon":
+ pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
+ elif self.config.prediction_type == "sample":
+ pred_original_sample = model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `sample`"
+ " for the UnCLIPScheduler."
+ )
+
+ # 3. Clip "predicted x_0"
+ if self.config.clip_sample:
+ pred_original_sample = torch.clamp(
+ pred_original_sample, -self.config.clip_sample_range, self.config.clip_sample_range
+ )
+
+ # 4. Compute coefficients for pred_original_sample x_0 and current sample x_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_original_sample_coeff = (alpha_prod_t_prev ** (0.5) * beta) / beta_prod_t
+ current_sample_coeff = alpha ** (0.5) * beta_prod_t_prev / beta_prod_t
+
+ # 5. Compute predicted previous sample µ_t
+ # See formula (7) from https://arxiv.org/pdf/2006.11239.pdf
+ pred_prev_sample = pred_original_sample_coeff * pred_original_sample + current_sample_coeff * sample
+
+ # 6. Add noise
+ variance = 0
+ if t > 0:
+ variance_noise = randn_tensor(
+ model_output.shape, dtype=model_output.dtype, generator=generator, device=model_output.device
+ )
+
+ variance = self._get_variance(
+ t,
+ predicted_variance=predicted_variance,
+ prev_timestep=prev_timestep,
+ )
+
+ if self.variance_type == "fixed_small_log":
+ variance = variance
+ elif self.variance_type == "learned_range":
+ variance = (0.5 * variance).exp()
+ else:
+ raise ValueError(
+ f"variance_type given as {self.variance_type} must be one of `fixed_small_log` or `learned_range`"
+ " for the UnCLIPScheduler."
+ )
+
+ variance = variance * variance_noise
+
+ pred_prev_sample = pred_prev_sample + variance
+
+ if not return_dict:
+ return (pred_prev_sample,)
+
+ return UnCLIPSchedulerOutput(prev_sample=pred_prev_sample, pred_original_sample=pred_original_sample)
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+ # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement
+ # for the subsequent add_noise calls
+ self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device)
+ alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype)
+ timesteps = timesteps.to(original_samples.device)
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
diff --git a/model/diffusers_c/schedulers/scheduling_unipc_multistep.py b/model/diffusers_c/schedulers/scheduling_unipc_multistep.py
new file mode 100644
index 0000000000000000000000000000000000000000..a5ee629f713a5412e800752dd639b54da1820350
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_unipc_multistep.py
@@ -0,0 +1,882 @@
+# Copyright 2024 TSAIL Team and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# DISCLAIMER: check https://arxiv.org/abs/2302.04867 and https://github.com/wl-zhao/UniPC for more info
+# The codebase is modified based on https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_dpmsolver_multistep.py
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import deprecate
+from .scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+ num_diffusion_timesteps,
+ max_beta=0.999,
+ alpha_transform_type="cosine",
+):
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+ alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+ Choose from `cosine` or `exp`
+
+ Returns:
+ betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+ if alpha_transform_type == "cosine":
+
+ def alpha_bar_fn(t):
+ return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ elif alpha_transform_type == "exp":
+
+ def alpha_bar_fn(t):
+ return math.exp(t * -12.0)
+
+ else:
+ raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+ return torch.tensor(betas, dtype=torch.float32)
+
+
+class UniPCMultistepScheduler(SchedulerMixin, ConfigMixin):
+ """
+ `UniPCMultistepScheduler` is a training-free framework designed for the fast sampling of diffusion models.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_train_timesteps (`int`, defaults to 1000):
+ The number of diffusion steps to train the model.
+ beta_start (`float`, defaults to 0.0001):
+ The starting `beta` value of inference.
+ beta_end (`float`, defaults to 0.02):
+ The final `beta` value.
+ beta_schedule (`str`, defaults to `"linear"`):
+ The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+ `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+ trained_betas (`np.ndarray`, *optional*):
+ Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+ solver_order (`int`, default `2`):
+ The UniPC order which can be any positive integer. The effective order of accuracy is `solver_order + 1`
+ due to the UniC. It is recommended to use `solver_order=2` for guided sampling, and `solver_order=3` for
+ unconditional sampling.
+ prediction_type (`str`, defaults to `epsilon`, *optional*):
+ Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+ `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+ Video](https://imagen.research.google/video/paper.pdf) paper).
+ thresholding (`bool`, defaults to `False`):
+ Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+ as Stable Diffusion.
+ dynamic_thresholding_ratio (`float`, defaults to 0.995):
+ The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+ sample_max_value (`float`, defaults to 1.0):
+ The threshold value for dynamic thresholding. Valid only when `thresholding=True` and `predict_x0=True`.
+ predict_x0 (`bool`, defaults to `True`):
+ Whether to use the updating algorithm on the predicted x0.
+ solver_type (`str`, default `bh2`):
+ Solver type for UniPC. It is recommended to use `bh1` for unconditional sampling when steps < 10, and `bh2`
+ otherwise.
+ lower_order_final (`bool`, default `True`):
+ Whether to use lower-order solvers in the final steps. Only valid for < 15 inference steps. This can
+ stabilize the sampling of DPMSolver for steps < 15, especially for steps <= 10.
+ disable_corrector (`list`, default `[]`):
+ Decides which step to disable the corrector to mitigate the misalignment between `epsilon_theta(x_t, c)`
+ and `epsilon_theta(x_t^c, c)` which can influence convergence for a large guidance scale. Corrector is
+ usually disabled during the first few steps.
+ solver_p (`SchedulerMixin`, default `None`):
+ Any other scheduler that if specified, the algorithm becomes `solver_p + UniC`.
+ use_karras_sigmas (`bool`, *optional*, defaults to `False`):
+ Whether to use Karras sigmas for step sizes in the noise schedule during the sampling process. If `True`,
+ the sigmas are determined according to a sequence of noise levels {σi}.
+ timestep_spacing (`str`, defaults to `"linspace"`):
+ The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+ Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+ steps_offset (`int`, defaults to 0):
+ An offset added to the inference steps. You can use a combination of `offset=1` and
+ `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+ Diffusion.
+ """
+
+ _compatibles = [e.name for e in KarrasDiffusionSchedulers]
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_train_timesteps: int = 1000,
+ beta_start: float = 0.0001,
+ beta_end: float = 0.02,
+ beta_schedule: str = "linear",
+ trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+ solver_order: int = 2,
+ prediction_type: str = "epsilon",
+ thresholding: bool = False,
+ dynamic_thresholding_ratio: float = 0.995,
+ sample_max_value: float = 1.0,
+ predict_x0: bool = True,
+ solver_type: str = "bh2",
+ lower_order_final: bool = True,
+ disable_corrector: List[int] = [],
+ solver_p: SchedulerMixin = None,
+ use_karras_sigmas: Optional[bool] = False,
+ timestep_spacing: str = "linspace",
+ steps_offset: int = 0,
+ ):
+ if trained_betas is not None:
+ self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+ elif beta_schedule == "linear":
+ self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+ elif beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+ elif beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ self.betas = betas_for_alpha_bar(num_train_timesteps)
+ else:
+ raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+ self.alphas = 1.0 - self.betas
+ self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+ # Currently we only support VP-type noise schedule
+ self.alpha_t = torch.sqrt(self.alphas_cumprod)
+ self.sigma_t = torch.sqrt(1 - self.alphas_cumprod)
+ self.lambda_t = torch.log(self.alpha_t) - torch.log(self.sigma_t)
+ self.sigmas = ((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5
+
+ # standard deviation of the initial noise distribution
+ self.init_noise_sigma = 1.0
+
+ if solver_type not in ["bh1", "bh2"]:
+ if solver_type in ["midpoint", "heun", "logrho"]:
+ self.register_to_config(solver_type="bh2")
+ else:
+ raise NotImplementedError(f"{solver_type} does is not implemented for {self.__class__}")
+
+ self.predict_x0 = predict_x0
+ # setable values
+ self.num_inference_steps = None
+ timesteps = np.linspace(0, num_train_timesteps - 1, num_train_timesteps, dtype=np.float32)[::-1].copy()
+ self.timesteps = torch.from_numpy(timesteps)
+ self.model_outputs = [None] * solver_order
+ self.timestep_list = [None] * solver_order
+ self.lower_order_nums = 0
+ self.disable_corrector = disable_corrector
+ self.solver_p = solver_p
+ self.last_sample = None
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ @property
+ def step_index(self):
+ """
+ The index counter for current timestep. It will increae 1 after each scheduler step.
+ """
+ return self._step_index
+
+ @property
+ def begin_index(self):
+ """
+ The index for the first timestep. It should be set from pipeline with `set_begin_index` method.
+ """
+ return self._begin_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.set_begin_index
+ def set_begin_index(self, begin_index: int = 0):
+ """
+ Sets the begin index for the scheduler. This function should be run from pipeline before the inference.
+
+ Args:
+ begin_index (`int`):
+ The begin index for the scheduler.
+ """
+ self._begin_index = begin_index
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ """
+ # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891
+ if self.config.timestep_spacing == "linspace":
+ timesteps = (
+ np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps + 1)
+ .round()[::-1][:-1]
+ .copy()
+ .astype(np.int64)
+ )
+ elif self.config.timestep_spacing == "leading":
+ step_ratio = self.config.num_train_timesteps // (num_inference_steps + 1)
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = (np.arange(0, num_inference_steps + 1) * step_ratio).round()[::-1][:-1].copy().astype(np.int64)
+ timesteps += self.config.steps_offset
+ elif self.config.timestep_spacing == "trailing":
+ step_ratio = self.config.num_train_timesteps / num_inference_steps
+ # creates integer timesteps by multiplying by ratio
+ # casting to int to avoid issues when num_inference_step is power of 3
+ timesteps = np.arange(self.config.num_train_timesteps, 0, -step_ratio).round().copy().astype(np.int64)
+ timesteps -= 1
+ else:
+ raise ValueError(
+ f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'."
+ )
+
+ sigmas = np.array(((1 - self.alphas_cumprod) / self.alphas_cumprod) ** 0.5)
+ if self.config.use_karras_sigmas:
+ log_sigmas = np.log(sigmas)
+ sigmas = np.flip(sigmas).copy()
+ sigmas = self._convert_to_karras(in_sigmas=sigmas, num_inference_steps=num_inference_steps)
+ timesteps = np.array([self._sigma_to_t(sigma, log_sigmas) for sigma in sigmas]).round()
+ sigmas = np.concatenate([sigmas, sigmas[-1:]]).astype(np.float32)
+ else:
+ sigmas = np.interp(timesteps, np.arange(0, len(sigmas)), sigmas)
+ sigma_last = ((1 - self.alphas_cumprod[0]) / self.alphas_cumprod[0]) ** 0.5
+ sigmas = np.concatenate([sigmas, [sigma_last]]).astype(np.float32)
+
+ self.sigmas = torch.from_numpy(sigmas)
+ self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.int64)
+
+ self.num_inference_steps = len(timesteps)
+
+ self.model_outputs = [
+ None,
+ ] * self.config.solver_order
+ self.lower_order_nums = 0
+ self.last_sample = None
+ if self.solver_p:
+ self.solver_p.set_timesteps(self.num_inference_steps, device=device)
+
+ # add an index counter for schedulers that allow duplicated timesteps
+ self._step_index = None
+ self._begin_index = None
+ self.sigmas = self.sigmas.to("cpu") # to avoid too much CPU/GPU communication
+
+ # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+ def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+ """
+ "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+ prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+ s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+ pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+ photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+ https://arxiv.org/abs/2205.11487
+ """
+ dtype = sample.dtype
+ batch_size, channels, *remaining_dims = sample.shape
+
+ if dtype not in (torch.float32, torch.float64):
+ sample = sample.float() # upcast for quantile calculation, and clamp not implemented for cpu half
+
+ # Flatten sample for doing quantile calculation along each image
+ sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+ abs_sample = sample.abs() # "a certain percentile absolute pixel value"
+
+ s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+ s = torch.clamp(
+ s, min=1, max=self.config.sample_max_value
+ ) # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+ s = s.unsqueeze(1) # (batch_size, 1) because clamp will broadcast along dim=0
+ sample = torch.clamp(sample, -s, s) / s # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+ sample = sample.reshape(batch_size, channels, *remaining_dims)
+ sample = sample.to(dtype)
+
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._sigma_to_t
+ def _sigma_to_t(self, sigma, log_sigmas):
+ # get log sigma
+ log_sigma = np.log(np.maximum(sigma, 1e-10))
+
+ # get distribution
+ dists = log_sigma - log_sigmas[:, np.newaxis]
+
+ # get sigmas range
+ low_idx = np.cumsum((dists >= 0), axis=0).argmax(axis=0).clip(max=log_sigmas.shape[0] - 2)
+ high_idx = low_idx + 1
+
+ low = log_sigmas[low_idx]
+ high = log_sigmas[high_idx]
+
+ # interpolate sigmas
+ w = (low - log_sigma) / (low - high)
+ w = np.clip(w, 0, 1)
+
+ # transform interpolation to time range
+ t = (1 - w) * low_idx + w * high_idx
+ t = t.reshape(sigma.shape)
+ return t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._sigma_to_alpha_sigma_t
+ def _sigma_to_alpha_sigma_t(self, sigma):
+ alpha_t = 1 / ((sigma**2 + 1) ** 0.5)
+ sigma_t = sigma * alpha_t
+
+ return alpha_t, sigma_t
+
+ # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._convert_to_karras
+ def _convert_to_karras(self, in_sigmas: torch.FloatTensor, num_inference_steps) -> torch.FloatTensor:
+ """Constructs the noise schedule of Karras et al. (2022)."""
+
+ # Hack to make sure that other schedulers which copy this function don't break
+ # TODO: Add this logic to the other schedulers
+ if hasattr(self.config, "sigma_min"):
+ sigma_min = self.config.sigma_min
+ else:
+ sigma_min = None
+
+ if hasattr(self.config, "sigma_max"):
+ sigma_max = self.config.sigma_max
+ else:
+ sigma_max = None
+
+ sigma_min = sigma_min if sigma_min is not None else in_sigmas[-1].item()
+ sigma_max = sigma_max if sigma_max is not None else in_sigmas[0].item()
+
+ rho = 7.0 # 7.0 is the value used in the paper
+ ramp = np.linspace(0, 1, num_inference_steps)
+ min_inv_rho = sigma_min ** (1 / rho)
+ max_inv_rho = sigma_max ** (1 / rho)
+ sigmas = (max_inv_rho + ramp * (min_inv_rho - max_inv_rho)) ** rho
+ return sigmas
+
+ def convert_model_output(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ r"""
+ Convert the model output to the corresponding type the UniPC algorithm needs.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+
+ Returns:
+ `torch.FloatTensor`:
+ The converted model output.
+ """
+ timestep = args[0] if len(args) > 0 else kwargs.pop("timestep", None)
+ if sample is None:
+ if len(args) > 1:
+ sample = args[1]
+ else:
+ raise ValueError("missing `sample` as a required keyward argument")
+ if timestep is not None:
+ deprecate(
+ "timesteps",
+ "1.0.0",
+ "Passing `timesteps` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ sigma = self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+
+ if self.predict_x0:
+ if self.config.prediction_type == "epsilon":
+ x0_pred = (sample - sigma_t * model_output) / alpha_t
+ elif self.config.prediction_type == "sample":
+ x0_pred = model_output
+ elif self.config.prediction_type == "v_prediction":
+ x0_pred = alpha_t * sample - sigma_t * model_output
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the UniPCMultistepScheduler."
+ )
+
+ if self.config.thresholding:
+ x0_pred = self._threshold_sample(x0_pred)
+
+ return x0_pred
+ else:
+ if self.config.prediction_type == "epsilon":
+ return model_output
+ elif self.config.prediction_type == "sample":
+ epsilon = (sample - alpha_t * model_output) / sigma_t
+ return epsilon
+ elif self.config.prediction_type == "v_prediction":
+ epsilon = alpha_t * model_output + sigma_t * sample
+ return epsilon
+ else:
+ raise ValueError(
+ f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample`, or"
+ " `v_prediction` for the UniPCMultistepScheduler."
+ )
+
+ def multistep_uni_p_bh_update(
+ self,
+ model_output: torch.FloatTensor,
+ *args,
+ sample: torch.FloatTensor = None,
+ order: int = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the UniP (B(h) version). Alternatively, `self.solver_p` is used if is specified.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from the learned diffusion model at the current timestep.
+ prev_timestep (`int`):
+ The previous discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ order (`int`):
+ The order of UniP at this timestep (corresponds to the *p* in UniPC-p).
+
+ Returns:
+ `torch.FloatTensor`:
+ The sample tensor at the previous timestep.
+ """
+ prev_timestep = args[0] if len(args) > 0 else kwargs.pop("prev_timestep", None)
+ if sample is None:
+ if len(args) > 1:
+ sample = args[1]
+ else:
+ raise ValueError(" missing `sample` as a required keyward argument")
+ if order is None:
+ if len(args) > 2:
+ order = args[2]
+ else:
+ raise ValueError(" missing `order` as a required keyward argument")
+ if prev_timestep is not None:
+ deprecate(
+ "prev_timestep",
+ "1.0.0",
+ "Passing `prev_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+ model_output_list = self.model_outputs
+
+ s0 = self.timestep_list[-1]
+ m0 = model_output_list[-1]
+ x = sample
+
+ if self.solver_p:
+ x_t = self.solver_p.step(model_output, s0, x).prev_sample
+ return x_t
+
+ sigma_t, sigma_s0 = self.sigmas[self.step_index + 1], self.sigmas[self.step_index]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+
+ h = lambda_t - lambda_s0
+ device = sample.device
+
+ rks = []
+ D1s = []
+ for i in range(1, order):
+ si = self.step_index - i
+ mi = model_output_list[-(i + 1)]
+ alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si])
+ lambda_si = torch.log(alpha_si) - torch.log(sigma_si)
+ rk = (lambda_si - lambda_s0) / h
+ rks.append(rk)
+ D1s.append((mi - m0) / rk)
+
+ rks.append(1.0)
+ rks = torch.tensor(rks, device=device)
+
+ R = []
+ b = []
+
+ hh = -h if self.predict_x0 else h
+ h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1
+ h_phi_k = h_phi_1 / hh - 1
+
+ factorial_i = 1
+
+ if self.config.solver_type == "bh1":
+ B_h = hh
+ elif self.config.solver_type == "bh2":
+ B_h = torch.expm1(hh)
+ else:
+ raise NotImplementedError()
+
+ for i in range(1, order + 1):
+ R.append(torch.pow(rks, i - 1))
+ b.append(h_phi_k * factorial_i / B_h)
+ factorial_i *= i + 1
+ h_phi_k = h_phi_k / hh - 1 / factorial_i
+
+ R = torch.stack(R)
+ b = torch.tensor(b, device=device)
+
+ if len(D1s) > 0:
+ D1s = torch.stack(D1s, dim=1) # (B, K)
+ # for order 2, we use a simplified version
+ if order == 2:
+ rhos_p = torch.tensor([0.5], dtype=x.dtype, device=device)
+ else:
+ rhos_p = torch.linalg.solve(R[:-1, :-1], b[:-1])
+ else:
+ D1s = None
+
+ if self.predict_x0:
+ x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0
+ if D1s is not None:
+ pred_res = torch.einsum("k,bkc...->bc...", rhos_p, D1s)
+ else:
+ pred_res = 0
+ x_t = x_t_ - alpha_t * B_h * pred_res
+ else:
+ x_t_ = alpha_t / alpha_s0 * x - sigma_t * h_phi_1 * m0
+ if D1s is not None:
+ pred_res = torch.einsum("k,bkc...->bc...", rhos_p, D1s)
+ else:
+ pred_res = 0
+ x_t = x_t_ - sigma_t * B_h * pred_res
+
+ x_t = x_t.to(x.dtype)
+ return x_t
+
+ def multistep_uni_c_bh_update(
+ self,
+ this_model_output: torch.FloatTensor,
+ *args,
+ last_sample: torch.FloatTensor = None,
+ this_sample: torch.FloatTensor = None,
+ order: int = None,
+ **kwargs,
+ ) -> torch.FloatTensor:
+ """
+ One step for the UniC (B(h) version).
+
+ Args:
+ this_model_output (`torch.FloatTensor`):
+ The model outputs at `x_t`.
+ this_timestep (`int`):
+ The current timestep `t`.
+ last_sample (`torch.FloatTensor`):
+ The generated sample before the last predictor `x_{t-1}`.
+ this_sample (`torch.FloatTensor`):
+ The generated sample after the last predictor `x_{t}`.
+ order (`int`):
+ The `p` of UniC-p at this step. The effective order of accuracy should be `order + 1`.
+
+ Returns:
+ `torch.FloatTensor`:
+ The corrected sample tensor at the current timestep.
+ """
+ this_timestep = args[0] if len(args) > 0 else kwargs.pop("this_timestep", None)
+ if last_sample is None:
+ if len(args) > 1:
+ last_sample = args[1]
+ else:
+ raise ValueError(" missing`last_sample` as a required keyward argument")
+ if this_sample is None:
+ if len(args) > 2:
+ this_sample = args[2]
+ else:
+ raise ValueError(" missing`this_sample` as a required keyward argument")
+ if order is None:
+ if len(args) > 3:
+ order = args[3]
+ else:
+ raise ValueError(" missing`order` as a required keyward argument")
+ if this_timestep is not None:
+ deprecate(
+ "this_timestep",
+ "1.0.0",
+ "Passing `this_timestep` is deprecated and has no effect as model output conversion is now handled via an internal counter `self.step_index`",
+ )
+
+ model_output_list = self.model_outputs
+
+ m0 = model_output_list[-1]
+ x = last_sample
+ x_t = this_sample
+ model_t = this_model_output
+
+ sigma_t, sigma_s0 = self.sigmas[self.step_index], self.sigmas[self.step_index - 1]
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma_t)
+ alpha_s0, sigma_s0 = self._sigma_to_alpha_sigma_t(sigma_s0)
+
+ lambda_t = torch.log(alpha_t) - torch.log(sigma_t)
+ lambda_s0 = torch.log(alpha_s0) - torch.log(sigma_s0)
+
+ h = lambda_t - lambda_s0
+ device = this_sample.device
+
+ rks = []
+ D1s = []
+ for i in range(1, order):
+ si = self.step_index - (i + 1)
+ mi = model_output_list[-(i + 1)]
+ alpha_si, sigma_si = self._sigma_to_alpha_sigma_t(self.sigmas[si])
+ lambda_si = torch.log(alpha_si) - torch.log(sigma_si)
+ rk = (lambda_si - lambda_s0) / h
+ rks.append(rk)
+ D1s.append((mi - m0) / rk)
+
+ rks.append(1.0)
+ rks = torch.tensor(rks, device=device)
+
+ R = []
+ b = []
+
+ hh = -h if self.predict_x0 else h
+ h_phi_1 = torch.expm1(hh) # h\phi_1(h) = e^h - 1
+ h_phi_k = h_phi_1 / hh - 1
+
+ factorial_i = 1
+
+ if self.config.solver_type == "bh1":
+ B_h = hh
+ elif self.config.solver_type == "bh2":
+ B_h = torch.expm1(hh)
+ else:
+ raise NotImplementedError()
+
+ for i in range(1, order + 1):
+ R.append(torch.pow(rks, i - 1))
+ b.append(h_phi_k * factorial_i / B_h)
+ factorial_i *= i + 1
+ h_phi_k = h_phi_k / hh - 1 / factorial_i
+
+ R = torch.stack(R)
+ b = torch.tensor(b, device=device)
+
+ if len(D1s) > 0:
+ D1s = torch.stack(D1s, dim=1)
+ else:
+ D1s = None
+
+ # for order 1, we use a simplified version
+ if order == 1:
+ rhos_c = torch.tensor([0.5], dtype=x.dtype, device=device)
+ else:
+ rhos_c = torch.linalg.solve(R, b)
+
+ if self.predict_x0:
+ x_t_ = sigma_t / sigma_s0 * x - alpha_t * h_phi_1 * m0
+ if D1s is not None:
+ corr_res = torch.einsum("k,bkc...->bc...", rhos_c[:-1], D1s)
+ else:
+ corr_res = 0
+ D1_t = model_t - m0
+ x_t = x_t_ - alpha_t * B_h * (corr_res + rhos_c[-1] * D1_t)
+ else:
+ x_t_ = alpha_t / alpha_s0 * x - sigma_t * h_phi_1 * m0
+ if D1s is not None:
+ corr_res = torch.einsum("k,bkc...->bc...", rhos_c[:-1], D1s)
+ else:
+ corr_res = 0
+ D1_t = model_t - m0
+ x_t = x_t_ - sigma_t * B_h * (corr_res + rhos_c[-1] * D1_t)
+ x_t = x_t.to(x.dtype)
+ return x_t
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.index_for_timestep
+ def index_for_timestep(self, timestep, schedule_timesteps=None):
+ if schedule_timesteps is None:
+ schedule_timesteps = self.timesteps
+
+ index_candidates = (schedule_timesteps == timestep).nonzero()
+
+ if len(index_candidates) == 0:
+ step_index = len(self.timesteps) - 1
+ # The sigma index that is taken for the **very** first `step`
+ # is always the second index (or the last index if there is only 1)
+ # This way we can ensure we don't accidentally skip a sigma in
+ # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+ elif len(index_candidates) > 1:
+ step_index = index_candidates[1].item()
+ else:
+ step_index = index_candidates[0].item()
+
+ return step_index
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler._init_step_index
+ def _init_step_index(self, timestep):
+ """
+ Initialize the step_index counter for the scheduler.
+ """
+
+ if self.begin_index is None:
+ if isinstance(timestep, torch.Tensor):
+ timestep = timestep.to(self.timesteps.device)
+ self._step_index = self.index_for_timestep(timestep)
+ else:
+ self._step_index = self._begin_index
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: int,
+ sample: torch.FloatTensor,
+ return_dict: bool = True,
+ ) -> Union[SchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with
+ the multistep UniPC.
+
+ Args:
+ model_output (`torch.FloatTensor`):
+ The direct output from learned diffusion model.
+ timestep (`int`):
+ The current discrete timestep in the diffusion chain.
+ sample (`torch.FloatTensor`):
+ A current instance of a sample created by the diffusion process.
+ return_dict (`bool`):
+ Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a
+ tuple is returned where the first element is the sample tensor.
+
+ """
+ if self.num_inference_steps is None:
+ raise ValueError(
+ "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler"
+ )
+
+ if self.step_index is None:
+ self._init_step_index(timestep)
+
+ use_corrector = (
+ self.step_index > 0 and self.step_index - 1 not in self.disable_corrector and self.last_sample is not None
+ )
+
+ model_output_convert = self.convert_model_output(model_output, sample=sample)
+ if use_corrector:
+ sample = self.multistep_uni_c_bh_update(
+ this_model_output=model_output_convert,
+ last_sample=self.last_sample,
+ this_sample=sample,
+ order=self.this_order,
+ )
+
+ for i in range(self.config.solver_order - 1):
+ self.model_outputs[i] = self.model_outputs[i + 1]
+ self.timestep_list[i] = self.timestep_list[i + 1]
+
+ self.model_outputs[-1] = model_output_convert
+ self.timestep_list[-1] = timestep
+
+ if self.config.lower_order_final:
+ this_order = min(self.config.solver_order, len(self.timesteps) - self.step_index)
+ else:
+ this_order = self.config.solver_order
+
+ self.this_order = min(this_order, self.lower_order_nums + 1) # warmup for multistep
+ assert self.this_order > 0
+
+ self.last_sample = sample
+ prev_sample = self.multistep_uni_p_bh_update(
+ model_output=model_output, # pass the original non-converted model output, in case solver-p is used
+ sample=sample,
+ order=self.this_order,
+ )
+
+ if self.lower_order_nums < self.config.solver_order:
+ self.lower_order_nums += 1
+
+ # upon completion increase step index by one
+ self._step_index += 1
+
+ if not return_dict:
+ return (prev_sample,)
+
+ return SchedulerOutput(prev_sample=prev_sample)
+
+ def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor:
+ """
+ Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+ current timestep.
+
+ Args:
+ sample (`torch.FloatTensor`):
+ The input sample.
+
+ Returns:
+ `torch.FloatTensor`:
+ A scaled input sample.
+ """
+ return sample
+
+ # Copied from diffusers.schedulers.scheduling_dpmsolver_multistep.DPMSolverMultistepScheduler.add_noise
+ def add_noise(
+ self,
+ original_samples: torch.FloatTensor,
+ noise: torch.FloatTensor,
+ timesteps: torch.IntTensor,
+ ) -> torch.FloatTensor:
+ # Make sure sigmas and timesteps have the same device and dtype as original_samples
+ sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
+ if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+ # mps does not support float64
+ schedule_timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+ timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+ else:
+ schedule_timesteps = self.timesteps.to(original_samples.device)
+ timesteps = timesteps.to(original_samples.device)
+
+ # begin_index is None when the scheduler is used for training
+ if self.begin_index is None:
+ step_indices = [self.index_for_timestep(t, schedule_timesteps) for t in timesteps]
+ else:
+ step_indices = [self.begin_index] * timesteps.shape[0]
+
+ sigma = sigmas[step_indices].flatten()
+ while len(sigma.shape) < len(original_samples.shape):
+ sigma = sigma.unsqueeze(-1)
+
+ alpha_t, sigma_t = self._sigma_to_alpha_sigma_t(sigma)
+ noisy_samples = alpha_t * original_samples + sigma_t * noise
+ return noisy_samples
+
+ def __len__(self):
+ return self.config.num_train_timesteps
diff --git a/model/diffusers_c/schedulers/scheduling_utils.py b/model/diffusers_c/schedulers/scheduling_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..5dbdb82884bca94824a01b07851d7ea4de810ddc
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_utils.py
@@ -0,0 +1,186 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib
+import os
+from dataclasses import dataclass
+from enum import Enum
+from typing import Optional, Union
+
+import torch
+from huggingface_hub.utils import validate_hf_hub_args
+
+from ..utils import BaseOutput, PushToHubMixin
+
+
+SCHEDULER_CONFIG_NAME = "scheduler_config.json"
+
+
+# NOTE: We make this type an enum because it simplifies usage in docs and prevents
+# circular imports when used for `_compatibles` within the schedulers module.
+# When it's used as a type in pipelines, it really is a Union because the actual
+# scheduler instance is passed in.
+class KarrasDiffusionSchedulers(Enum):
+ DDIMScheduler = 1
+ DDPMScheduler = 2
+ PNDMScheduler = 3
+ LMSDiscreteScheduler = 4
+ EulerDiscreteScheduler = 5
+ HeunDiscreteScheduler = 6
+ EulerAncestralDiscreteScheduler = 7
+ DPMSolverMultistepScheduler = 8
+ DPMSolverSinglestepScheduler = 9
+ KDPM2DiscreteScheduler = 10
+ KDPM2AncestralDiscreteScheduler = 11
+ DEISMultistepScheduler = 12
+ UniPCMultistepScheduler = 13
+ DPMSolverSDEScheduler = 14
+ EDMEulerScheduler = 15
+
+
+@dataclass
+class SchedulerOutput(BaseOutput):
+ """
+ Base class for the output of a scheduler's `step` function.
+
+ Args:
+ prev_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample `(x_{t-1})` of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ """
+
+ prev_sample: torch.FloatTensor
+
+
+class SchedulerMixin(PushToHubMixin):
+ """
+ Base class for all schedulers.
+
+ [`SchedulerMixin`] contains common functions shared by all schedulers such as general loading and saving
+ functionalities.
+
+ [`ConfigMixin`] takes care of storing the configuration attributes (like `num_train_timesteps`) that are passed to
+ the scheduler's `__init__` function, and the attributes can be accessed by `scheduler.config.num_train_timesteps`.
+
+ Class attributes:
+ - **_compatibles** (`List[str]`) -- A list of scheduler classes that are compatible with the parent scheduler
+ class. Use [`~ConfigMixin.from_config`] to load a different compatible scheduler class (should be overridden
+ by parent class).
+ """
+
+ config_name = SCHEDULER_CONFIG_NAME
+ _compatibles = []
+ has_compatibles = True
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(
+ cls,
+ pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None,
+ subfolder: Optional[str] = None,
+ return_unused_kwargs=False,
+ **kwargs,
+ ):
+ r"""
+ Instantiate a scheduler from a pre-defined JSON configuration file in a local directory or Hub repository.
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+ the Hub.
+ - A path to a *directory* (for example `./my_model_directory`) containing the scheduler
+ configuration saved with [`~SchedulerMixin.save_pretrained`].
+ subfolder (`str`, *optional*):
+ The subfolder location of a model file within a larger model repository on the Hub or locally.
+ return_unused_kwargs (`bool`, *optional*, defaults to `False`):
+ Whether kwargs that are not consumed by the Python class should be returned or not.
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+ is not used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+ incompletely downloaded files are deleted.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only(`bool`, *optional*, defaults to `False`):
+ Whether to only load local model weights and configuration files or not. If set to `True`, the model
+ won't be downloaded from the Hub.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+ `diffusers-cli login` (stored in `~/.huggingface`) is used.
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+ allowed by Git.
+
+
+
+ To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
+ `huggingface-cli login`. You can also activate the special
+ ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a
+ firewalled environment.
+
+
+
+ """
+ config, kwargs, commit_hash = cls.load_config(
+ pretrained_model_name_or_path=pretrained_model_name_or_path,
+ subfolder=subfolder,
+ return_unused_kwargs=True,
+ return_commit_hash=True,
+ **kwargs,
+ )
+ return cls.from_config(config, return_unused_kwargs=return_unused_kwargs, **kwargs)
+
+ def save_pretrained(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs):
+ """
+ Save a scheduler configuration object to a directory so that it can be reloaded using the
+ [`~SchedulerMixin.from_pretrained`] class method.
+
+ Args:
+ save_directory (`str` or `os.PathLike`):
+ Directory where the configuration JSON file will be saved (will be created if it does not exist).
+ push_to_hub (`bool`, *optional*, defaults to `False`):
+ Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the
+ repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+ namespace).
+ kwargs (`Dict[str, Any]`, *optional*):
+ Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+ """
+ self.save_config(save_directory=save_directory, push_to_hub=push_to_hub, **kwargs)
+
+ @property
+ def compatibles(self):
+ """
+ Returns all schedulers that are compatible with this scheduler
+
+ Returns:
+ `List[SchedulerMixin]`: List of compatible schedulers
+ """
+ return self._get_compatibles()
+
+ @classmethod
+ def _get_compatibles(cls):
+ compatible_classes_str = list(set([cls.__name__] + cls._compatibles))
+ diffusers_library = importlib.import_module(__name__.split(".")[0])
+ compatible_classes = [
+ getattr(diffusers_library, c) for c in compatible_classes_str if hasattr(diffusers_library, c)
+ ]
+ return compatible_classes
diff --git a/model/diffusers_c/schedulers/scheduling_utils_flax.py b/model/diffusers_c/schedulers/scheduling_utils_flax.py
new file mode 100644
index 0000000000000000000000000000000000000000..a1d471f910e5576ccadbe426636a9ba0418da396
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_utils_flax.py
@@ -0,0 +1,293 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib
+import math
+import os
+from dataclasses import dataclass
+from enum import Enum
+from typing import Optional, Tuple, Union
+
+import flax
+import jax.numpy as jnp
+from huggingface_hub.utils import validate_hf_hub_args
+
+from ..utils import BaseOutput, PushToHubMixin
+
+
+SCHEDULER_CONFIG_NAME = "scheduler_config.json"
+
+
+# NOTE: We make this type an enum because it simplifies usage in docs and prevents
+# circular imports when used for `_compatibles` within the schedulers module.
+# When it's used as a type in pipelines, it really is a Union because the actual
+# scheduler instance is passed in.
+class FlaxKarrasDiffusionSchedulers(Enum):
+ FlaxDDIMScheduler = 1
+ FlaxDDPMScheduler = 2
+ FlaxPNDMScheduler = 3
+ FlaxLMSDiscreteScheduler = 4
+ FlaxDPMSolverMultistepScheduler = 5
+ FlaxEulerDiscreteScheduler = 6
+
+
+@dataclass
+class FlaxSchedulerOutput(BaseOutput):
+ """
+ Base class for the scheduler's step function output.
+
+ Args:
+ prev_sample (`jnp.ndarray` of shape `(batch_size, num_channels, height, width)` for images):
+ Computed sample (x_{t-1}) of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ """
+
+ prev_sample: jnp.ndarray
+
+
+class FlaxSchedulerMixin(PushToHubMixin):
+ """
+ Mixin containing common functions for the schedulers.
+
+ Class attributes:
+ - **_compatibles** (`List[str]`) -- A list of classes that are compatible with the parent class, so that
+ `from_config` can be used from a class different than the one used to save the config (should be overridden
+ by parent class).
+ """
+
+ config_name = SCHEDULER_CONFIG_NAME
+ ignore_for_config = ["dtype"]
+ _compatibles = []
+ has_compatibles = True
+
+ @classmethod
+ @validate_hf_hub_args
+ def from_pretrained(
+ cls,
+ pretrained_model_name_or_path: Optional[Union[str, os.PathLike]] = None,
+ subfolder: Optional[str] = None,
+ return_unused_kwargs=False,
+ **kwargs,
+ ):
+ r"""
+ Instantiate a Scheduler class from a pre-defined JSON-file.
+
+ Parameters:
+ pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+ Can be either:
+
+ - A string, the *model id* of a model repo on huggingface.co. Valid model ids should have an
+ organization name, like `google/ddpm-celebahq-256`.
+ - A path to a *directory* containing model weights saved using [`~SchedulerMixin.save_pretrained`],
+ e.g., `./my_model_directory/`.
+ subfolder (`str`, *optional*):
+ In case the relevant files are located inside a subfolder of the model repo (either remote in
+ huggingface.co or downloaded locally), you can specify the folder name here.
+ return_unused_kwargs (`bool`, *optional*, defaults to `False`):
+ Whether kwargs that are not consumed by the Python class should be returned or not.
+
+ cache_dir (`Union[str, os.PathLike]`, *optional*):
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the
+ standard cache should not be used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+ cached versions if they exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to delete incompletely received files. Will attempt to resume the download if such a
+ file exists.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+ output_loading_info(`bool`, *optional*, defaults to `False`):
+ Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+ local_files_only(`bool`, *optional*, defaults to `False`):
+ Whether or not to only look at local files (i.e., do not try to download the model).
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+ when running `transformers-cli login` (stored in `~/.huggingface`).
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+ git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+ identifier allowed by git.
+
+
+
+ It is required to be logged in (`huggingface-cli login`) when you want to use private or [gated
+ models](https://huggingface.co/docs/hub/models-gated#gated-models).
+
+
+
+
+
+ Activate the special ["offline-mode"](https://huggingface.co/transformers/installation.html#offline-mode) to
+ use this method in a firewalled environment.
+
+
+
+ """
+ config, kwargs = cls.load_config(
+ pretrained_model_name_or_path=pretrained_model_name_or_path,
+ subfolder=subfolder,
+ return_unused_kwargs=True,
+ **kwargs,
+ )
+ scheduler, unused_kwargs = cls.from_config(config, return_unused_kwargs=True, **kwargs)
+
+ if hasattr(scheduler, "create_state") and getattr(scheduler, "has_state", False):
+ state = scheduler.create_state()
+
+ if return_unused_kwargs:
+ return scheduler, state, unused_kwargs
+
+ return scheduler, state
+
+ def save_pretrained(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs):
+ """
+ Save a scheduler configuration object to the directory `save_directory`, so that it can be re-loaded using the
+ [`~FlaxSchedulerMixin.from_pretrained`] class method.
+
+ Args:
+ save_directory (`str` or `os.PathLike`):
+ Directory where the configuration JSON file will be saved (will be created if it does not exist).
+ push_to_hub (`bool`, *optional*, defaults to `False`):
+ Whether or not to push your model to the Hugging Face Hub after saving it. You can specify the
+ repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
+ namespace).
+ kwargs (`Dict[str, Any]`, *optional*):
+ Additional keyword arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
+ """
+ self.save_config(save_directory=save_directory, push_to_hub=push_to_hub, **kwargs)
+
+ @property
+ def compatibles(self):
+ """
+ Returns all schedulers that are compatible with this scheduler
+
+ Returns:
+ `List[SchedulerMixin]`: List of compatible schedulers
+ """
+ return self._get_compatibles()
+
+ @classmethod
+ def _get_compatibles(cls):
+ compatible_classes_str = list(set([cls.__name__] + cls._compatibles))
+ diffusers_library = importlib.import_module(__name__.split(".")[0])
+ compatible_classes = [
+ getattr(diffusers_library, c) for c in compatible_classes_str if hasattr(diffusers_library, c)
+ ]
+ return compatible_classes
+
+
+def broadcast_to_shape_from_left(x: jnp.ndarray, shape: Tuple[int]) -> jnp.ndarray:
+ assert len(shape) >= x.ndim
+ return jnp.broadcast_to(x.reshape(x.shape + (1,) * (len(shape) - x.ndim)), shape)
+
+
+def betas_for_alpha_bar(num_diffusion_timesteps: int, max_beta=0.999, dtype=jnp.float32) -> jnp.ndarray:
+ """
+ Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+ (1-beta) over time from t = [0,1].
+
+ Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+ to that part of the diffusion process.
+
+
+ Args:
+ num_diffusion_timesteps (`int`): the number of betas to produce.
+ max_beta (`float`): the maximum beta to use; use values lower than 1 to
+ prevent singularities.
+
+ Returns:
+ betas (`jnp.ndarray`): the betas used by the scheduler to step the model outputs
+ """
+
+ def alpha_bar(time_step):
+ return math.cos((time_step + 0.008) / 1.008 * math.pi / 2) ** 2
+
+ betas = []
+ for i in range(num_diffusion_timesteps):
+ t1 = i / num_diffusion_timesteps
+ t2 = (i + 1) / num_diffusion_timesteps
+ betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta))
+ return jnp.array(betas, dtype=dtype)
+
+
+@flax.struct.dataclass
+class CommonSchedulerState:
+ alphas: jnp.ndarray
+ betas: jnp.ndarray
+ alphas_cumprod: jnp.ndarray
+
+ @classmethod
+ def create(cls, scheduler):
+ config = scheduler.config
+
+ if config.trained_betas is not None:
+ betas = jnp.asarray(config.trained_betas, dtype=scheduler.dtype)
+ elif config.beta_schedule == "linear":
+ betas = jnp.linspace(config.beta_start, config.beta_end, config.num_train_timesteps, dtype=scheduler.dtype)
+ elif config.beta_schedule == "scaled_linear":
+ # this schedule is very specific to the latent diffusion model.
+ betas = (
+ jnp.linspace(
+ config.beta_start**0.5, config.beta_end**0.5, config.num_train_timesteps, dtype=scheduler.dtype
+ )
+ ** 2
+ )
+ elif config.beta_schedule == "squaredcos_cap_v2":
+ # Glide cosine schedule
+ betas = betas_for_alpha_bar(config.num_train_timesteps, dtype=scheduler.dtype)
+ else:
+ raise NotImplementedError(
+ f"beta_schedule {config.beta_schedule} is not implemented for scheduler {scheduler.__class__.__name__}"
+ )
+
+ alphas = 1.0 - betas
+
+ alphas_cumprod = jnp.cumprod(alphas, axis=0)
+
+ return cls(
+ alphas=alphas,
+ betas=betas,
+ alphas_cumprod=alphas_cumprod,
+ )
+
+
+def get_sqrt_alpha_prod(
+ state: CommonSchedulerState, original_samples: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray
+):
+ alphas_cumprod = state.alphas_cumprod
+
+ sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+ sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+ sqrt_alpha_prod = broadcast_to_shape_from_left(sqrt_alpha_prod, original_samples.shape)
+
+ sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+ sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+ sqrt_one_minus_alpha_prod = broadcast_to_shape_from_left(sqrt_one_minus_alpha_prod, original_samples.shape)
+
+ return sqrt_alpha_prod, sqrt_one_minus_alpha_prod
+
+
+def add_noise_common(
+ state: CommonSchedulerState, original_samples: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray
+):
+ sqrt_alpha_prod, sqrt_one_minus_alpha_prod = get_sqrt_alpha_prod(state, original_samples, noise, timesteps)
+ noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+ return noisy_samples
+
+
+def get_velocity_common(state: CommonSchedulerState, sample: jnp.ndarray, noise: jnp.ndarray, timesteps: jnp.ndarray):
+ sqrt_alpha_prod, sqrt_one_minus_alpha_prod = get_sqrt_alpha_prod(state, sample, noise, timesteps)
+ velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+ return velocity
diff --git a/model/diffusers_c/schedulers/scheduling_vq_diffusion.py b/model/diffusers_c/schedulers/scheduling_vq_diffusion.py
new file mode 100644
index 0000000000000000000000000000000000000000..03ba95cad68b7340e0cfef77a3c2c5b545959cc3
--- /dev/null
+++ b/model/diffusers_c/schedulers/scheduling_vq_diffusion.py
@@ -0,0 +1,467 @@
+# Copyright 2024 Microsoft and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput
+from .scheduling_utils import SchedulerMixin
+
+
+@dataclass
+class VQDiffusionSchedulerOutput(BaseOutput):
+ """
+ Output class for the scheduler's step function output.
+
+ Args:
+ prev_sample (`torch.LongTensor` of shape `(batch size, num latent pixels)`):
+ Computed sample x_{t-1} of previous timestep. `prev_sample` should be used as next model input in the
+ denoising loop.
+ """
+
+ prev_sample: torch.LongTensor
+
+
+def index_to_log_onehot(x: torch.LongTensor, num_classes: int) -> torch.FloatTensor:
+ """
+ Convert batch of vector of class indices into batch of log onehot vectors
+
+ Args:
+ x (`torch.LongTensor` of shape `(batch size, vector length)`):
+ Batch of class indices
+
+ num_classes (`int`):
+ number of classes to be used for the onehot vectors
+
+ Returns:
+ `torch.FloatTensor` of shape `(batch size, num classes, vector length)`:
+ Log onehot vectors
+ """
+ x_onehot = F.one_hot(x, num_classes)
+ x_onehot = x_onehot.permute(0, 2, 1)
+ log_x = torch.log(x_onehot.float().clamp(min=1e-30))
+ return log_x
+
+
+def gumbel_noised(logits: torch.FloatTensor, generator: Optional[torch.Generator]) -> torch.FloatTensor:
+ """
+ Apply gumbel noise to `logits`
+ """
+ uniform = torch.rand(logits.shape, device=logits.device, generator=generator)
+ gumbel_noise = -torch.log(-torch.log(uniform + 1e-30) + 1e-30)
+ noised = gumbel_noise + logits
+ return noised
+
+
+def alpha_schedules(num_diffusion_timesteps: int, alpha_cum_start=0.99999, alpha_cum_end=0.000009):
+ """
+ Cumulative and non-cumulative alpha schedules.
+
+ See section 4.1.
+ """
+ att = (
+ np.arange(0, num_diffusion_timesteps) / (num_diffusion_timesteps - 1) * (alpha_cum_end - alpha_cum_start)
+ + alpha_cum_start
+ )
+ att = np.concatenate(([1], att))
+ at = att[1:] / att[:-1]
+ att = np.concatenate((att[1:], [1]))
+ return at, att
+
+
+def gamma_schedules(num_diffusion_timesteps: int, gamma_cum_start=0.000009, gamma_cum_end=0.99999):
+ """
+ Cumulative and non-cumulative gamma schedules.
+
+ See section 4.1.
+ """
+ ctt = (
+ np.arange(0, num_diffusion_timesteps) / (num_diffusion_timesteps - 1) * (gamma_cum_end - gamma_cum_start)
+ + gamma_cum_start
+ )
+ ctt = np.concatenate(([0], ctt))
+ one_minus_ctt = 1 - ctt
+ one_minus_ct = one_minus_ctt[1:] / one_minus_ctt[:-1]
+ ct = 1 - one_minus_ct
+ ctt = np.concatenate((ctt[1:], [0]))
+ return ct, ctt
+
+
+class VQDiffusionScheduler(SchedulerMixin, ConfigMixin):
+ """
+ A scheduler for vector quantized diffusion.
+
+ This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic
+ methods the library implements for all schedulers such as loading and saving.
+
+ Args:
+ num_vec_classes (`int`):
+ The number of classes of the vector embeddings of the latent pixels. Includes the class for the masked
+ latent pixel.
+ num_train_timesteps (`int`, defaults to 100):
+ The number of diffusion steps to train the model.
+ alpha_cum_start (`float`, defaults to 0.99999):
+ The starting cumulative alpha value.
+ alpha_cum_end (`float`, defaults to 0.00009):
+ The ending cumulative alpha value.
+ gamma_cum_start (`float`, defaults to 0.00009):
+ The starting cumulative gamma value.
+ gamma_cum_end (`float`, defaults to 0.99999):
+ The ending cumulative gamma value.
+ """
+
+ order = 1
+
+ @register_to_config
+ def __init__(
+ self,
+ num_vec_classes: int,
+ num_train_timesteps: int = 100,
+ alpha_cum_start: float = 0.99999,
+ alpha_cum_end: float = 0.000009,
+ gamma_cum_start: float = 0.000009,
+ gamma_cum_end: float = 0.99999,
+ ):
+ self.num_embed = num_vec_classes
+
+ # By convention, the index for the mask class is the last class index
+ self.mask_class = self.num_embed - 1
+
+ at, att = alpha_schedules(num_train_timesteps, alpha_cum_start=alpha_cum_start, alpha_cum_end=alpha_cum_end)
+ ct, ctt = gamma_schedules(num_train_timesteps, gamma_cum_start=gamma_cum_start, gamma_cum_end=gamma_cum_end)
+
+ num_non_mask_classes = self.num_embed - 1
+ bt = (1 - at - ct) / num_non_mask_classes
+ btt = (1 - att - ctt) / num_non_mask_classes
+
+ at = torch.tensor(at.astype("float64"))
+ bt = torch.tensor(bt.astype("float64"))
+ ct = torch.tensor(ct.astype("float64"))
+ log_at = torch.log(at)
+ log_bt = torch.log(bt)
+ log_ct = torch.log(ct)
+
+ att = torch.tensor(att.astype("float64"))
+ btt = torch.tensor(btt.astype("float64"))
+ ctt = torch.tensor(ctt.astype("float64"))
+ log_cumprod_at = torch.log(att)
+ log_cumprod_bt = torch.log(btt)
+ log_cumprod_ct = torch.log(ctt)
+
+ self.log_at = log_at.float()
+ self.log_bt = log_bt.float()
+ self.log_ct = log_ct.float()
+ self.log_cumprod_at = log_cumprod_at.float()
+ self.log_cumprod_bt = log_cumprod_bt.float()
+ self.log_cumprod_ct = log_cumprod_ct.float()
+
+ # setable values
+ self.num_inference_steps = None
+ self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy())
+
+ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None):
+ """
+ Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+ Args:
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps and diffusion process parameters (alpha, beta, gamma) should be moved
+ to.
+ """
+ self.num_inference_steps = num_inference_steps
+ timesteps = np.arange(0, self.num_inference_steps)[::-1].copy()
+ self.timesteps = torch.from_numpy(timesteps).to(device)
+
+ self.log_at = self.log_at.to(device)
+ self.log_bt = self.log_bt.to(device)
+ self.log_ct = self.log_ct.to(device)
+ self.log_cumprod_at = self.log_cumprod_at.to(device)
+ self.log_cumprod_bt = self.log_cumprod_bt.to(device)
+ self.log_cumprod_ct = self.log_cumprod_ct.to(device)
+
+ def step(
+ self,
+ model_output: torch.FloatTensor,
+ timestep: torch.long,
+ sample: torch.LongTensor,
+ generator: Optional[torch.Generator] = None,
+ return_dict: bool = True,
+ ) -> Union[VQDiffusionSchedulerOutput, Tuple]:
+ """
+ Predict the sample from the previous timestep by the reverse transition distribution. See
+ [`~VQDiffusionScheduler.q_posterior`] for more details about how the distribution is computer.
+
+ Args:
+ log_p_x_0: (`torch.FloatTensor` of shape `(batch size, num classes - 1, num latent pixels)`):
+ The log probabilities for the predicted classes of the initial latent pixels. Does not include a
+ prediction for the masked class as the initial unnoised image cannot be masked.
+ t (`torch.long`):
+ The timestep that determines which transition matrices are used.
+ x_t (`torch.LongTensor` of shape `(batch size, num latent pixels)`):
+ The classes of each latent pixel at time `t`.
+ generator (`torch.Generator`, or `None`):
+ A random number generator for the noise applied to `p(x_{t-1} | x_t)` before it is sampled from.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] or
+ `tuple`.
+
+ Returns:
+ [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] or `tuple`:
+ If return_dict is `True`, [`~schedulers.scheduling_vq_diffusion.VQDiffusionSchedulerOutput`] is
+ returned, otherwise a tuple is returned where the first element is the sample tensor.
+ """
+ if timestep == 0:
+ log_p_x_t_min_1 = model_output
+ else:
+ log_p_x_t_min_1 = self.q_posterior(model_output, sample, timestep)
+
+ log_p_x_t_min_1 = gumbel_noised(log_p_x_t_min_1, generator)
+
+ x_t_min_1 = log_p_x_t_min_1.argmax(dim=1)
+
+ if not return_dict:
+ return (x_t_min_1,)
+
+ return VQDiffusionSchedulerOutput(prev_sample=x_t_min_1)
+
+ def q_posterior(self, log_p_x_0, x_t, t):
+ """
+ Calculates the log probabilities for the predicted classes of the image at timestep `t-1`:
+
+ ```
+ p(x_{t-1} | x_t) = sum( q(x_t | x_{t-1}) * q(x_{t-1} | x_0) * p(x_0) / q(x_t | x_0) )
+ ```
+
+ Args:
+ log_p_x_0 (`torch.FloatTensor` of shape `(batch size, num classes - 1, num latent pixels)`):
+ The log probabilities for the predicted classes of the initial latent pixels. Does not include a
+ prediction for the masked class as the initial unnoised image cannot be masked.
+ x_t (`torch.LongTensor` of shape `(batch size, num latent pixels)`):
+ The classes of each latent pixel at time `t`.
+ t (`torch.Long`):
+ The timestep that determines which transition matrix is used.
+
+ Returns:
+ `torch.FloatTensor` of shape `(batch size, num classes, num latent pixels)`:
+ The log probabilities for the predicted classes of the image at timestep `t-1`.
+ """
+ log_onehot_x_t = index_to_log_onehot(x_t, self.num_embed)
+
+ log_q_x_t_given_x_0 = self.log_Q_t_transitioning_to_known_class(
+ t=t, x_t=x_t, log_onehot_x_t=log_onehot_x_t, cumulative=True
+ )
+
+ log_q_t_given_x_t_min_1 = self.log_Q_t_transitioning_to_known_class(
+ t=t, x_t=x_t, log_onehot_x_t=log_onehot_x_t, cumulative=False
+ )
+
+ # p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) ... p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0)
+ # . . .
+ # . . .
+ # . . .
+ # p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) ... p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1})
+ q = log_p_x_0 - log_q_x_t_given_x_0
+
+ # sum_0 = p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) + ... + p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}), ... ,
+ # sum_n = p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) + ... + p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1})
+ q_log_sum_exp = torch.logsumexp(q, dim=1, keepdim=True)
+
+ # p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0 ... p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n
+ # . . .
+ # . . .
+ # . . .
+ # p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0 ... p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n
+ q = q - q_log_sum_exp
+
+ # (p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1} ... (p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1}
+ # . . .
+ # . . .
+ # . . .
+ # (p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1} ... (p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1}
+ # c_cumulative_{t-1} ... c_cumulative_{t-1}
+ q = self.apply_cumulative_transitions(q, t - 1)
+
+ # ((p_0(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_0) * sum_0 ... ((p_n(x_0=C_0 | x_t) / q(x_t | x_0=C_0) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_0) * sum_n
+ # . . .
+ # . . .
+ # . . .
+ # ((p_0(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_0) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_{k-1}) * sum_0 ... ((p_n(x_0=C_{k-1} | x_t) / q(x_t | x_0=C_{k-1}) / sum_n) * a_cumulative_{t-1} + b_cumulative_{t-1}) * q(x_t | x_{t-1}=C_{k-1}) * sum_n
+ # c_cumulative_{t-1} * q(x_t | x_{t-1}=C_k) * sum_0 ... c_cumulative_{t-1} * q(x_t | x_{t-1}=C_k) * sum_0
+ log_p_x_t_min_1 = q + log_q_t_given_x_t_min_1 + q_log_sum_exp
+
+ # For each column, there are two possible cases.
+ #
+ # Where:
+ # - sum(p_n(x_0))) is summing over all classes for x_0
+ # - C_i is the class transitioning from (not to be confused with c_t and c_cumulative_t being used for gamma's)
+ # - C_j is the class transitioning to
+ #
+ # 1. x_t is masked i.e. x_t = c_k
+ #
+ # Simplifying the expression, the column vector is:
+ # .
+ # .
+ # .
+ # (c_t / c_cumulative_t) * (a_cumulative_{t-1} * p_n(x_0 = C_i | x_t) + b_cumulative_{t-1} * sum(p_n(x_0)))
+ # .
+ # .
+ # .
+ # (c_cumulative_{t-1} / c_cumulative_t) * sum(p_n(x_0))
+ #
+ # From equation (11) stated in terms of forward probabilities, the last row is trivially verified.
+ #
+ # For the other rows, we can state the equation as ...
+ #
+ # (c_t / c_cumulative_t) * [b_cumulative_{t-1} * p(x_0=c_0) + ... + (a_cumulative_{t-1} + b_cumulative_{t-1}) * p(x_0=C_i) + ... + b_cumulative_{k-1} * p(x_0=c_{k-1})]
+ #
+ # This verifies the other rows.
+ #
+ # 2. x_t is not masked
+ #
+ # Simplifying the expression, there are two cases for the rows of the column vector, where C_j = C_i and where C_j != C_i:
+ # .
+ # .
+ # .
+ # C_j != C_i: b_t * ((b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_0) + ... + ((a_cumulative_{t-1} + b_cumulative_{t-1}) / b_cumulative_t) * p_n(x_0 = C_i) + ... + (b_cumulative_{t-1} / (a_cumulative_t + b_cumulative_t)) * p_n(c_0=C_j) + ... + (b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_{k-1}))
+ # .
+ # .
+ # .
+ # C_j = C_i: (a_t + b_t) * ((b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_0) + ... + ((a_cumulative_{t-1} + b_cumulative_{t-1}) / (a_cumulative_t + b_cumulative_t)) * p_n(x_0 = C_i = C_j) + ... + (b_cumulative_{t-1} / b_cumulative_t) * p_n(x_0 = c_{k-1}))
+ # .
+ # .
+ # .
+ # 0
+ #
+ # The last row is trivially verified. The other rows can be verified by directly expanding equation (11) stated in terms of forward probabilities.
+ return log_p_x_t_min_1
+
+ def log_Q_t_transitioning_to_known_class(
+ self, *, t: torch.int, x_t: torch.LongTensor, log_onehot_x_t: torch.FloatTensor, cumulative: bool
+ ):
+ """
+ Calculates the log probabilities of the rows from the (cumulative or non-cumulative) transition matrix for each
+ latent pixel in `x_t`.
+
+ Args:
+ t (`torch.Long`):
+ The timestep that determines which transition matrix is used.
+ x_t (`torch.LongTensor` of shape `(batch size, num latent pixels)`):
+ The classes of each latent pixel at time `t`.
+ log_onehot_x_t (`torch.FloatTensor` of shape `(batch size, num classes, num latent pixels)`):
+ The log one-hot vectors of `x_t`.
+ cumulative (`bool`):
+ If cumulative is `False`, the single step transition matrix `t-1`->`t` is used. If cumulative is
+ `True`, the cumulative transition matrix `0`->`t` is used.
+
+ Returns:
+ `torch.FloatTensor` of shape `(batch size, num classes - 1, num latent pixels)`:
+ Each _column_ of the returned matrix is a _row_ of log probabilities of the complete probability
+ transition matrix.
+
+ When non cumulative, returns `self.num_classes - 1` rows because the initial latent pixel cannot be
+ masked.
+
+ Where:
+ - `q_n` is the probability distribution for the forward process of the `n`th latent pixel.
+ - C_0 is a class of a latent pixel embedding
+ - C_k is the class of the masked latent pixel
+
+ non-cumulative result (omitting logarithms):
+ ```
+ q_0(x_t | x_{t-1} = C_0) ... q_n(x_t | x_{t-1} = C_0)
+ . . .
+ . . .
+ . . .
+ q_0(x_t | x_{t-1} = C_k) ... q_n(x_t | x_{t-1} = C_k)
+ ```
+
+ cumulative result (omitting logarithms):
+ ```
+ q_0_cumulative(x_t | x_0 = C_0) ... q_n_cumulative(x_t | x_0 = C_0)
+ . . .
+ . . .
+ . . .
+ q_0_cumulative(x_t | x_0 = C_{k-1}) ... q_n_cumulative(x_t | x_0 = C_{k-1})
+ ```
+ """
+ if cumulative:
+ a = self.log_cumprod_at[t]
+ b = self.log_cumprod_bt[t]
+ c = self.log_cumprod_ct[t]
+ else:
+ a = self.log_at[t]
+ b = self.log_bt[t]
+ c = self.log_ct[t]
+
+ if not cumulative:
+ # The values in the onehot vector can also be used as the logprobs for transitioning
+ # from masked latent pixels. If we are not calculating the cumulative transitions,
+ # we need to save these vectors to be re-appended to the final matrix so the values
+ # aren't overwritten.
+ #
+ # `P(x_t!=mask|x_{t-1=mask}) = 0` and 0 will be the value of the last row of the onehot vector
+ # if x_t is not masked
+ #
+ # `P(x_t=mask|x_{t-1=mask}) = 1` and 1 will be the value of the last row of the onehot vector
+ # if x_t is masked
+ log_onehot_x_t_transitioning_from_masked = log_onehot_x_t[:, -1, :].unsqueeze(1)
+
+ # `index_to_log_onehot` will add onehot vectors for masked pixels,
+ # so the default one hot matrix has one too many rows. See the doc string
+ # for an explanation of the dimensionality of the returned matrix.
+ log_onehot_x_t = log_onehot_x_t[:, :-1, :]
+
+ # this is a cheeky trick to produce the transition probabilities using log one-hot vectors.
+ #
+ # Don't worry about what values this sets in the columns that mark transitions
+ # to masked latent pixels. They are overwrote later with the `mask_class_mask`.
+ #
+ # Looking at the below logspace formula in non-logspace, each value will evaluate to either
+ # `1 * a + b = a + b` where `log_Q_t` has the one hot value in the column
+ # or
+ # `0 * a + b = b` where `log_Q_t` has the 0 values in the column.
+ #
+ # See equation 7 for more details.
+ log_Q_t = (log_onehot_x_t + a).logaddexp(b)
+
+ # The whole column of each masked pixel is `c`
+ mask_class_mask = x_t == self.mask_class
+ mask_class_mask = mask_class_mask.unsqueeze(1).expand(-1, self.num_embed - 1, -1)
+ log_Q_t[mask_class_mask] = c
+
+ if not cumulative:
+ log_Q_t = torch.cat((log_Q_t, log_onehot_x_t_transitioning_from_masked), dim=1)
+
+ return log_Q_t
+
+ def apply_cumulative_transitions(self, q, t):
+ bsz = q.shape[0]
+ a = self.log_cumprod_at[t]
+ b = self.log_cumprod_bt[t]
+ c = self.log_cumprod_ct[t]
+
+ num_latent_pixels = q.shape[2]
+ c = c.expand(bsz, 1, num_latent_pixels)
+
+ q = (q + a).logaddexp(b)
+ q = torch.cat((q, c), dim=1)
+
+ return q
diff --git a/model/diffusers_c/training_utils.py b/model/diffusers_c/training_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..25e02a3d14925e8116ee8034812a11c6d5cc339b
--- /dev/null
+++ b/model/diffusers_c/training_utils.py
@@ -0,0 +1,453 @@
+import contextlib
+import copy
+import random
+from typing import Any, Dict, Iterable, List, Optional, Union
+
+import numpy as np
+import torch
+
+from .models import UNet2DConditionModel
+from .utils import (
+ convert_state_dict_to_diffusers,
+ convert_state_dict_to_peft,
+ deprecate,
+ is_peft_available,
+ is_torch_npu_available,
+ is_torchvision_available,
+ is_transformers_available,
+)
+
+
+if is_transformers_available():
+ import transformers
+
+if is_peft_available():
+ from peft import set_peft_model_state_dict
+
+if is_torchvision_available():
+ from torchvision import transforms
+
+if is_torch_npu_available():
+ import torch_npu # noqa: F401
+
+
+def set_seed(seed: int):
+ """
+ Args:
+ Helper function for reproducible behavior to set the seed in `random`, `numpy`, `torch`.
+ seed (`int`): The seed to set.
+ """
+ random.seed(seed)
+ np.random.seed(seed)
+ torch.manual_seed(seed)
+ if is_torch_npu_available():
+ torch.npu.manual_seed_all(seed)
+ else:
+ torch.cuda.manual_seed_all(seed)
+ # ^^ safe to call this function even if cuda is not available
+
+
+def compute_snr(noise_scheduler, timesteps):
+ """
+ Computes SNR as per
+ https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L847-L849
+ """
+ alphas_cumprod = noise_scheduler.alphas_cumprod
+ sqrt_alphas_cumprod = alphas_cumprod**0.5
+ sqrt_one_minus_alphas_cumprod = (1.0 - alphas_cumprod) ** 0.5
+
+ # Expand the tensors.
+ # Adapted from https://github.com/TiankaiHang/Min-SNR-Diffusion-Training/blob/521b624bd70c67cee4bdf49225915f5945a872e3/guided_diffusion/gaussian_diffusion.py#L1026
+ sqrt_alphas_cumprod = sqrt_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
+ while len(sqrt_alphas_cumprod.shape) < len(timesteps.shape):
+ sqrt_alphas_cumprod = sqrt_alphas_cumprod[..., None]
+ alpha = sqrt_alphas_cumprod.expand(timesteps.shape)
+
+ sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod.to(device=timesteps.device)[timesteps].float()
+ while len(sqrt_one_minus_alphas_cumprod.shape) < len(timesteps.shape):
+ sqrt_one_minus_alphas_cumprod = sqrt_one_minus_alphas_cumprod[..., None]
+ sigma = sqrt_one_minus_alphas_cumprod.expand(timesteps.shape)
+
+ # Compute SNR.
+ snr = (alpha / sigma) ** 2
+ return snr
+
+
+def resolve_interpolation_mode(interpolation_type: str):
+ """
+ Maps a string describing an interpolation function to the corresponding torchvision `InterpolationMode` enum. The
+ full list of supported enums is documented at
+ https://pytorch.org/vision/0.9/transforms.html#torchvision.transforms.functional.InterpolationMode.
+
+ Args:
+ interpolation_type (`str`):
+ A string describing an interpolation method. Currently, `bilinear`, `bicubic`, `box`, `nearest`,
+ `nearest_exact`, `hamming`, and `lanczos` are supported, corresponding to the supported interpolation modes
+ in torchvision.
+
+ Returns:
+ `torchvision.transforms.InterpolationMode`: an `InterpolationMode` enum used by torchvision's `resize`
+ transform.
+ """
+ if not is_torchvision_available():
+ raise ImportError(
+ "Please make sure to install `torchvision` to be able to use the `resolve_interpolation_mode()` function."
+ )
+
+ if interpolation_type == "bilinear":
+ interpolation_mode = transforms.InterpolationMode.BILINEAR
+ elif interpolation_type == "bicubic":
+ interpolation_mode = transforms.InterpolationMode.BICUBIC
+ elif interpolation_type == "box":
+ interpolation_mode = transforms.InterpolationMode.BOX
+ elif interpolation_type == "nearest":
+ interpolation_mode = transforms.InterpolationMode.NEAREST
+ elif interpolation_type == "nearest_exact":
+ interpolation_mode = transforms.InterpolationMode.NEAREST_EXACT
+ elif interpolation_type == "hamming":
+ interpolation_mode = transforms.InterpolationMode.HAMMING
+ elif interpolation_type == "lanczos":
+ interpolation_mode = transforms.InterpolationMode.LANCZOS
+ else:
+ raise ValueError(
+ f"The given interpolation mode {interpolation_type} is not supported. Currently supported interpolation"
+ f" modes are `bilinear`, `bicubic`, `box`, `nearest`, `nearest_exact`, `hamming`, and `lanczos`."
+ )
+
+ return interpolation_mode
+
+
+def unet_lora_state_dict(unet: UNet2DConditionModel) -> Dict[str, torch.Tensor]:
+ r"""
+ Returns:
+ A state dict containing just the LoRA parameters.
+ """
+ lora_state_dict = {}
+
+ for name, module in unet.named_modules():
+ if hasattr(module, "set_lora_layer"):
+ lora_layer = getattr(module, "lora_layer")
+ if lora_layer is not None:
+ current_lora_layer_sd = lora_layer.state_dict()
+ for lora_layer_matrix_name, lora_param in current_lora_layer_sd.items():
+ # The matrix name can either be "down" or "up".
+ lora_state_dict[f"{name}.lora.{lora_layer_matrix_name}"] = lora_param
+
+ return lora_state_dict
+
+
+def cast_training_params(model: Union[torch.nn.Module, List[torch.nn.Module]], dtype=torch.float32):
+ if not isinstance(model, list):
+ model = [model]
+ for m in model:
+ for param in m.parameters():
+ # only upcast trainable parameters into fp32
+ if param.requires_grad:
+ param.data = param.to(dtype)
+
+
+def _set_state_dict_into_text_encoder(
+ lora_state_dict: Dict[str, torch.Tensor], prefix: str, text_encoder: torch.nn.Module
+):
+ """
+ Sets the `lora_state_dict` into `text_encoder` coming from `transformers`.
+
+ Args:
+ lora_state_dict: The state dictionary to be set.
+ prefix: String identifier to retrieve the portion of the state dict that belongs to `text_encoder`.
+ text_encoder: Where the `lora_state_dict` is to be set.
+ """
+
+ text_encoder_state_dict = {
+ f'{k.replace(prefix, "")}': v for k, v in lora_state_dict.items() if k.startswith(prefix)
+ }
+ text_encoder_state_dict = convert_state_dict_to_peft(convert_state_dict_to_diffusers(text_encoder_state_dict))
+ set_peft_model_state_dict(text_encoder, text_encoder_state_dict, adapter_name="default")
+
+
+# Adapted from torch-ema https://github.com/fadel/pytorch_ema/blob/master/torch_ema/ema.py#L14
+class EMAModel:
+ """
+ Exponential Moving Average of models weights
+ """
+
+ def __init__(
+ self,
+ parameters: Iterable[torch.nn.Parameter],
+ decay: float = 0.9999,
+ min_decay: float = 0.0,
+ update_after_step: int = 0,
+ use_ema_warmup: bool = False,
+ inv_gamma: Union[float, int] = 1.0,
+ power: Union[float, int] = 2 / 3,
+ model_cls: Optional[Any] = None,
+ model_config: Dict[str, Any] = None,
+ **kwargs,
+ ):
+ """
+ Args:
+ parameters (Iterable[torch.nn.Parameter]): The parameters to track.
+ decay (float): The decay factor for the exponential moving average.
+ min_decay (float): The minimum decay factor for the exponential moving average.
+ update_after_step (int): The number of steps to wait before starting to update the EMA weights.
+ use_ema_warmup (bool): Whether to use EMA warmup.
+ inv_gamma (float):
+ Inverse multiplicative factor of EMA warmup. Default: 1. Only used if `use_ema_warmup` is True.
+ power (float): Exponential factor of EMA warmup. Default: 2/3. Only used if `use_ema_warmup` is True.
+ device (Optional[Union[str, torch.device]]): The device to store the EMA weights on. If None, the EMA
+ weights will be stored on CPU.
+
+ @crowsonkb's notes on EMA Warmup:
+ If gamma=1 and power=1, implements a simple average. gamma=1, power=2/3 are good values for models you plan
+ to train for a million or more steps (reaches decay factor 0.999 at 31.6K steps, 0.9999 at 1M steps),
+ gamma=1, power=3/4 for models you plan to train for less (reaches decay factor 0.999 at 10K steps, 0.9999
+ at 215.4k steps).
+ """
+
+ if isinstance(parameters, torch.nn.Module):
+ deprecation_message = (
+ "Passing a `torch.nn.Module` to `ExponentialMovingAverage` is deprecated. "
+ "Please pass the parameters of the module instead."
+ )
+ deprecate(
+ "passing a `torch.nn.Module` to `ExponentialMovingAverage`",
+ "1.0.0",
+ deprecation_message,
+ standard_warn=False,
+ )
+ parameters = parameters.parameters()
+
+ # set use_ema_warmup to True if a torch.nn.Module is passed for backwards compatibility
+ use_ema_warmup = True
+
+ if kwargs.get("max_value", None) is not None:
+ deprecation_message = "The `max_value` argument is deprecated. Please use `decay` instead."
+ deprecate("max_value", "1.0.0", deprecation_message, standard_warn=False)
+ decay = kwargs["max_value"]
+
+ if kwargs.get("min_value", None) is not None:
+ deprecation_message = "The `min_value` argument is deprecated. Please use `min_decay` instead."
+ deprecate("min_value", "1.0.0", deprecation_message, standard_warn=False)
+ min_decay = kwargs["min_value"]
+
+ parameters = list(parameters)
+ self.shadow_params = [p.clone().detach() for p in parameters]
+
+ if kwargs.get("device", None) is not None:
+ deprecation_message = "The `device` argument is deprecated. Please use `to` instead."
+ deprecate("device", "1.0.0", deprecation_message, standard_warn=False)
+ self.to(device=kwargs["device"])
+
+ self.temp_stored_params = None
+
+ self.decay = decay
+ self.min_decay = min_decay
+ self.update_after_step = update_after_step
+ self.use_ema_warmup = use_ema_warmup
+ self.inv_gamma = inv_gamma
+ self.power = power
+ self.optimization_step = 0
+ self.cur_decay_value = None # set in `step()`
+
+ self.model_cls = model_cls
+ self.model_config = model_config
+
+ @classmethod
+ def from_pretrained(cls, path, model_cls) -> "EMAModel":
+ _, ema_kwargs = model_cls.load_config(path, return_unused_kwargs=True)
+ model = model_cls.from_pretrained(path)
+
+ ema_model = cls(model.parameters(), model_cls=model_cls, model_config=model.config)
+
+ ema_model.load_state_dict(ema_kwargs)
+ return ema_model
+
+ def save_pretrained(self, path):
+ if self.model_cls is None:
+ raise ValueError("`save_pretrained` can only be used if `model_cls` was defined at __init__.")
+
+ if self.model_config is None:
+ raise ValueError("`save_pretrained` can only be used if `model_config` was defined at __init__.")
+
+ model = self.model_cls.from_config(self.model_config)
+ state_dict = self.state_dict()
+ state_dict.pop("shadow_params", None)
+
+ model.register_to_config(**state_dict)
+ self.copy_to(model.parameters())
+ model.save_pretrained(path)
+
+ def get_decay(self, optimization_step: int) -> float:
+ """
+ Compute the decay factor for the exponential moving average.
+ """
+ step = max(0, optimization_step - self.update_after_step - 1)
+
+ if step <= 0:
+ return 0.0
+
+ if self.use_ema_warmup:
+ cur_decay_value = 1 - (1 + step / self.inv_gamma) ** -self.power
+ else:
+ cur_decay_value = (1 + step) / (10 + step)
+
+ cur_decay_value = min(cur_decay_value, self.decay)
+ # make sure decay is not smaller than min_decay
+ cur_decay_value = max(cur_decay_value, self.min_decay)
+ return cur_decay_value
+
+ @torch.no_grad()
+ def step(self, parameters: Iterable[torch.nn.Parameter]):
+ if isinstance(parameters, torch.nn.Module):
+ deprecation_message = (
+ "Passing a `torch.nn.Module` to `ExponentialMovingAverage.step` is deprecated. "
+ "Please pass the parameters of the module instead."
+ )
+ deprecate(
+ "passing a `torch.nn.Module` to `ExponentialMovingAverage.step`",
+ "1.0.0",
+ deprecation_message,
+ standard_warn=False,
+ )
+ parameters = parameters.parameters()
+
+ parameters = list(parameters)
+
+ self.optimization_step += 1
+
+ # Compute the decay factor for the exponential moving average.
+ decay = self.get_decay(self.optimization_step)
+ self.cur_decay_value = decay
+ one_minus_decay = 1 - decay
+
+ context_manager = contextlib.nullcontext
+ if is_transformers_available() and transformers.deepspeed.is_deepspeed_zero3_enabled():
+ import deepspeed
+
+ for s_param, param in zip(self.shadow_params, parameters):
+ if is_transformers_available() and transformers.deepspeed.is_deepspeed_zero3_enabled():
+ context_manager = deepspeed.zero.GatheredParameters(param, modifier_rank=None)
+
+ with context_manager():
+ if param.requires_grad:
+ s_param.sub_(one_minus_decay * (s_param - param))
+ else:
+ s_param.copy_(param)
+
+ def copy_to(self, parameters: Iterable[torch.nn.Parameter]) -> None:
+ """
+ Copy current averaged parameters into given collection of parameters.
+
+ Args:
+ parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+ updated with the stored moving averages. If `None`, the parameters with which this
+ `ExponentialMovingAverage` was initialized will be used.
+ """
+ parameters = list(parameters)
+ for s_param, param in zip(self.shadow_params, parameters):
+ param.data.copy_(s_param.to(param.device).data)
+
+ def to(self, device=None, dtype=None) -> None:
+ r"""Move internal buffers of the ExponentialMovingAverage to `device`.
+
+ Args:
+ device: like `device` argument to `torch.Tensor.to`
+ """
+ # .to() on the tensors handles None correctly
+ self.shadow_params = [
+ p.to(device=device, dtype=dtype) if p.is_floating_point() else p.to(device=device)
+ for p in self.shadow_params
+ ]
+
+ def state_dict(self) -> dict:
+ r"""
+ Returns the state of the ExponentialMovingAverage as a dict. This method is used by accelerate during
+ checkpointing to save the ema state dict.
+ """
+ # Following PyTorch conventions, references to tensors are returned:
+ # "returns a reference to the state and not its copy!" -
+ # https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict
+ return {
+ "decay": self.decay,
+ "min_decay": self.min_decay,
+ "optimization_step": self.optimization_step,
+ "update_after_step": self.update_after_step,
+ "use_ema_warmup": self.use_ema_warmup,
+ "inv_gamma": self.inv_gamma,
+ "power": self.power,
+ "shadow_params": self.shadow_params,
+ }
+
+ def store(self, parameters: Iterable[torch.nn.Parameter]) -> None:
+ r"""
+ Args:
+ Save the current parameters for restoring later.
+ parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+ temporarily stored.
+ """
+ self.temp_stored_params = [param.detach().cpu().clone() for param in parameters]
+
+ def restore(self, parameters: Iterable[torch.nn.Parameter]) -> None:
+ r"""
+ Args:
+ Restore the parameters stored with the `store` method. Useful to validate the model with EMA parameters without:
+ affecting the original optimization process. Store the parameters before the `copy_to()` method. After
+ validation (or model saving), use this to restore the former parameters.
+ parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+ updated with the stored parameters. If `None`, the parameters with which this
+ `ExponentialMovingAverage` was initialized will be used.
+ """
+ if self.temp_stored_params is None:
+ raise RuntimeError("This ExponentialMovingAverage has no `store()`ed weights " "to `restore()`")
+ for c_param, param in zip(self.temp_stored_params, parameters):
+ param.data.copy_(c_param.data)
+
+ # Better memory-wise.
+ self.temp_stored_params = None
+
+ def load_state_dict(self, state_dict: dict) -> None:
+ r"""
+ Args:
+ Loads the ExponentialMovingAverage state. This method is used by accelerate during checkpointing to save the
+ ema state dict.
+ state_dict (dict): EMA state. Should be an object returned
+ from a call to :meth:`state_dict`.
+ """
+ # deepcopy, to be consistent with module API
+ state_dict = copy.deepcopy(state_dict)
+
+ self.decay = state_dict.get("decay", self.decay)
+ if self.decay < 0.0 or self.decay > 1.0:
+ raise ValueError("Decay must be between 0 and 1")
+
+ self.min_decay = state_dict.get("min_decay", self.min_decay)
+ if not isinstance(self.min_decay, float):
+ raise ValueError("Invalid min_decay")
+
+ self.optimization_step = state_dict.get("optimization_step", self.optimization_step)
+ if not isinstance(self.optimization_step, int):
+ raise ValueError("Invalid optimization_step")
+
+ self.update_after_step = state_dict.get("update_after_step", self.update_after_step)
+ if not isinstance(self.update_after_step, int):
+ raise ValueError("Invalid update_after_step")
+
+ self.use_ema_warmup = state_dict.get("use_ema_warmup", self.use_ema_warmup)
+ if not isinstance(self.use_ema_warmup, bool):
+ raise ValueError("Invalid use_ema_warmup")
+
+ self.inv_gamma = state_dict.get("inv_gamma", self.inv_gamma)
+ if not isinstance(self.inv_gamma, (float, int)):
+ raise ValueError("Invalid inv_gamma")
+
+ self.power = state_dict.get("power", self.power)
+ if not isinstance(self.power, (float, int)):
+ raise ValueError("Invalid power")
+
+ shadow_params = state_dict.get("shadow_params", None)
+ if shadow_params is not None:
+ self.shadow_params = shadow_params
+ if not isinstance(self.shadow_params, list):
+ raise ValueError("shadow_params must be a list")
+ if not all(isinstance(p, torch.Tensor) for p in self.shadow_params):
+ raise ValueError("shadow_params must all be Tensors")
diff --git a/model/diffusers_c/utils/__init__.py b/model/diffusers_c/utils/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..4e2f07f2ba890e466851d0263b2ae745ec734a27
--- /dev/null
+++ b/model/diffusers_c/utils/__init__.py
@@ -0,0 +1,124 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import os
+
+from packaging import version
+
+from .. import __version__
+from .constants import (
+ CONFIG_NAME,
+ DEPRECATED_REVISION_ARGS,
+ DIFFUSERS_DYNAMIC_MODULE_NAME,
+ FLAX_WEIGHTS_NAME,
+ HF_MODULES_CACHE,
+ HUGGINGFACE_CO_RESOLVE_ENDPOINT,
+ MIN_PEFT_VERSION,
+ ONNX_EXTERNAL_WEIGHTS_NAME,
+ ONNX_WEIGHTS_NAME,
+ SAFETENSORS_FILE_EXTENSION,
+ SAFETENSORS_WEIGHTS_NAME,
+ USE_PEFT_BACKEND,
+ WEIGHTS_NAME,
+)
+from .deprecation_utils import deprecate
+from .doc_utils import replace_example_docstring
+from .dynamic_modules_utils import get_class_from_dynamic_module
+from .export_utils import export_to_gif, export_to_obj, export_to_ply, export_to_video
+from .hub_utils import (
+ PushToHubMixin,
+ _add_variant,
+ _get_model_file,
+ extract_commit_hash,
+ http_user_agent,
+)
+from .import_utils import (
+ BACKENDS_MAPPING,
+ DIFFUSERS_SLOW_IMPORT,
+ ENV_VARS_TRUE_AND_AUTO_VALUES,
+ ENV_VARS_TRUE_VALUES,
+ USE_JAX,
+ USE_TF,
+ USE_TORCH,
+ DummyObject,
+ OptionalDependencyNotAvailable,
+ _LazyModule,
+ get_objects_from_module,
+ is_accelerate_available,
+ is_accelerate_version,
+ is_bs4_available,
+ is_flax_available,
+ is_ftfy_available,
+ is_inflect_available,
+ is_invisible_watermark_available,
+ is_k_diffusion_available,
+ is_k_diffusion_version,
+ is_librosa_available,
+ is_note_seq_available,
+ is_onnx_available,
+ is_peft_available,
+ is_scipy_available,
+ is_tensorboard_available,
+ is_torch_available,
+ is_torch_npu_available,
+ is_torch_version,
+ is_torch_xla_available,
+ is_torchsde_available,
+ is_torchvision_available,
+ is_transformers_available,
+ is_transformers_version,
+ is_unidecode_available,
+ is_wandb_available,
+ is_xformers_available,
+ requires_backends,
+)
+from .loading_utils import load_image
+from .logging import get_logger
+from .outputs import BaseOutput
+from .peft_utils import (
+ check_peft_version,
+ delete_adapter_layers,
+ get_adapter_name,
+ get_peft_kwargs,
+ recurse_remove_peft_layers,
+ scale_lora_layers,
+ set_adapter_layers,
+ set_weights_and_activate_adapters,
+ unscale_lora_layers,
+)
+from .pil_utils import PIL_INTERPOLATION, make_image_grid, numpy_to_pil, pt_to_pil
+from .state_dict_utils import (
+ convert_all_state_dict_to_peft,
+ convert_state_dict_to_diffusers,
+ convert_state_dict_to_kohya,
+ convert_state_dict_to_peft,
+ convert_unet_state_dict_to_peft,
+)
+
+
+logger = get_logger(__name__)
+
+
+def check_min_version(min_version):
+ if version.parse(__version__) < version.parse(min_version):
+ if "dev" in min_version:
+ error_message = (
+ "This example requires a source install from HuggingFace diffusers (see "
+ "`https://huggingface.co/docs/diffusers/installation#install-from-source`),"
+ )
+ else:
+ error_message = f"This example requires a minimum version of {min_version},"
+ error_message += f" but the version found is {__version__}.\n"
+ raise ImportError(error_message)
diff --git a/model/diffusers_c/utils/__pycache__/__init__.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/__init__.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c3381bf723f10533fddc4775b1a0d3799c128ae1
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/__init__.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/__init__.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/__init__.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..507dcd39300e6029f618eb490640a09c94a70db2
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/__init__.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/accelerate_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/accelerate_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..437675efa72d84bf69191a825f7fb4b06b75f9b3
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/accelerate_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/accelerate_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/accelerate_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..25be1a5e2217504dcd726d8a70d6a3a098ca4b6f
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/accelerate_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/constants.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/constants.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..924c797af90d401ca864cd40c4e755b7bed92cc0
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/constants.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/constants.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/constants.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..c3acaa7219d4746d95dcdbcb93dc2ec331e4cd80
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/constants.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/deprecation_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/deprecation_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b748a9929c093f077c8559d63ac03cb404338bd2
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/deprecation_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/deprecation_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/deprecation_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f6500746c32c2351b02e4ba3131a6e8792182080
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/deprecation_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/doc_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/doc_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..df8cd57a3954f6f80313abe4fb7ef4c813e7457a
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/doc_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/doc_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/doc_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0744f23fa32c9deb2afc522f47b29e4608395f34
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/doc_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..da32ee1ce95ed3d1c0195790fadb687313de2449
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..84aba4dde4709a2af69b7a63d77d3ca67863403b
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_flax_and_transformers_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_flax_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_flax_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f3474abefec930d2199078bc508db336e2495848
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_flax_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_flax_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_flax_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..b9f610e436a8adb8645a46b2f7196e3c0b871870
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_flax_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_note_seq_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_note_seq_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6e32b3482bdc3fa546b6a78846adf6bfb92e5feb
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_note_seq_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_note_seq_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_note_seq_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f584eef5a54b46c6e6651c530328e6f835f5e65a
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_note_seq_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_onnx_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_onnx_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..88e17d5e7ffafd07ec830d8bb89984023d5078f7
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_onnx_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_onnx_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_onnx_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..e6593d7c044d29807dd0bb339c5da9d0f543d8d0
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_onnx_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..747d3e14c05e760fc2a97b06488374575e447afb
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5e484b6a0bcd7f19914fa1b58d59195c2b658cde
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_torch_and_librosa_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d29a4863153c11b215fa0dd089e4c92cc1a18d0f
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..59aa58b3e54722c322e0a3176c5881bf1900da75
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_torch_and_torchsde_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..16ea35c19408832ba9b4d3d80a1e9a73823b288b
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..4e6876128310452177ae1cd6d1ae5e1c075768ae
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_k_diffusion_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..43be5aca35a38213ad76b0e41984980247448083
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..03e062eef856fed5b21b3d640716d30fefcdf8e1
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_torch_and_transformers_and_onnx_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..a61173b66ed37f696bb46febdbe760114f48af90
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..1efdc5b88a696dc375116060e1ae32180f40146c
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dummy_transformers_and_torch_and_note_seq_objects.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dynamic_modules_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/dynamic_modules_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d07d9df7a36fa3924473810b435c9b2dc0d67847
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dynamic_modules_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/dynamic_modules_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/dynamic_modules_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0904609e378bf8c288b1fad0414a575f8251d09e
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/dynamic_modules_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/export_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/export_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..45c0af881f061415eefe29c558bdcd707f9487a4
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/export_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/export_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/export_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f564a23d7ff766bea74006941c5556115734c088
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/export_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/hub_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/hub_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..0c95c517d727adf8f55e8563bce49055646c8621
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/hub_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/hub_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/hub_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..00daae9d8cb8f167ced9a1044d086d75e9f7453e
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/hub_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/import_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/import_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..8c82afeccbb0fcdc81927a90dab91b7c781291aa
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/import_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/import_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/import_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..94f24fdf9db041bc9b033be28d38d5d77d7a9d43
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/import_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/loading_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/loading_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..d5a968fe090333a6ac96b13c7132c609c9e8dfc3
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/loading_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/loading_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/loading_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f38f886d9ea7edb28c8611868dbc8bb3a1111c2b
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/loading_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/logging.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/logging.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2a563483fa1b7bdf79ef95c135e8ec3a0243355d
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/logging.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/logging.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/logging.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..5b53262e60d51585cb455e1e301093487f2a881f
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/logging.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/outputs.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/outputs.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2c1495dd31aa15478f9321a869efd94631026362
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/outputs.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/outputs.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/outputs.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..519ec531f2d1d96855d0739a2c5f20c07bde7c60
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/outputs.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/peft_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/peft_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3a187ac59c96f57264d139c3942942cce83b8ac4
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/peft_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/peft_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/peft_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..3efd04f2d5492b01fa915b94b8e2b69e712b7bea
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/peft_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/pil_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/pil_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..571ffeef3f69c4edb91fad73576fd333eee84413
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/pil_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/pil_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/pil_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6969617b809e851e41ae0929ad613c1e93036139
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/pil_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/state_dict_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/state_dict_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..16bd3063116e9f7b45cca880d253432128339360
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/state_dict_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/state_dict_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/state_dict_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2d641713b905e3e311f8c6431147023b685eeed6
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/state_dict_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/torch_utils.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/torch_utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f647a912b443a31635d383266c5f5eb8725375c0
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/torch_utils.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/torch_utils.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/torch_utils.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..f634552e0e57fed35d71ea040519ac82fc9bfca6
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/torch_utils.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/versions.cpython-310.pyc b/model/diffusers_c/utils/__pycache__/versions.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..2d523571519cd34b2657e58e847c2b7bd741745b
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/versions.cpython-310.pyc differ
diff --git a/model/diffusers_c/utils/__pycache__/versions.cpython-39.pyc b/model/diffusers_c/utils/__pycache__/versions.cpython-39.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..87ffe367129aca222d5e4c89ee722ae00bf3dc80
Binary files /dev/null and b/model/diffusers_c/utils/__pycache__/versions.cpython-39.pyc differ
diff --git a/model/diffusers_c/utils/accelerate_utils.py b/model/diffusers_c/utils/accelerate_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..99a8b3a47c256cc80dbd5de7d14101dcc7a9ff9a
--- /dev/null
+++ b/model/diffusers_c/utils/accelerate_utils.py
@@ -0,0 +1,48 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Accelerate utilities: Utilities related to accelerate
+"""
+
+from packaging import version
+
+from .import_utils import is_accelerate_available
+
+
+if is_accelerate_available():
+ import accelerate
+
+
+def apply_forward_hook(method):
+ """
+ Decorator that applies a registered CpuOffload hook to an arbitrary function rather than `forward`. This is useful
+ for cases where a PyTorch module provides functions other than `forward` that should trigger a move to the
+ appropriate acceleration device. This is the case for `encode` and `decode` in [`AutoencoderKL`].
+
+ This decorator looks inside the internal `_hf_hook` property to find a registered offload hook.
+
+ :param method: The method to decorate. This method should be a method of a PyTorch module.
+ """
+ if not is_accelerate_available():
+ return method
+ accelerate_version = version.parse(accelerate.__version__).base_version
+ if version.parse(accelerate_version) < version.parse("0.17.0"):
+ return method
+
+ def wrapper(self, *args, **kwargs):
+ if hasattr(self, "_hf_hook") and hasattr(self._hf_hook, "pre_forward"):
+ self._hf_hook.pre_forward(self)
+ return method(self, *args, **kwargs)
+
+ return wrapper
diff --git a/model/diffusers_c/utils/constants.py b/model/diffusers_c/utils/constants.py
new file mode 100644
index 0000000000000000000000000000000000000000..bc4268a32ac5b72b059d0c39c715a0e69a371494
--- /dev/null
+++ b/model/diffusers_c/utils/constants.py
@@ -0,0 +1,55 @@
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib
+import os
+
+from huggingface_hub.constants import HF_HOME
+from packaging import version
+
+from ..dependency_versions_check import dep_version_check
+from .import_utils import ENV_VARS_TRUE_VALUES, is_peft_available, is_transformers_available
+
+
+MIN_PEFT_VERSION = "0.6.0"
+MIN_TRANSFORMERS_VERSION = "4.34.0"
+_CHECK_PEFT = os.environ.get("_CHECK_PEFT", "1") in ENV_VARS_TRUE_VALUES
+
+
+CONFIG_NAME = "config.json"
+WEIGHTS_NAME = "diffusion_pytorch_model.bin"
+FLAX_WEIGHTS_NAME = "diffusion_flax_model.msgpack"
+ONNX_WEIGHTS_NAME = "model.onnx"
+SAFETENSORS_WEIGHTS_NAME = "diffusion_pytorch_model.safetensors"
+SAFETENSORS_FILE_EXTENSION = "safetensors"
+ONNX_EXTERNAL_WEIGHTS_NAME = "weights.pb"
+HUGGINGFACE_CO_RESOLVE_ENDPOINT = os.environ.get("HF_ENDPOINT", "https://huggingface.co")
+DIFFUSERS_DYNAMIC_MODULE_NAME = "diffusers_modules"
+HF_MODULES_CACHE = os.getenv("HF_MODULES_CACHE", os.path.join(HF_HOME, "modules"))
+DEPRECATED_REVISION_ARGS = ["fp16", "non-ema"]
+
+# Below should be `True` if the current version of `peft` and `transformers` are compatible with
+# PEFT backend. Will automatically fall back to PEFT backend if the correct versions of the libraries are
+# available.
+# For PEFT it is has to be greater than or equal to 0.6.0 and for transformers it has to be greater than or equal to 4.34.0.
+_required_peft_version = is_peft_available() and version.parse(
+ version.parse(importlib.metadata.version("peft")).base_version
+) >= version.parse(MIN_PEFT_VERSION)
+_required_transformers_version = is_transformers_available() and version.parse(
+ version.parse(importlib.metadata.version("transformers")).base_version
+) >= version.parse(MIN_TRANSFORMERS_VERSION)
+
+USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version
+
+if USE_PEFT_BACKEND and _CHECK_PEFT:
+ dep_version_check("peft")
diff --git a/model/diffusers_c/utils/deprecation_utils.py b/model/diffusers_c/utils/deprecation_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..f482deddd2f46b8d2e29d5229faa0e9a21f2fd98
--- /dev/null
+++ b/model/diffusers_c/utils/deprecation_utils.py
@@ -0,0 +1,49 @@
+import inspect
+import warnings
+from typing import Any, Dict, Optional, Union
+
+from packaging import version
+
+
+def deprecate(*args, take_from: Optional[Union[Dict, Any]] = None, standard_warn=True, stacklevel=2):
+ from .. import __version__
+
+ deprecated_kwargs = take_from
+ values = ()
+ if not isinstance(args[0], tuple):
+ args = (args,)
+
+ for attribute, version_name, message in args:
+ if version.parse(version.parse(__version__).base_version) >= version.parse(version_name):
+ raise ValueError(
+ f"The deprecation tuple {(attribute, version_name, message)} should be removed since diffusers'"
+ f" version {__version__} is >= {version_name}"
+ )
+
+ warning = None
+ if isinstance(deprecated_kwargs, dict) and attribute in deprecated_kwargs:
+ values += (deprecated_kwargs.pop(attribute),)
+ warning = f"The `{attribute}` argument is deprecated and will be removed in version {version_name}."
+ elif hasattr(deprecated_kwargs, attribute):
+ values += (getattr(deprecated_kwargs, attribute),)
+ warning = f"The `{attribute}` attribute is deprecated and will be removed in version {version_name}."
+ elif deprecated_kwargs is None:
+ warning = f"`{attribute}` is deprecated and will be removed in version {version_name}."
+
+ if warning is not None:
+ warning = warning + " " if standard_warn else ""
+ warnings.warn(warning + message, FutureWarning, stacklevel=stacklevel)
+
+ if isinstance(deprecated_kwargs, dict) and len(deprecated_kwargs) > 0:
+ call_frame = inspect.getouterframes(inspect.currentframe())[1]
+ filename = call_frame.filename
+ line_number = call_frame.lineno
+ function = call_frame.function
+ key, value = next(iter(deprecated_kwargs.items()))
+ raise TypeError(f"{function} in {filename} line {line_number-1} got an unexpected keyword argument `{key}`")
+
+ if len(values) == 0:
+ return
+ elif len(values) == 1:
+ return values[0]
+ return values
diff --git a/model/diffusers_c/utils/doc_utils.py b/model/diffusers_c/utils/doc_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..fe633e6836422f22618548b3f6d946d6b073650f
--- /dev/null
+++ b/model/diffusers_c/utils/doc_utils.py
@@ -0,0 +1,39 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Doc utilities: Utilities related to documentation
+"""
+
+import re
+
+
+def replace_example_docstring(example_docstring):
+ def docstring_decorator(fn):
+ func_doc = fn.__doc__
+ lines = func_doc.split("\n")
+ i = 0
+ while i < len(lines) and re.search(r"^\s*Examples?:\s*$", lines[i]) is None:
+ i += 1
+ if i < len(lines):
+ lines[i] = example_docstring
+ func_doc = "\n".join(lines)
+ else:
+ raise ValueError(
+ f"The function {fn} should have an empty 'Examples:' in its docstring as placeholder, "
+ f"current docstring is:\n{func_doc}"
+ )
+ fn.__doc__ = func_doc
+ return fn
+
+ return docstring_decorator
diff --git a/model/diffusers_c/utils/dummy_flax_and_transformers_objects.py b/model/diffusers_c/utils/dummy_flax_and_transformers_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..5e65e5349bb0a6a0bac62cddf0ce0fad64237c68
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_flax_and_transformers_objects.py
@@ -0,0 +1,77 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class FlaxStableDiffusionControlNetPipeline(metaclass=DummyObject):
+ _backends = ["flax", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+
+class FlaxStableDiffusionImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["flax", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+
+class FlaxStableDiffusionInpaintPipeline(metaclass=DummyObject):
+ _backends = ["flax", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+
+class FlaxStableDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["flax", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+
+class FlaxStableDiffusionXLPipeline(metaclass=DummyObject):
+ _backends = ["flax", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax", "transformers"])
diff --git a/model/diffusers_c/utils/dummy_flax_objects.py b/model/diffusers_c/utils/dummy_flax_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..5fa8dbc819316e96f7483addba43f90b9d8f397b
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_flax_objects.py
@@ -0,0 +1,212 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class FlaxControlNetModel(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxModelMixin(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxUNet2DConditionModel(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxAutoencoderKL(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxDDIMScheduler(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxDDPMScheduler(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxDPMSolverMultistepScheduler(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxEulerDiscreteScheduler(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxKarrasVeScheduler(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxLMSDiscreteScheduler(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxPNDMScheduler(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxSchedulerMixin(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+
+class FlaxScoreSdeVeScheduler(metaclass=DummyObject):
+ _backends = ["flax"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["flax"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["flax"])
diff --git a/model/diffusers_c/utils/dummy_note_seq_objects.py b/model/diffusers_c/utils/dummy_note_seq_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..c02d0b015aedc37c01fb3b843bc79547aae5da68
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_note_seq_objects.py
@@ -0,0 +1,17 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class MidiProcessor(metaclass=DummyObject):
+ _backends = ["note_seq"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["note_seq"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["note_seq"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["note_seq"])
diff --git a/model/diffusers_c/utils/dummy_onnx_objects.py b/model/diffusers_c/utils/dummy_onnx_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..bde5f6ad0793e2d81bc638600b46ff81748d09ee
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_onnx_objects.py
@@ -0,0 +1,17 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class OnnxRuntimeModel(metaclass=DummyObject):
+ _backends = ["onnx"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["onnx"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["onnx"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["onnx"])
diff --git a/model/diffusers_c/utils/dummy_pt_objects.py b/model/diffusers_c/utils/dummy_pt_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..14947848a43f566e4af15c4365fc2405b5dcc84c
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_pt_objects.py
@@ -0,0 +1,1170 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class AsymmetricAutoencoderKL(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class AutoencoderKL(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class AutoencoderKLTemporalDecoder(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class AutoencoderTiny(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class ConsistencyDecoderVAE(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class ControlNetModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class I2VGenXLUNet(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class Kandinsky3UNet(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class ModelMixin(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class MotionAdapter(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class MultiAdapter(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class PriorTransformer(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class T2IAdapter(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class T5FilmDecoder(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class Transformer2DModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UNet1DModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UNet2DConditionModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UNet2DModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UNet3DConditionModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UNetMotionModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UNetSpatioTemporalConditionModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UVit2DModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class VQModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+def get_constant_schedule(*args, **kwargs):
+ requires_backends(get_constant_schedule, ["torch"])
+
+
+def get_constant_schedule_with_warmup(*args, **kwargs):
+ requires_backends(get_constant_schedule_with_warmup, ["torch"])
+
+
+def get_cosine_schedule_with_warmup(*args, **kwargs):
+ requires_backends(get_cosine_schedule_with_warmup, ["torch"])
+
+
+def get_cosine_with_hard_restarts_schedule_with_warmup(*args, **kwargs):
+ requires_backends(get_cosine_with_hard_restarts_schedule_with_warmup, ["torch"])
+
+
+def get_linear_schedule_with_warmup(*args, **kwargs):
+ requires_backends(get_linear_schedule_with_warmup, ["torch"])
+
+
+def get_polynomial_decay_schedule_with_warmup(*args, **kwargs):
+ requires_backends(get_polynomial_decay_schedule_with_warmup, ["torch"])
+
+
+def get_scheduler(*args, **kwargs):
+ requires_backends(get_scheduler, ["torch"])
+
+
+class AudioPipelineOutput(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class AutoPipelineForImage2Image(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class AutoPipelineForInpainting(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class AutoPipelineForText2Image(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class BlipDiffusionControlNetPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class BlipDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class CLIPImageProjection(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class ConsistencyModelPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DanceDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DDIMPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DDPMPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DiTPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class ImagePipelineOutput(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class KarrasVePipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class LDMPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class LDMSuperResolutionPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class PNDMPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class RePaintPipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class ScoreSdeVePipeline(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class StableDiffusionMixin(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class AmusedScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class CMStochasticIterativeScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DDIMInverseScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DDIMParallelScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DDIMScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DDPMParallelScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DDPMScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DDPMWuerstchenScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DEISMultistepScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DPMSolverMultistepInverseScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DPMSolverMultistepScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class DPMSolverSinglestepScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class EDMDPMSolverMultistepScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class EDMEulerScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class EulerAncestralDiscreteScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class EulerDiscreteScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class HeunDiscreteScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class IPNDMScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class KarrasVeScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class KDPM2AncestralDiscreteScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class KDPM2DiscreteScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class LCMScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class PNDMScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class RePaintScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class SASolverScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class SchedulerMixin(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class ScoreSdeVeScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class TCDScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UnCLIPScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class UniPCMultistepScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class VQDiffusionScheduler(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
+class EMAModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
diff --git a/model/diffusers_c/utils/dummy_torch_and_librosa_objects.py b/model/diffusers_c/utils/dummy_torch_and_librosa_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..2088bc4a744198284f22fe54e6f1055cf3568566
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_torch_and_librosa_objects.py
@@ -0,0 +1,32 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class AudioDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "librosa"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "librosa"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "librosa"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "librosa"])
+
+
+class Mel(metaclass=DummyObject):
+ _backends = ["torch", "librosa"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "librosa"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "librosa"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "librosa"])
diff --git a/model/diffusers_c/utils/dummy_torch_and_scipy_objects.py b/model/diffusers_c/utils/dummy_torch_and_scipy_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..a1ff25863822b04971d2c6dfdc17f5b28774cf05
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_torch_and_scipy_objects.py
@@ -0,0 +1,17 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class LMSDiscreteScheduler(metaclass=DummyObject):
+ _backends = ["torch", "scipy"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "scipy"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "scipy"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "scipy"])
diff --git a/model/diffusers_c/utils/dummy_torch_and_torchsde_objects.py b/model/diffusers_c/utils/dummy_torch_and_torchsde_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..a81bbb316f32267c31b06598519f1eef9ddde643
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_torch_and_torchsde_objects.py
@@ -0,0 +1,17 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class DPMSolverSDEScheduler(metaclass=DummyObject):
+ _backends = ["torch", "torchsde"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "torchsde"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "torchsde"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "torchsde"])
diff --git a/model/diffusers_c/utils/dummy_torch_and_transformers_and_k_diffusion_objects.py b/model/diffusers_c/utils/dummy_torch_and_transformers_and_k_diffusion_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..2ab00c54ce2d7ca78f674085f0fe294d1f05c1e0
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_torch_and_transformers_and_k_diffusion_objects.py
@@ -0,0 +1,32 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class StableDiffusionKDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers", "k_diffusion"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers", "k_diffusion"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "k_diffusion"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "k_diffusion"])
+
+
+class StableDiffusionXLKDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers", "k_diffusion"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers", "k_diffusion"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "k_diffusion"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "k_diffusion"])
diff --git a/model/diffusers_c/utils/dummy_torch_and_transformers_and_onnx_objects.py b/model/diffusers_c/utils/dummy_torch_and_transformers_and_onnx_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..b7afad8226b87292100270e3e7daad6885be0e7f
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_torch_and_transformers_and_onnx_objects.py
@@ -0,0 +1,92 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class OnnxStableDiffusionImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers", "onnx"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+
+class OnnxStableDiffusionInpaintPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers", "onnx"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+
+class OnnxStableDiffusionInpaintPipelineLegacy(metaclass=DummyObject):
+ _backends = ["torch", "transformers", "onnx"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+
+class OnnxStableDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers", "onnx"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+
+class OnnxStableDiffusionUpscalePipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers", "onnx"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+
+class StableDiffusionOnnxPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers", "onnx"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers", "onnx"])
diff --git a/model/diffusers_c/utils/dummy_torch_and_transformers_objects.py b/model/diffusers_c/utils/dummy_torch_and_transformers_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..3d54398786488448d8009896974852ac6735c40e
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_torch_and_transformers_objects.py
@@ -0,0 +1,1577 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class AltDiffusionImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AltDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AmusedImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AmusedInpaintPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AmusedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AnimateDiffPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AnimateDiffVideoToVideoPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AudioLDM2Pipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AudioLDM2ProjectionModel(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AudioLDM2UNet2DConditionModel(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class AudioLDMPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class CLIPImageProjection(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class CycleDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class I2VGenXLPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class IFImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class IFImg2ImgSuperResolutionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class IFInpaintingPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class IFInpaintingSuperResolutionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class IFPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class IFSuperResolutionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class ImageTextPipelineOutput(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class Kandinsky3Img2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class Kandinsky3Pipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyCombinedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyImg2ImgCombinedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyInpaintCombinedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyInpaintPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyPriorPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22CombinedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22ControlnetImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22ControlnetPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22Img2ImgCombinedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22Img2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22InpaintCombinedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22InpaintPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22Pipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22PriorEmb2EmbPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class KandinskyV22PriorPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class LatentConsistencyModelImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class LatentConsistencyModelPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class LDMTextToImagePipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class MusicLDMPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class PaintByExamplePipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class PIAPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class PixArtAlphaPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class SemanticStableDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class ShapEImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class ShapEPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableCascadeCombinedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableCascadeDecoderPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableCascadePriorPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionAdapterPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionAttendAndExcitePipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionControlNetImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionControlNetInpaintPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionControlNetPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionDepth2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionDiffEditPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionGLIGENPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionGLIGENTextImagePipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionImageVariationPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionInpaintPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionInpaintPipelineLegacy(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionInstructPix2PixPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionLatentUpscalePipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionLDM3DPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionModelEditingPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionPanoramaPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionParadigmsPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionPipelineSafe(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionPix2PixZeroPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionSAGPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionUpscalePipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLAdapterPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLControlNetImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLControlNetInpaintPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLControlNetPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLInpaintPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLInstructPix2PixPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableDiffusionXLPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableUnCLIPImg2ImgPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableUnCLIPPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class StableVideoDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class TextToVideoSDPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class TextToVideoZeroPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class TextToVideoZeroSDXLPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class UnCLIPImageVariationPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class UnCLIPPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class UniDiffuserModel(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class UniDiffuserPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class UniDiffuserTextDecoder(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class VersatileDiffusionDualGuidedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class VersatileDiffusionImageVariationPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class VersatileDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class VersatileDiffusionTextToImagePipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class VideoToVideoSDPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class VQDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class WuerstchenCombinedPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class WuerstchenDecoderPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
+class WuerstchenPriorPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
diff --git a/model/diffusers_c/utils/dummy_transformers_and_torch_and_note_seq_objects.py b/model/diffusers_c/utils/dummy_transformers_and_torch_and_note_seq_objects.py
new file mode 100644
index 0000000000000000000000000000000000000000..fbde04e33f0abd86d12f3dee048a4f0585c9f19d
--- /dev/null
+++ b/model/diffusers_c/utils/dummy_transformers_and_torch_and_note_seq_objects.py
@@ -0,0 +1,17 @@
+# This file is autogenerated by the command `make fix-copies`, do not edit.
+from ..utils import DummyObject, requires_backends
+
+
+class SpectrogramDiffusionPipeline(metaclass=DummyObject):
+ _backends = ["transformers", "torch", "note_seq"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["transformers", "torch", "note_seq"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["transformers", "torch", "note_seq"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["transformers", "torch", "note_seq"])
diff --git a/model/diffusers_c/utils/dynamic_modules_utils.py b/model/diffusers_c/utils/dynamic_modules_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..5b0937638aa0431ceafc06d0ab63ebf6224fd22f
--- /dev/null
+++ b/model/diffusers_c/utils/dynamic_modules_utils.py
@@ -0,0 +1,451 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Utilities to dynamically load objects from the Hub."""
+
+import importlib
+import inspect
+import json
+import os
+import re
+import shutil
+import sys
+from pathlib import Path
+from typing import Dict, Optional, Union
+from urllib import request
+
+from huggingface_hub import cached_download, hf_hub_download, model_info
+from huggingface_hub.utils import validate_hf_hub_args
+from packaging import version
+
+from .. import __version__
+from . import DIFFUSERS_DYNAMIC_MODULE_NAME, HF_MODULES_CACHE, logging
+
+
+COMMUNITY_PIPELINES_URL = (
+ "https://raw.githubusercontent.com/huggingface/diffusers/{revision}/examples/community/{pipeline}.py"
+)
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def get_diffusers_versions():
+ url = "https://pypi.org/pypi/diffusers/json"
+ releases = json.loads(request.urlopen(url).read())["releases"].keys()
+ return sorted(releases, key=lambda x: version.Version(x))
+
+
+def init_hf_modules():
+ """
+ Creates the cache directory for modules with an init, and adds it to the Python path.
+ """
+ # This function has already been executed if HF_MODULES_CACHE already is in the Python path.
+ if HF_MODULES_CACHE in sys.path:
+ return
+
+ sys.path.append(HF_MODULES_CACHE)
+ os.makedirs(HF_MODULES_CACHE, exist_ok=True)
+ init_path = Path(HF_MODULES_CACHE) / "__init__.py"
+ if not init_path.exists():
+ init_path.touch()
+
+
+def create_dynamic_module(name: Union[str, os.PathLike]):
+ """
+ Creates a dynamic module in the cache directory for modules.
+ """
+ init_hf_modules()
+ dynamic_module_path = Path(HF_MODULES_CACHE) / name
+ # If the parent module does not exist yet, recursively create it.
+ if not dynamic_module_path.parent.exists():
+ create_dynamic_module(dynamic_module_path.parent)
+ os.makedirs(dynamic_module_path, exist_ok=True)
+ init_path = dynamic_module_path / "__init__.py"
+ if not init_path.exists():
+ init_path.touch()
+
+
+def get_relative_imports(module_file):
+ """
+ Get the list of modules that are relatively imported in a module file.
+
+ Args:
+ module_file (`str` or `os.PathLike`): The module file to inspect.
+ """
+ with open(module_file, "r", encoding="utf-8") as f:
+ content = f.read()
+
+ # Imports of the form `import .xxx`
+ relative_imports = re.findall(r"^\s*import\s+\.(\S+)\s*$", content, flags=re.MULTILINE)
+ # Imports of the form `from .xxx import yyy`
+ relative_imports += re.findall(r"^\s*from\s+\.(\S+)\s+import", content, flags=re.MULTILINE)
+ # Unique-ify
+ return list(set(relative_imports))
+
+
+def get_relative_import_files(module_file):
+ """
+ Get the list of all files that are needed for a given module. Note that this function recurses through the relative
+ imports (if a imports b and b imports c, it will return module files for b and c).
+
+ Args:
+ module_file (`str` or `os.PathLike`): The module file to inspect.
+ """
+ no_change = False
+ files_to_check = [module_file]
+ all_relative_imports = []
+
+ # Let's recurse through all relative imports
+ while not no_change:
+ new_imports = []
+ for f in files_to_check:
+ new_imports.extend(get_relative_imports(f))
+
+ module_path = Path(module_file).parent
+ new_import_files = [str(module_path / m) for m in new_imports]
+ new_import_files = [f for f in new_import_files if f not in all_relative_imports]
+ files_to_check = [f"{f}.py" for f in new_import_files]
+
+ no_change = len(new_import_files) == 0
+ all_relative_imports.extend(files_to_check)
+
+ return all_relative_imports
+
+
+def check_imports(filename):
+ """
+ Check if the current Python environment contains all the libraries that are imported in a file.
+ """
+ with open(filename, "r", encoding="utf-8") as f:
+ content = f.read()
+
+ # Imports of the form `import xxx`
+ imports = re.findall(r"^\s*import\s+(\S+)\s*$", content, flags=re.MULTILINE)
+ # Imports of the form `from xxx import yyy`
+ imports += re.findall(r"^\s*from\s+(\S+)\s+import", content, flags=re.MULTILINE)
+ # Only keep the top-level module
+ imports = [imp.split(".")[0] for imp in imports if not imp.startswith(".")]
+
+ # Unique-ify and test we got them all
+ imports = list(set(imports))
+ missing_packages = []
+ for imp in imports:
+ try:
+ importlib.import_module(imp)
+ except ImportError:
+ missing_packages.append(imp)
+
+ if len(missing_packages) > 0:
+ raise ImportError(
+ "This modeling file requires the following packages that were not found in your environment: "
+ f"{', '.join(missing_packages)}. Run `pip install {' '.join(missing_packages)}`"
+ )
+
+ return get_relative_imports(filename)
+
+
+def get_class_in_module(class_name, module_path):
+ """
+ Import a module on the cache directory for modules and extract a class from it.
+ """
+ module_path = module_path.replace(os.path.sep, ".")
+ module = importlib.import_module(module_path)
+
+ if class_name is None:
+ return find_pipeline_class(module)
+ return getattr(module, class_name)
+
+
+def find_pipeline_class(loaded_module):
+ """
+ Retrieve pipeline class that inherits from `DiffusionPipeline`. Note that there has to be exactly one class
+ inheriting from `DiffusionPipeline`.
+ """
+ from ..pipelines import DiffusionPipeline
+
+ cls_members = dict(inspect.getmembers(loaded_module, inspect.isclass))
+
+ pipeline_class = None
+ for cls_name, cls in cls_members.items():
+ if (
+ cls_name != DiffusionPipeline.__name__
+ and issubclass(cls, DiffusionPipeline)
+ and cls.__module__.split(".")[0] != "diffusers"
+ ):
+ if pipeline_class is not None:
+ raise ValueError(
+ f"Multiple classes that inherit from {DiffusionPipeline.__name__} have been found:"
+ f" {pipeline_class.__name__}, and {cls_name}. Please make sure to define only one in"
+ f" {loaded_module}."
+ )
+ pipeline_class = cls
+
+ return pipeline_class
+
+
+@validate_hf_hub_args
+def get_cached_module_file(
+ pretrained_model_name_or_path: Union[str, os.PathLike],
+ module_file: str,
+ cache_dir: Optional[Union[str, os.PathLike]] = None,
+ force_download: bool = False,
+ resume_download: bool = False,
+ proxies: Optional[Dict[str, str]] = None,
+ token: Optional[Union[bool, str]] = None,
+ revision: Optional[str] = None,
+ local_files_only: bool = False,
+):
+ """
+ Prepares Downloads a module from a local folder or a distant repo and returns its path inside the cached
+ Transformers module.
+
+ Args:
+ pretrained_model_name_or_path (`str` or `os.PathLike`):
+ This can be either:
+
+ - a string, the *model id* of a pretrained model configuration hosted inside a model repo on
+ huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced
+ under a user or organization name, like `dbmdz/bert-base-german-cased`.
+ - a path to a *directory* containing a configuration file saved using the
+ [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
+
+ module_file (`str`):
+ The name of the module file containing the class to look for.
+ cache_dir (`str` or `os.PathLike`, *optional*):
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
+ cache should not be used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force to (re-)download the configuration files and override the cached versions if they
+ exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to delete incompletely received file. Attempts to resume the download if such a file exists.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+ token (`str` or *bool*, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+ when running `transformers-cli login` (stored in `~/.huggingface`).
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+ git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+ identifier allowed by git.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ If `True`, will only try to load the tokenizer configuration from local files.
+
+
+
+ You may pass a token in `token` if you are not logged in (`huggingface-cli login`) and want to use private
+ or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models).
+
+
+
+ Returns:
+ `str`: The path to the module inside the cache.
+ """
+ # Download and cache module_file from the repo `pretrained_model_name_or_path` of grab it if it's a local file.
+ pretrained_model_name_or_path = str(pretrained_model_name_or_path)
+
+ module_file_or_url = os.path.join(pretrained_model_name_or_path, module_file)
+
+ if os.path.isfile(module_file_or_url):
+ resolved_module_file = module_file_or_url
+ submodule = "local"
+ elif pretrained_model_name_or_path.count("/") == 0:
+ available_versions = get_diffusers_versions()
+ # cut ".dev0"
+ latest_version = "v" + ".".join(__version__.split(".")[:3])
+
+ # retrieve github version that matches
+ if revision is None:
+ revision = latest_version if latest_version[1:] in available_versions else "main"
+ logger.info(f"Defaulting to latest_version: {revision}.")
+ elif revision in available_versions:
+ revision = f"v{revision}"
+ elif revision == "main":
+ revision = revision
+ else:
+ raise ValueError(
+ f"`custom_revision`: {revision} does not exist. Please make sure to choose one of"
+ f" {', '.join(available_versions + ['main'])}."
+ )
+
+ # community pipeline on GitHub
+ github_url = COMMUNITY_PIPELINES_URL.format(revision=revision, pipeline=pretrained_model_name_or_path)
+ try:
+ resolved_module_file = cached_download(
+ github_url,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ proxies=proxies,
+ resume_download=resume_download,
+ local_files_only=local_files_only,
+ token=False,
+ )
+ submodule = "git"
+ module_file = pretrained_model_name_or_path + ".py"
+ except EnvironmentError:
+ logger.error(f"Could not locate the {module_file} inside {pretrained_model_name_or_path}.")
+ raise
+ else:
+ try:
+ # Load from URL or cache if already cached
+ resolved_module_file = hf_hub_download(
+ pretrained_model_name_or_path,
+ module_file,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ proxies=proxies,
+ resume_download=resume_download,
+ local_files_only=local_files_only,
+ token=token,
+ )
+ submodule = os.path.join("local", "--".join(pretrained_model_name_or_path.split("/")))
+ except EnvironmentError:
+ logger.error(f"Could not locate the {module_file} inside {pretrained_model_name_or_path}.")
+ raise
+
+ # Check we have all the requirements in our environment
+ modules_needed = check_imports(resolved_module_file)
+
+ # Now we move the module inside our cached dynamic modules.
+ full_submodule = DIFFUSERS_DYNAMIC_MODULE_NAME + os.path.sep + submodule
+ create_dynamic_module(full_submodule)
+ submodule_path = Path(HF_MODULES_CACHE) / full_submodule
+ if submodule == "local" or submodule == "git":
+ # We always copy local files (we could hash the file to see if there was a change, and give them the name of
+ # that hash, to only copy when there is a modification but it seems overkill for now).
+ # The only reason we do the copy is to avoid putting too many folders in sys.path.
+ shutil.copy(resolved_module_file, submodule_path / module_file)
+ for module_needed in modules_needed:
+ module_needed = f"{module_needed}.py"
+ shutil.copy(os.path.join(pretrained_model_name_or_path, module_needed), submodule_path / module_needed)
+ else:
+ # Get the commit hash
+ # TODO: we will get this info in the etag soon, so retrieve it from there and not here.
+ commit_hash = model_info(pretrained_model_name_or_path, revision=revision, token=token).sha
+
+ # The module file will end up being placed in a subfolder with the git hash of the repo. This way we get the
+ # benefit of versioning.
+ submodule_path = submodule_path / commit_hash
+ full_submodule = full_submodule + os.path.sep + commit_hash
+ create_dynamic_module(full_submodule)
+
+ if not (submodule_path / module_file).exists():
+ shutil.copy(resolved_module_file, submodule_path / module_file)
+ # Make sure we also have every file with relative
+ for module_needed in modules_needed:
+ if not (submodule_path / module_needed).exists():
+ get_cached_module_file(
+ pretrained_model_name_or_path,
+ f"{module_needed}.py",
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ token=token,
+ revision=revision,
+ local_files_only=local_files_only,
+ )
+ return os.path.join(full_submodule, module_file)
+
+
+@validate_hf_hub_args
+def get_class_from_dynamic_module(
+ pretrained_model_name_or_path: Union[str, os.PathLike],
+ module_file: str,
+ class_name: Optional[str] = None,
+ cache_dir: Optional[Union[str, os.PathLike]] = None,
+ force_download: bool = False,
+ resume_download: bool = False,
+ proxies: Optional[Dict[str, str]] = None,
+ token: Optional[Union[bool, str]] = None,
+ revision: Optional[str] = None,
+ local_files_only: bool = False,
+ **kwargs,
+):
+ """
+ Extracts a class from a module file, present in the local folder or repository of a model.
+
+
+
+ Calling this function will execute the code in the module file found locally or downloaded from the Hub. It should
+ therefore only be called on trusted repos.
+
+
+
+ Args:
+ pretrained_model_name_or_path (`str` or `os.PathLike`):
+ This can be either:
+
+ - a string, the *model id* of a pretrained model configuration hosted inside a model repo on
+ huggingface.co. Valid model ids can be located at the root-level, like `bert-base-uncased`, or namespaced
+ under a user or organization name, like `dbmdz/bert-base-german-cased`.
+ - a path to a *directory* containing a configuration file saved using the
+ [`~PreTrainedTokenizer.save_pretrained`] method, e.g., `./my_model_directory/`.
+
+ module_file (`str`):
+ The name of the module file containing the class to look for.
+ class_name (`str`):
+ The name of the class to import in the module.
+ cache_dir (`str` or `os.PathLike`, *optional*):
+ Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
+ cache should not be used.
+ force_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to force to (re-)download the configuration files and override the cached versions if they
+ exist.
+ resume_download (`bool`, *optional*, defaults to `False`):
+ Whether or not to delete incompletely received file. Attempts to resume the download if such a file exists.
+ proxies (`Dict[str, str]`, *optional*):
+ A dictionary of proxy servers to use by protocol or endpoint, e.g., `{'http': 'foo.bar:3128',
+ 'http://hostname': 'foo.bar:4012'}.` The proxies are used on each request.
+ token (`str` or `bool`, *optional*):
+ The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+ when running `transformers-cli login` (stored in `~/.huggingface`).
+ revision (`str`, *optional*, defaults to `"main"`):
+ The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+ git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+ identifier allowed by git.
+ local_files_only (`bool`, *optional*, defaults to `False`):
+ If `True`, will only try to load the tokenizer configuration from local files.
+
+
+
+ You may pass a token in `token` if you are not logged in (`huggingface-cli login`) and want to use private
+ or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models).
+
+
+
+ Returns:
+ `type`: The class, dynamically imported from the module.
+
+ Examples:
+
+ ```python
+ # Download module `modeling.py` from huggingface.co and cache then extract the class `MyBertModel` from this
+ # module.
+ cls = get_class_from_dynamic_module("sgugger/my-bert-model", "modeling.py", "MyBertModel")
+ ```"""
+ # And lastly we get the class inside our newly created module
+ final_module = get_cached_module_file(
+ pretrained_model_name_or_path,
+ module_file,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ resume_download=resume_download,
+ proxies=proxies,
+ token=token,
+ revision=revision,
+ local_files_only=local_files_only,
+ )
+ return get_class_in_module(class_name, final_module.replace(".py", ""))
diff --git a/model/diffusers_c/utils/export_utils.py b/model/diffusers_c/utils/export_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..24cce07ab8663a1416c145e89aaaf319e6806635
--- /dev/null
+++ b/model/diffusers_c/utils/export_utils.py
@@ -0,0 +1,140 @@
+import io
+import random
+import struct
+import tempfile
+from contextlib import contextmanager
+from typing import List, Union
+
+import numpy as np
+import PIL.Image
+import PIL.ImageOps
+
+from .import_utils import (
+ BACKENDS_MAPPING,
+ is_opencv_available,
+)
+from .logging import get_logger
+
+
+global_rng = random.Random()
+
+logger = get_logger(__name__)
+
+
+@contextmanager
+def buffered_writer(raw_f):
+ f = io.BufferedWriter(raw_f)
+ yield f
+ f.flush()
+
+
+def export_to_gif(image: List[PIL.Image.Image], output_gif_path: str = None, fps: int = 10) -> str:
+ if output_gif_path is None:
+ output_gif_path = tempfile.NamedTemporaryFile(suffix=".gif").name
+
+ image[0].save(
+ output_gif_path,
+ save_all=True,
+ append_images=image[1:],
+ optimize=False,
+ duration=1000 // fps,
+ loop=0,
+ )
+ return output_gif_path
+
+
+def export_to_ply(mesh, output_ply_path: str = None):
+ """
+ Write a PLY file for a mesh.
+ """
+ if output_ply_path is None:
+ output_ply_path = tempfile.NamedTemporaryFile(suffix=".ply").name
+
+ coords = mesh.verts.detach().cpu().numpy()
+ faces = mesh.faces.cpu().numpy()
+ rgb = np.stack([mesh.vertex_channels[x].detach().cpu().numpy() for x in "RGB"], axis=1)
+
+ with buffered_writer(open(output_ply_path, "wb")) as f:
+ f.write(b"ply\n")
+ f.write(b"format binary_little_endian 1.0\n")
+ f.write(bytes(f"element vertex {len(coords)}\n", "ascii"))
+ f.write(b"property float x\n")
+ f.write(b"property float y\n")
+ f.write(b"property float z\n")
+ if rgb is not None:
+ f.write(b"property uchar red\n")
+ f.write(b"property uchar green\n")
+ f.write(b"property uchar blue\n")
+ if faces is not None:
+ f.write(bytes(f"element face {len(faces)}\n", "ascii"))
+ f.write(b"property list uchar int vertex_index\n")
+ f.write(b"end_header\n")
+
+ if rgb is not None:
+ rgb = (rgb * 255.499).round().astype(int)
+ vertices = [
+ (*coord, *rgb)
+ for coord, rgb in zip(
+ coords.tolist(),
+ rgb.tolist(),
+ )
+ ]
+ format = struct.Struct("<3f3B")
+ for item in vertices:
+ f.write(format.pack(*item))
+ else:
+ format = struct.Struct("<3f")
+ for vertex in coords.tolist():
+ f.write(format.pack(*vertex))
+
+ if faces is not None:
+ format = struct.Struct(" str:
+ if is_opencv_available():
+ import cv2
+ else:
+ raise ImportError(BACKENDS_MAPPING["opencv"][1].format("export_to_video"))
+ if output_video_path is None:
+ output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name
+
+ if isinstance(video_frames[0], np.ndarray):
+ video_frames = [(frame * 255).astype(np.uint8) for frame in video_frames]
+
+ elif isinstance(video_frames[0], PIL.Image.Image):
+ video_frames = [np.array(frame) for frame in video_frames]
+
+ fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+ h, w, c = video_frames[0].shape
+ video_writer = cv2.VideoWriter(output_video_path, fourcc, fps=fps, frameSize=(w, h))
+ for i in range(len(video_frames)):
+ img = cv2.cvtColor(video_frames[i], cv2.COLOR_RGB2BGR)
+ video_writer.write(img)
+ return output_video_path
diff --git a/model/diffusers_c/utils/hub_utils.py b/model/diffusers_c/utils/hub_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..666ac3406759476eff92a8e458788d6fe6ceb26c
--- /dev/null
+++ b/model/diffusers_c/utils/hub_utils.py
@@ -0,0 +1,492 @@
+# Copyright 2024 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import os
+import re
+import sys
+import tempfile
+import traceback
+import warnings
+from pathlib import Path
+from typing import Dict, List, Optional, Union
+from uuid import uuid4
+
+from huggingface_hub import (
+ ModelCard,
+ ModelCardData,
+ create_repo,
+ hf_hub_download,
+ upload_folder,
+)
+from huggingface_hub.constants import HF_HUB_CACHE, HF_HUB_DISABLE_TELEMETRY, HF_HUB_OFFLINE
+from huggingface_hub.file_download import REGEX_COMMIT_HASH
+from huggingface_hub.utils import (
+ EntryNotFoundError,
+ RepositoryNotFoundError,
+ RevisionNotFoundError,
+ is_jinja_available,
+ validate_hf_hub_args,
+)
+from packaging import version
+from requests import HTTPError
+
+from .. import __version__
+from .constants import (
+ DEPRECATED_REVISION_ARGS,
+ HUGGINGFACE_CO_RESOLVE_ENDPOINT,
+ SAFETENSORS_WEIGHTS_NAME,
+ WEIGHTS_NAME,
+)
+from .import_utils import (
+ ENV_VARS_TRUE_VALUES,
+ _flax_version,
+ _jax_version,
+ _onnxruntime_version,
+ _torch_version,
+ is_flax_available,
+ is_onnx_available,
+ is_torch_available,
+)
+from .logging import get_logger
+
+
+logger = get_logger(__name__)
+
+MODEL_CARD_TEMPLATE_PATH = Path(__file__).parent / "model_card_template.md"
+SESSION_ID = uuid4().hex
+
+
+def http_user_agent(user_agent: Union[Dict, str, None] = None) -> str:
+ """
+ Formats a user-agent string with basic info about a request.
+ """
+ ua = f"diffusers/{__version__}; python/{sys.version.split()[0]}; session_id/{SESSION_ID}"
+ if HF_HUB_DISABLE_TELEMETRY or HF_HUB_OFFLINE:
+ return ua + "; telemetry/off"
+ if is_torch_available():
+ ua += f"; torch/{_torch_version}"
+ if is_flax_available():
+ ua += f"; jax/{_jax_version}"
+ ua += f"; flax/{_flax_version}"
+ if is_onnx_available():
+ ua += f"; onnxruntime/{_onnxruntime_version}"
+ # CI will set this value to True
+ if os.environ.get("DIFFUSERS_IS_CI", "").upper() in ENV_VARS_TRUE_VALUES:
+ ua += "; is_ci/true"
+ if isinstance(user_agent, dict):
+ ua += "; " + "; ".join(f"{k}/{v}" for k, v in user_agent.items())
+ elif isinstance(user_agent, str):
+ ua += "; " + user_agent
+ return ua
+
+
+def load_or_create_model_card(
+ repo_id_or_path: str = None,
+ token: Optional[str] = None,
+ is_pipeline: bool = False,
+ from_training: bool = False,
+ model_description: Optional[str] = None,
+ base_model: str = None,
+ prompt: Optional[str] = None,
+ license: Optional[str] = None,
+ widget: Optional[List[dict]] = None,
+ inference: Optional[bool] = None,
+) -> ModelCard:
+ """
+ Loads or creates a model card.
+
+ Args:
+ repo_id_or_path (`str`):
+ The repo id (e.g., "runwayml/stable-diffusion-v1-5") or local path where to look for the model card.
+ token (`str`, *optional*):
+ Authentication token. Will default to the stored token. See https://huggingface.co/settings/token for more details.
+ is_pipeline (`bool`):
+ Boolean to indicate if we're adding tag to a [`DiffusionPipeline`].
+ from_training: (`bool`): Boolean flag to denote if the model card is being created from a training script.
+ model_description (`str`, *optional*): Model description to add to the model card. Helpful when using
+ `load_or_create_model_card` from a training script.
+ base_model (`str`): Base model identifier (e.g., "stabilityai/stable-diffusion-xl-base-1.0"). Useful
+ for DreamBooth-like training.
+ prompt (`str`, *optional*): Prompt used for training. Useful for DreamBooth-like training.
+ license: (`str`, *optional*): License of the output artifact. Helpful when using
+ `load_or_create_model_card` from a training script.
+ widget (`List[dict]`, *optional*): Widget to accompany a gallery template.
+ inference: (`bool`, optional): Whether to turn on inference widget. Helpful when using
+ `load_or_create_model_card` from a training script.
+ """
+ if not is_jinja_available():
+ raise ValueError(
+ "Modelcard rendering is based on Jinja templates."
+ " Please make sure to have `jinja` installed before using `load_or_create_model_card`."
+ " To install it, please run `pip install Jinja2`."
+ )
+
+ try:
+ # Check if the model card is present on the remote repo
+ model_card = ModelCard.load(repo_id_or_path, token=token)
+ except (EntryNotFoundError, RepositoryNotFoundError):
+ # Otherwise create a model card from template
+ if from_training:
+ model_card = ModelCard.from_template(
+ card_data=ModelCardData( # Card metadata object that will be converted to YAML block
+ license=license,
+ library_name="diffusers",
+ inference=inference,
+ base_model=base_model,
+ instance_prompt=prompt,
+ widget=widget,
+ ),
+ template_path=MODEL_CARD_TEMPLATE_PATH,
+ model_description=model_description,
+ )
+ else:
+ card_data = ModelCardData()
+ component = "pipeline" if is_pipeline else "model"
+ if model_description is None:
+ model_description = f"This is the model card of a 🧨 diffusers {component} that has been pushed on the Hub. This model card has been automatically generated."
+ model_card = ModelCard.from_template(card_data, model_description=model_description)
+
+ return model_card
+
+
+def populate_model_card(model_card: ModelCard, tags: Union[str, List[str]] = None) -> ModelCard:
+ """Populates the `model_card` with library name and optional tags."""
+ if model_card.data.library_name is None:
+ model_card.data.library_name = "diffusers"
+
+ if tags is not None:
+ if isinstance(tags, str):
+ tags = [tags]
+ if model_card.data.tags is None:
+ model_card.data.tags = []
+ for tag in tags:
+ model_card.data.tags.append(tag)
+
+ return model_card
+
+
+def extract_commit_hash(resolved_file: Optional[str], commit_hash: Optional[str] = None):
+ """
+ Extracts the commit hash from a resolved filename toward a cache file.
+ """
+ if resolved_file is None or commit_hash is not None:
+ return commit_hash
+ resolved_file = str(Path(resolved_file).as_posix())
+ search = re.search(r"snapshots/([^/]+)/", resolved_file)
+ if search is None:
+ return None
+ commit_hash = search.groups()[0]
+ return commit_hash if REGEX_COMMIT_HASH.match(commit_hash) else None
+
+
+# Old default cache path, potentially to be migrated.
+# This logic was more or less taken from `transformers`, with the following differences:
+# - Diffusers doesn't use custom environment variables to specify the cache path.
+# - There is no need to migrate the cache format, just move the files to the new location.
+hf_cache_home = os.path.expanduser(
+ os.getenv("HF_HOME", os.path.join(os.getenv("XDG_CACHE_HOME", "~/.cache"), "huggingface"))
+)
+old_diffusers_cache = os.path.join(hf_cache_home, "diffusers")
+
+
+def move_cache(old_cache_dir: Optional[str] = None, new_cache_dir: Optional[str] = None) -> None:
+ if new_cache_dir is None:
+ new_cache_dir = HF_HUB_CACHE
+ if old_cache_dir is None:
+ old_cache_dir = old_diffusers_cache
+
+ old_cache_dir = Path(old_cache_dir).expanduser()
+ new_cache_dir = Path(new_cache_dir).expanduser()
+ for old_blob_path in old_cache_dir.glob("**/blobs/*"):
+ if old_blob_path.is_file() and not old_blob_path.is_symlink():
+ new_blob_path = new_cache_dir / old_blob_path.relative_to(old_cache_dir)
+ new_blob_path.parent.mkdir(parents=True, exist_ok=True)
+ os.replace(old_blob_path, new_blob_path)
+ try:
+ os.symlink(new_blob_path, old_blob_path)
+ except OSError:
+ logger.warning(
+ "Could not create symlink between old cache and new cache. If you use an older version of diffusers again, files will be re-downloaded."
+ )
+ # At this point, old_cache_dir contains symlinks to the new cache (it can still be used).
+
+
+cache_version_file = os.path.join(HF_HUB_CACHE, "version_diffusers_cache.txt")
+if not os.path.isfile(cache_version_file):
+ cache_version = 0
+else:
+ with open(cache_version_file) as f:
+ try:
+ cache_version = int(f.read())
+ except ValueError:
+ cache_version = 0
+
+if cache_version < 1:
+ old_cache_is_not_empty = os.path.isdir(old_diffusers_cache) and len(os.listdir(old_diffusers_cache)) > 0
+ if old_cache_is_not_empty:
+ logger.warning(
+ "The cache for model files in Diffusers v0.14.0 has moved to a new location. Moving your "
+ "existing cached models. This is a one-time operation, you can interrupt it or run it "
+ "later by calling `diffusers.utils.hub_utils.move_cache()`."
+ )
+ try:
+ move_cache()
+ except Exception as e:
+ trace = "\n".join(traceback.format_tb(e.__traceback__))
+ logger.error(
+ f"There was a problem when trying to move your cache:\n\n{trace}\n{e.__class__.__name__}: {e}\n\nPlease "
+ "file an issue at https://github.com/huggingface/diffusers/issues/new/choose, copy paste this whole "
+ "message and we will do our best to help."
+ )
+
+if cache_version < 1:
+ try:
+ os.makedirs(HF_HUB_CACHE, exist_ok=True)
+ with open(cache_version_file, "w") as f:
+ f.write("1")
+ except Exception:
+ logger.warning(
+ f"There was a problem when trying to write in your cache folder ({HF_HUB_CACHE}). Please, ensure "
+ "the directory exists and can be written to."
+ )
+
+
+def _add_variant(weights_name: str, variant: Optional[str] = None) -> str:
+ if variant is not None:
+ splits = weights_name.split(".")
+ splits = splits[:-1] + [variant] + splits[-1:]
+ weights_name = ".".join(splits)
+
+ return weights_name
+
+
+@validate_hf_hub_args
+def _get_model_file(
+ pretrained_model_name_or_path: Union[str, Path],
+ *,
+ weights_name: str,
+ subfolder: Optional[str] = None,
+ cache_dir: Optional[str] = None,
+ force_download: bool = False,
+ proxies: Optional[Dict] = None,
+ resume_download: bool = False,
+ local_files_only: bool = False,
+ token: Optional[str] = None,
+ user_agent: Optional[Union[Dict, str]] = None,
+ revision: Optional[str] = None,
+ commit_hash: Optional[str] = None,
+):
+ pretrained_model_name_or_path = str(pretrained_model_name_or_path)
+ if os.path.isfile(pretrained_model_name_or_path):
+ return pretrained_model_name_or_path
+ elif os.path.isdir(pretrained_model_name_or_path):
+ if os.path.isfile(os.path.join(pretrained_model_name_or_path, weights_name)):
+ # Load from a PyTorch checkpoint
+ model_file = os.path.join(pretrained_model_name_or_path, weights_name)
+ return model_file
+ elif subfolder is not None and os.path.isfile(
+ os.path.join(pretrained_model_name_or_path, subfolder, weights_name)
+ ):
+ model_file = os.path.join(pretrained_model_name_or_path, subfolder, weights_name)
+ return model_file
+ else:
+ raise EnvironmentError(
+ f"Error no file named {weights_name} found in directory {pretrained_model_name_or_path}."
+ )
+ else:
+ # 1. First check if deprecated way of loading from branches is used
+ if (
+ revision in DEPRECATED_REVISION_ARGS
+ and (weights_name == WEIGHTS_NAME or weights_name == SAFETENSORS_WEIGHTS_NAME)
+ and version.parse(version.parse(__version__).base_version) >= version.parse("0.22.0")
+ ):
+ try:
+ model_file = hf_hub_download(
+ pretrained_model_name_or_path,
+ filename=_add_variant(weights_name, revision),
+ cache_dir=cache_dir,
+ force_download=force_download,
+ proxies=proxies,
+ resume_download=resume_download,
+ local_files_only=local_files_only,
+ token=token,
+ user_agent=user_agent,
+ subfolder=subfolder,
+ revision=revision or commit_hash,
+ )
+ warnings.warn(
+ f"Loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'` is deprecated. Loading instead from `revision='main'` with `variant={revision}`. Loading model variants via `revision='{revision}'` will be removed in diffusers v1. Please use `variant='{revision}'` instead.",
+ FutureWarning,
+ )
+ return model_file
+ except: # noqa: E722
+ warnings.warn(
+ f"You are loading the variant {revision} from {pretrained_model_name_or_path} via `revision='{revision}'`. This behavior is deprecated and will be removed in diffusers v1. One should use `variant='{revision}'` instead. However, it appears that {pretrained_model_name_or_path} currently does not have a {_add_variant(weights_name, revision)} file in the 'main' branch of {pretrained_model_name_or_path}. \n The Diffusers team and community would be very grateful if you could open an issue: https://github.com/huggingface/diffusers/issues/new with the title '{pretrained_model_name_or_path} is missing {_add_variant(weights_name, revision)}' so that the correct variant file can be added.",
+ FutureWarning,
+ )
+ try:
+ # 2. Load model file as usual
+ model_file = hf_hub_download(
+ pretrained_model_name_or_path,
+ filename=weights_name,
+ cache_dir=cache_dir,
+ force_download=force_download,
+ proxies=proxies,
+ resume_download=resume_download,
+ local_files_only=local_files_only,
+ token=token,
+ user_agent=user_agent,
+ subfolder=subfolder,
+ revision=revision or commit_hash,
+ )
+ return model_file
+
+ except RepositoryNotFoundError:
+ raise EnvironmentError(
+ f"{pretrained_model_name_or_path} is not a local folder and is not a valid model identifier "
+ "listed on 'https://huggingface.co/models'\nIf this is a private repository, make sure to pass a "
+ "token having permission to this repo with `token` or log in with `huggingface-cli "
+ "login`."
+ )
+ except RevisionNotFoundError:
+ raise EnvironmentError(
+ f"{revision} is not a valid git identifier (branch name, tag name or commit id) that exists for "
+ "this model name. Check the model page at "
+ f"'https://huggingface.co/{pretrained_model_name_or_path}' for available revisions."
+ )
+ except EntryNotFoundError:
+ raise EnvironmentError(
+ f"{pretrained_model_name_or_path} does not appear to have a file named {weights_name}."
+ )
+ except HTTPError as err:
+ raise EnvironmentError(
+ f"There was a specific connection error when trying to load {pretrained_model_name_or_path}:\n{err}"
+ )
+ except ValueError:
+ raise EnvironmentError(
+ f"We couldn't connect to '{HUGGINGFACE_CO_RESOLVE_ENDPOINT}' to load this model, couldn't find it"
+ f" in the cached files and it looks like {pretrained_model_name_or_path} is not the path to a"
+ f" directory containing a file named {weights_name} or"
+ " \nCheckout your internet connection or see how to run the library in"
+ " offline mode at 'https://huggingface.co/docs/diffusers/installation#offline-mode'."
+ )
+ except EnvironmentError:
+ raise EnvironmentError(
+ f"Can't load the model for '{pretrained_model_name_or_path}'. If you were trying to load it from "
+ "'https://huggingface.co/models', make sure you don't have a local directory with the same name. "
+ f"Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a directory "
+ f"containing a file named {weights_name}"
+ )
+
+
+class PushToHubMixin:
+ """
+ A Mixin to push a model, scheduler, or pipeline to the Hugging Face Hub.
+ """
+
+ def _upload_folder(
+ self,
+ working_dir: Union[str, os.PathLike],
+ repo_id: str,
+ token: Optional[str] = None,
+ commit_message: Optional[str] = None,
+ create_pr: bool = False,
+ ):
+ """
+ Uploads all files in `working_dir` to `repo_id`.
+ """
+ if commit_message is None:
+ if "Model" in self.__class__.__name__:
+ commit_message = "Upload model"
+ elif "Scheduler" in self.__class__.__name__:
+ commit_message = "Upload scheduler"
+ else:
+ commit_message = f"Upload {self.__class__.__name__}"
+
+ logger.info(f"Uploading the files of {working_dir} to {repo_id}.")
+ return upload_folder(
+ repo_id=repo_id, folder_path=working_dir, token=token, commit_message=commit_message, create_pr=create_pr
+ )
+
+ def push_to_hub(
+ self,
+ repo_id: str,
+ commit_message: Optional[str] = None,
+ private: Optional[bool] = None,
+ token: Optional[str] = None,
+ create_pr: bool = False,
+ safe_serialization: bool = True,
+ variant: Optional[str] = None,
+ ) -> str:
+ """
+ Upload model, scheduler, or pipeline files to the 🤗 Hugging Face Hub.
+
+ Parameters:
+ repo_id (`str`):
+ The name of the repository you want to push your model, scheduler, or pipeline files to. It should
+ contain your organization name when pushing to an organization. `repo_id` can also be a path to a local
+ directory.
+ commit_message (`str`, *optional*):
+ Message to commit while pushing. Default to `"Upload {object}"`.
+ private (`bool`, *optional*):
+ Whether or not the repository created should be private.
+ token (`str`, *optional*):
+ The token to use as HTTP bearer authorization for remote files. The token generated when running
+ `huggingface-cli login` (stored in `~/.huggingface`).
+ create_pr (`bool`, *optional*, defaults to `False`):
+ Whether or not to create a PR with the uploaded files or directly commit.
+ safe_serialization (`bool`, *optional*, defaults to `True`):
+ Whether or not to convert the model weights to the `safetensors` format.
+ variant (`str`, *optional*):
+ If specified, weights are saved in the format `pytorch_model..bin`.
+
+ Examples:
+
+ ```python
+ from diffusers import UNet2DConditionModel
+
+ unet = UNet2DConditionModel.from_pretrained("stabilityai/stable-diffusion-2", subfolder="unet")
+
+ # Push the `unet` to your namespace with the name "my-finetuned-unet".
+ unet.push_to_hub("my-finetuned-unet")
+
+ # Push the `unet` to an organization with the name "my-finetuned-unet".
+ unet.push_to_hub("your-org/my-finetuned-unet")
+ ```
+ """
+ repo_id = create_repo(repo_id, private=private, token=token, exist_ok=True).repo_id
+
+ # Create a new empty model card and eventually tag it
+ model_card = load_or_create_model_card(repo_id, token=token)
+ model_card = populate_model_card(model_card)
+
+ # Save all files.
+ save_kwargs = {"safe_serialization": safe_serialization}
+ if "Scheduler" not in self.__class__.__name__:
+ save_kwargs.update({"variant": variant})
+
+ with tempfile.TemporaryDirectory() as tmpdir:
+ self.save_pretrained(tmpdir, **save_kwargs)
+
+ # Update model card if needed:
+ model_card.save(os.path.join(tmpdir, "README.md"))
+
+ return self._upload_folder(
+ tmpdir,
+ repo_id,
+ token=token,
+ commit_message=commit_message,
+ create_pr=create_pr,
+ )
diff --git a/model/diffusers_c/utils/import_utils.py b/model/diffusers_c/utils/import_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..a3ee31c91c8c66deb73dda05e4a064310ccca2ea
--- /dev/null
+++ b/model/diffusers_c/utils/import_utils.py
@@ -0,0 +1,726 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Import utilities: Utilities related to imports and our lazy inits.
+"""
+
+import importlib.util
+import operator as op
+import os
+import sys
+from collections import OrderedDict
+from itertools import chain
+from types import ModuleType
+from typing import Any, Union
+
+from huggingface_hub.utils import is_jinja_available # noqa: F401
+from packaging import version
+from packaging.version import Version, parse
+
+from . import logging
+
+
+# The package importlib_metadata is in a different place, depending on the python version.
+if sys.version_info < (3, 8):
+ import importlib_metadata
+else:
+ import importlib.metadata as importlib_metadata
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+ENV_VARS_TRUE_VALUES = {"1", "ON", "YES", "TRUE"}
+ENV_VARS_TRUE_AND_AUTO_VALUES = ENV_VARS_TRUE_VALUES.union({"AUTO"})
+
+USE_TF = os.environ.get("USE_TF", "AUTO").upper()
+USE_TORCH = os.environ.get("USE_TORCH", "AUTO").upper()
+USE_JAX = os.environ.get("USE_FLAX", "AUTO").upper()
+USE_SAFETENSORS = os.environ.get("USE_SAFETENSORS", "AUTO").upper()
+DIFFUSERS_SLOW_IMPORT = os.environ.get("DIFFUSERS_SLOW_IMPORT", "FALSE").upper()
+DIFFUSERS_SLOW_IMPORT = DIFFUSERS_SLOW_IMPORT in ENV_VARS_TRUE_VALUES
+
+STR_OPERATION_TO_FUNC = {">": op.gt, ">=": op.ge, "==": op.eq, "!=": op.ne, "<=": op.le, "<": op.lt}
+
+_torch_version = "N/A"
+if USE_TORCH in ENV_VARS_TRUE_AND_AUTO_VALUES and USE_TF not in ENV_VARS_TRUE_VALUES:
+ _torch_available = importlib.util.find_spec("torch") is not None
+ if _torch_available:
+ try:
+ _torch_version = importlib_metadata.version("torch")
+ logger.info(f"PyTorch version {_torch_version} available.")
+ except importlib_metadata.PackageNotFoundError:
+ _torch_available = False
+else:
+ logger.info("Disabling PyTorch because USE_TORCH is set")
+ _torch_available = False
+
+_torch_xla_available = importlib.util.find_spec("torch_xla") is not None
+if _torch_xla_available:
+ try:
+ _torch_xla_version = importlib_metadata.version("torch_xla")
+ logger.info(f"PyTorch XLA version {_torch_xla_version} available.")
+ except ImportError:
+ _torch_xla_available = False
+
+# check whether torch_npu is available
+_torch_npu_available = importlib.util.find_spec("torch_npu") is not None
+if _torch_npu_available:
+ try:
+ _torch_npu_version = importlib_metadata.version("torch_npu")
+ logger.info(f"torch_npu version {_torch_npu_version} available.")
+ except ImportError:
+ _torch_npu_available = False
+
+_jax_version = "N/A"
+_flax_version = "N/A"
+if USE_JAX in ENV_VARS_TRUE_AND_AUTO_VALUES:
+ _flax_available = importlib.util.find_spec("jax") is not None and importlib.util.find_spec("flax") is not None
+ if _flax_available:
+ try:
+ _jax_version = importlib_metadata.version("jax")
+ _flax_version = importlib_metadata.version("flax")
+ logger.info(f"JAX version {_jax_version}, Flax version {_flax_version} available.")
+ except importlib_metadata.PackageNotFoundError:
+ _flax_available = False
+else:
+ _flax_available = False
+
+if USE_SAFETENSORS in ENV_VARS_TRUE_AND_AUTO_VALUES:
+ _safetensors_available = importlib.util.find_spec("safetensors") is not None
+ if _safetensors_available:
+ try:
+ _safetensors_version = importlib_metadata.version("safetensors")
+ logger.info(f"Safetensors version {_safetensors_version} available.")
+ except importlib_metadata.PackageNotFoundError:
+ _safetensors_available = False
+else:
+ logger.info("Disabling Safetensors because USE_TF is set")
+ _safetensors_available = False
+
+_transformers_available = importlib.util.find_spec("transformers") is not None
+try:
+ _transformers_version = importlib_metadata.version("transformers")
+ logger.debug(f"Successfully imported transformers version {_transformers_version}")
+except importlib_metadata.PackageNotFoundError:
+ _transformers_available = False
+
+
+_inflect_available = importlib.util.find_spec("inflect") is not None
+try:
+ _inflect_version = importlib_metadata.version("inflect")
+ logger.debug(f"Successfully imported inflect version {_inflect_version}")
+except importlib_metadata.PackageNotFoundError:
+ _inflect_available = False
+
+
+_unidecode_available = importlib.util.find_spec("unidecode") is not None
+try:
+ _unidecode_version = importlib_metadata.version("unidecode")
+ logger.debug(f"Successfully imported unidecode version {_unidecode_version}")
+except importlib_metadata.PackageNotFoundError:
+ _unidecode_available = False
+
+
+_onnxruntime_version = "N/A"
+_onnx_available = importlib.util.find_spec("onnxruntime") is not None
+if _onnx_available:
+ candidates = (
+ "onnxruntime",
+ "onnxruntime-gpu",
+ "ort_nightly_gpu",
+ "onnxruntime-directml",
+ "onnxruntime-openvino",
+ "ort_nightly_directml",
+ "onnxruntime-rocm",
+ "onnxruntime-training",
+ )
+ _onnxruntime_version = None
+ # For the metadata, we have to look for both onnxruntime and onnxruntime-gpu
+ for pkg in candidates:
+ try:
+ _onnxruntime_version = importlib_metadata.version(pkg)
+ break
+ except importlib_metadata.PackageNotFoundError:
+ pass
+ _onnx_available = _onnxruntime_version is not None
+ if _onnx_available:
+ logger.debug(f"Successfully imported onnxruntime version {_onnxruntime_version}")
+
+# (sayakpaul): importlib.util.find_spec("opencv-python") returns None even when it's installed.
+# _opencv_available = importlib.util.find_spec("opencv-python") is not None
+try:
+ candidates = (
+ "opencv-python",
+ "opencv-contrib-python",
+ "opencv-python-headless",
+ "opencv-contrib-python-headless",
+ )
+ _opencv_version = None
+ for pkg in candidates:
+ try:
+ _opencv_version = importlib_metadata.version(pkg)
+ break
+ except importlib_metadata.PackageNotFoundError:
+ pass
+ _opencv_available = _opencv_version is not None
+ if _opencv_available:
+ logger.debug(f"Successfully imported cv2 version {_opencv_version}")
+except importlib_metadata.PackageNotFoundError:
+ _opencv_available = False
+
+_scipy_available = importlib.util.find_spec("scipy") is not None
+try:
+ _scipy_version = importlib_metadata.version("scipy")
+ logger.debug(f"Successfully imported scipy version {_scipy_version}")
+except importlib_metadata.PackageNotFoundError:
+ _scipy_available = False
+
+_librosa_available = importlib.util.find_spec("librosa") is not None
+try:
+ _librosa_version = importlib_metadata.version("librosa")
+ logger.debug(f"Successfully imported librosa version {_librosa_version}")
+except importlib_metadata.PackageNotFoundError:
+ _librosa_available = False
+
+_accelerate_available = importlib.util.find_spec("accelerate") is not None
+try:
+ _accelerate_version = importlib_metadata.version("accelerate")
+ logger.debug(f"Successfully imported accelerate version {_accelerate_version}")
+except importlib_metadata.PackageNotFoundError:
+ _accelerate_available = False
+
+_xformers_available = importlib.util.find_spec("xformers") is not None
+try:
+ _xformers_version = importlib_metadata.version("xformers")
+ if _torch_available:
+ _torch_version = importlib_metadata.version("torch")
+ if version.Version(_torch_version) < version.Version("1.12"):
+ raise ValueError("xformers is installed in your environment and requires PyTorch >= 1.12")
+
+ logger.debug(f"Successfully imported xformers version {_xformers_version}")
+except importlib_metadata.PackageNotFoundError:
+ _xformers_available = False
+
+_k_diffusion_available = importlib.util.find_spec("k_diffusion") is not None
+try:
+ _k_diffusion_version = importlib_metadata.version("k_diffusion")
+ logger.debug(f"Successfully imported k-diffusion version {_k_diffusion_version}")
+except importlib_metadata.PackageNotFoundError:
+ _k_diffusion_available = False
+
+_note_seq_available = importlib.util.find_spec("note_seq") is not None
+try:
+ _note_seq_version = importlib_metadata.version("note_seq")
+ logger.debug(f"Successfully imported note-seq version {_note_seq_version}")
+except importlib_metadata.PackageNotFoundError:
+ _note_seq_available = False
+
+_wandb_available = importlib.util.find_spec("wandb") is not None
+try:
+ _wandb_version = importlib_metadata.version("wandb")
+ logger.debug(f"Successfully imported wandb version {_wandb_version }")
+except importlib_metadata.PackageNotFoundError:
+ _wandb_available = False
+
+
+_tensorboard_available = importlib.util.find_spec("tensorboard")
+try:
+ _tensorboard_version = importlib_metadata.version("tensorboard")
+ logger.debug(f"Successfully imported tensorboard version {_tensorboard_version}")
+except importlib_metadata.PackageNotFoundError:
+ _tensorboard_available = False
+
+
+_compel_available = importlib.util.find_spec("compel")
+try:
+ _compel_version = importlib_metadata.version("compel")
+ logger.debug(f"Successfully imported compel version {_compel_version}")
+except importlib_metadata.PackageNotFoundError:
+ _compel_available = False
+
+
+_ftfy_available = importlib.util.find_spec("ftfy") is not None
+try:
+ _ftfy_version = importlib_metadata.version("ftfy")
+ logger.debug(f"Successfully imported ftfy version {_ftfy_version}")
+except importlib_metadata.PackageNotFoundError:
+ _ftfy_available = False
+
+
+_bs4_available = importlib.util.find_spec("bs4") is not None
+try:
+ # importlib metadata under different name
+ _bs4_version = importlib_metadata.version("beautifulsoup4")
+ logger.debug(f"Successfully imported ftfy version {_bs4_version}")
+except importlib_metadata.PackageNotFoundError:
+ _bs4_available = False
+
+_torchsde_available = importlib.util.find_spec("torchsde") is not None
+try:
+ _torchsde_version = importlib_metadata.version("torchsde")
+ logger.debug(f"Successfully imported torchsde version {_torchsde_version}")
+except importlib_metadata.PackageNotFoundError:
+ _torchsde_available = False
+
+_invisible_watermark_available = importlib.util.find_spec("imwatermark") is not None
+try:
+ _invisible_watermark_version = importlib_metadata.version("invisible-watermark")
+ logger.debug(f"Successfully imported invisible-watermark version {_invisible_watermark_version}")
+except importlib_metadata.PackageNotFoundError:
+ _invisible_watermark_available = False
+
+
+_peft_available = importlib.util.find_spec("peft") is not None
+try:
+ _peft_version = importlib_metadata.version("peft")
+ logger.debug(f"Successfully imported peft version {_peft_version}")
+except importlib_metadata.PackageNotFoundError:
+ _peft_available = False
+
+_torchvision_available = importlib.util.find_spec("torchvision") is not None
+try:
+ _torchvision_version = importlib_metadata.version("torchvision")
+ logger.debug(f"Successfully imported torchvision version {_torchvision_version}")
+except importlib_metadata.PackageNotFoundError:
+ _torchvision_available = False
+
+
+def is_torch_available():
+ return _torch_available
+
+
+def is_torch_xla_available():
+ return _torch_xla_available
+
+
+def is_torch_npu_available():
+ return _torch_npu_available
+
+
+def is_flax_available():
+ return _flax_available
+
+
+def is_transformers_available():
+ return _transformers_available
+
+
+def is_inflect_available():
+ return _inflect_available
+
+
+def is_unidecode_available():
+ return _unidecode_available
+
+
+def is_onnx_available():
+ return _onnx_available
+
+
+def is_opencv_available():
+ return _opencv_available
+
+
+def is_scipy_available():
+ return _scipy_available
+
+
+def is_librosa_available():
+ return _librosa_available
+
+
+def is_xformers_available():
+ return _xformers_available
+
+
+def is_accelerate_available():
+ return _accelerate_available
+
+
+def is_k_diffusion_available():
+ return _k_diffusion_available
+
+
+def is_note_seq_available():
+ return _note_seq_available
+
+
+def is_wandb_available():
+ return _wandb_available
+
+
+def is_tensorboard_available():
+ return _tensorboard_available
+
+
+def is_compel_available():
+ return _compel_available
+
+
+def is_ftfy_available():
+ return _ftfy_available
+
+
+def is_bs4_available():
+ return _bs4_available
+
+
+def is_torchsde_available():
+ return _torchsde_available
+
+
+def is_invisible_watermark_available():
+ return _invisible_watermark_available
+
+
+def is_peft_available():
+ return _peft_available
+
+
+def is_torchvision_available():
+ return _torchvision_available
+
+
+# docstyle-ignore
+FLAX_IMPORT_ERROR = """
+{0} requires the FLAX library but it was not found in your environment. Checkout the instructions on the
+installation page: https://github.com/google/flax and follow the ones that match your environment.
+"""
+
+# docstyle-ignore
+INFLECT_IMPORT_ERROR = """
+{0} requires the inflect library but it was not found in your environment. You can install it with pip: `pip install
+inflect`
+"""
+
+# docstyle-ignore
+PYTORCH_IMPORT_ERROR = """
+{0} requires the PyTorch library but it was not found in your environment. Checkout the instructions on the
+installation page: https://pytorch.org/get-started/locally/ and follow the ones that match your environment.
+"""
+
+# docstyle-ignore
+ONNX_IMPORT_ERROR = """
+{0} requires the onnxruntime library but it was not found in your environment. You can install it with pip: `pip
+install onnxruntime`
+"""
+
+# docstyle-ignore
+OPENCV_IMPORT_ERROR = """
+{0} requires the OpenCV library but it was not found in your environment. You can install it with pip: `pip
+install opencv-python`
+"""
+
+# docstyle-ignore
+SCIPY_IMPORT_ERROR = """
+{0} requires the scipy library but it was not found in your environment. You can install it with pip: `pip install
+scipy`
+"""
+
+# docstyle-ignore
+LIBROSA_IMPORT_ERROR = """
+{0} requires the librosa library but it was not found in your environment. Checkout the instructions on the
+installation page: https://librosa.org/doc/latest/install.html and follow the ones that match your environment.
+"""
+
+# docstyle-ignore
+TRANSFORMERS_IMPORT_ERROR = """
+{0} requires the transformers library but it was not found in your environment. You can install it with pip: `pip
+install transformers`
+"""
+
+# docstyle-ignore
+UNIDECODE_IMPORT_ERROR = """
+{0} requires the unidecode library but it was not found in your environment. You can install it with pip: `pip install
+Unidecode`
+"""
+
+# docstyle-ignore
+K_DIFFUSION_IMPORT_ERROR = """
+{0} requires the k-diffusion library but it was not found in your environment. You can install it with pip: `pip
+install k-diffusion`
+"""
+
+# docstyle-ignore
+NOTE_SEQ_IMPORT_ERROR = """
+{0} requires the note-seq library but it was not found in your environment. You can install it with pip: `pip
+install note-seq`
+"""
+
+# docstyle-ignore
+WANDB_IMPORT_ERROR = """
+{0} requires the wandb library but it was not found in your environment. You can install it with pip: `pip
+install wandb`
+"""
+
+# docstyle-ignore
+TENSORBOARD_IMPORT_ERROR = """
+{0} requires the tensorboard library but it was not found in your environment. You can install it with pip: `pip
+install tensorboard`
+"""
+
+
+# docstyle-ignore
+COMPEL_IMPORT_ERROR = """
+{0} requires the compel library but it was not found in your environment. You can install it with pip: `pip install compel`
+"""
+
+# docstyle-ignore
+BS4_IMPORT_ERROR = """
+{0} requires the Beautiful Soup library but it was not found in your environment. You can install it with pip:
+`pip install beautifulsoup4`. Please note that you may need to restart your runtime after installation.
+"""
+
+# docstyle-ignore
+FTFY_IMPORT_ERROR = """
+{0} requires the ftfy library but it was not found in your environment. Checkout the instructions on the
+installation section: https://github.com/rspeer/python-ftfy/tree/master#installing and follow the ones
+that match your environment. Please note that you may need to restart your runtime after installation.
+"""
+
+# docstyle-ignore
+TORCHSDE_IMPORT_ERROR = """
+{0} requires the torchsde library but it was not found in your environment. You can install it with pip: `pip install torchsde`
+"""
+
+# docstyle-ignore
+INVISIBLE_WATERMARK_IMPORT_ERROR = """
+{0} requires the invisible-watermark library but it was not found in your environment. You can install it with pip: `pip install invisible-watermark>=0.2.0`
+"""
+
+
+BACKENDS_MAPPING = OrderedDict(
+ [
+ ("bs4", (is_bs4_available, BS4_IMPORT_ERROR)),
+ ("flax", (is_flax_available, FLAX_IMPORT_ERROR)),
+ ("inflect", (is_inflect_available, INFLECT_IMPORT_ERROR)),
+ ("onnx", (is_onnx_available, ONNX_IMPORT_ERROR)),
+ ("opencv", (is_opencv_available, OPENCV_IMPORT_ERROR)),
+ ("scipy", (is_scipy_available, SCIPY_IMPORT_ERROR)),
+ ("torch", (is_torch_available, PYTORCH_IMPORT_ERROR)),
+ ("transformers", (is_transformers_available, TRANSFORMERS_IMPORT_ERROR)),
+ ("unidecode", (is_unidecode_available, UNIDECODE_IMPORT_ERROR)),
+ ("librosa", (is_librosa_available, LIBROSA_IMPORT_ERROR)),
+ ("k_diffusion", (is_k_diffusion_available, K_DIFFUSION_IMPORT_ERROR)),
+ ("note_seq", (is_note_seq_available, NOTE_SEQ_IMPORT_ERROR)),
+ ("wandb", (is_wandb_available, WANDB_IMPORT_ERROR)),
+ ("tensorboard", (is_tensorboard_available, TENSORBOARD_IMPORT_ERROR)),
+ ("compel", (is_compel_available, COMPEL_IMPORT_ERROR)),
+ ("ftfy", (is_ftfy_available, FTFY_IMPORT_ERROR)),
+ ("torchsde", (is_torchsde_available, TORCHSDE_IMPORT_ERROR)),
+ ("invisible_watermark", (is_invisible_watermark_available, INVISIBLE_WATERMARK_IMPORT_ERROR)),
+ ]
+)
+
+
+def requires_backends(obj, backends):
+ if not isinstance(backends, (list, tuple)):
+ backends = [backends]
+
+ name = obj.__name__ if hasattr(obj, "__name__") else obj.__class__.__name__
+ checks = (BACKENDS_MAPPING[backend] for backend in backends)
+ failed = [msg.format(name) for available, msg in checks if not available()]
+ if failed:
+ raise ImportError("".join(failed))
+
+ if name in [
+ "VersatileDiffusionTextToImagePipeline",
+ "VersatileDiffusionPipeline",
+ "VersatileDiffusionDualGuidedPipeline",
+ "StableDiffusionImageVariationPipeline",
+ "UnCLIPPipeline",
+ ] and is_transformers_version("<", "4.25.0"):
+ raise ImportError(
+ f"You need to install `transformers>=4.25` in order to use {name}: \n```\n pip install"
+ " --upgrade transformers \n```"
+ )
+
+ if name in ["StableDiffusionDepth2ImgPipeline", "StableDiffusionPix2PixZeroPipeline"] and is_transformers_version(
+ "<", "4.26.0"
+ ):
+ raise ImportError(
+ f"You need to install `transformers>=4.26` in order to use {name}: \n```\n pip install"
+ " --upgrade transformers \n```"
+ )
+
+
+class DummyObject(type):
+ """
+ Metaclass for the dummy objects. Any class inheriting from it will return the ImportError generated by
+ `requires_backend` each time a user tries to access any method of that class.
+ """
+
+ def __getattr__(cls, key):
+ if key.startswith("_") and key not in ["_load_connected_pipes", "_is_onnx"]:
+ return super().__getattr__(cls, key)
+ requires_backends(cls, cls._backends)
+
+
+# This function was copied from: https://github.com/huggingface/accelerate/blob/874c4967d94badd24f893064cc3bef45f57cadf7/src/accelerate/utils/versions.py#L319
+def compare_versions(library_or_version: Union[str, Version], operation: str, requirement_version: str):
+ """
+ Args:
+ Compares a library version to some requirement using a given operation.
+ library_or_version (`str` or `packaging.version.Version`):
+ A library name or a version to check.
+ operation (`str`):
+ A string representation of an operator, such as `">"` or `"<="`.
+ requirement_version (`str`):
+ The version to compare the library version against
+ """
+ if operation not in STR_OPERATION_TO_FUNC.keys():
+ raise ValueError(f"`operation` must be one of {list(STR_OPERATION_TO_FUNC.keys())}, received {operation}")
+ operation = STR_OPERATION_TO_FUNC[operation]
+ if isinstance(library_or_version, str):
+ library_or_version = parse(importlib_metadata.version(library_or_version))
+ return operation(library_or_version, parse(requirement_version))
+
+
+# This function was copied from: https://github.com/huggingface/accelerate/blob/874c4967d94badd24f893064cc3bef45f57cadf7/src/accelerate/utils/versions.py#L338
+def is_torch_version(operation: str, version: str):
+ """
+ Args:
+ Compares the current PyTorch version to a given reference with an operation.
+ operation (`str`):
+ A string representation of an operator, such as `">"` or `"<="`
+ version (`str`):
+ A string version of PyTorch
+ """
+ return compare_versions(parse(_torch_version), operation, version)
+
+
+def is_transformers_version(operation: str, version: str):
+ """
+ Args:
+ Compares the current Transformers version to a given reference with an operation.
+ operation (`str`):
+ A string representation of an operator, such as `">"` or `"<="`
+ version (`str`):
+ A version string
+ """
+ if not _transformers_available:
+ return False
+ return compare_versions(parse(_transformers_version), operation, version)
+
+
+def is_accelerate_version(operation: str, version: str):
+ """
+ Args:
+ Compares the current Accelerate version to a given reference with an operation.
+ operation (`str`):
+ A string representation of an operator, such as `">"` or `"<="`
+ version (`str`):
+ A version string
+ """
+ if not _accelerate_available:
+ return False
+ return compare_versions(parse(_accelerate_version), operation, version)
+
+
+def is_k_diffusion_version(operation: str, version: str):
+ """
+ Args:
+ Compares the current k-diffusion version to a given reference with an operation.
+ operation (`str`):
+ A string representation of an operator, such as `">"` or `"<="`
+ version (`str`):
+ A version string
+ """
+ if not _k_diffusion_available:
+ return False
+ return compare_versions(parse(_k_diffusion_version), operation, version)
+
+
+def get_objects_from_module(module):
+ """
+ Args:
+ Returns a dict of object names and values in a module, while skipping private/internal objects
+ module (ModuleType):
+ Module to extract the objects from.
+
+ Returns:
+ dict: Dictionary of object names and corresponding values
+ """
+
+ objects = {}
+ for name in dir(module):
+ if name.startswith("_"):
+ continue
+ objects[name] = getattr(module, name)
+
+ return objects
+
+
+class OptionalDependencyNotAvailable(BaseException):
+ """An error indicating that an optional dependency of Diffusers was not found in the environment."""
+
+
+class _LazyModule(ModuleType):
+ """
+ Module class that surfaces all objects but only performs associated imports when the objects are requested.
+ """
+
+ # Very heavily inspired by optuna.integration._IntegrationModule
+ # https://github.com/optuna/optuna/blob/master/optuna/integration/__init__.py
+ def __init__(self, name, module_file, import_structure, module_spec=None, extra_objects=None):
+ super().__init__(name)
+ self._modules = set(import_structure.keys())
+ self._class_to_module = {}
+ for key, values in import_structure.items():
+ for value in values:
+ self._class_to_module[value] = key
+ # Needed for autocompletion in an IDE
+ self.__all__ = list(import_structure.keys()) + list(chain(*import_structure.values()))
+ self.__file__ = module_file
+ self.__spec__ = module_spec
+ self.__path__ = [os.path.dirname(module_file)]
+ self._objects = {} if extra_objects is None else extra_objects
+ self._name = name
+ self._import_structure = import_structure
+
+ # Needed for autocompletion in an IDE
+ def __dir__(self):
+ result = super().__dir__()
+ # The elements of self.__all__ that are submodules may or may not be in the dir already, depending on whether
+ # they have been accessed or not. So we only add the elements of self.__all__ that are not already in the dir.
+ for attr in self.__all__:
+ if attr not in result:
+ result.append(attr)
+ return result
+
+ def __getattr__(self, name: str) -> Any:
+ if name in self._objects:
+ return self._objects[name]
+ if name in self._modules:
+ value = self._get_module(name)
+ elif name in self._class_to_module.keys():
+ module = self._get_module(self._class_to_module[name])
+ value = getattr(module, name)
+ else:
+ raise AttributeError(f"module {self.__name__} has no attribute {name}")
+
+ setattr(self, name, value)
+ return value
+
+ def _get_module(self, module_name: str):
+ try:
+ return importlib.import_module("." + module_name, self.__name__)
+ except Exception as e:
+ raise RuntimeError(
+ f"Failed to import {self.__name__}.{module_name} because of the following error (look up to see its"
+ f" traceback):\n{e}"
+ ) from e
+
+ def __reduce__(self):
+ return (self.__class__, (self._name, self.__file__, self._import_structure))
diff --git a/model/diffusers_c/utils/loading_utils.py b/model/diffusers_c/utils/loading_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..18f6ead64c4e631c4adbffbcccff1f3e4b72f3fd
--- /dev/null
+++ b/model/diffusers_c/utils/loading_utils.py
@@ -0,0 +1,49 @@
+import os
+from typing import Callable, Union
+
+import PIL.Image
+import PIL.ImageOps
+import requests
+
+
+def load_image(
+ image: Union[str, PIL.Image.Image], convert_method: Callable[[PIL.Image.Image], PIL.Image.Image] = None
+) -> PIL.Image.Image:
+ """
+ Loads `image` to a PIL Image.
+
+ Args:
+ image (`str` or `PIL.Image.Image`):
+ The image to convert to the PIL Image format.
+ convert_method (Callable[[PIL.Image.Image], PIL.Image.Image], optional):
+ A conversion method to apply to the image after loading it.
+ When set to `None` the image will be converted "RGB".
+
+ Returns:
+ `PIL.Image.Image`:
+ A PIL Image.
+ """
+ if isinstance(image, str):
+ if image.startswith("http://") or image.startswith("https://"):
+ image = PIL.Image.open(requests.get(image, stream=True).raw)
+ elif os.path.isfile(image):
+ image = PIL.Image.open(image)
+ else:
+ raise ValueError(
+ f"Incorrect path or URL. URLs must start with `http://` or `https://`, and {image} is not a valid path."
+ )
+ elif isinstance(image, PIL.Image.Image):
+ image = image
+ else:
+ raise ValueError(
+ "Incorrect format used for the image. Should be a URL linking to an image, a local path, or a PIL image."
+ )
+
+ image = PIL.ImageOps.exif_transpose(image)
+
+ if convert_method is not None:
+ image = convert_method(image)
+ else:
+ image = image.convert("RGB")
+
+ return image
diff --git a/model/diffusers_c/utils/logging.py b/model/diffusers_c/utils/logging.py
new file mode 100644
index 0000000000000000000000000000000000000000..2098d51462bff4d6a4717b86cea76b3232d99d37
--- /dev/null
+++ b/model/diffusers_c/utils/logging.py
@@ -0,0 +1,338 @@
+# Copyright 2024 Optuna, Hugging Face
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Logging utilities."""
+
+import logging
+import os
+import sys
+import threading
+from logging import (
+ CRITICAL, # NOQA
+ DEBUG, # NOQA
+ ERROR, # NOQA
+ FATAL, # NOQA
+ INFO, # NOQA
+ NOTSET, # NOQA
+ WARN, # NOQA
+ WARNING, # NOQA
+)
+from typing import Dict, Optional
+
+from tqdm import auto as tqdm_lib
+
+
+_lock = threading.Lock()
+_default_handler: Optional[logging.Handler] = None
+
+log_levels = {
+ "debug": logging.DEBUG,
+ "info": logging.INFO,
+ "warning": logging.WARNING,
+ "error": logging.ERROR,
+ "critical": logging.CRITICAL,
+}
+
+_default_log_level = logging.WARNING
+
+_tqdm_active = True
+
+
+def _get_default_logging_level() -> int:
+ """
+ If DIFFUSERS_VERBOSITY env var is set to one of the valid choices return that as the new default level. If it is
+ not - fall back to `_default_log_level`
+ """
+ env_level_str = os.getenv("DIFFUSERS_VERBOSITY", None)
+ if env_level_str:
+ if env_level_str in log_levels:
+ return log_levels[env_level_str]
+ else:
+ logging.getLogger().warning(
+ f"Unknown option DIFFUSERS_VERBOSITY={env_level_str}, "
+ f"has to be one of: { ', '.join(log_levels.keys()) }"
+ )
+ return _default_log_level
+
+
+def _get_library_name() -> str:
+ return __name__.split(".")[0]
+
+
+def _get_library_root_logger() -> logging.Logger:
+ return logging.getLogger(_get_library_name())
+
+
+def _configure_library_root_logger() -> None:
+ global _default_handler
+
+ with _lock:
+ if _default_handler:
+ # This library has already configured the library root logger.
+ return
+ _default_handler = logging.StreamHandler() # Set sys.stderr as stream.
+ _default_handler.flush = sys.stderr.flush
+
+ # Apply our default configuration to the library root logger.
+ library_root_logger = _get_library_root_logger()
+ library_root_logger.addHandler(_default_handler)
+ library_root_logger.setLevel(_get_default_logging_level())
+ library_root_logger.propagate = False
+
+
+def _reset_library_root_logger() -> None:
+ global _default_handler
+
+ with _lock:
+ if not _default_handler:
+ return
+
+ library_root_logger = _get_library_root_logger()
+ library_root_logger.removeHandler(_default_handler)
+ library_root_logger.setLevel(logging.NOTSET)
+ _default_handler = None
+
+
+def get_log_levels_dict() -> Dict[str, int]:
+ return log_levels
+
+
+def get_logger(name: Optional[str] = None) -> logging.Logger:
+ """
+ Return a logger with the specified name.
+
+ This function is not supposed to be directly accessed unless you are writing a custom diffusers module.
+ """
+
+ if name is None:
+ name = _get_library_name()
+
+ _configure_library_root_logger()
+ return logging.getLogger(name)
+
+
+def get_verbosity() -> int:
+ """
+ Return the current level for the 🤗 Diffusers' root logger as an `int`.
+
+ Returns:
+ `int`:
+ Logging level integers which can be one of:
+
+ - `50`: `diffusers.logging.CRITICAL` or `diffusers.logging.FATAL`
+ - `40`: `diffusers.logging.ERROR`
+ - `30`: `diffusers.logging.WARNING` or `diffusers.logging.WARN`
+ - `20`: `diffusers.logging.INFO`
+ - `10`: `diffusers.logging.DEBUG`
+
+ """
+
+ _configure_library_root_logger()
+ return _get_library_root_logger().getEffectiveLevel()
+
+
+def set_verbosity(verbosity: int) -> None:
+ """
+ Set the verbosity level for the 🤗 Diffusers' root logger.
+
+ Args:
+ verbosity (`int`):
+ Logging level which can be one of:
+
+ - `diffusers.logging.CRITICAL` or `diffusers.logging.FATAL`
+ - `diffusers.logging.ERROR`
+ - `diffusers.logging.WARNING` or `diffusers.logging.WARN`
+ - `diffusers.logging.INFO`
+ - `diffusers.logging.DEBUG`
+ """
+
+ _configure_library_root_logger()
+ _get_library_root_logger().setLevel(verbosity)
+
+
+def set_verbosity_info() -> None:
+ """Set the verbosity to the `INFO` level."""
+ return set_verbosity(INFO)
+
+
+def set_verbosity_warning() -> None:
+ """Set the verbosity to the `WARNING` level."""
+ return set_verbosity(WARNING)
+
+
+def set_verbosity_debug() -> None:
+ """Set the verbosity to the `DEBUG` level."""
+ return set_verbosity(DEBUG)
+
+
+def set_verbosity_error() -> None:
+ """Set the verbosity to the `ERROR` level."""
+ return set_verbosity(ERROR)
+
+
+def disable_default_handler() -> None:
+ """Disable the default handler of the 🤗 Diffusers' root logger."""
+
+ _configure_library_root_logger()
+
+ assert _default_handler is not None
+ _get_library_root_logger().removeHandler(_default_handler)
+
+
+def enable_default_handler() -> None:
+ """Enable the default handler of the 🤗 Diffusers' root logger."""
+
+ _configure_library_root_logger()
+
+ assert _default_handler is not None
+ _get_library_root_logger().addHandler(_default_handler)
+
+
+def add_handler(handler: logging.Handler) -> None:
+ """adds a handler to the HuggingFace Diffusers' root logger."""
+
+ _configure_library_root_logger()
+
+ assert handler is not None
+ _get_library_root_logger().addHandler(handler)
+
+
+def remove_handler(handler: logging.Handler) -> None:
+ """removes given handler from the HuggingFace Diffusers' root logger."""
+
+ _configure_library_root_logger()
+
+ assert handler is not None and handler in _get_library_root_logger().handlers
+ _get_library_root_logger().removeHandler(handler)
+
+
+def disable_propagation() -> None:
+ """
+ Disable propagation of the library log outputs. Note that log propagation is disabled by default.
+ """
+
+ _configure_library_root_logger()
+ _get_library_root_logger().propagate = False
+
+
+def enable_propagation() -> None:
+ """
+ Enable propagation of the library log outputs. Please disable the HuggingFace Diffusers' default handler to prevent
+ double logging if the root logger has been configured.
+ """
+
+ _configure_library_root_logger()
+ _get_library_root_logger().propagate = True
+
+
+def enable_explicit_format() -> None:
+ """
+ Enable explicit formatting for every 🤗 Diffusers' logger. The explicit formatter is as follows:
+ ```
+ [LEVELNAME|FILENAME|LINE NUMBER] TIME >> MESSAGE
+ ```
+ All handlers currently bound to the root logger are affected by this method.
+ """
+ handlers = _get_library_root_logger().handlers
+
+ for handler in handlers:
+ formatter = logging.Formatter("[%(levelname)s|%(filename)s:%(lineno)s] %(asctime)s >> %(message)s")
+ handler.setFormatter(formatter)
+
+
+def reset_format() -> None:
+ """
+ Resets the formatting for 🤗 Diffusers' loggers.
+
+ All handlers currently bound to the root logger are affected by this method.
+ """
+ handlers = _get_library_root_logger().handlers
+
+ for handler in handlers:
+ handler.setFormatter(None)
+
+
+def warning_advice(self, *args, **kwargs) -> None:
+ """
+ This method is identical to `logger.warning()`, but if env var DIFFUSERS_NO_ADVISORY_WARNINGS=1 is set, this
+ warning will not be printed
+ """
+ no_advisory_warnings = os.getenv("DIFFUSERS_NO_ADVISORY_WARNINGS", False)
+ if no_advisory_warnings:
+ return
+ self.warning(*args, **kwargs)
+
+
+logging.Logger.warning_advice = warning_advice
+
+
+class EmptyTqdm:
+ """Dummy tqdm which doesn't do anything."""
+
+ def __init__(self, *args, **kwargs): # pylint: disable=unused-argument
+ self._iterator = args[0] if args else None
+
+ def __iter__(self):
+ return iter(self._iterator)
+
+ def __getattr__(self, _):
+ """Return empty function."""
+
+ def empty_fn(*args, **kwargs): # pylint: disable=unused-argument
+ return
+
+ return empty_fn
+
+ def __enter__(self):
+ return self
+
+ def __exit__(self, type_, value, traceback):
+ return
+
+
+class _tqdm_cls:
+ def __call__(self, *args, **kwargs):
+ if _tqdm_active:
+ return tqdm_lib.tqdm(*args, **kwargs)
+ else:
+ return EmptyTqdm(*args, **kwargs)
+
+ def set_lock(self, *args, **kwargs):
+ self._lock = None
+ if _tqdm_active:
+ return tqdm_lib.tqdm.set_lock(*args, **kwargs)
+
+ def get_lock(self):
+ if _tqdm_active:
+ return tqdm_lib.tqdm.get_lock()
+
+
+tqdm = _tqdm_cls()
+
+
+def is_progress_bar_enabled() -> bool:
+ """Return a boolean indicating whether tqdm progress bars are enabled."""
+ global _tqdm_active
+ return bool(_tqdm_active)
+
+
+def enable_progress_bar() -> None:
+ """Enable tqdm progress bar."""
+ global _tqdm_active
+ _tqdm_active = True
+
+
+def disable_progress_bar() -> None:
+ """Disable tqdm progress bar."""
+ global _tqdm_active
+ _tqdm_active = False
diff --git a/model/diffusers_c/utils/model_card_template.md b/model/diffusers_c/utils/model_card_template.md
new file mode 100644
index 0000000000000000000000000000000000000000..f41b71e24e2081425d3049ae51b50036d5d28b6a
--- /dev/null
+++ b/model/diffusers_c/utils/model_card_template.md
@@ -0,0 +1,24 @@
+---
+{{ card_data }}
+---
+
+
+
+{{ model_description }}
+
+## Intended uses & limitations
+
+#### How to use
+
+```python
+# TODO: add an example code snippet for running this diffusion pipeline
+```
+
+#### Limitations and bias
+
+[TODO: provide examples of latent issues and potential remediations]
+
+## Training details
+
+[TODO: describe the data used to train the model]
diff --git a/model/diffusers_c/utils/outputs.py b/model/diffusers_c/utils/outputs.py
new file mode 100644
index 0000000000000000000000000000000000000000..6080a86b871aec2218471591bccac408a45d82b2
--- /dev/null
+++ b/model/diffusers_c/utils/outputs.py
@@ -0,0 +1,137 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Generic utilities
+"""
+
+from collections import OrderedDict
+from dataclasses import fields, is_dataclass
+from typing import Any, Tuple
+
+import numpy as np
+
+from .import_utils import is_torch_available, is_torch_version
+
+
+def is_tensor(x) -> bool:
+ """
+ Tests if `x` is a `torch.Tensor` or `np.ndarray`.
+ """
+ if is_torch_available():
+ import torch
+
+ if isinstance(x, torch.Tensor):
+ return True
+
+ return isinstance(x, np.ndarray)
+
+
+class BaseOutput(OrderedDict):
+ """
+ Base class for all model outputs as dataclass. Has a `__getitem__` that allows indexing by integer or slice (like a
+ tuple) or strings (like a dictionary) that will ignore the `None` attributes. Otherwise behaves like a regular
+ Python dictionary.
+
+
+
+ You can't unpack a [`BaseOutput`] directly. Use the [`~utils.BaseOutput.to_tuple`] method to convert it to a tuple
+ first.
+
+
+ """
+
+ def __init_subclass__(cls) -> None:
+ """Register subclasses as pytree nodes.
+
+ This is necessary to synchronize gradients when using `torch.nn.parallel.DistributedDataParallel` with
+ `static_graph=True` with modules that output `ModelOutput` subclasses.
+ """
+ if is_torch_available():
+ import torch.utils._pytree
+
+ if is_torch_version("<", "2.2"):
+ torch.utils._pytree._register_pytree_node(
+ cls,
+ torch.utils._pytree._dict_flatten,
+ lambda values, context: cls(**torch.utils._pytree._dict_unflatten(values, context)),
+ )
+ else:
+ torch.utils._pytree.register_pytree_node(
+ cls,
+ torch.utils._pytree._dict_flatten,
+ lambda values, context: cls(**torch.utils._pytree._dict_unflatten(values, context)),
+ )
+
+ def __post_init__(self) -> None:
+ class_fields = fields(self)
+
+ # Safety and consistency checks
+ if not len(class_fields):
+ raise ValueError(f"{self.__class__.__name__} has no fields.")
+
+ first_field = getattr(self, class_fields[0].name)
+ other_fields_are_none = all(getattr(self, field.name) is None for field in class_fields[1:])
+
+ if other_fields_are_none and isinstance(first_field, dict):
+ for key, value in first_field.items():
+ self[key] = value
+ else:
+ for field in class_fields:
+ v = getattr(self, field.name)
+ if v is not None:
+ self[field.name] = v
+
+ def __delitem__(self, *args, **kwargs):
+ raise Exception(f"You cannot use ``__delitem__`` on a {self.__class__.__name__} instance.")
+
+ def setdefault(self, *args, **kwargs):
+ raise Exception(f"You cannot use ``setdefault`` on a {self.__class__.__name__} instance.")
+
+ def pop(self, *args, **kwargs):
+ raise Exception(f"You cannot use ``pop`` on a {self.__class__.__name__} instance.")
+
+ def update(self, *args, **kwargs):
+ raise Exception(f"You cannot use ``update`` on a {self.__class__.__name__} instance.")
+
+ def __getitem__(self, k: Any) -> Any:
+ if isinstance(k, str):
+ inner_dict = dict(self.items())
+ return inner_dict[k]
+ else:
+ return self.to_tuple()[k]
+
+ def __setattr__(self, name: Any, value: Any) -> None:
+ if name in self.keys() and value is not None:
+ # Don't call self.__setitem__ to avoid recursion errors
+ super().__setitem__(name, value)
+ super().__setattr__(name, value)
+
+ def __setitem__(self, key, value):
+ # Will raise a KeyException if needed
+ super().__setitem__(key, value)
+ # Don't call self.__setattr__ to avoid recursion errors
+ super().__setattr__(key, value)
+
+ def __reduce__(self):
+ if not is_dataclass(self):
+ return super().__reduce__()
+ callable, _args, *remaining = super().__reduce__()
+ args = tuple(getattr(self, field.name) for field in fields(self))
+ return callable, args, *remaining
+
+ def to_tuple(self) -> Tuple[Any, ...]:
+ """
+ Convert self to a tuple containing all the attributes/keys that are not `None`.
+ """
+ return tuple(self[k] for k in self.keys())
diff --git a/model/diffusers_c/utils/peft_utils.py b/model/diffusers_c/utils/peft_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..fb37545eb472c3c1592ee9596c173312cf4d24a8
--- /dev/null
+++ b/model/diffusers_c/utils/peft_utils.py
@@ -0,0 +1,269 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+PEFT utilities: Utilities related to peft library
+"""
+
+import collections
+import importlib
+from typing import Optional
+
+from packaging import version
+
+from .import_utils import is_peft_available, is_torch_available
+
+
+if is_torch_available():
+ import torch
+
+
+def recurse_remove_peft_layers(model):
+ r"""
+ Recursively replace all instances of `LoraLayer` with corresponding new layers in `model`.
+ """
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ has_base_layer_pattern = False
+ for module in model.modules():
+ if isinstance(module, BaseTunerLayer):
+ has_base_layer_pattern = hasattr(module, "base_layer")
+ break
+
+ if has_base_layer_pattern:
+ from peft.utils import _get_submodules
+
+ key_list = [key for key, _ in model.named_modules() if "lora" not in key]
+ for key in key_list:
+ try:
+ parent, target, target_name = _get_submodules(model, key)
+ except AttributeError:
+ continue
+ if hasattr(target, "base_layer"):
+ setattr(parent, target_name, target.get_base_layer())
+ else:
+ # This is for backwards compatibility with PEFT <= 0.6.2.
+ # TODO can be removed once that PEFT version is no longer supported.
+ from peft.tuners.lora import LoraLayer
+
+ for name, module in model.named_children():
+ if len(list(module.children())) > 0:
+ ## compound module, go inside it
+ recurse_remove_peft_layers(module)
+
+ module_replaced = False
+
+ if isinstance(module, LoraLayer) and isinstance(module, torch.nn.Linear):
+ new_module = torch.nn.Linear(module.in_features, module.out_features, bias=module.bias is not None).to(
+ module.weight.device
+ )
+ new_module.weight = module.weight
+ if module.bias is not None:
+ new_module.bias = module.bias
+
+ module_replaced = True
+ elif isinstance(module, LoraLayer) and isinstance(module, torch.nn.Conv2d):
+ new_module = torch.nn.Conv2d(
+ module.in_channels,
+ module.out_channels,
+ module.kernel_size,
+ module.stride,
+ module.padding,
+ module.dilation,
+ module.groups,
+ ).to(module.weight.device)
+
+ new_module.weight = module.weight
+ if module.bias is not None:
+ new_module.bias = module.bias
+
+ module_replaced = True
+
+ if module_replaced:
+ setattr(model, name, new_module)
+ del module
+
+ if torch.cuda.is_available():
+ torch.cuda.empty_cache()
+ return model
+
+
+def scale_lora_layers(model, weight):
+ """
+ Adjust the weightage given to the LoRA layers of the model.
+
+ Args:
+ model (`torch.nn.Module`):
+ The model to scale.
+ weight (`float`):
+ The weight to be given to the LoRA layers.
+ """
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ for module in model.modules():
+ if isinstance(module, BaseTunerLayer):
+ module.scale_layer(weight)
+
+
+def unscale_lora_layers(model, weight: Optional[float] = None):
+ """
+ Removes the previously passed weight given to the LoRA layers of the model.
+
+ Args:
+ model (`torch.nn.Module`):
+ The model to scale.
+ weight (`float`, *optional*):
+ The weight to be given to the LoRA layers. If no scale is passed the scale of the lora layer will be
+ re-initialized to the correct value. If 0.0 is passed, we will re-initialize the scale with the correct
+ value.
+ """
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ for module in model.modules():
+ if isinstance(module, BaseTunerLayer):
+ if weight is not None and weight != 0:
+ module.unscale_layer(weight)
+ elif weight is not None and weight == 0:
+ for adapter_name in module.active_adapters:
+ # if weight == 0 unscale should re-set the scale to the original value.
+ module.set_scale(adapter_name, 1.0)
+
+
+def get_peft_kwargs(rank_dict, network_alpha_dict, peft_state_dict, is_unet=True):
+ rank_pattern = {}
+ alpha_pattern = {}
+ r = lora_alpha = list(rank_dict.values())[0]
+
+ if len(set(rank_dict.values())) > 1:
+ # get the rank occuring the most number of times
+ r = collections.Counter(rank_dict.values()).most_common()[0][0]
+
+ # for modules with rank different from the most occuring rank, add it to the `rank_pattern`
+ rank_pattern = dict(filter(lambda x: x[1] != r, rank_dict.items()))
+ rank_pattern = {k.split(".lora_B.")[0]: v for k, v in rank_pattern.items()}
+
+ if network_alpha_dict is not None and len(network_alpha_dict) > 0:
+ if len(set(network_alpha_dict.values())) > 1:
+ # get the alpha occuring the most number of times
+ lora_alpha = collections.Counter(network_alpha_dict.values()).most_common()[0][0]
+
+ # for modules with alpha different from the most occuring alpha, add it to the `alpha_pattern`
+ alpha_pattern = dict(filter(lambda x: x[1] != lora_alpha, network_alpha_dict.items()))
+ if is_unet:
+ alpha_pattern = {
+ ".".join(k.split(".lora_A.")[0].split(".")).replace(".alpha", ""): v
+ for k, v in alpha_pattern.items()
+ }
+ else:
+ alpha_pattern = {".".join(k.split(".down.")[0].split(".")[:-1]): v for k, v in alpha_pattern.items()}
+ else:
+ lora_alpha = set(network_alpha_dict.values()).pop()
+
+ # layer names without the Diffusers specific
+ target_modules = list({name.split(".lora")[0] for name in peft_state_dict.keys()})
+
+ lora_config_kwargs = {
+ "r": r,
+ "lora_alpha": lora_alpha,
+ "rank_pattern": rank_pattern,
+ "alpha_pattern": alpha_pattern,
+ "target_modules": target_modules,
+ }
+ return lora_config_kwargs
+
+
+def get_adapter_name(model):
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ for module in model.modules():
+ if isinstance(module, BaseTunerLayer):
+ return f"default_{len(module.r)}"
+ return "default_0"
+
+
+def set_adapter_layers(model, enabled=True):
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ for module in model.modules():
+ if isinstance(module, BaseTunerLayer):
+ # The recent version of PEFT needs to call `enable_adapters` instead
+ if hasattr(module, "enable_adapters"):
+ module.enable_adapters(enabled=enabled)
+ else:
+ module.disable_adapters = not enabled
+
+
+def delete_adapter_layers(model, adapter_name):
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ for module in model.modules():
+ if isinstance(module, BaseTunerLayer):
+ if hasattr(module, "delete_adapter"):
+ module.delete_adapter(adapter_name)
+ else:
+ raise ValueError(
+ "The version of PEFT you are using is not compatible, please use a version that is greater than 0.6.1"
+ )
+
+ # For transformers integration - we need to pop the adapter from the config
+ if getattr(model, "_hf_peft_config_loaded", False) and hasattr(model, "peft_config"):
+ model.peft_config.pop(adapter_name, None)
+ # In case all adapters are deleted, we need to delete the config
+ # and make sure to set the flag to False
+ if len(model.peft_config) == 0:
+ del model.peft_config
+ model._hf_peft_config_loaded = None
+
+
+def set_weights_and_activate_adapters(model, adapter_names, weights):
+ from peft.tuners.tuners_utils import BaseTunerLayer
+
+ # iterate over each adapter, make it active and set the corresponding scaling weight
+ for adapter_name, weight in zip(adapter_names, weights):
+ for module in model.modules():
+ if isinstance(module, BaseTunerLayer):
+ # For backward compatbility with previous PEFT versions
+ if hasattr(module, "set_adapter"):
+ module.set_adapter(adapter_name)
+ else:
+ module.active_adapter = adapter_name
+ module.set_scale(adapter_name, weight)
+
+ # set multiple active adapters
+ for module in model.modules():
+ if isinstance(module, BaseTunerLayer):
+ # For backward compatbility with previous PEFT versions
+ if hasattr(module, "set_adapter"):
+ module.set_adapter(adapter_names)
+ else:
+ module.active_adapter = adapter_names
+
+
+def check_peft_version(min_version: str) -> None:
+ r"""
+ Checks if the version of PEFT is compatible.
+
+ Args:
+ version (`str`):
+ The version of PEFT to check against.
+ """
+ if not is_peft_available():
+ raise ValueError("PEFT is not installed. Please install it with `pip install peft`")
+
+ is_peft_version_compatible = version.parse(importlib.metadata.version("peft")) > version.parse(min_version)
+
+ if not is_peft_version_compatible:
+ raise ValueError(
+ f"The version of PEFT you are using is not compatible, please use a version that is greater"
+ f" than {min_version}"
+ )
diff --git a/model/diffusers_c/utils/pil_utils.py b/model/diffusers_c/utils/pil_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..76678070b697c7d87fc3691d9bc5bb3bea83c5b1
--- /dev/null
+++ b/model/diffusers_c/utils/pil_utils.py
@@ -0,0 +1,67 @@
+from typing import List
+
+import PIL.Image
+import PIL.ImageOps
+from packaging import version
+from PIL import Image
+
+
+if version.parse(version.parse(PIL.__version__).base_version) >= version.parse("9.1.0"):
+ PIL_INTERPOLATION = {
+ "linear": PIL.Image.Resampling.BILINEAR,
+ "bilinear": PIL.Image.Resampling.BILINEAR,
+ "bicubic": PIL.Image.Resampling.BICUBIC,
+ "lanczos": PIL.Image.Resampling.LANCZOS,
+ "nearest": PIL.Image.Resampling.NEAREST,
+ }
+else:
+ PIL_INTERPOLATION = {
+ "linear": PIL.Image.LINEAR,
+ "bilinear": PIL.Image.BILINEAR,
+ "bicubic": PIL.Image.BICUBIC,
+ "lanczos": PIL.Image.LANCZOS,
+ "nearest": PIL.Image.NEAREST,
+ }
+
+
+def pt_to_pil(images):
+ """
+ Convert a torch image to a PIL image.
+ """
+ images = (images / 2 + 0.5).clamp(0, 1)
+ images = images.cpu().permute(0, 2, 3, 1).float().numpy()
+ images = numpy_to_pil(images)
+ return images
+
+
+def numpy_to_pil(images):
+ """
+ Convert a numpy image or a batch of images to a PIL image.
+ """
+ if images.ndim == 3:
+ images = images[None, ...]
+ images = (images * 255).round().astype("uint8")
+ if images.shape[-1] == 1:
+ # special case for grayscale (single channel) images
+ pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
+ else:
+ pil_images = [Image.fromarray(image) for image in images]
+
+ return pil_images
+
+
+def make_image_grid(images: List[PIL.Image.Image], rows: int, cols: int, resize: int = None) -> PIL.Image.Image:
+ """
+ Prepares a single grid of images. Useful for visualization purposes.
+ """
+ assert len(images) == rows * cols
+
+ if resize is not None:
+ images = [img.resize((resize, resize)) for img in images]
+
+ w, h = images[0].size
+ grid = Image.new("RGB", size=(cols * w, rows * h))
+
+ for i, img in enumerate(images):
+ grid.paste(img, box=(i % cols * w, i // cols * h))
+ return grid
diff --git a/model/diffusers_c/utils/state_dict_utils.py b/model/diffusers_c/utils/state_dict_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..00b96495caf12c8c9312263ff1d8b394baef94eb
--- /dev/null
+++ b/model/diffusers_c/utils/state_dict_utils.py
@@ -0,0 +1,325 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+State dict utilities: utility methods for converting state dicts easily
+"""
+
+import enum
+
+from .logging import get_logger
+
+
+logger = get_logger(__name__)
+
+
+class StateDictType(enum.Enum):
+ """
+ The mode to use when converting state dicts.
+ """
+
+ DIFFUSERS_OLD = "diffusers_old"
+ KOHYA_SS = "kohya_ss"
+ PEFT = "peft"
+ DIFFUSERS = "diffusers"
+
+
+# We need to define a proper mapping for Unet since it uses different output keys than text encoder
+# e.g. to_q_lora -> q_proj / to_q
+UNET_TO_DIFFUSERS = {
+ ".to_out_lora.up": ".to_out.0.lora_B",
+ ".to_out_lora.down": ".to_out.0.lora_A",
+ ".to_q_lora.down": ".to_q.lora_A",
+ ".to_q_lora.up": ".to_q.lora_B",
+ ".to_k_lora.down": ".to_k.lora_A",
+ ".to_k_lora.up": ".to_k.lora_B",
+ ".to_v_lora.down": ".to_v.lora_A",
+ ".to_v_lora.up": ".to_v.lora_B",
+ ".lora.up": ".lora_B",
+ ".lora.down": ".lora_A",
+}
+
+
+DIFFUSERS_TO_PEFT = {
+ ".q_proj.lora_linear_layer.up": ".q_proj.lora_B",
+ ".q_proj.lora_linear_layer.down": ".q_proj.lora_A",
+ ".k_proj.lora_linear_layer.up": ".k_proj.lora_B",
+ ".k_proj.lora_linear_layer.down": ".k_proj.lora_A",
+ ".v_proj.lora_linear_layer.up": ".v_proj.lora_B",
+ ".v_proj.lora_linear_layer.down": ".v_proj.lora_A",
+ ".out_proj.lora_linear_layer.up": ".out_proj.lora_B",
+ ".out_proj.lora_linear_layer.down": ".out_proj.lora_A",
+ ".lora_linear_layer.up": ".lora_B",
+ ".lora_linear_layer.down": ".lora_A",
+}
+
+DIFFUSERS_OLD_TO_PEFT = {
+ ".to_q_lora.up": ".q_proj.lora_B",
+ ".to_q_lora.down": ".q_proj.lora_A",
+ ".to_k_lora.up": ".k_proj.lora_B",
+ ".to_k_lora.down": ".k_proj.lora_A",
+ ".to_v_lora.up": ".v_proj.lora_B",
+ ".to_v_lora.down": ".v_proj.lora_A",
+ ".to_out_lora.up": ".out_proj.lora_B",
+ ".to_out_lora.down": ".out_proj.lora_A",
+ ".lora_linear_layer.up": ".lora_B",
+ ".lora_linear_layer.down": ".lora_A",
+}
+
+PEFT_TO_DIFFUSERS = {
+ ".q_proj.lora_B": ".q_proj.lora_linear_layer.up",
+ ".q_proj.lora_A": ".q_proj.lora_linear_layer.down",
+ ".k_proj.lora_B": ".k_proj.lora_linear_layer.up",
+ ".k_proj.lora_A": ".k_proj.lora_linear_layer.down",
+ ".v_proj.lora_B": ".v_proj.lora_linear_layer.up",
+ ".v_proj.lora_A": ".v_proj.lora_linear_layer.down",
+ ".out_proj.lora_B": ".out_proj.lora_linear_layer.up",
+ ".out_proj.lora_A": ".out_proj.lora_linear_layer.down",
+ "to_k.lora_A": "to_k.lora.down",
+ "to_k.lora_B": "to_k.lora.up",
+ "to_q.lora_A": "to_q.lora.down",
+ "to_q.lora_B": "to_q.lora.up",
+ "to_v.lora_A": "to_v.lora.down",
+ "to_v.lora_B": "to_v.lora.up",
+ "to_out.0.lora_A": "to_out.0.lora.down",
+ "to_out.0.lora_B": "to_out.0.lora.up",
+}
+
+DIFFUSERS_OLD_TO_DIFFUSERS = {
+ ".to_q_lora.up": ".q_proj.lora_linear_layer.up",
+ ".to_q_lora.down": ".q_proj.lora_linear_layer.down",
+ ".to_k_lora.up": ".k_proj.lora_linear_layer.up",
+ ".to_k_lora.down": ".k_proj.lora_linear_layer.down",
+ ".to_v_lora.up": ".v_proj.lora_linear_layer.up",
+ ".to_v_lora.down": ".v_proj.lora_linear_layer.down",
+ ".to_out_lora.up": ".out_proj.lora_linear_layer.up",
+ ".to_out_lora.down": ".out_proj.lora_linear_layer.down",
+}
+
+PEFT_TO_KOHYA_SS = {
+ "lora_A": "lora_down",
+ "lora_B": "lora_up",
+ # This is not a comprehensive dict as kohya format requires replacing `.` with `_` in keys,
+ # adding prefixes and adding alpha values
+ # Check `convert_state_dict_to_kohya` for more
+}
+
+PEFT_STATE_DICT_MAPPINGS = {
+ StateDictType.DIFFUSERS_OLD: DIFFUSERS_OLD_TO_PEFT,
+ StateDictType.DIFFUSERS: DIFFUSERS_TO_PEFT,
+}
+
+DIFFUSERS_STATE_DICT_MAPPINGS = {
+ StateDictType.DIFFUSERS_OLD: DIFFUSERS_OLD_TO_DIFFUSERS,
+ StateDictType.PEFT: PEFT_TO_DIFFUSERS,
+}
+
+KOHYA_STATE_DICT_MAPPINGS = {StateDictType.PEFT: PEFT_TO_KOHYA_SS}
+
+KEYS_TO_ALWAYS_REPLACE = {
+ ".processor.": ".",
+}
+
+
+def convert_state_dict(state_dict, mapping):
+ r"""
+ Simply iterates over the state dict and replaces the patterns in `mapping` with the corresponding values.
+
+ Args:
+ state_dict (`dict[str, torch.Tensor]`):
+ The state dict to convert.
+ mapping (`dict[str, str]`):
+ The mapping to use for conversion, the mapping should be a dictionary with the following structure:
+ - key: the pattern to replace
+ - value: the pattern to replace with
+
+ Returns:
+ converted_state_dict (`dict`)
+ The converted state dict.
+ """
+ converted_state_dict = {}
+ for k, v in state_dict.items():
+ # First, filter out the keys that we always want to replace
+ for pattern in KEYS_TO_ALWAYS_REPLACE.keys():
+ if pattern in k:
+ new_pattern = KEYS_TO_ALWAYS_REPLACE[pattern]
+ k = k.replace(pattern, new_pattern)
+
+ for pattern in mapping.keys():
+ if pattern in k:
+ new_pattern = mapping[pattern]
+ k = k.replace(pattern, new_pattern)
+ break
+ converted_state_dict[k] = v
+ return converted_state_dict
+
+
+def convert_state_dict_to_peft(state_dict, original_type=None, **kwargs):
+ r"""
+ Converts a state dict to the PEFT format The state dict can be from previous diffusers format (`OLD_DIFFUSERS`), or
+ new diffusers format (`DIFFUSERS`). The method only supports the conversion from diffusers old/new to PEFT for now.
+
+ Args:
+ state_dict (`dict[str, torch.Tensor]`):
+ The state dict to convert.
+ original_type (`StateDictType`, *optional*):
+ The original type of the state dict, if not provided, the method will try to infer it automatically.
+ """
+ if original_type is None:
+ # Old diffusers to PEFT
+ if any("to_out_lora" in k for k in state_dict.keys()):
+ original_type = StateDictType.DIFFUSERS_OLD
+ elif any("lora_linear_layer" in k for k in state_dict.keys()):
+ original_type = StateDictType.DIFFUSERS
+ else:
+ raise ValueError("Could not automatically infer state dict type")
+
+ if original_type not in PEFT_STATE_DICT_MAPPINGS.keys():
+ raise ValueError(f"Original type {original_type} is not supported")
+
+ mapping = PEFT_STATE_DICT_MAPPINGS[original_type]
+ return convert_state_dict(state_dict, mapping)
+
+
+def convert_state_dict_to_diffusers(state_dict, original_type=None, **kwargs):
+ r"""
+ Converts a state dict to new diffusers format. The state dict can be from previous diffusers format
+ (`OLD_DIFFUSERS`), or PEFT format (`PEFT`) or new diffusers format (`DIFFUSERS`). In the last case the method will
+ return the state dict as is.
+
+ The method only supports the conversion from diffusers old, PEFT to diffusers new for now.
+
+ Args:
+ state_dict (`dict[str, torch.Tensor]`):
+ The state dict to convert.
+ original_type (`StateDictType`, *optional*):
+ The original type of the state dict, if not provided, the method will try to infer it automatically.
+ kwargs (`dict`, *args*):
+ Additional arguments to pass to the method.
+
+ - **adapter_name**: For example, in case of PEFT, some keys will be pre-pended
+ with the adapter name, therefore needs a special handling. By default PEFT also takes care of that in
+ `get_peft_model_state_dict` method:
+ https://github.com/huggingface/peft/blob/ba0477f2985b1ba311b83459d29895c809404e99/src/peft/utils/save_and_load.py#L92
+ but we add it here in case we don't want to rely on that method.
+ """
+ peft_adapter_name = kwargs.pop("adapter_name", None)
+ if peft_adapter_name is not None:
+ peft_adapter_name = "." + peft_adapter_name
+ else:
+ peft_adapter_name = ""
+
+ if original_type is None:
+ # Old diffusers to PEFT
+ if any("to_out_lora" in k for k in state_dict.keys()):
+ original_type = StateDictType.DIFFUSERS_OLD
+ elif any(f".lora_A{peft_adapter_name}.weight" in k for k in state_dict.keys()):
+ original_type = StateDictType.PEFT
+ elif any("lora_linear_layer" in k for k in state_dict.keys()):
+ # nothing to do
+ return state_dict
+ else:
+ raise ValueError("Could not automatically infer state dict type")
+
+ if original_type not in DIFFUSERS_STATE_DICT_MAPPINGS.keys():
+ raise ValueError(f"Original type {original_type} is not supported")
+
+ mapping = DIFFUSERS_STATE_DICT_MAPPINGS[original_type]
+ return convert_state_dict(state_dict, mapping)
+
+
+def convert_unet_state_dict_to_peft(state_dict):
+ r"""
+ Converts a state dict from UNet format to diffusers format - i.e. by removing some keys
+ """
+ mapping = UNET_TO_DIFFUSERS
+ return convert_state_dict(state_dict, mapping)
+
+
+def convert_all_state_dict_to_peft(state_dict):
+ r"""
+ Attempts to first `convert_state_dict_to_peft`, and if it doesn't detect `lora_linear_layer`
+ for a valid `DIFFUSERS` LoRA for example, attempts to exclusively convert the Unet `convert_unet_state_dict_to_peft`
+ """
+ try:
+ peft_dict = convert_state_dict_to_peft(state_dict)
+ except Exception as e:
+ if str(e) == "Could not automatically infer state dict type":
+ peft_dict = convert_unet_state_dict_to_peft(state_dict)
+ else:
+ raise
+
+ if not any("lora_A" in key or "lora_B" in key for key in peft_dict.keys()):
+ raise ValueError("Your LoRA was not converted to PEFT")
+
+ return peft_dict
+
+
+def convert_state_dict_to_kohya(state_dict, original_type=None, **kwargs):
+ r"""
+ Converts a `PEFT` state dict to `Kohya` format that can be used in AUTOMATIC1111, ComfyUI, SD.Next, InvokeAI, etc.
+ The method only supports the conversion from PEFT to Kohya for now.
+
+ Args:
+ state_dict (`dict[str, torch.Tensor]`):
+ The state dict to convert.
+ original_type (`StateDictType`, *optional*):
+ The original type of the state dict, if not provided, the method will try to infer it automatically.
+ kwargs (`dict`, *args*):
+ Additional arguments to pass to the method.
+
+ - **adapter_name**: For example, in case of PEFT, some keys will be pre-pended
+ with the adapter name, therefore needs a special handling. By default PEFT also takes care of that in
+ `get_peft_model_state_dict` method:
+ https://github.com/huggingface/peft/blob/ba0477f2985b1ba311b83459d29895c809404e99/src/peft/utils/save_and_load.py#L92
+ but we add it here in case we don't want to rely on that method.
+ """
+ try:
+ import torch
+ except ImportError:
+ logger.error("Converting PEFT state dicts to Kohya requires torch to be installed.")
+ raise
+
+ peft_adapter_name = kwargs.pop("adapter_name", None)
+ if peft_adapter_name is not None:
+ peft_adapter_name = "." + peft_adapter_name
+ else:
+ peft_adapter_name = ""
+
+ if original_type is None:
+ if any(f".lora_A{peft_adapter_name}.weight" in k for k in state_dict.keys()):
+ original_type = StateDictType.PEFT
+
+ if original_type not in KOHYA_STATE_DICT_MAPPINGS.keys():
+ raise ValueError(f"Original type {original_type} is not supported")
+
+ # Use the convert_state_dict function with the appropriate mapping
+ kohya_ss_partial_state_dict = convert_state_dict(state_dict, KOHYA_STATE_DICT_MAPPINGS[StateDictType.PEFT])
+ kohya_ss_state_dict = {}
+
+ # Additional logic for replacing header, alpha parameters `.` with `_` in all keys
+ for kohya_key, weight in kohya_ss_partial_state_dict.items():
+ if "text_encoder_2." in kohya_key:
+ kohya_key = kohya_key.replace("text_encoder_2.", "lora_te2.")
+ elif "text_encoder." in kohya_key:
+ kohya_key = kohya_key.replace("text_encoder.", "lora_te1.")
+ elif "unet" in kohya_key:
+ kohya_key = kohya_key.replace("unet", "lora_unet")
+ kohya_key = kohya_key.replace(".", "_", kohya_key.count(".") - 2)
+ kohya_key = kohya_key.replace(peft_adapter_name, "") # Kohya doesn't take names
+ kohya_ss_state_dict[kohya_key] = weight
+ if "lora_down" in kohya_key:
+ alpha_key = f'{kohya_key.split(".")[0]}.alpha'
+ kohya_ss_state_dict[alpha_key] = torch.tensor(len(weight))
+
+ return kohya_ss_state_dict
diff --git a/model/diffusers_c/utils/testing_utils.py b/model/diffusers_c/utils/testing_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..edbf6f31a833a2005d9476489f74cd5562a9c2ec
--- /dev/null
+++ b/model/diffusers_c/utils/testing_utils.py
@@ -0,0 +1,967 @@
+import functools
+import importlib
+import inspect
+import io
+import logging
+import multiprocessing
+import os
+import random
+import re
+import struct
+import sys
+import tempfile
+import time
+import unittest
+import urllib.parse
+from contextlib import contextmanager
+from distutils.util import strtobool
+from io import BytesIO, StringIO
+from pathlib import Path
+from typing import Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import PIL.ImageOps
+import requests
+from numpy.linalg import norm
+from packaging import version
+
+from .import_utils import (
+ BACKENDS_MAPPING,
+ is_compel_available,
+ is_flax_available,
+ is_note_seq_available,
+ is_onnx_available,
+ is_opencv_available,
+ is_peft_available,
+ is_torch_available,
+ is_torch_version,
+ is_torchsde_available,
+ is_transformers_available,
+)
+from .logging import get_logger
+
+
+global_rng = random.Random()
+
+logger = get_logger(__name__)
+
+_required_peft_version = is_peft_available() and version.parse(
+ version.parse(importlib.metadata.version("peft")).base_version
+) > version.parse("0.5")
+_required_transformers_version = is_transformers_available() and version.parse(
+ version.parse(importlib.metadata.version("transformers")).base_version
+) > version.parse("4.33")
+
+USE_PEFT_BACKEND = _required_peft_version and _required_transformers_version
+
+if is_torch_available():
+ import torch
+
+ # Set a backend environment variable for any extra module import required for a custom accelerator
+ if "DIFFUSERS_TEST_BACKEND" in os.environ:
+ backend = os.environ["DIFFUSERS_TEST_BACKEND"]
+ try:
+ _ = importlib.import_module(backend)
+ except ModuleNotFoundError as e:
+ raise ModuleNotFoundError(
+ f"Failed to import `DIFFUSERS_TEST_BACKEND` '{backend}'! This should be the name of an installed module \
+ to enable a specified backend.):\n{e}"
+ ) from e
+
+ if "DIFFUSERS_TEST_DEVICE" in os.environ:
+ torch_device = os.environ["DIFFUSERS_TEST_DEVICE"]
+ try:
+ # try creating device to see if provided device is valid
+ _ = torch.device(torch_device)
+ except RuntimeError as e:
+ raise RuntimeError(
+ f"Unknown testing device specified by environment variable `DIFFUSERS_TEST_DEVICE`: {torch_device}"
+ ) from e
+ logger.info(f"torch_device overrode to {torch_device}")
+ else:
+ torch_device = "cuda" if torch.cuda.is_available() else "cpu"
+ is_torch_higher_equal_than_1_12 = version.parse(
+ version.parse(torch.__version__).base_version
+ ) >= version.parse("1.12")
+
+ if is_torch_higher_equal_than_1_12:
+ # Some builds of torch 1.12 don't have the mps backend registered. See #892 for more details
+ mps_backend_registered = hasattr(torch.backends, "mps")
+ torch_device = "mps" if (mps_backend_registered and torch.backends.mps.is_available()) else torch_device
+
+
+def torch_all_close(a, b, *args, **kwargs):
+ if not is_torch_available():
+ raise ValueError("PyTorch needs to be installed to use this function.")
+ if not torch.allclose(a, b, *args, **kwargs):
+ assert False, f"Max diff is absolute {(a - b).abs().max()}. Diff tensor is {(a - b).abs()}."
+ return True
+
+
+def numpy_cosine_similarity_distance(a, b):
+ similarity = np.dot(a, b) / (norm(a) * norm(b))
+ distance = 1.0 - similarity.mean()
+
+ return distance
+
+
+def print_tensor_test(tensor, filename="test_corrections.txt", expected_tensor_name="expected_slice"):
+ test_name = os.environ.get("PYTEST_CURRENT_TEST")
+ if not torch.is_tensor(tensor):
+ tensor = torch.from_numpy(tensor)
+
+ tensor_str = str(tensor.detach().cpu().flatten().to(torch.float32)).replace("\n", "")
+ # format is usually:
+ # expected_slice = np.array([-0.5713, -0.3018, -0.9814, 0.04663, -0.879, 0.76, -1.734, 0.1044, 1.161])
+ output_str = tensor_str.replace("tensor", f"{expected_tensor_name} = np.array")
+ test_file, test_class, test_fn = test_name.split("::")
+ test_fn = test_fn.split()[0]
+ with open(filename, "a") as f:
+ print(";".join([test_file, test_class, test_fn, output_str]), file=f)
+
+
+def get_tests_dir(append_path=None):
+ """
+ Args:
+ append_path: optional path to append to the tests dir path
+ Return:
+ The full path to the `tests` dir, so that the tests can be invoked from anywhere. Optionally `append_path` is
+ joined after the `tests` dir the former is provided.
+ """
+ # this function caller's __file__
+ caller__file__ = inspect.stack()[1][1]
+ tests_dir = os.path.abspath(os.path.dirname(caller__file__))
+
+ while not tests_dir.endswith("tests"):
+ tests_dir = os.path.dirname(tests_dir)
+
+ if append_path:
+ return Path(tests_dir, append_path).as_posix()
+ else:
+ return tests_dir
+
+
+def parse_flag_from_env(key, default=False):
+ try:
+ value = os.environ[key]
+ except KeyError:
+ # KEY isn't set, default to `default`.
+ _value = default
+ else:
+ # KEY is set, convert it to True or False.
+ try:
+ _value = strtobool(value)
+ except ValueError:
+ # More values are supported, but let's keep the message simple.
+ raise ValueError(f"If set, {key} must be yes or no.")
+ return _value
+
+
+_run_slow_tests = parse_flag_from_env("RUN_SLOW", default=False)
+_run_nightly_tests = parse_flag_from_env("RUN_NIGHTLY", default=False)
+
+
+def floats_tensor(shape, scale=1.0, rng=None, name=None):
+ """Creates a random float32 tensor"""
+ if rng is None:
+ rng = global_rng
+
+ total_dims = 1
+ for dim in shape:
+ total_dims *= dim
+
+ values = []
+ for _ in range(total_dims):
+ values.append(rng.random() * scale)
+
+ return torch.tensor(data=values, dtype=torch.float).view(shape).contiguous()
+
+
+def slow(test_case):
+ """
+ Decorator marking a test as slow.
+
+ Slow tests are skipped by default. Set the RUN_SLOW environment variable to a truthy value to run them.
+
+ """
+ return unittest.skipUnless(_run_slow_tests, "test is slow")(test_case)
+
+
+def nightly(test_case):
+ """
+ Decorator marking a test that runs nightly in the diffusers CI.
+
+ Slow tests are skipped by default. Set the RUN_NIGHTLY environment variable to a truthy value to run them.
+
+ """
+ return unittest.skipUnless(_run_nightly_tests, "test is nightly")(test_case)
+
+
+def require_torch(test_case):
+ """
+ Decorator marking a test that requires PyTorch. These tests are skipped when PyTorch isn't installed.
+ """
+ return unittest.skipUnless(is_torch_available(), "test requires PyTorch")(test_case)
+
+
+def require_torch_2(test_case):
+ """
+ Decorator marking a test that requires PyTorch 2. These tests are skipped when it isn't installed.
+ """
+ return unittest.skipUnless(is_torch_available() and is_torch_version(">=", "2.0.0"), "test requires PyTorch 2")(
+ test_case
+ )
+
+
+def require_torch_gpu(test_case):
+ """Decorator marking a test that requires CUDA and PyTorch."""
+ return unittest.skipUnless(is_torch_available() and torch_device == "cuda", "test requires PyTorch+CUDA")(
+ test_case
+ )
+
+
+# These decorators are for accelerator-specific behaviours that are not GPU-specific
+def require_torch_accelerator(test_case):
+ """Decorator marking a test that requires an accelerator backend and PyTorch."""
+ return unittest.skipUnless(is_torch_available() and torch_device != "cpu", "test requires accelerator+PyTorch")(
+ test_case
+ )
+
+
+def require_torch_accelerator_with_fp16(test_case):
+ """Decorator marking a test that requires an accelerator with support for the FP16 data type."""
+ return unittest.skipUnless(_is_torch_fp16_available(torch_device), "test requires accelerator with fp16 support")(
+ test_case
+ )
+
+
+def require_torch_accelerator_with_fp64(test_case):
+ """Decorator marking a test that requires an accelerator with support for the FP64 data type."""
+ return unittest.skipUnless(_is_torch_fp64_available(torch_device), "test requires accelerator with fp64 support")(
+ test_case
+ )
+
+
+def require_torch_accelerator_with_training(test_case):
+ """Decorator marking a test that requires an accelerator with support for training."""
+ return unittest.skipUnless(
+ is_torch_available() and backend_supports_training(torch_device),
+ "test requires accelerator with training support",
+ )(test_case)
+
+
+def skip_mps(test_case):
+ """Decorator marking a test to skip if torch_device is 'mps'"""
+ return unittest.skipUnless(torch_device != "mps", "test requires non 'mps' device")(test_case)
+
+
+def require_flax(test_case):
+ """
+ Decorator marking a test that requires JAX & Flax. These tests are skipped when one / both are not installed
+ """
+ return unittest.skipUnless(is_flax_available(), "test requires JAX & Flax")(test_case)
+
+
+def require_compel(test_case):
+ """
+ Decorator marking a test that requires compel: https://github.com/damian0815/compel. These tests are skipped when
+ the library is not installed.
+ """
+ return unittest.skipUnless(is_compel_available(), "test requires compel")(test_case)
+
+
+def require_onnxruntime(test_case):
+ """
+ Decorator marking a test that requires onnxruntime. These tests are skipped when onnxruntime isn't installed.
+ """
+ return unittest.skipUnless(is_onnx_available(), "test requires onnxruntime")(test_case)
+
+
+def require_note_seq(test_case):
+ """
+ Decorator marking a test that requires note_seq. These tests are skipped when note_seq isn't installed.
+ """
+ return unittest.skipUnless(is_note_seq_available(), "test requires note_seq")(test_case)
+
+
+def require_torchsde(test_case):
+ """
+ Decorator marking a test that requires torchsde. These tests are skipped when torchsde isn't installed.
+ """
+ return unittest.skipUnless(is_torchsde_available(), "test requires torchsde")(test_case)
+
+
+def require_peft_backend(test_case):
+ """
+ Decorator marking a test that requires PEFT backend, this would require some specific versions of PEFT and
+ transformers.
+ """
+ return unittest.skipUnless(USE_PEFT_BACKEND, "test requires PEFT backend")(test_case)
+
+
+def require_peft_version_greater(peft_version):
+ """
+ Decorator marking a test that requires PEFT backend with a specific version, this would require some specific
+ versions of PEFT and transformers.
+ """
+
+ def decorator(test_case):
+ correct_peft_version = is_peft_available() and version.parse(
+ version.parse(importlib.metadata.version("peft")).base_version
+ ) > version.parse(peft_version)
+ return unittest.skipUnless(
+ correct_peft_version, f"test requires PEFT backend with the version greater than {peft_version}"
+ )(test_case)
+
+ return decorator
+
+
+def deprecate_after_peft_backend(test_case):
+ """
+ Decorator marking a test that will be skipped after PEFT backend
+ """
+ return unittest.skipUnless(not USE_PEFT_BACKEND, "test skipped in favor of PEFT backend")(test_case)
+
+
+def require_python39_or_higher(test_case):
+ def python39_available():
+ sys_info = sys.version_info
+ major, minor = sys_info.major, sys_info.minor
+ return major == 3 and minor >= 9
+
+ return unittest.skipUnless(python39_available(), "test requires Python 3.9 or higher")(test_case)
+
+
+def load_numpy(arry: Union[str, np.ndarray], local_path: Optional[str] = None) -> np.ndarray:
+ if isinstance(arry, str):
+ if local_path is not None:
+ # local_path can be passed to correct images of tests
+ return Path(local_path, arry.split("/")[-5], arry.split("/")[-2], arry.split("/")[-1]).as_posix()
+ elif arry.startswith("http://") or arry.startswith("https://"):
+ response = requests.get(arry)
+ response.raise_for_status()
+ arry = np.load(BytesIO(response.content))
+ elif os.path.isfile(arry):
+ arry = np.load(arry)
+ else:
+ raise ValueError(
+ f"Incorrect path or url, URLs must start with `http://` or `https://`, and {arry} is not a valid path"
+ )
+ elif isinstance(arry, np.ndarray):
+ pass
+ else:
+ raise ValueError(
+ "Incorrect format used for numpy ndarray. Should be an url linking to an image, a local path, or a"
+ " ndarray."
+ )
+
+ return arry
+
+
+def load_pt(url: str):
+ response = requests.get(url)
+ response.raise_for_status()
+ arry = torch.load(BytesIO(response.content))
+ return arry
+
+
+def load_image(image: Union[str, PIL.Image.Image]) -> PIL.Image.Image:
+ """
+ Loads `image` to a PIL Image.
+
+ Args:
+ image (`str` or `PIL.Image.Image`):
+ The image to convert to the PIL Image format.
+ Returns:
+ `PIL.Image.Image`:
+ A PIL Image.
+ """
+ if isinstance(image, str):
+ if image.startswith("http://") or image.startswith("https://"):
+ image = PIL.Image.open(requests.get(image, stream=True).raw)
+ elif os.path.isfile(image):
+ image = PIL.Image.open(image)
+ else:
+ raise ValueError(
+ f"Incorrect path or url, URLs must start with `http://` or `https://`, and {image} is not a valid path"
+ )
+ elif isinstance(image, PIL.Image.Image):
+ image = image
+ else:
+ raise ValueError(
+ "Incorrect format used for image. Should be an url linking to an image, a local path, or a PIL image."
+ )
+ image = PIL.ImageOps.exif_transpose(image)
+ image = image.convert("RGB")
+ return image
+
+
+def preprocess_image(image: PIL.Image, batch_size: int):
+ w, h = image.size
+ w, h = (x - x % 8 for x in (w, h)) # resize to integer multiple of 8
+ image = image.resize((w, h), resample=PIL.Image.LANCZOS)
+ image = np.array(image).astype(np.float32) / 255.0
+ image = np.vstack([image[None].transpose(0, 3, 1, 2)] * batch_size)
+ image = torch.from_numpy(image)
+ return 2.0 * image - 1.0
+
+
+def export_to_gif(image: List[PIL.Image.Image], output_gif_path: str = None) -> str:
+ if output_gif_path is None:
+ output_gif_path = tempfile.NamedTemporaryFile(suffix=".gif").name
+
+ image[0].save(
+ output_gif_path,
+ save_all=True,
+ append_images=image[1:],
+ optimize=False,
+ duration=100,
+ loop=0,
+ )
+ return output_gif_path
+
+
+@contextmanager
+def buffered_writer(raw_f):
+ f = io.BufferedWriter(raw_f)
+ yield f
+ f.flush()
+
+
+def export_to_ply(mesh, output_ply_path: str = None):
+ """
+ Write a PLY file for a mesh.
+ """
+ if output_ply_path is None:
+ output_ply_path = tempfile.NamedTemporaryFile(suffix=".ply").name
+
+ coords = mesh.verts.detach().cpu().numpy()
+ faces = mesh.faces.cpu().numpy()
+ rgb = np.stack([mesh.vertex_channels[x].detach().cpu().numpy() for x in "RGB"], axis=1)
+
+ with buffered_writer(open(output_ply_path, "wb")) as f:
+ f.write(b"ply\n")
+ f.write(b"format binary_little_endian 1.0\n")
+ f.write(bytes(f"element vertex {len(coords)}\n", "ascii"))
+ f.write(b"property float x\n")
+ f.write(b"property float y\n")
+ f.write(b"property float z\n")
+ if rgb is not None:
+ f.write(b"property uchar red\n")
+ f.write(b"property uchar green\n")
+ f.write(b"property uchar blue\n")
+ if faces is not None:
+ f.write(bytes(f"element face {len(faces)}\n", "ascii"))
+ f.write(b"property list uchar int vertex_index\n")
+ f.write(b"end_header\n")
+
+ if rgb is not None:
+ rgb = (rgb * 255.499).round().astype(int)
+ vertices = [
+ (*coord, *rgb)
+ for coord, rgb in zip(
+ coords.tolist(),
+ rgb.tolist(),
+ )
+ ]
+ format = struct.Struct("<3f3B")
+ for item in vertices:
+ f.write(format.pack(*item))
+ else:
+ format = struct.Struct("<3f")
+ for vertex in coords.tolist():
+ f.write(format.pack(*vertex))
+
+ if faces is not None:
+ format = struct.Struct(" str:
+ if is_opencv_available():
+ import cv2
+ else:
+ raise ImportError(BACKENDS_MAPPING["opencv"][1].format("export_to_video"))
+ if output_video_path is None:
+ output_video_path = tempfile.NamedTemporaryFile(suffix=".mp4").name
+
+ fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+ h, w, c = video_frames[0].shape
+ video_writer = cv2.VideoWriter(output_video_path, fourcc, fps=8, frameSize=(w, h))
+ for i in range(len(video_frames)):
+ img = cv2.cvtColor(video_frames[i], cv2.COLOR_RGB2BGR)
+ video_writer.write(img)
+ return output_video_path
+
+
+def load_hf_numpy(path) -> np.ndarray:
+ base_url = "https://huggingface.co/datasets/fusing/diffusers-testing/resolve/main"
+
+ if not path.startswith("http://") and not path.startswith("https://"):
+ path = os.path.join(base_url, urllib.parse.quote(path))
+
+ return load_numpy(path)
+
+
+# --- pytest conf functions --- #
+
+# to avoid multiple invocation from tests/conftest.py and examples/conftest.py - make sure it's called only once
+pytest_opt_registered = {}
+
+
+def pytest_addoption_shared(parser):
+ """
+ This function is to be called from `conftest.py` via `pytest_addoption` wrapper that has to be defined there.
+
+ It allows loading both `conftest.py` files at once without causing a failure due to adding the same `pytest`
+ option.
+
+ """
+ option = "--make-reports"
+ if option not in pytest_opt_registered:
+ parser.addoption(
+ option,
+ action="store",
+ default=False,
+ help="generate report files. The value of this option is used as a prefix to report names",
+ )
+ pytest_opt_registered[option] = 1
+
+
+def pytest_terminal_summary_main(tr, id):
+ """
+ Generate multiple reports at the end of test suite run - each report goes into a dedicated file in the current
+ directory. The report files are prefixed with the test suite name.
+
+ This function emulates --duration and -rA pytest arguments.
+
+ This function is to be called from `conftest.py` via `pytest_terminal_summary` wrapper that has to be defined
+ there.
+
+ Args:
+ - tr: `terminalreporter` passed from `conftest.py`
+ - id: unique id like `tests` or `examples` that will be incorporated into the final reports filenames - this is
+ needed as some jobs have multiple runs of pytest, so we can't have them overwrite each other.
+
+ NB: this functions taps into a private _pytest API and while unlikely, it could break should
+ pytest do internal changes - also it calls default internal methods of terminalreporter which
+ can be hijacked by various `pytest-` plugins and interfere.
+
+ """
+ from _pytest.config import create_terminal_writer
+
+ if not len(id):
+ id = "tests"
+
+ config = tr.config
+ orig_writer = config.get_terminal_writer()
+ orig_tbstyle = config.option.tbstyle
+ orig_reportchars = tr.reportchars
+
+ dir = "reports"
+ Path(dir).mkdir(parents=True, exist_ok=True)
+ report_files = {
+ k: f"{dir}/{id}_{k}.txt"
+ for k in [
+ "durations",
+ "errors",
+ "failures_long",
+ "failures_short",
+ "failures_line",
+ "passes",
+ "stats",
+ "summary_short",
+ "warnings",
+ ]
+ }
+
+ # custom durations report
+ # note: there is no need to call pytest --durations=XX to get this separate report
+ # adapted from https://github.com/pytest-dev/pytest/blob/897f151e/src/_pytest/runner.py#L66
+ dlist = []
+ for replist in tr.stats.values():
+ for rep in replist:
+ if hasattr(rep, "duration"):
+ dlist.append(rep)
+ if dlist:
+ dlist.sort(key=lambda x: x.duration, reverse=True)
+ with open(report_files["durations"], "w") as f:
+ durations_min = 0.05 # sec
+ f.write("slowest durations\n")
+ for i, rep in enumerate(dlist):
+ if rep.duration < durations_min:
+ f.write(f"{len(dlist)-i} durations < {durations_min} secs were omitted")
+ break
+ f.write(f"{rep.duration:02.2f}s {rep.when:<8} {rep.nodeid}\n")
+
+ def summary_failures_short(tr):
+ # expecting that the reports were --tb=long (default) so we chop them off here to the last frame
+ reports = tr.getreports("failed")
+ if not reports:
+ return
+ tr.write_sep("=", "FAILURES SHORT STACK")
+ for rep in reports:
+ msg = tr._getfailureheadline(rep)
+ tr.write_sep("_", msg, red=True, bold=True)
+ # chop off the optional leading extra frames, leaving only the last one
+ longrepr = re.sub(r".*_ _ _ (_ ){10,}_ _ ", "", rep.longreprtext, 0, re.M | re.S)
+ tr._tw.line(longrepr)
+ # note: not printing out any rep.sections to keep the report short
+
+ # use ready-made report funcs, we are just hijacking the filehandle to log to a dedicated file each
+ # adapted from https://github.com/pytest-dev/pytest/blob/897f151e/src/_pytest/terminal.py#L814
+ # note: some pytest plugins may interfere by hijacking the default `terminalreporter` (e.g.
+ # pytest-instafail does that)
+
+ # report failures with line/short/long styles
+ config.option.tbstyle = "auto" # full tb
+ with open(report_files["failures_long"], "w") as f:
+ tr._tw = create_terminal_writer(config, f)
+ tr.summary_failures()
+
+ # config.option.tbstyle = "short" # short tb
+ with open(report_files["failures_short"], "w") as f:
+ tr._tw = create_terminal_writer(config, f)
+ summary_failures_short(tr)
+
+ config.option.tbstyle = "line" # one line per error
+ with open(report_files["failures_line"], "w") as f:
+ tr._tw = create_terminal_writer(config, f)
+ tr.summary_failures()
+
+ with open(report_files["errors"], "w") as f:
+ tr._tw = create_terminal_writer(config, f)
+ tr.summary_errors()
+
+ with open(report_files["warnings"], "w") as f:
+ tr._tw = create_terminal_writer(config, f)
+ tr.summary_warnings() # normal warnings
+ tr.summary_warnings() # final warnings
+
+ tr.reportchars = "wPpsxXEf" # emulate -rA (used in summary_passes() and short_test_summary())
+ with open(report_files["passes"], "w") as f:
+ tr._tw = create_terminal_writer(config, f)
+ tr.summary_passes()
+
+ with open(report_files["summary_short"], "w") as f:
+ tr._tw = create_terminal_writer(config, f)
+ tr.short_test_summary()
+
+ with open(report_files["stats"], "w") as f:
+ tr._tw = create_terminal_writer(config, f)
+ tr.summary_stats()
+
+ # restore:
+ tr._tw = orig_writer
+ tr.reportchars = orig_reportchars
+ config.option.tbstyle = orig_tbstyle
+
+
+# Copied from https://github.com/huggingface/transformers/blob/000e52aec8850d3fe2f360adc6fd256e5b47fe4c/src/transformers/testing_utils.py#L1905
+def is_flaky(max_attempts: int = 5, wait_before_retry: Optional[float] = None, description: Optional[str] = None):
+ """
+ To decorate flaky tests. They will be retried on failures.
+
+ Args:
+ max_attempts (`int`, *optional*, defaults to 5):
+ The maximum number of attempts to retry the flaky test.
+ wait_before_retry (`float`, *optional*):
+ If provided, will wait that number of seconds before retrying the test.
+ description (`str`, *optional*):
+ A string to describe the situation (what / where / why is flaky, link to GH issue/PR comments, errors,
+ etc.)
+ """
+
+ def decorator(test_func_ref):
+ @functools.wraps(test_func_ref)
+ def wrapper(*args, **kwargs):
+ retry_count = 1
+
+ while retry_count < max_attempts:
+ try:
+ return test_func_ref(*args, **kwargs)
+
+ except Exception as err:
+ print(f"Test failed with {err} at try {retry_count}/{max_attempts}.", file=sys.stderr)
+ if wait_before_retry is not None:
+ time.sleep(wait_before_retry)
+ retry_count += 1
+
+ return test_func_ref(*args, **kwargs)
+
+ return wrapper
+
+ return decorator
+
+
+# Taken from: https://github.com/huggingface/transformers/blob/3658488ff77ff8d45101293e749263acf437f4d5/src/transformers/testing_utils.py#L1787
+def run_test_in_subprocess(test_case, target_func, inputs=None, timeout=None):
+ """
+ To run a test in a subprocess. In particular, this can avoid (GPU) memory issue.
+
+ Args:
+ test_case (`unittest.TestCase`):
+ The test that will run `target_func`.
+ target_func (`Callable`):
+ The function implementing the actual testing logic.
+ inputs (`dict`, *optional*, defaults to `None`):
+ The inputs that will be passed to `target_func` through an (input) queue.
+ timeout (`int`, *optional*, defaults to `None`):
+ The timeout (in seconds) that will be passed to the input and output queues. If not specified, the env.
+ variable `PYTEST_TIMEOUT` will be checked. If still `None`, its value will be set to `600`.
+ """
+ if timeout is None:
+ timeout = int(os.environ.get("PYTEST_TIMEOUT", 600))
+
+ start_methohd = "spawn"
+ ctx = multiprocessing.get_context(start_methohd)
+
+ input_queue = ctx.Queue(1)
+ output_queue = ctx.JoinableQueue(1)
+
+ # We can't send `unittest.TestCase` to the child, otherwise we get issues regarding pickle.
+ input_queue.put(inputs, timeout=timeout)
+
+ process = ctx.Process(target=target_func, args=(input_queue, output_queue, timeout))
+ process.start()
+ # Kill the child process if we can't get outputs from it in time: otherwise, the hanging subprocess prevents
+ # the test to exit properly.
+ try:
+ results = output_queue.get(timeout=timeout)
+ output_queue.task_done()
+ except Exception as e:
+ process.terminate()
+ test_case.fail(e)
+ process.join(timeout=timeout)
+
+ if results["error"] is not None:
+ test_case.fail(f'{results["error"]}')
+
+
+class CaptureLogger:
+ """
+ Args:
+ Context manager to capture `logging` streams
+ logger: 'logging` logger object
+ Returns:
+ The captured output is available via `self.out`
+ Example:
+ ```python
+ >>> from diffusers import logging
+ >>> from diffusers.testing_utils import CaptureLogger
+
+ >>> msg = "Testing 1, 2, 3"
+ >>> logging.set_verbosity_info()
+ >>> logger = logging.get_logger("diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.py")
+ >>> with CaptureLogger(logger) as cl:
+ ... logger.info(msg)
+ >>> assert cl.out, msg + "\n"
+ ```
+ """
+
+ def __init__(self, logger):
+ self.logger = logger
+ self.io = StringIO()
+ self.sh = logging.StreamHandler(self.io)
+ self.out = ""
+
+ def __enter__(self):
+ self.logger.addHandler(self.sh)
+ return self
+
+ def __exit__(self, *exc):
+ self.logger.removeHandler(self.sh)
+ self.out = self.io.getvalue()
+
+ def __repr__(self):
+ return f"captured: {self.out}\n"
+
+
+def enable_full_determinism():
+ """
+ Helper function for reproducible behavior during distributed training. See
+ - https://pytorch.org/docs/stable/notes/randomness.html for pytorch
+ """
+ # Enable PyTorch deterministic mode. This potentially requires either the environment
+ # variable 'CUDA_LAUNCH_BLOCKING' or 'CUBLAS_WORKSPACE_CONFIG' to be set,
+ # depending on the CUDA version, so we set them both here
+ os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
+ os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":16:8"
+ torch.use_deterministic_algorithms(True)
+
+ # Enable CUDNN deterministic mode
+ torch.backends.cudnn.deterministic = True
+ torch.backends.cudnn.benchmark = False
+ torch.backends.cuda.matmul.allow_tf32 = False
+
+
+def disable_full_determinism():
+ os.environ["CUDA_LAUNCH_BLOCKING"] = "0"
+ os.environ["CUBLAS_WORKSPACE_CONFIG"] = ""
+ torch.use_deterministic_algorithms(False)
+
+
+# Utils for custom and alternative accelerator devices
+def _is_torch_fp16_available(device):
+ if not is_torch_available():
+ return False
+
+ import torch
+
+ device = torch.device(device)
+
+ try:
+ x = torch.zeros((2, 2), dtype=torch.float16).to(device)
+ _ = torch.mul(x, x)
+ return True
+
+ except Exception as e:
+ if device.type == "cuda":
+ raise ValueError(
+ f"You have passed a device of type 'cuda' which should work with 'fp16', but 'cuda' does not seem to be correctly installed on your machine: {e}"
+ )
+
+ return False
+
+
+def _is_torch_fp64_available(device):
+ if not is_torch_available():
+ return False
+
+ import torch
+
+ device = torch.device(device)
+
+ try:
+ x = torch.zeros((2, 2), dtype=torch.float64).to(device)
+ _ = torch.mul(x, x)
+ return True
+
+ except Exception as e:
+ if device.type == "cuda":
+ raise ValueError(
+ f"You have passed a device of type 'cuda' which should work with 'fp64', but 'cuda' does not seem to be correctly installed on your machine: {e}"
+ )
+
+ return False
+
+
+# Guard these lookups for when Torch is not used - alternative accelerator support is for PyTorch
+if is_torch_available():
+ # Behaviour flags
+ BACKEND_SUPPORTS_TRAINING = {"cuda": True, "cpu": True, "mps": False, "default": True}
+
+ # Function definitions
+ BACKEND_EMPTY_CACHE = {"cuda": torch.cuda.empty_cache, "cpu": None, "mps": None, "default": None}
+ BACKEND_DEVICE_COUNT = {"cuda": torch.cuda.device_count, "cpu": lambda: 0, "mps": lambda: 0, "default": 0}
+ BACKEND_MANUAL_SEED = {"cuda": torch.cuda.manual_seed, "cpu": torch.manual_seed, "default": torch.manual_seed}
+
+
+# This dispatches a defined function according to the accelerator from the function definitions.
+def _device_agnostic_dispatch(device: str, dispatch_table: Dict[str, Callable], *args, **kwargs):
+ if device not in dispatch_table:
+ return dispatch_table["default"](*args, **kwargs)
+
+ fn = dispatch_table[device]
+
+ # Some device agnostic functions return values. Need to guard against 'None' instead at
+ # user level
+ if fn is None:
+ return None
+
+ return fn(*args, **kwargs)
+
+
+# These are callables which automatically dispatch the function specific to the accelerator
+def backend_manual_seed(device: str, seed: int):
+ return _device_agnostic_dispatch(device, BACKEND_MANUAL_SEED, seed)
+
+
+def backend_empty_cache(device: str):
+ return _device_agnostic_dispatch(device, BACKEND_EMPTY_CACHE)
+
+
+def backend_device_count(device: str):
+ return _device_agnostic_dispatch(device, BACKEND_DEVICE_COUNT)
+
+
+# These are callables which return boolean behaviour flags and can be used to specify some
+# device agnostic alternative where the feature is unsupported.
+def backend_supports_training(device: str):
+ if not is_torch_available():
+ return False
+
+ if device not in BACKEND_SUPPORTS_TRAINING:
+ device = "default"
+
+ return BACKEND_SUPPORTS_TRAINING[device]
+
+
+# Guard for when Torch is not available
+if is_torch_available():
+ # Update device function dict mapping
+ def update_mapping_from_spec(device_fn_dict: Dict[str, Callable], attribute_name: str):
+ try:
+ # Try to import the function directly
+ spec_fn = getattr(device_spec_module, attribute_name)
+ device_fn_dict[torch_device] = spec_fn
+ except AttributeError as e:
+ # If the function doesn't exist, and there is no default, throw an error
+ if "default" not in device_fn_dict:
+ raise AttributeError(
+ f"`{attribute_name}` not found in '{device_spec_path}' and no default fallback function found."
+ ) from e
+
+ if "DIFFUSERS_TEST_DEVICE_SPEC" in os.environ:
+ device_spec_path = os.environ["DIFFUSERS_TEST_DEVICE_SPEC"]
+ if not Path(device_spec_path).is_file():
+ raise ValueError(f"Specified path to device specification file is not found. Received {device_spec_path}")
+
+ try:
+ import_name = device_spec_path[: device_spec_path.index(".py")]
+ except ValueError as e:
+ raise ValueError(f"Provided device spec file is not a Python file! Received {device_spec_path}") from e
+
+ device_spec_module = importlib.import_module(import_name)
+
+ try:
+ device_name = device_spec_module.DEVICE_NAME
+ except AttributeError:
+ raise AttributeError("Device spec file did not contain `DEVICE_NAME`")
+
+ if "DIFFUSERS_TEST_DEVICE" in os.environ and torch_device != device_name:
+ msg = f"Mismatch between environment variable `DIFFUSERS_TEST_DEVICE` '{torch_device}' and device found in spec '{device_name}'\n"
+ msg += "Either unset `DIFFUSERS_TEST_DEVICE` or ensure it matches device spec name."
+ raise ValueError(msg)
+
+ torch_device = device_name
+
+ # Add one entry here for each `BACKEND_*` dictionary.
+ update_mapping_from_spec(BACKEND_MANUAL_SEED, "MANUAL_SEED_FN")
+ update_mapping_from_spec(BACKEND_EMPTY_CACHE, "EMPTY_CACHE_FN")
+ update_mapping_from_spec(BACKEND_DEVICE_COUNT, "DEVICE_COUNT_FN")
+ update_mapping_from_spec(BACKEND_SUPPORTS_TRAINING, "SUPPORTS_TRAINING")
diff --git a/model/diffusers_c/utils/torch_utils.py b/model/diffusers_c/utils/torch_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..0cf75b4fad4e17e416e5a3f72a25b1f2b3702fe1
--- /dev/null
+++ b/model/diffusers_c/utils/torch_utils.py
@@ -0,0 +1,148 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+PyTorch utilities: Utilities related to PyTorch
+"""
+
+from typing import List, Optional, Tuple, Union
+
+from . import logging
+from .import_utils import is_torch_available, is_torch_version
+
+
+if is_torch_available():
+ import torch
+ from torch.fft import fftn, fftshift, ifftn, ifftshift
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+try:
+ from torch._dynamo import allow_in_graph as maybe_allow_in_graph
+except (ImportError, ModuleNotFoundError):
+
+ def maybe_allow_in_graph(cls):
+ return cls
+
+
+def randn_tensor(
+ shape: Union[Tuple, List],
+ generator: Optional[Union[List["torch.Generator"], "torch.Generator"]] = None,
+ device: Optional["torch.device"] = None,
+ dtype: Optional["torch.dtype"] = None,
+ layout: Optional["torch.layout"] = None,
+):
+ """A helper function to create random tensors on the desired `device` with the desired `dtype`. When
+ passing a list of generators, you can seed each batch size individually. If CPU generators are passed, the tensor
+ is always created on the CPU.
+ """
+ # device on which tensor is created defaults to device
+ rand_device = device
+ batch_size = shape[0]
+
+ layout = layout or torch.strided
+ device = device or torch.device("cpu")
+
+ if generator is not None:
+ gen_device_type = generator.device.type if not isinstance(generator, list) else generator[0].device.type
+ if gen_device_type != device.type and gen_device_type == "cpu":
+ rand_device = "cpu"
+ if device != "mps":
+ logger.info(
+ f"The passed generator was created on 'cpu' even though a tensor on {device} was expected."
+ f" Tensors will be created on 'cpu' and then moved to {device}. Note that one can probably"
+ f" slighly speed up this function by passing a generator that was created on the {device} device."
+ )
+ elif gen_device_type != device.type and gen_device_type == "cuda":
+ raise ValueError(f"Cannot generate a {device} tensor from a generator of type {gen_device_type}.")
+
+ # make sure generator list of length 1 is treated like a non-list
+ if isinstance(generator, list) and len(generator) == 1:
+ generator = generator[0]
+
+ if isinstance(generator, list):
+ shape = (1,) + shape[1:]
+ latents = [
+ torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype, layout=layout)
+ for i in range(batch_size)
+ ]
+ latents = torch.cat(latents, dim=0).to(device)
+ else:
+ latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype, layout=layout).to(device)
+
+ return latents
+
+
+def is_compiled_module(module) -> bool:
+ """Check whether the module was compiled with torch.compile()"""
+ if is_torch_version("<", "2.0.0") or not hasattr(torch, "_dynamo"):
+ return False
+ return isinstance(module, torch._dynamo.eval_frame.OptimizedModule)
+
+
+def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.Tensor":
+ """Fourier filter as introduced in FreeU (https://arxiv.org/abs/2309.11497).
+
+ This version of the method comes from here:
+ https://github.com/huggingface/diffusers/pull/5164#issuecomment-1732638706
+ """
+ x = x_in
+ B, C, H, W = x.shape
+
+ # Non-power of 2 images must be float32
+ if (W & (W - 1)) != 0 or (H & (H - 1)) != 0:
+ x = x.to(dtype=torch.float32)
+
+ # FFT
+ x_freq = fftn(x, dim=(-2, -1))
+ x_freq = fftshift(x_freq, dim=(-2, -1))
+
+ B, C, H, W = x_freq.shape
+ mask = torch.ones((B, C, H, W), device=x.device)
+
+ crow, ccol = H // 2, W // 2
+ mask[..., crow - threshold : crow + threshold, ccol - threshold : ccol + threshold] = scale
+ x_freq = x_freq * mask
+
+ # IFFT
+ x_freq = ifftshift(x_freq, dim=(-2, -1))
+ x_filtered = ifftn(x_freq, dim=(-2, -1)).real
+
+ return x_filtered.to(dtype=x_in.dtype)
+
+
+def apply_freeu(
+ resolution_idx: int, hidden_states: "torch.Tensor", res_hidden_states: "torch.Tensor", **freeu_kwargs
+) -> Tuple["torch.Tensor", "torch.Tensor"]:
+ """Applies the FreeU mechanism as introduced in https:
+ //arxiv.org/abs/2309.11497. Adapted from the official code repository: https://github.com/ChenyangSi/FreeU.
+
+ Args:
+ resolution_idx (`int`): Integer denoting the UNet block where FreeU is being applied.
+ hidden_states (`torch.Tensor`): Inputs to the underlying block.
+ res_hidden_states (`torch.Tensor`): Features from the skip block corresponding to the underlying block.
+ s1 (`float`): Scaling factor for stage 1 to attenuate the contributions of the skip features.
+ s2 (`float`): Scaling factor for stage 2 to attenuate the contributions of the skip features.
+ b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+ b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+ """
+ if resolution_idx == 0:
+ num_half_channels = hidden_states.shape[1] // 2
+ hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b1"]
+ res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s1"])
+ if resolution_idx == 1:
+ num_half_channels = hidden_states.shape[1] // 2
+ hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b2"]
+ res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s2"])
+
+ return hidden_states, res_hidden_states
diff --git a/model/diffusers_c/utils/versions.py b/model/diffusers_c/utils/versions.py
new file mode 100644
index 0000000000000000000000000000000000000000..945a3977ce62a9a55307862193e4be6f12c3c17f
--- /dev/null
+++ b/model/diffusers_c/utils/versions.py
@@ -0,0 +1,117 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Utilities for working with package versions
+"""
+
+import importlib.metadata
+import operator
+import re
+import sys
+from typing import Optional
+
+from packaging import version
+
+
+ops = {
+ "<": operator.lt,
+ "<=": operator.le,
+ "==": operator.eq,
+ "!=": operator.ne,
+ ">=": operator.ge,
+ ">": operator.gt,
+}
+
+
+def _compare_versions(op, got_ver, want_ver, requirement, pkg, hint):
+ if got_ver is None or want_ver is None:
+ raise ValueError(
+ f"Unable to compare versions for {requirement}: need={want_ver} found={got_ver}. This is unusual. Consider"
+ f" reinstalling {pkg}."
+ )
+ if not ops[op](version.parse(got_ver), version.parse(want_ver)):
+ raise ImportError(
+ f"{requirement} is required for a normal functioning of this module, but found {pkg}=={got_ver}.{hint}"
+ )
+
+
+def require_version(requirement: str, hint: Optional[str] = None) -> None:
+ """
+ Perform a runtime check of the dependency versions, using the exact same syntax used by pip.
+
+ The installed module version comes from the *site-packages* dir via *importlib.metadata*.
+
+ Args:
+ requirement (`str`): pip style definition, e.g., "tokenizers==0.9.4", "tqdm>=4.27", "numpy"
+ hint (`str`, *optional*): what suggestion to print in case of requirements not being met
+
+ Example:
+
+ ```python
+ require_version("pandas>1.1.2")
+ require_version("numpy>1.18.5", "this is important to have for whatever reason")
+ ```"""
+
+ hint = f"\n{hint}" if hint is not None else ""
+
+ # non-versioned check
+ if re.match(r"^[\w_\-\d]+$", requirement):
+ pkg, op, want_ver = requirement, None, None
+ else:
+ match = re.findall(r"^([^!=<>\s]+)([\s!=<>]{1,2}.+)", requirement)
+ if not match:
+ raise ValueError(
+ "requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23, but"
+ f" got {requirement}"
+ )
+ pkg, want_full = match[0]
+ want_range = want_full.split(",") # there could be multiple requirements
+ wanted = {}
+ for w in want_range:
+ match = re.findall(r"^([\s!=<>]{1,2})(.+)", w)
+ if not match:
+ raise ValueError(
+ "requirement needs to be in the pip package format, .e.g., package_a==1.23, or package_b>=1.23,"
+ f" but got {requirement}"
+ )
+ op, want_ver = match[0]
+ wanted[op] = want_ver
+ if op not in ops:
+ raise ValueError(f"{requirement}: need one of {list(ops.keys())}, but got {op}")
+
+ # special case
+ if pkg == "python":
+ got_ver = ".".join([str(x) for x in sys.version_info[:3]])
+ for op, want_ver in wanted.items():
+ _compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
+ return
+
+ # check if any version is installed
+ try:
+ got_ver = importlib.metadata.version(pkg)
+ except importlib.metadata.PackageNotFoundError:
+ raise importlib.metadata.PackageNotFoundError(
+ f"The '{requirement}' distribution was not found and is required by this application. {hint}"
+ )
+
+ # check that the right version is installed if version number or a range was provided
+ if want_ver is not None:
+ for op, want_ver in wanted.items():
+ _compare_versions(op, got_ver, want_ver, requirement, pkg, hint)
+
+
+def require_version_core(requirement):
+ """require_version wrapper which emits a core-specific hint on failure"""
+ hint = "Try: pip install transformers -U or pip install -e '.[dev]' if you're working with git main"
+ return require_version(requirement, hint)
diff --git a/pipeline/__pycache__/pipeline_PowerPaint_Brushnet_CA.cpython-310.pyc b/pipeline/__pycache__/pipeline_PowerPaint_Brushnet_CA.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..6373b7f5c212295c45d1745e20cf3b2c8a398d7d
Binary files /dev/null and b/pipeline/__pycache__/pipeline_PowerPaint_Brushnet_CA.cpython-310.pyc differ
diff --git a/pipeline/pipeline_PowerPaint.py b/pipeline/pipeline_PowerPaint.py
new file mode 100644
index 0000000000000000000000000000000000000000..c389122b34210510d7d51a88445de25dbd74d925
--- /dev/null
+++ b/pipeline/pipeline_PowerPaint.py
@@ -0,0 +1,1077 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from diffusers.models import AsymmetricAutoencoderKL, AutoencoderKL, UNet2DConditionModel
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import deprecate, is_accelerate_available, is_accelerate_version, logging
+from diffusers.utils.torch_utils import randn_tensor
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+def prepare_mask_and_masked_image(image, mask, height, width, return_image: bool = False):
+ """
+ Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
+ converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
+ ``image`` and ``1`` for the ``mask``.
+
+ The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+ binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+ Args:
+ image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+ ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+ mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+ ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+
+
+ Raises:
+ ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+ should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+ TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+ (ot the other way around).
+
+ Returns:
+ tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
+ dimensions: ``batch x channels x height x width``.
+ """
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ if mask is None:
+ raise ValueError("`mask_image` input cannot be undefined.")
+
+ if isinstance(image, torch.Tensor):
+ if not isinstance(mask, torch.Tensor):
+ raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
+
+ # Batch single image
+ if image.ndim == 3:
+ assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
+ image = image.unsqueeze(0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Single batched mask, no channel dim or single mask not batched but channel dim
+ if mask.shape[0] == 1:
+ mask = mask.unsqueeze(0)
+
+ # Batched masks no channel dim
+ else:
+ mask = mask.unsqueeze(1)
+
+ assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+ assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+ assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+
+ # Check image is in [-1, 1]
+ if image.min() < -1 or image.max() > 1:
+ raise ValueError("Image should be in [-1, 1] range")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("Mask should be in [0, 1] range")
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ # Image as float32
+ image = image.to(dtype=torch.float32)
+ elif isinstance(mask, torch.Tensor):
+ raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+ else:
+ # preprocess image
+ if isinstance(image, (PIL.Image.Image, np.ndarray)):
+ image = [image]
+ if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+ # resize all images w.r.t passed height an width
+ image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
+ image = [np.array(i.convert("RGB"))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+ image = np.concatenate([i[None, :] for i in image], axis=0)
+
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ # preprocess mask
+ if isinstance(mask, (PIL.Image.Image, np.ndarray)):
+ mask = [mask]
+
+ if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
+ mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
+ mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+ mask = mask.astype(np.float32) / 255.0
+ elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+ mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+ mask = torch.from_numpy(mask)
+
+ masked_image = image * (mask < 0.5)
+
+ # n.b. ensure backwards compatibility as old function does not return image
+ if return_image:
+ return mask, masked_image, image
+
+ return mask, masked_image
+
+
+class StableDiffusionInpaintPipeline(
+ DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
+ r"""
+ Pipeline for text-guided image inpainting using Stable Diffusion.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+
+ Args:
+ vae ([`AutoencoderKL`, `AsymmetricAutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ _optional_components = ["safety_checker", "feature_extractor"]
+
+ def __init__(
+ self,
+ vae: Union[AutoencoderKL, AsymmetricAutoencoderKL],
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+ f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+ "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
+ " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
+ " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
+ " file"
+ )
+ deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["steps_offset"] = 1
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if hasattr(scheduler.config, "skip_prk_steps") and scheduler.config.skip_prk_steps is False:
+ deprecation_message = (
+ f"The configuration file of this scheduler: {scheduler} has not set the configuration"
+ " `skip_prk_steps`. `skip_prk_steps` should be set to True in the configuration file. Please make"
+ " sure to update the config accordingly as not setting `skip_prk_steps` in the config might lead to"
+ " incorrect results in future versions. If you have downloaded this checkpoint from the Hugging Face"
+ " Hub, it would be very nice if you could open a Pull request for the"
+ " `scheduler/scheduler_config.json` file"
+ )
+ deprecate("skip_prk_steps not set", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(scheduler.config)
+ new_config["skip_prk_steps"] = True
+ scheduler._internal_dict = FrozenDict(new_config)
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+ version.parse(unet.config._diffusers_version).base_version
+ ) < version.parse("0.9.0.dev0")
+ is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+ if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+ deprecation_message = (
+ "The configuration file of the unet has set the default `sample_size` to smaller than"
+ " 64 which seems highly unlikely .If you're checkpoint is a fine-tuned version of any of the"
+ " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+ " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+ " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+ " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+ " in the config might lead to incorrect results in future versions. If you have downloaded this"
+ " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
+ " the `unet/config.json` file"
+ )
+ deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+ new_config = dict(unet.config)
+ new_config["sample_size"] = 64
+ unet._internal_dict = FrozenDict(new_config)
+
+ # Check shapes, assume num_channels_latents == 4, num_channels_mask == 1, num_channels_masked == 4
+ if unet.config.in_channels != 9:
+ logger.info(f"You have loaded a UNet with {unet.config.in_channels} input channels which.")
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_model_cpu_offload
+ def enable_model_cpu_offload(self, gpu_id=0):
+ r"""
+ Offload all models to CPU to reduce memory usage with a low impact on performance. Moves one whole model at a
+ time to the GPU when its `forward` method is called, and the model remains in GPU until the next model runs.
+ Memory savings are lower than using `enable_sequential_cpu_offload`, but performance is much better due to the
+ iterative execution of the `unet`.
+ """
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+ from accelerate import cpu_offload_with_hook
+ else:
+ raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+ device = torch.device(f"cuda:{gpu_id}")
+
+ if self.device.type != "cpu":
+ self.to("cpu", silence_dtype_warnings=True)
+ torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist)
+
+ hook = None
+ for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+ _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+ if self.safety_checker is not None:
+ _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+ # We'll offload the last model manually.
+ self.final_offload_hook = hook
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ promptA,
+ promptB,
+ t,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_promptA=None,
+ negative_promptB=None,
+ t_nag=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ prompt = promptA
+ negative_prompt = negative_promptA
+
+ if promptA is not None and isinstance(promptA, str):
+ batch_size = 1
+ elif promptA is not None and isinstance(promptA, list):
+ batch_size = len(promptA)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: procecss multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ promptA = self.maybe_convert_prompt(promptA, self.tokenizer)
+
+ text_inputsA = self.tokenizer(
+ promptA,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_inputsB = self.tokenizer(
+ promptB,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_idsA = text_inputsA.input_ids
+ text_input_idsB = text_inputsB.input_ids
+ untruncated_ids = self.tokenizer(promptA, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_idsA.shape[-1] and not torch.equal(
+ text_input_idsA, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputsA.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ # print("text_input_idsA: ",text_input_idsA)
+ # print("text_input_idsB: ",text_input_idsB)
+ # print('t: ',t)
+
+ prompt_embedsA = self.text_encoder(
+ text_input_idsA.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embedsA = prompt_embedsA[0]
+
+ prompt_embedsB = self.text_encoder(
+ text_input_idsB.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embedsB = prompt_embedsB[0]
+ prompt_embeds = prompt_embedsA * (t) + (1 - t) * prompt_embedsB
+ # print("prompt_embeds: ",prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokensA: List[str]
+ uncond_tokensB: List[str]
+ if negative_prompt is None:
+ uncond_tokensA = [""] * batch_size
+ uncond_tokensB = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokensA = [negative_promptA]
+ uncond_tokensB = [negative_promptB]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokensA = negative_promptA
+ uncond_tokensB = negative_promptB
+
+ # textual inversion: procecss multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokensA = self.maybe_convert_prompt(uncond_tokensA, self.tokenizer)
+ uncond_tokensB = self.maybe_convert_prompt(uncond_tokensB, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_inputA = self.tokenizer(
+ uncond_tokensA,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_inputB = self.tokenizer(
+ uncond_tokensB,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_inputA.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embedsA = self.text_encoder(
+ uncond_inputA.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embedsB = self.text_encoder(
+ uncond_inputB.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embedsA[0] * (t_nag) + (1 - t_nag) * negative_prompt_embedsB[0]
+
+ # negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ # print("prompt_embeds: ",prompt_embeds)
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ strength,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ):
+ if strength < 0 or strength > 1:
+ raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}")
+
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ def prepare_latents(
+ self,
+ batch_size,
+ num_channels_latents,
+ height,
+ width,
+ dtype,
+ device,
+ generator,
+ latents=None,
+ image=None,
+ timestep=None,
+ is_strength_max=True,
+ return_noise=False,
+ return_image_latents=False,
+ ):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if (image is None or timestep is None) and not is_strength_max:
+ raise ValueError(
+ "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
+ "However, either the image or the noise timestep has not been provided."
+ )
+
+ if return_image_latents or (latents is None and not is_strength_max):
+ image = image.to(device=device, dtype=dtype)
+ image_latents = self._encode_vae_image(image=image, generator=generator)
+
+ if latents is None:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ # if strength is 1. then initialise the latents to noise, else initial to image + noise
+ latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
+ # if pure noise then scale the initial latents by the Scheduler's init sigma
+ latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+ else:
+ noise = latents.to(device)
+ latents = noise * self.scheduler.init_noise_sigma
+
+ outputs = (latents,)
+
+ if return_noise:
+ outputs += (noise,)
+
+ if return_image_latents:
+ outputs += (image_latents,)
+
+ return outputs
+
+ def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+ if isinstance(generator, list):
+ image_latents = [
+ self.vae.encode(image[i : i + 1]).latent_dist.sample(generator=generator[i])
+ for i in range(image.shape[0])
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = self.vae.encode(image).latent_dist.sample(generator=generator)
+
+ image_latents = self.vae.config.scaling_factor * image_latents
+
+ return image_latents
+
+ def prepare_mask_latents(
+ self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+ ):
+ # resize the mask to latents shape as we concatenate the mask to the latents
+ # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+ # and half precision
+ mask = torch.nn.functional.interpolate(
+ mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+ )
+ mask = mask.to(device=device, dtype=dtype)
+
+ masked_image = masked_image.to(device=device, dtype=dtype)
+ masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+
+ # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+ if mask.shape[0] < batch_size:
+ if not batch_size % mask.shape[0] == 0:
+ raise ValueError(
+ "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+ f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+ " of masks that you pass is divisible by the total requested batch size."
+ )
+ mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+ if masked_image_latents.shape[0] < batch_size:
+ if not batch_size % masked_image_latents.shape[0] == 0:
+ raise ValueError(
+ "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+ f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+ " Make sure the number of images that you pass is divisible by the total requested batch size."
+ )
+ masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)
+
+ mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+ masked_image_latents = (
+ torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+ )
+
+ # aligning device to prevent device errors when concating it with the latent model input
+ masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+ return mask, masked_image_latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+ return timesteps, num_inference_steps - t_start
+
+ @torch.no_grad()
+ def __call__(
+ self,
+ promptA: Union[str, List[str]] = None,
+ promptB: Union[str, List[str]] = None,
+ image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ strength: float = 1.0,
+ tradoff: float = 1.0,
+ tradoff_nag: float = 1.0,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_promptA: Optional[Union[str, List[str]]] = None,
+ negative_promptB: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ task_class: Union[torch.Tensor, float, int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`PIL.Image.Image`):
+ `Image` or tensor representing an image batch to be inpainted (which parts of the image to be masked
+ out with `mask_image` and repainted according to `prompt`).
+ mask_image (`PIL.Image.Image`):
+ `Image` or tensor representing an image batch to mask `image`. White pixels in the mask are repainted
+ while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a single channel
+ (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, so the
+ expected shape would be `(B, H, W, 1)`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ strength (`float`, *optional*, defaults to 1.0):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+
+ Examples:
+
+ ```py
+ >>> import PIL
+ >>> import requests
+ >>> import torch
+ >>> from io import BytesIO
+
+ >>> from diffusers import StableDiffusionInpaintPipeline
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+ >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+ >>> init_image = download_image(img_url).resize((512, 512))
+ >>> mask_image = download_image(mask_url).resize((512, 512))
+
+ >>> pipe = StableDiffusionInpaintPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+ >>> image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+ prompt = promptA
+ negative_prompt = negative_promptA
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ strength,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds = self._encode_prompt(
+ promptA,
+ promptB,
+ tradoff,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_promptA,
+ negative_promptB,
+ tradoff_nag,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ )
+
+ # 4. set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(
+ num_inference_steps=num_inference_steps, strength=strength, device=device
+ )
+ # check that number of inference steps is not < 1 - as this doesn't make sense
+ if num_inference_steps < 1:
+ raise ValueError(
+ f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+ f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+ )
+ # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+ is_strength_max = strength == 1.0
+
+ # 5. Preprocess mask and image
+ mask, masked_image, init_image = prepare_mask_and_masked_image(
+ image, mask_image, height, width, return_image=True
+ )
+ mask_condition = mask.clone()
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ num_channels_unet = self.unet.config.in_channels
+ return_image_latents = num_channels_unet == 4
+
+ latents_outputs = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ image=init_image,
+ timestep=latent_timestep,
+ is_strength_max=is_strength_max,
+ return_noise=True,
+ return_image_latents=return_image_latents,
+ )
+
+ if return_image_latents:
+ latents, noise, image_latents = latents_outputs
+ else:
+ latents, noise = latents_outputs
+
+ # 7. Prepare mask latent variables
+ mask, masked_image_latents = self.prepare_mask_latents(
+ mask,
+ masked_image,
+ batch_size * num_images_per_prompt,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ do_classifier_free_guidance,
+ )
+
+ # 8. Check that sizes of mask, masked image and latents match
+ if num_channels_unet == 9:
+ # default case for runwayml/stable-diffusion-inpainting
+ num_channels_mask = mask.shape[1]
+ num_channels_masked_image = masked_image_latents.shape[1]
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+ elif num_channels_unet != 4:
+ raise ValueError(
+ f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
+ )
+
+ # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 10. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ if num_channels_unet == 9:
+ latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+ # predict the noise residual
+ if task_class is not None:
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ task_class=task_class,
+ )[0]
+ else:
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if num_channels_unet == 4:
+ init_latents_proper = image_latents[:1]
+ init_mask = mask[:1]
+
+ if i < len(timesteps) - 1:
+ noise_timestep = timesteps[i + 1]
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_proper, noise, torch.tensor([noise_timestep])
+ )
+
+ latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, latents)
+
+ if not output_type == "latent":
+ condition_kwargs = {}
+ if isinstance(self.vae, AsymmetricAutoencoderKL):
+ init_image = init_image.to(device=device, dtype=masked_image_latents.dtype)
+ init_image_condition = init_image.clone()
+ init_image = self._encode_vae_image(init_image, generator=generator)
+ mask_condition = mask_condition.to(device=device, dtype=masked_image_latents.dtype)
+ condition_kwargs = {"image": init_image_condition, "mask": mask_condition}
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, **condition_kwargs)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload last model to CPU
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.final_offload_hook.offload()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/pipeline/pipeline_PowerPaint_Brushnet_CA.py b/pipeline/pipeline_PowerPaint_Brushnet_CA.py
new file mode 100644
index 0000000000000000000000000000000000000000..71501340fbcfcf850aa32e9a3bb8134dc951b5f3
--- /dev/null
+++ b/pipeline/pipeline_PowerPaint_Brushnet_CA.py
@@ -0,0 +1,1501 @@
+import inspect
+import sys
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+
+
+sys.path.append("..")
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from diffusers.models import AutoencoderKL
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+ USE_PEFT_BACKEND,
+ deprecate,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from diffusers.utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor
+from model.BrushNet_CA import BrushNetModel
+from model.diffusers_c.models import ImageProjection, UNet2DConditionModel
+from model.diffusers_c.pipelines.pipeline_utils import StableDiffusionMixin
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ from diffusers import StableDiffusionBrushNetPipeline, BrushNetModel, UniPCMultistepScheduler
+ from diffusers.utils import load_image
+ import torch
+ import cv2
+ import numpy as np
+ from PIL import Image
+
+ base_model_path = "runwayml/stable-diffusion-v1-5"
+ brushnet_path = "ckpt_path"
+
+ brushnet = BrushNetModel.from_pretrained(brushnet_path, torch_dtype=torch.float16)
+ pipe = StableDiffusionBrushNetPipeline.from_pretrained(
+ base_model_path, brushnet=brushnet, torch_dtype=torch.float16, low_cpu_mem_usage=False
+ )
+
+ # speed up diffusion process with faster scheduler and memory optimization
+ pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
+ # remove following line if xformers is not installed or when using Torch 2.0.
+ # pipe.enable_xformers_memory_efficient_attention()
+ # memory optimization.
+ pipe.enable_model_cpu_offload()
+
+ image_path="examples/brushnet/src/test_image.jpg"
+ mask_path="examples/brushnet/src/test_mask.jpg"
+ caption="A cake on the table."
+
+ init_image = cv2.imread(image_path)
+ mask_image = 1.*(cv2.imread(mask_path).sum(-1)>255)[:,:,np.newaxis]
+ init_image = init_image * (1-mask_image)
+
+ init_image = Image.fromarray(init_image.astype(np.uint8)).convert("RGB")
+ mask_image = Image.fromarray(mask_image.astype(np.uint8).repeat(3,-1)*255).convert("RGB")
+
+ generator = torch.Generator("cuda").manual_seed(1234)
+
+ image = pipe(
+ caption,
+ init_image,
+ mask_image,
+ num_inference_steps=50,
+ generator=generator,
+ paintingnet_conditioning_scale=1.0
+ ).images[0]
+ image.save("output.png")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
+def retrieve_timesteps(
+ scheduler,
+ num_inference_steps: Optional[int] = None,
+ device: Optional[Union[str, torch.device]] = None,
+ timesteps: Optional[List[int]] = None,
+ **kwargs,
+):
+ """
+ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+ custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+ Args:
+ scheduler (`SchedulerMixin`):
+ The scheduler to get timesteps from.
+ num_inference_steps (`int`):
+ The number of diffusion steps used when generating samples with a pre-trained model. If used,
+ `timesteps` must be `None`.
+ device (`str` or `torch.device`, *optional*):
+ The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+ timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+ must be `None`.
+
+ Returns:
+ `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+ second element is the number of inference steps.
+ """
+ if timesteps is not None:
+ accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+ if not accepts_timesteps:
+ raise ValueError(
+ f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+ f" timestep schedules. Please check whether you are using the correct scheduler."
+ )
+ scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ num_inference_steps = len(timesteps)
+ else:
+ scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+ timesteps = scheduler.timesteps
+ return timesteps, num_inference_steps
+
+
+class StableDiffusionPowerPaintBrushNetPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ LoraLoaderMixin,
+ IPAdapterMixin,
+ FromSingleFileMixin,
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with BrushNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+ text_encoder ([`~transformers.CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer ([`~transformers.CLIPTokenizer`]):
+ A `CLIPTokenizer` to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A `UNet2DConditionModel` to denoise the encoded image latents.
+ brushnet ([`BrushNetModel`]`):
+ Provides additional conditioning to the `unet` during the denoising process.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+ about a model's potential harms.
+ feature_extractor ([`~transformers.CLIPImageProcessor`]):
+ A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+ _exclude_from_cpu_offload = ["safety_checker"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ text_encoder_brushnet: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ brushnet: BrushNetModel,
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ text_encoder_brushnet=text_encoder_brushnet,
+ tokenizer=tokenizer,
+ unet=unet,
+ brushnet=brushnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True)
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ promptA,
+ promptB,
+ t,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_promptA=None,
+ negative_promptB=None,
+ t_nag=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ prompt = promptA
+ negative_prompt = negative_promptA
+
+ if promptA is not None and isinstance(promptA, str):
+ batch_size = 1
+ elif promptA is not None and isinstance(promptA, list):
+ batch_size = len(promptA)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: procecss multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ promptA = self.maybe_convert_prompt(promptA, self.tokenizer)
+
+ text_inputsA = self.tokenizer(
+ promptA,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_inputsB = self.tokenizer(
+ promptB,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_idsA = text_inputsA.input_ids
+ text_input_idsB = text_inputsB.input_ids
+ untruncated_ids = self.tokenizer(promptA, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_idsA.shape[-1] and not torch.equal(
+ text_input_idsA, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if (
+ hasattr(self.text_encoder_brushnet.config, "use_attention_mask")
+ and self.text_encoder_brushnet.config.use_attention_mask
+ ):
+ attention_mask = text_inputsA.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ # print("text_input_idsA: ",text_input_idsA)
+ # print("text_input_idsB: ",text_input_idsB)
+ # print('t: ',t)
+
+ prompt_embedsA = self.text_encoder_brushnet(
+ text_input_idsA.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embedsA = prompt_embedsA[0]
+
+ prompt_embedsB = self.text_encoder_brushnet(
+ text_input_idsB.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embedsB = prompt_embedsB[0]
+ prompt_embeds = prompt_embedsA * (t) + (1 - t) * prompt_embedsB
+ # print("prompt_embeds: ",prompt_embeds)
+
+ if self.text_encoder_brushnet is not None:
+ prompt_embeds_dtype = self.text_encoder_brushnet.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokensA: List[str]
+ uncond_tokensB: List[str]
+ if negative_prompt is None:
+ uncond_tokensA = [""] * batch_size
+ uncond_tokensB = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokensA = [negative_promptA]
+ uncond_tokensB = [negative_promptB]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokensA = negative_promptA
+ uncond_tokensB = negative_promptB
+
+ # textual inversion: procecss multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokensA = self.maybe_convert_prompt(uncond_tokensA, self.tokenizer)
+ uncond_tokensB = self.maybe_convert_prompt(uncond_tokensB, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_inputA = self.tokenizer(
+ uncond_tokensA,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_inputB = self.tokenizer(
+ uncond_tokensB,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if (
+ hasattr(self.text_encoder_brushnet.config, "use_attention_mask")
+ and self.text_encoder_brushnet.config.use_attention_mask
+ ):
+ attention_mask = uncond_inputA.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embedsA = self.text_encoder_brushnet(
+ uncond_inputA.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embedsB = self.text_encoder_brushnet(
+ uncond_inputB.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embedsA[0] * (t_nag) + (1 - t_nag) * negative_prompt_embedsB[0]
+
+ # negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ # print("prompt_embeds: ",prompt_embeds)
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ # print('1 ',prompt,negative_prompt)
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+ # print('2 ',prompt,negative_prompt)
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+ # print('3 ',prompt,negative_prompt)
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ # print('4 ',prompt,negative_prompt)
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ # print('5 ',prompt,negative_prompt)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ # print(prompt, text_input_ids)
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ # print("neg: ", uncond_input.input_ids)
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+ image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ if ip_adapter_image_embeds is None:
+ if not isinstance(ip_adapter_image, list):
+ ip_adapter_image = [ip_adapter_image]
+
+ if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+ raise ValueError(
+ f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ image_embeds = []
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+ single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+ single_negative_image_embeds = torch.stack(
+ [single_negative_image_embeds] * num_images_per_prompt, dim=0
+ )
+
+ if do_classifier_free_guidance:
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ single_image_embeds = single_image_embeds.to(device)
+
+ image_embeds.append(single_image_embeds)
+ else:
+ repeat_dims = [1]
+ image_embeds = []
+ for single_image_embeds in ip_adapter_image_embeds:
+ if do_classifier_free_guidance:
+ single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ single_negative_image_embeds = single_negative_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+ )
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+ else:
+ single_image_embeds = single_image_embeds.repeat(
+ num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+ )
+ image_embeds.append(single_image_embeds)
+
+ return image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+ deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ mask,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ brushnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ callback_on_step_end_tensor_inputs=None,
+ ):
+ if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.brushnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+ if (
+ isinstance(self.brushnet, BrushNetModel)
+ or is_compiled
+ and isinstance(self.brushnet._orig_mod, BrushNetModel)
+ ):
+ self.check_image(image, mask, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `brushnet_conditioning_scale`
+ if (
+ isinstance(self.brushnet, BrushNetModel)
+ or is_compiled
+ and isinstance(self.brushnet._orig_mod, BrushNetModel)
+ ):
+ if not isinstance(brushnet_conditioning_scale, float):
+ raise TypeError("For single brushnet: `brushnet_conditioning_scale` must be type `float`.")
+ else:
+ assert False
+
+ if not isinstance(control_guidance_start, (tuple, list)):
+ control_guidance_start = [control_guidance_start]
+
+ if not isinstance(control_guidance_end, (tuple, list)):
+ control_guidance_end = [control_guidance_end]
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ def check_image(self, image, mask, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ mask_is_pil = isinstance(mask, PIL.Image.Image)
+ mask_is_tensor = isinstance(mask, torch.Tensor)
+ mask_is_np = isinstance(mask, np.ndarray)
+ mask_is_pil_list = isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image)
+ mask_is_tensor_list = isinstance(mask, list) and isinstance(mask[0], torch.Tensor)
+ mask_is_np_list = isinstance(mask, list) and isinstance(mask[0], np.ndarray)
+
+ if (
+ not mask_is_pil
+ and not mask_is_tensor
+ and not mask_is_np
+ and not mask_is_pil_list
+ and not mask_is_tensor_list
+ and not mask_is_np_list
+ ):
+ raise TypeError(
+ f"mask must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(mask)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ def prepare_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image.to(device=device, dtype=dtype)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+ def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ noise = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = noise * self.scheduler.init_noise_sigma
+ return latents, noise
+
+ # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+ def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+ """
+ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+ Args:
+ timesteps (`torch.Tensor`):
+ generate embedding vectors at these timesteps
+ embedding_dim (`int`, *optional*, defaults to 512):
+ dimension of the embeddings to generate
+ dtype:
+ data type of the generated embeddings
+
+ Returns:
+ `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+ """
+ assert len(w.shape) == 1
+ w = w * 1000.0
+
+ half_dim = embedding_dim // 2
+ emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+ emb = w.to(dtype)[:, None] * emb[None, :]
+ emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+ if embedding_dim % 2 == 1: # zero pad
+ emb = torch.nn.functional.pad(emb, (0, 1))
+ assert emb.shape == (w.shape[0], embedding_dim)
+ return emb
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ promptA: Union[str, List[str]] = None,
+ promptB: Union[str, List[str]] = None,
+ promptU: Union[str, List[str]] = None,
+ tradoff: float = 1.0,
+ tradoff_nag: float = 1.0,
+ image: PipelineImageInput = None,
+ mask: PipelineImageInput = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_inference_steps: int = 50,
+ timesteps: List[int] = None,
+ guidance_scale: float = 7.5,
+ negative_promptA: Optional[Union[str, List[str]]] = None,
+ negative_promptB: Optional[Union[str, List[str]]] = None,
+ negative_promptU: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ brushnet_conditioning_scale: Union[float, List[float]] = 1.0,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ **kwargs,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The BrushNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to BrushNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple BrushNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single BrushNet. When `prompt` is a list, and if a list of images is passed for a single BrushNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple BrushNets,
+ where a list of image lists can be passed to batch for each prompt and each BrushNet.
+ mask (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,:
+ `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`):
+ The BrushNet input condition to provide guidance to the `unet` for generation. If the type is
+ specified as `torch.FloatTensor`, it is passed to BrushNet as is. `PIL.Image.Image` can also be
+ accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height
+ and/or width are passed, `image` is resized accordingly. If multiple BrushNets are specified in
+ `init`, images must be passed as a list such that each element of the list can be correctly batched for
+ input to a single BrushNet. When `prompt` is a list, and if a list of images is passed for a single BrushNet,
+ each will be paired with each prompt in the `prompt` list. This also applies to multiple BrushNets,
+ where a list of image lists can be passed to batch for each prompt and each BrushNet.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ timesteps (`List[int]`, *optional*):
+ Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+ in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+ passed will be used. Must be in descending order.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+ Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+ if `do_classifier_free_guidance` is set to `True`.
+ If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ brushnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the BrushNet are multiplied by `brushnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple BrushNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ The BrushNet encoder tries to recognize the content of the input image even if you remove all
+ prompts. A `guidance_scale` value between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the BrushNet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the BrushNet stops applying.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeine class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+
+ callback = kwargs.pop("callback", None)
+ callback_steps = kwargs.pop("callback_steps", None)
+
+ if callback is not None:
+ deprecate(
+ "callback",
+ "1.0.0",
+ "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+ if callback_steps is not None:
+ deprecate(
+ "callback_steps",
+ "1.0.0",
+ "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+ )
+
+ brushnet = self.brushnet._orig_mod if is_compiled_module(self.brushnet) else self.brushnet
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ control_guidance_start, control_guidance_end = (
+ [control_guidance_start],
+ [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ prompt = promptA
+ negative_prompt = negative_promptA
+ self.check_inputs(
+ prompt,
+ image,
+ mask,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ brushnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ callback_on_step_end_tensor_inputs,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ global_pool_conditions = (
+ brushnet.config.global_pool_conditions
+ if isinstance(brushnet, BrushNetModel)
+ else brushnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+
+ prompt_embeds = self._encode_prompt(
+ promptA,
+ promptB,
+ tradoff,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_promptA,
+ negative_promptB,
+ tradoff_nag,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ )
+ prompt_embedsU = None
+ negative_prompt_embedsU = None
+ prompt_embedsU = self.encode_prompt(
+ promptU,
+ device,
+ num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_promptU,
+ prompt_embeds=prompt_embedsU,
+ negative_prompt_embeds=negative_prompt_embedsU,
+ lora_scale=text_encoder_lora_scale,
+ )
+
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_images_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 4. Prepare image
+ if isinstance(brushnet, BrushNetModel):
+ image = self.prepare_image(
+ image=image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=brushnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ original_mask = self.prepare_image(
+ image=mask,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=brushnet.dtype,
+ do_classifier_free_guidance=self.do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ original_mask = (original_mask.sum(1)[:, None, :, :] < 0).to(image.dtype)
+ height, width = image.shape[-2:]
+ else:
+ assert False
+
+ # 5. Prepare timesteps
+ timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+ self._num_timesteps = len(timesteps)
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents, noise = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 6.1 prepare condition latents
+ # mask_i = transforms.ToPILImage()(image[0:1,:,:,:].squeeze(0))
+ # mask_i.save('_mask.png')
+ # print(brushnet.dtype)
+ conditioning_latents = (
+ self.vae.encode(image.to(device=device, dtype=brushnet.dtype)).latent_dist.sample()
+ * self.vae.config.scaling_factor
+ )
+ mask = torch.nn.functional.interpolate(
+ original_mask, size=(conditioning_latents.shape[-2], conditioning_latents.shape[-1])
+ )
+ conditioning_latents = torch.concat([conditioning_latents, mask], 1)
+ # image = self.vae.decode(conditioning_latents[:1,:4,:,:] / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]
+ # from torchvision import transforms
+ # mask_i = transforms.ToPILImage()(image[0:1,:,:,:].squeeze(0)/2+0.5)
+ # mask_i.save(str(timesteps[0]) +'_C.png')
+
+ # 6.5 Optionally get Guidance Scale Embedding
+ timestep_cond = None
+ if self.unet.config.time_cond_proj_dim is not None:
+ guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+ timestep_cond = self.get_guidance_scale_embedding(
+ guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+ ).to(device=device, dtype=latents.dtype)
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ # 7.2 Create tensor stating which brushnets to keep
+ brushnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ brushnet_keep.append(keeps[0] if isinstance(brushnet, BrushNetModel) else keeps)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ is_unet_compiled = is_compiled_module(self.unet)
+ is_brushnet_compiled = is_compiled_module(self.brushnet)
+ is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1")
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # Relevant thread:
+ # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428
+ if (is_unet_compiled and is_brushnet_compiled) and is_torch_higher_equal_2_1:
+ torch._inductor.cudagraph_mark_step_begin()
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # brushnet(s) inference
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer BrushNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ brushnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ brushnet_prompt_embeds = prompt_embeds
+
+ if isinstance(brushnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(brushnet_conditioning_scale, brushnet_keep[i])]
+ else:
+ brushnet_cond_scale = brushnet_conditioning_scale
+ if isinstance(brushnet_cond_scale, list):
+ brushnet_cond_scale = brushnet_cond_scale[0]
+ cond_scale = brushnet_cond_scale * brushnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample, up_block_res_samples = self.brushnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=brushnet_prompt_embeds,
+ brushnet_cond=conditioning_latents,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infered BrushNet only for the conditional batch.
+ # To apply the output of BrushNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+ up_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in up_block_res_samples]
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embedsU,
+ timestep_cond=timestep_cond,
+ cross_attention_kwargs=self.cross_attention_kwargs,
+ down_block_add_samples=down_block_res_samples,
+ mid_block_add_sample=mid_block_res_sample,
+ up_block_add_samples=up_block_res_samples,
+ added_cond_kwargs=added_cond_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ step_idx = i // getattr(self.scheduler, "order", 1)
+ callback(step_idx, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and brushnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.brushnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+ 0
+ ]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/pipeline/pipeline_PowerPaint_ControlNet.py b/pipeline/pipeline_PowerPaint_ControlNet.py
new file mode 100644
index 0000000000000000000000000000000000000000..07262b000e9491d6f11737497a841cd85552543f
--- /dev/null
+++ b/pipeline/pipeline_PowerPaint_ControlNet.py
@@ -0,0 +1,1774 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# This model implementation is heavily inspired by https://github.com/haofanwang/ControlNet-for-Diffusers/
+
+import inspect
+import warnings
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL.Image
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer
+
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin
+from diffusers.models import AutoencoderKL, ControlNetModel, UNet2DConditionModel
+from diffusers.pipelines.controlnet import MultiControlNetModel
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline
+from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+ is_accelerate_available,
+ is_accelerate_version,
+ logging,
+ replace_example_docstring,
+)
+from diffusers.utils.torch_utils import is_compiled_module, randn_tensor
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```py
+ >>> # !pip install transformers accelerate
+ >>> from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel, DDIMScheduler
+ >>> from diffusers.utils import load_image
+ >>> import numpy as np
+ >>> import torch
+
+ >>> init_image = load_image(
+ ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy.png"
+ ... )
+ >>> init_image = init_image.resize((512, 512))
+
+ >>> generator = torch.Generator(device="cpu").manual_seed(1)
+
+ >>> mask_image = load_image(
+ ... "https://huggingface.co/datasets/diffusers/test-arrays/resolve/main/stable_diffusion_inpaint/boy_mask.png"
+ ... )
+ >>> mask_image = mask_image.resize((512, 512))
+
+
+ >>> def make_inpaint_condition(image, image_mask):
+ ... image = np.array(image.convert("RGB")).astype(np.float32) / 255.0
+ ... image_mask = np.array(image_mask.convert("L")).astype(np.float32) / 255.0
+
+ ... assert image.shape[0:1] == image_mask.shape[0:1], "image and image_mask must have the same image size"
+ ... image[image_mask > 0.5] = -1.0 # set as masked pixel
+ ... image = np.expand_dims(image, 0).transpose(0, 3, 1, 2)
+ ... image = torch.from_numpy(image)
+ ... return image
+
+
+ >>> control_image = make_inpaint_condition(init_image, mask_image)
+
+ >>> controlnet = ControlNetModel.from_pretrained(
+ ... "lllyasviel/control_v11p_sd15_inpaint", torch_dtype=torch.float16
+ ... )
+ >>> pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
+ ... )
+
+ >>> pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+ >>> pipe.enable_model_cpu_offload()
+
+ >>> # generate image
+ >>> image = pipe(
+ ... "a handsome man with ray-ban sunglasses",
+ ... num_inference_steps=20,
+ ... generator=generator,
+ ... eta=1.0,
+ ... image=init_image,
+ ... mask_image=mask_image,
+ ... control_image=control_image,
+ ... ).images[0]
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.prepare_mask_and_masked_image
+def prepare_mask_and_masked_image(image, mask, height, width, return_image=False):
+ """
+ Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be
+ converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the
+ ``image`` and ``1`` for the ``mask``.
+
+ The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be
+ binarized (``mask > 0.5``) and cast to ``torch.float32`` too.
+
+ Args:
+ image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width``
+ ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``.
+ mask (_type_): The mask to apply to the image, i.e. regions to inpaint.
+ It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width``
+ ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``.
+
+
+ Raises:
+ ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask
+ should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions.
+ TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not
+ (ot the other way around).
+
+ Returns:
+ tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4
+ dimensions: ``batch x channels x height x width``.
+ """
+
+ if image is None:
+ raise ValueError("`image` input cannot be undefined.")
+
+ if mask is None:
+ raise ValueError("`mask_image` input cannot be undefined.")
+
+ if isinstance(image, torch.Tensor):
+ if not isinstance(mask, torch.Tensor):
+ raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not")
+
+ # Batch single image
+ if image.ndim == 3:
+ assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)"
+ image = image.unsqueeze(0)
+
+ # Batch and add channel dim for single mask
+ if mask.ndim == 2:
+ mask = mask.unsqueeze(0).unsqueeze(0)
+
+ # Batch single mask or add channel dim
+ if mask.ndim == 3:
+ # Single batched mask, no channel dim or single mask not batched but channel dim
+ if mask.shape[0] == 1:
+ mask = mask.unsqueeze(0)
+
+ # Batched masks no channel dim
+ else:
+ mask = mask.unsqueeze(1)
+
+ assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions"
+ assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions"
+ assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size"
+
+ # Check image is in [-1, 1]
+ if image.min() < -1 or image.max() > 1:
+ raise ValueError("Image should be in [-1, 1] range")
+
+ # Check mask is in [0, 1]
+ if mask.min() < 0 or mask.max() > 1:
+ raise ValueError("Mask should be in [0, 1] range")
+
+ # Binarize mask
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+
+ # Image as float32
+ image = image.to(dtype=torch.float32)
+ elif isinstance(mask, torch.Tensor):
+ raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not")
+ else:
+ # preprocess image
+ if isinstance(image, (PIL.Image.Image, np.ndarray)):
+ image = [image]
+ if isinstance(image, list) and isinstance(image[0], PIL.Image.Image):
+ # resize all images w.r.t passed height an width
+ image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image]
+ image = [np.array(i.convert("RGB"))[None, :] for i in image]
+ image = np.concatenate(image, axis=0)
+ elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+ image = np.concatenate([i[None, :] for i in image], axis=0)
+
+ image = image.transpose(0, 3, 1, 2)
+ image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+ # preprocess mask
+ if isinstance(mask, (PIL.Image.Image, np.ndarray)):
+ mask = [mask]
+
+ if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image):
+ mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask]
+ mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0)
+ mask = mask.astype(np.float32) / 255.0
+ elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+ mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+ mask[mask < 0.5] = 0
+ mask[mask >= 0.5] = 1
+ mask = torch.from_numpy(mask)
+
+ masked_image = image * (mask < 0.5)
+
+ # n.b. ensure backwards compatibility as old function does not return image
+ if return_image:
+ return mask, masked_image, image
+
+ return mask, masked_image
+
+
+class StableDiffusionControlNetInpaintPipeline(
+ DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, FromSingleFileMixin
+):
+ r"""
+ Pipeline for text-to-image generation using Stable Diffusion with ControlNet guidance.
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+ library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+ In addition the pipeline inherits the following loading methods:
+ - *Textual-Inversion*: [`loaders.TextualInversionLoaderMixin.load_textual_inversion`]
+
+
+
+ This pipeline can be used both with checkpoints that have been specifically fine-tuned for inpainting, such as
+ [runwayml/stable-diffusion-inpainting](https://huggingface.co/runwayml/stable-diffusion-inpainting)
+ as well as default text-to-image stable diffusion checkpoints, such as
+ [runwayml/stable-diffusion-v1-5](https://huggingface.co/runwayml/stable-diffusion-v1-5).
+ Default text-to-image stable diffusion checkpoints might be preferable for controlnets that have been fine-tuned on
+ those, such as [lllyasviel/control_v11p_sd15_inpaint](https://huggingface.co/lllyasviel/control_v11p_sd15_inpaint).
+
+
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder. Stable Diffusion uses the text portion of
+ [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
+ the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
+ tokenizer (`CLIPTokenizer`):
+ Tokenizer of class
+ [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+ unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+ controlnet ([`ControlNetModel`] or `List[ControlNetModel]`):
+ Provides additional conditioning to the unet during the denoising process. If you set multiple ControlNets
+ as a list, the outputs from each ControlNet are added together to create one combined additional
+ conditioning.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ safety_checker ([`StableDiffusionSafetyChecker`]):
+ Classification module that estimates whether generated images could be considered offensive or harmful.
+ Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details.
+ feature_extractor ([`CLIPImageProcessor`]):
+ Model that extracts features from generated images to be used as inputs for the `safety_checker`.
+ """
+
+ _optional_components = ["safety_checker", "feature_extractor"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: UNet2DConditionModel,
+ controlnet: Union[ControlNetModel, List[ControlNetModel], Tuple[ControlNetModel], MultiControlNetModel],
+ scheduler: KarrasDiffusionSchedulers,
+ safety_checker: StableDiffusionSafetyChecker,
+ feature_extractor: CLIPImageProcessor,
+ requires_safety_checker: bool = True,
+ ):
+ super().__init__()
+
+ if safety_checker is None and requires_safety_checker:
+ logger.warning(
+ f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+ " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+ " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+ " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+ " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+ " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+ )
+
+ if safety_checker is not None and feature_extractor is None:
+ raise ValueError(
+ "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+ " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+ )
+
+ if isinstance(controlnet, (list, tuple)):
+ controlnet = MultiControlNetModel(controlnet)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ safety_checker=safety_checker,
+ feature_extractor=feature_extractor,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+ self.control_image_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+ )
+ self.register_to_config(requires_safety_checker=requires_safety_checker)
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+ def enable_vae_slicing(self):
+ r"""
+ Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
+ compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
+ """
+ self.vae.enable_slicing()
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+ def disable_vae_slicing(self):
+ r"""
+ Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
+ computing decoding in one step.
+ """
+ self.vae.disable_slicing()
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling
+ def enable_vae_tiling(self):
+ r"""
+ Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
+ compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
+ processing larger images.
+ """
+ self.vae.enable_tiling()
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling
+ def disable_vae_tiling(self):
+ r"""
+ Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
+ computing decoding in one step.
+ """
+ self.vae.disable_tiling()
+
+ def enable_model_cpu_offload(self, gpu_id=0):
+ r"""
+ Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
+ to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
+ method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with
+ `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`.
+ """
+ if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"):
+ from accelerate import cpu_offload_with_hook
+ else:
+ raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
+
+ device = torch.device(f"cuda:{gpu_id}")
+
+ hook = None
+ for cpu_offloaded_model in [self.text_encoder, self.unet, self.vae]:
+ _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook)
+
+ if self.safety_checker is not None:
+ # the safety checker can offload the vae again
+ _, hook = cpu_offload_with_hook(self.safety_checker, device, prev_module_hook=hook)
+
+ # control net hook has be manually offloaded as it alternates with unet
+ cpu_offload_with_hook(self.controlnet, device)
+
+ # We'll offload the last model manually.
+ self.final_offload_hook = hook
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+ def _encode_prompt(
+ self,
+ promptA,
+ promptB,
+ t,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_promptA=None,
+ negative_promptB=None,
+ t_nag=None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ lora_scale: Optional[float] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ prompt = promptA
+ negative_prompt = negative_promptA
+
+ if promptA is not None and isinstance(promptA, str):
+ batch_size = 1
+ elif promptA is not None and isinstance(promptA, list):
+ batch_size = len(promptA)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: procecss multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ promptA = self.maybe_convert_prompt(promptA, self.tokenizer)
+
+ text_inputsA = self.tokenizer(
+ promptA,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_inputsB = self.tokenizer(
+ promptB,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_idsA = text_inputsA.input_ids
+ text_input_idsB = text_inputsB.input_ids
+ untruncated_ids = self.tokenizer(promptA, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_idsA.shape[-1] and not torch.equal(
+ text_input_idsA, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputsA.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ # print("text_input_idsA: ",text_input_idsA)
+ # print("text_input_idsB: ",text_input_idsB)
+ # print('t: ',t)
+
+ prompt_embedsA = self.text_encoder(
+ text_input_idsA.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embedsA = prompt_embedsA[0]
+
+ prompt_embedsB = self.text_encoder(
+ text_input_idsB.to(device),
+ attention_mask=attention_mask,
+ )
+ prompt_embedsB = prompt_embedsB[0]
+ prompt_embeds = prompt_embedsA * (t) + (1 - t) * prompt_embedsB
+ # print("prompt_embeds: ",prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokensA: List[str]
+ uncond_tokensB: List[str]
+ if negative_prompt is None:
+ uncond_tokensA = [""] * batch_size
+ uncond_tokensB = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokensA = [negative_promptA]
+ uncond_tokensB = [negative_promptB]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokensA = negative_promptA
+ uncond_tokensB = negative_promptB
+
+ # textual inversion: procecss multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokensA = self.maybe_convert_prompt(uncond_tokensA, self.tokenizer)
+ uncond_tokensB = self.maybe_convert_prompt(uncond_tokensB, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_inputA = self.tokenizer(
+ uncond_tokensA,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ uncond_inputB = self.tokenizer(
+ uncond_tokensB,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_inputA.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embedsA = self.text_encoder(
+ uncond_inputA.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embedsB = self.text_encoder(
+ uncond_inputB.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embedsA[0] * (t_nag) + (1 - t_nag) * negative_prompt_embedsB[0]
+
+ # negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ # print("prompt_embeds: ",prompt_embeds)
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ return prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
+ def run_safety_checker(self, image, device, dtype):
+ if self.safety_checker is None:
+ has_nsfw_concept = None
+ else:
+ if torch.is_tensor(image):
+ feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+ else:
+ feature_extractor_input = self.image_processor.numpy_to_pil(image)
+ safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+ image, has_nsfw_concept = self.safety_checker(
+ images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+ )
+ return image, has_nsfw_concept
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+ def decode_latents(self, latents):
+ warnings.warn(
+ "The decode_latents method is deprecated and will be removed in a future version. Please"
+ " use VaeImageProcessor instead",
+ FutureWarning,
+ )
+ latents = 1 / self.vae.config.scaling_factor * latents
+ image = self.vae.decode(latents, return_dict=False)[0]
+ image = (image / 2 + 0.5).clamp(0, 1)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.StableDiffusionImg2ImgPipeline.get_timesteps
+ def get_timesteps(self, num_inference_steps, strength, device):
+ # get the original timestep using init_timestep
+ init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+ t_start = max(num_inference_steps - init_timestep, 0)
+ timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
+
+ return timesteps, num_inference_steps - t_start
+
+ def check_inputs(
+ self,
+ prompt,
+ image,
+ height,
+ width,
+ callback_steps,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ controlnet_conditioning_scale=1.0,
+ control_guidance_start=0.0,
+ control_guidance_end=1.0,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if (callback_steps is None) or (
+ callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+ ):
+ raise ValueError(
+ f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+ f" {type(callback_steps)}."
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ # `prompt` needs more sophisticated handling when there are multiple
+ # conditionings.
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if isinstance(prompt, list):
+ logger.warning(
+ f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}"
+ " prompts. The conditionings will be fixed across the prompts."
+ )
+
+ # Check `image`
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ self.check_image(image, prompt, prompt_embeds)
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if not isinstance(image, list):
+ raise TypeError("For multiple controlnets: `image` must be type `list`")
+
+ # When `image` is a nested list:
+ # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]])
+ elif any(isinstance(i, list) for i in image):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif len(image) != len(self.controlnet.nets):
+ raise ValueError(
+ f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets."
+ )
+
+ for image_ in image:
+ self.check_image(image_, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `controlnet_conditioning_scale`
+ if (
+ isinstance(self.controlnet, ControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, ControlNetModel)
+ ):
+ if not isinstance(controlnet_conditioning_scale, float):
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+ elif (
+ isinstance(self.controlnet, MultiControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, MultiControlNetModel)
+ ):
+ if isinstance(controlnet_conditioning_scale, list):
+ if any(isinstance(i, list) for i in controlnet_conditioning_scale):
+ raise ValueError("A single batch of multiple conditionings are supported at the moment.")
+ elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len(
+ self.controlnet.nets
+ ):
+ raise ValueError(
+ "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have"
+ " the same length as the number of controlnets"
+ )
+ else:
+ assert False
+
+ if len(control_guidance_start) != len(control_guidance_end):
+ raise ValueError(
+ f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list."
+ )
+
+ if isinstance(self.controlnet, MultiControlNetModel):
+ if len(control_guidance_start) != len(self.controlnet.nets):
+ raise ValueError(
+ f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}."
+ )
+
+ for start, end in zip(control_guidance_start, control_guidance_end):
+ if start >= end:
+ raise ValueError(
+ f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}."
+ )
+ if start < 0.0:
+ raise ValueError(f"control guidance start: {start} can't be smaller than 0.")
+ if end > 1.0:
+ raise ValueError(f"control guidance end: {end} can't be larger than 1.0.")
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
+ def check_image(self, image, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image
+ def prepare_control_image(
+ self,
+ image,
+ width,
+ height,
+ batch_size,
+ num_images_per_prompt,
+ device,
+ dtype,
+ do_classifier_free_guidance=False,
+ guess_mode=False,
+ ):
+ image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32)
+ image_batch_size = image.shape[0]
+
+ if image_batch_size == 1:
+ repeat_by = batch_size
+ else:
+ # image batch size is the same as prompt batch size
+ repeat_by = num_images_per_prompt
+
+ image = image.repeat_interleave(repeat_by, dim=0)
+
+ image = image.to(device=device, dtype=dtype)
+
+ if do_classifier_free_guidance and not guess_mode:
+ image = torch.cat([image] * 2)
+
+ return image
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_latents
+ def prepare_latents(
+ self,
+ batch_size,
+ num_channels_latents,
+ height,
+ width,
+ dtype,
+ device,
+ generator,
+ latents=None,
+ image=None,
+ timestep=None,
+ is_strength_max=True,
+ return_noise=False,
+ return_image_latents=False,
+ ):
+ shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if (image is None or timestep is None) and not is_strength_max:
+ raise ValueError(
+ "Since strength < 1. initial latents are to be initialised as a combination of Image + Noise."
+ "However, either the image or the noise timestep has not been provided."
+ )
+
+ if return_image_latents or (latents is None and not is_strength_max):
+ image = image.to(device=device, dtype=dtype)
+ image_latents = self._encode_vae_image(image=image, generator=generator)
+
+ if latents is None:
+ noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ # if strength is 1. then initialise the latents to noise, else initial to image + noise
+ latents = noise if is_strength_max else self.scheduler.add_noise(image_latents, noise, timestep)
+ # if pure noise then scale the initial latents by the Scheduler's init sigma
+ latents = latents * self.scheduler.init_noise_sigma if is_strength_max else latents
+ else:
+ noise = latents.to(device)
+ latents = noise * self.scheduler.init_noise_sigma
+
+ outputs = (latents,)
+
+ if return_noise:
+ outputs += (noise,)
+
+ if return_image_latents:
+ outputs += (image_latents,)
+
+ return outputs
+
+ def _default_height_width(self, height, width, image):
+ # NOTE: It is possible that a list of images have different
+ # dimensions for each image, so just checking the first image
+ # is not _exactly_ correct, but it is simple.
+ while isinstance(image, list):
+ image = image[0]
+
+ if height is None:
+ if isinstance(image, PIL.Image.Image):
+ height = image.height
+ elif isinstance(image, torch.Tensor):
+ height = image.shape[2]
+
+ height = (height // 8) * 8 # round down to nearest multiple of 8
+
+ if width is None:
+ if isinstance(image, PIL.Image.Image):
+ width = image.width
+ elif isinstance(image, torch.Tensor):
+ width = image.shape[3]
+
+ width = (width // 8) * 8 # round down to nearest multiple of 8
+
+ return height, width
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline.prepare_mask_latents
+ def prepare_mask_latents(
+ self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance
+ ):
+ # resize the mask to latents shape as we concatenate the mask to the latents
+ # we do that before converting to dtype to avoid breaking in case we're using cpu_offload
+ # and half precision
+ mask = torch.nn.functional.interpolate(
+ mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor)
+ )
+ mask = mask.to(device=device, dtype=dtype)
+
+ masked_image = masked_image.to(device=device, dtype=dtype)
+ masked_image_latents = self._encode_vae_image(masked_image, generator=generator)
+
+ # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+ if mask.shape[0] < batch_size:
+ if not batch_size % mask.shape[0] == 0:
+ raise ValueError(
+ "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+ f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+ " of masks that you pass is divisible by the total requested batch size."
+ )
+ mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+ if masked_image_latents.shape[0] < batch_size:
+ if not batch_size % masked_image_latents.shape[0] == 0:
+ raise ValueError(
+ "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+ f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+ " Make sure the number of images that you pass is divisible by the total requested batch size."
+ )
+ masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)
+
+ mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask
+ masked_image_latents = (
+ torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents
+ )
+
+ # aligning device to prevent device errors when concating it with the latent model input
+ masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
+ return mask, masked_image_latents
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.StableDiffusionInpaintPipeline._encode_vae_image
+ def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
+ if isinstance(generator, list):
+ image_latents = [
+ self.vae.encode(image[i : i + 1]).latent_dist.sample(generator=generator[i])
+ for i in range(image.shape[0])
+ ]
+ image_latents = torch.cat(image_latents, dim=0)
+ else:
+ image_latents = self.vae.encode(image).latent_dist.sample(generator=generator)
+
+ image_latents = self.vae.config.scaling_factor * image_latents
+
+ return image_latents
+
+ @torch.no_grad()
+ def predict_woControl(
+ self,
+ promptA: Union[str, List[str]] = None,
+ promptB: Union[str, List[str]] = None,
+ image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ mask_image: Union[torch.FloatTensor, PIL.Image.Image] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ strength: float = 1.0,
+ tradoff: float = 1.0,
+ tradoff_nag: float = 1.0,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_promptA: Optional[Union[str, List[str]]] = None,
+ negative_promptB: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ task_class: Union[torch.Tensor, float, int] = None,
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ image (`PIL.Image.Image`):
+ `Image` or tensor representing an image batch to be inpainted (which parts of the image to be masked
+ out with `mask_image` and repainted according to `prompt`).
+ mask_image (`PIL.Image.Image`):
+ `Image` or tensor representing an image batch to mask `image`. White pixels in the mask are repainted
+ while black pixels are preserved. If `mask_image` is a PIL image, it is converted to a single channel
+ (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, so the
+ expected shape would be `(B, H, W, 1)`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated image.
+ strength (`float`, *optional*, defaults to 1.0):
+ Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
+ starting point and more noise is added the higher the `strength`. The number of denoising steps depends
+ on the amount of noise initially added. When `strength` is 1, added noise is maximum and the denoising
+ process runs for the full number of iterations specified in `num_inference_steps`. A value of 1
+ essentially ignores `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference. This parameter is modulated by `strength`.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that calls every `callback_steps` steps during inference. The function is called with the
+ following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function is called. If not specified, the callback is called at
+ every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+
+ Examples:
+
+ ```py
+ >>> import PIL
+ >>> import requests
+ >>> import torch
+ >>> from io import BytesIO
+
+ >>> from diffusers import StableDiffusionInpaintPipeline
+
+
+ >>> def download_image(url):
+ ... response = requests.get(url)
+ ... return PIL.Image.open(BytesIO(response.content)).convert("RGB")
+
+
+ >>> img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+ >>> mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+ >>> init_image = download_image(img_url).resize((512, 512))
+ >>> mask_image = download_image(mask_url).resize((512, 512))
+
+ >>> pipe = StableDiffusionInpaintPipeline.from_pretrained(
+ ... "runwayml/stable-diffusion-inpainting", torch_dtype=torch.float16
+ ... )
+ >>> pipe = pipe.to("cuda")
+
+ >>> prompt = "Face of a yellow cat, high resolution, sitting on a park bench"
+ >>> image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
+ ```
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+ otherwise a `tuple` is returned where the first element is a list with the generated images and the
+ second element is a list of `bool`s indicating whether the corresponding generated image contains
+ "not-safe-for-work" (nsfw) content.
+ """
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+ prompt = promptA
+ negative_prompt = negative_promptA
+ # 1. Check inputs
+ self.check_inputs(
+ prompt,
+ height,
+ width,
+ strength,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds = self._encode_prompt(
+ promptA,
+ promptB,
+ tradoff,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_promptA,
+ negative_promptB,
+ tradoff_nag,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ )
+
+ # 4. set timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(
+ num_inference_steps=num_inference_steps, strength=strength, device=device
+ )
+ # check that number of inference steps is not < 1 - as this doesn't make sense
+ if num_inference_steps < 1:
+ raise ValueError(
+ f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline"
+ f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
+ )
+ # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+ is_strength_max = strength == 1.0
+
+ # 5. Preprocess mask and image
+ mask, masked_image, init_image = prepare_mask_and_masked_image(
+ image, mask_image, height, width, return_image=True
+ )
+ mask_condition = mask.clone()
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ num_channels_unet = self.unet.config.in_channels
+ return_image_latents = num_channels_unet == 4
+
+ latents_outputs = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ image=init_image,
+ timestep=latent_timestep,
+ is_strength_max=is_strength_max,
+ return_noise=True,
+ return_image_latents=return_image_latents,
+ )
+
+ if return_image_latents:
+ latents, noise, image_latents = latents_outputs
+ else:
+ latents, noise = latents_outputs
+
+ # 7. Prepare mask latent variables
+ mask, masked_image_latents = self.prepare_mask_latents(
+ mask,
+ masked_image,
+ batch_size * num_images_per_prompt,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ do_classifier_free_guidance,
+ )
+
+ # 8. Check that sizes of mask, masked image and latents match
+ if num_channels_unet == 9:
+ # default case for runwayml/stable-diffusion-inpainting
+ num_channels_mask = mask.shape[1]
+ num_channels_masked_image = masked_image_latents.shape[1]
+ if num_channels_latents + num_channels_mask + num_channels_masked_image != self.unet.config.in_channels:
+ raise ValueError(
+ f"Incorrect configuration settings! The config of `pipeline.unet`: {self.unet.config} expects"
+ f" {self.unet.config.in_channels} but received `num_channels_latents`: {num_channels_latents} +"
+ f" `num_channels_mask`: {num_channels_mask} + `num_channels_masked_image`: {num_channels_masked_image}"
+ f" = {num_channels_latents+num_channels_masked_image+num_channels_mask}. Please verify the config of"
+ " `pipeline.unet` or your `mask_image` or `image` input."
+ )
+ elif num_channels_unet != 4:
+ raise ValueError(
+ f"The unet {self.unet.__class__} should have either 4 or 9 input channels, not {self.unet.config.in_channels}."
+ )
+
+ # 9. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 10. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+
+ # concat latents, mask, masked_image_latents in the channel dimension
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ if num_channels_unet == 9:
+ latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+ # predict the noise residual
+ if task_class is not None:
+ noise_pred = self.unet(
+ sample=latent_model_input,
+ timestep=t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ task_class=task_class,
+ )[0]
+ else:
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if num_channels_unet == 4:
+ init_latents_proper = image_latents[:1]
+ init_mask = mask[:1]
+
+ if i < len(timesteps) - 1:
+ noise_timestep = timesteps[i + 1]
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_proper, noise, torch.tensor([noise_timestep])
+ )
+
+ latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, latents)
+
+ if not output_type == "latent":
+ condition_kwargs = {}
+ if isinstance(self.vae, AsymmetricAutoencoderKL):
+ init_image = init_image.to(device=device, dtype=masked_image_latents.dtype)
+ init_image_condition = init_image.clone()
+ init_image = self._encode_vae_image(init_image, generator=generator)
+ mask_condition = mask_condition.to(device=device, dtype=masked_image_latents.dtype)
+ condition_kwargs = {"image": init_image_condition, "mask": mask_condition}
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, **condition_kwargs)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload last model to CPU
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.final_offload_hook.offload()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ promptA: Union[str, List[str]] = None,
+ promptB: Union[str, List[str]] = None,
+ image: Union[torch.Tensor, PIL.Image.Image] = None,
+ mask_image: Union[torch.Tensor, PIL.Image.Image] = None,
+ control_image: Union[
+ torch.FloatTensor,
+ PIL.Image.Image,
+ np.ndarray,
+ List[torch.FloatTensor],
+ List[PIL.Image.Image],
+ List[np.ndarray],
+ ] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ strength: float = 1.0,
+ tradoff: float = 1.0,
+ tradoff_nag: float = 1.0,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_promptA: Optional[Union[str, List[str]]] = None,
+ negative_promptB: Optional[Union[str, List[str]]] = None,
+ num_images_per_prompt: Optional[int] = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.FloatTensor] = None,
+ prompt_embeds: Optional[torch.FloatTensor] = None,
+ negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+ output_type: Optional[str] = "pil",
+ return_dict: bool = True,
+ callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+ callback_steps: int = 1,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ controlnet_conditioning_scale: Union[float, List[float]] = 0.5,
+ guess_mode: bool = False,
+ control_guidance_start: Union[float, List[float]] = 0.0,
+ control_guidance_end: Union[float, List[float]] = 1.0,
+ ):
+ r"""
+ Function invoked when calling the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+ instead.
+ image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`,
+ `List[List[torch.FloatTensor]]`, or `List[List[PIL.Image.Image]]`):
+ The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. If
+ the type is specified as `Torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can
+ also be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If
+ height and/or width are passed, `image` is resized according to them. If multiple ControlNets are
+ specified in init, images must be passed as a list such that each element of the list can be correctly
+ batched for input to a single controlnet.
+ height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The height in pixels of the generated image.
+ width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+ The width in pixels of the generated image.
+ strength (`float`, *optional*, defaults to 1.):
+ Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be
+ between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the
+ `strength`. The number of denoising steps depends on the amount of noise initially added. When
+ `strength` is 1, added noise will be maximum and the denoising process will run for the full number of
+ iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores the masked
+ portion of the reference `image`.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+ `guidance_scale` is defined as `w` of equation 2. of [Imagen
+ Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+ 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+ usually at the expense of lower image quality.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ num_images_per_prompt (`int`, *optional*, defaults to 1):
+ The number of images to generate per prompt.
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+ [`schedulers.DDIMScheduler`], will be ignored for others.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+ to make generation deterministic.
+ latents (`torch.FloatTensor`, *optional*):
+ Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor will ge generated by sampling using the supplied random `generator`.
+ prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generate image. Choose between
+ [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+ plain tuple.
+ callback (`Callable`, *optional*):
+ A function that will be called every `callback_steps` steps during inference. The function will be
+ called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+ callback_steps (`int`, *optional*, defaults to 1):
+ The frequency at which the `callback` function will be called. If not specified, the callback will be
+ called at every step.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+ `self.processor` in
+ [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 0.5):
+ The outputs of the controlnet are multiplied by `controlnet_conditioning_scale` before they are added
+ to the residual in the original unet. If multiple ControlNets are specified in init, you can set the
+ corresponding scale as a list. Note that by default, we use a smaller conditioning scale for inpainting
+ than for [`~StableDiffusionControlNetPipeline.__call__`].
+ guess_mode (`bool`, *optional*, defaults to `False`):
+ In this mode, the ControlNet encoder will try best to recognize the content of the input image even if
+ you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended.
+ control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0):
+ The percentage of total steps at which the controlnet starts applying.
+ control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The percentage of total steps at which the controlnet stops applying.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+ [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple.
+ When returning a tuple, the first element is a list with the generated images, and the second element is a
+ list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work"
+ (nsfw) content, according to the `safety_checker`.
+ """
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+ # 0. Default height and width to unet
+ height, width = self._default_height_width(height, width, image)
+
+ prompt = promptA
+ negative_prompt = negative_promptA
+
+ # align format for control guidance
+ if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list):
+ control_guidance_start = len(control_guidance_end) * [control_guidance_start]
+ elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list):
+ control_guidance_end = len(control_guidance_start) * [control_guidance_end]
+ elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list):
+ mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1
+ control_guidance_start, control_guidance_end = (
+ mult * [control_guidance_start],
+ mult * [control_guidance_end],
+ )
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt,
+ control_image,
+ height,
+ width,
+ callback_steps,
+ negative_prompt,
+ prompt_embeds,
+ negative_prompt_embeds,
+ controlnet_conditioning_scale,
+ control_guidance_start,
+ control_guidance_end,
+ )
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ do_classifier_free_guidance = guidance_scale > 1.0
+
+ if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float):
+ controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets)
+
+ global_pool_conditions = (
+ controlnet.config.global_pool_conditions
+ if isinstance(controlnet, ControlNetModel)
+ else controlnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+ )
+ prompt_embeds = self._encode_prompt(
+ promptA,
+ promptB,
+ tradoff,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_promptA,
+ negative_promptB,
+ tradoff_nag,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ )
+
+ # 4. Prepare image
+ if isinstance(controlnet, ControlNetModel):
+ control_image = self.prepare_control_image(
+ image=control_image,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+ elif isinstance(controlnet, MultiControlNetModel):
+ control_images = []
+
+ for control_image_ in control_image:
+ control_image_ = self.prepare_control_image(
+ image=control_image_,
+ width=width,
+ height=height,
+ batch_size=batch_size * num_images_per_prompt,
+ num_images_per_prompt=num_images_per_prompt,
+ device=device,
+ dtype=controlnet.dtype,
+ do_classifier_free_guidance=do_classifier_free_guidance,
+ guess_mode=guess_mode,
+ )
+
+ control_images.append(control_image_)
+
+ control_image = control_images
+ else:
+ assert False
+
+ # 4. Preprocess mask and image - resizes image and mask w.r.t height and width
+ mask, masked_image, init_image = prepare_mask_and_masked_image(
+ image, mask_image, height, width, return_image=True
+ )
+
+ # 5. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps, num_inference_steps = self.get_timesteps(
+ num_inference_steps=num_inference_steps, strength=strength, device=device
+ )
+ # at which timestep to set the initial noise (n.b. 50% if strength is 0.5)
+ latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
+ # create a boolean to check if the strength is set to 1. if so then initialise the latents with pure noise
+ is_strength_max = strength == 1.0
+
+ # 6. Prepare latent variables
+ num_channels_latents = self.vae.config.latent_channels
+ num_channels_unet = self.unet.config.in_channels
+ return_image_latents = num_channels_unet == 4
+ latents_outputs = self.prepare_latents(
+ batch_size * num_images_per_prompt,
+ num_channels_latents,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ image=init_image,
+ timestep=latent_timestep,
+ is_strength_max=is_strength_max,
+ return_noise=True,
+ return_image_latents=return_image_latents,
+ )
+
+ if return_image_latents:
+ latents, noise, image_latents = latents_outputs
+ else:
+ latents, noise = latents_outputs
+
+ # 7. Prepare mask latent variables
+ mask, masked_image_latents = self.prepare_mask_latents(
+ mask,
+ masked_image,
+ batch_size * num_images_per_prompt,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ do_classifier_free_guidance,
+ )
+
+ # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 7.1 Create tensor stating which controlnets to keep
+ controlnet_keep = []
+ for i in range(len(timesteps)):
+ keeps = [
+ 1.0 - float(i / len(timesteps) < s or (i + 1) / len(timesteps) > e)
+ for s, e in zip(control_guidance_start, control_guidance_end)
+ ]
+ controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps)
+
+ # 8. Denoising loop
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+ with self.progress_bar(total=num_inference_steps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ # controlnet(s) inference
+ if guess_mode and do_classifier_free_guidance:
+ # Infer ControlNet only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ controlnet_prompt_embeds = prompt_embeds
+
+ if isinstance(controlnet_keep[i], list):
+ cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])]
+ else:
+ controlnet_cond_scale = controlnet_conditioning_scale
+ if isinstance(controlnet_cond_scale, list):
+ controlnet_cond_scale = controlnet_cond_scale[0]
+ cond_scale = controlnet_cond_scale * controlnet_keep[i]
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=controlnet_prompt_embeds,
+ controlnet_cond=control_image,
+ conditioning_scale=cond_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ if guess_mode and do_classifier_free_guidance:
+ # Infered ControlNet only for the conditional batch.
+ # To apply the output of ControlNet to both the unconditional and conditional batches,
+ # add 0 to the unconditional batch to keep it unchanged.
+ down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples]
+ mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample])
+
+ # predict the noise residual
+ if num_channels_unet == 9:
+ latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1)
+
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ return_dict=False,
+ )[0]
+
+ # perform guidance
+ if do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+ if num_channels_unet == 4:
+ init_latents_proper = image_latents[:1]
+ init_mask = mask[:1]
+
+ if i < len(timesteps) - 1:
+ noise_timestep = timesteps[i + 1]
+ init_latents_proper = self.scheduler.add_noise(
+ init_latents_proper, noise, torch.tensor([noise_timestep])
+ )
+
+ latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+ if callback is not None and i % callback_steps == 0:
+ callback(i, t, latents)
+
+ # If we do sequential model offloading, let's offload unet and controlnet
+ # manually for max memory savings
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.unet.to("cpu")
+ self.controlnet.to("cpu")
+ torch.cuda.empty_cache()
+
+ if not output_type == "latent":
+ image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+ image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+ else:
+ image = latents
+ has_nsfw_concept = None
+
+ if has_nsfw_concept is None:
+ do_denormalize = [True] * image.shape[0]
+ else:
+ do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+ image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+ # Offload last model to CPU
+ if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+ self.final_offload_hook.offload()
+
+ if not return_dict:
+ return (image, has_nsfw_concept)
+
+ return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)
diff --git a/pyproject.toml b/pyproject.toml
new file mode 100644
index 0000000000000000000000000000000000000000..0612f2f9e059e4f62e8d393aab01b6c2d73f7108
--- /dev/null
+++ b/pyproject.toml
@@ -0,0 +1,27 @@
+[tool.ruff]
+# Never enforce `E501` (line length violations).
+ignore = ["C901", "E501", "E741", "F402", "F823"]
+select = ["C", "E", "F", "I", "W"]
+line-length = 119
+
+# Ignore import violations in all `__init__.py` files.
+[tool.ruff.per-file-ignores]
+"__init__.py" = ["E402", "F401", "F403", "F811"]
+"src/diffusers/utils/dummy_*.py" = ["F401"]
+
+[tool.ruff.isort]
+lines-after-imports = 2
+known-first-party = ["diffusers"]
+
+[tool.ruff.format]
+# Like Black, use double quotes for strings.
+quote-style = "double"
+
+# Like Black, indent with spaces, rather than tabs.
+indent-style = "space"
+
+# Like Black, respect magic trailing commas.
+skip-magic-trailing-comma = false
+
+# Like Black, automatically detect the appropriate line ending.
+line-ending = "auto"
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..f51e7810340131a8509f7834cbae98e1ceccb7c7
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,9 @@
+accelerate
+controlnet-aux==0.0.3
+diffusers==0.24.0
+gradio==3.41.0
+mmengine
+opencv-python
+torch
+torchvision
+transformers==4.28.0
diff --git a/utils/__pycache__/utils.cpython-310.pyc b/utils/__pycache__/utils.cpython-310.pyc
new file mode 100644
index 0000000000000000000000000000000000000000..051685920e3b29dd9a40e7636fdaeef80f3f5afe
Binary files /dev/null and b/utils/__pycache__/utils.cpython-310.pyc differ
diff --git a/utils/utils.py b/utils/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..34e931a11fe1a6107d0d3a44cd05b88be64e88f9
--- /dev/null
+++ b/utils/utils.py
@@ -0,0 +1,529 @@
+import copy
+import os
+import random
+from logging import WARNING
+from typing import Any, List, Optional, Union
+
+import torch
+import torch.nn as nn
+import transformers
+# from mmagic.utils import try_import
+from mmengine import print_log
+
+
+class TokenizerWrapper:
+ """Tokenizer wrapper for CLIPTokenizer. Only support CLIPTokenizer
+ currently. This wrapper is modified from https://github.com/huggingface/dif
+ fusers/blob/e51f19aee82c8dd874b715a09dbc521d88835d68/src/diffusers/loaders.
+ py#L358 # noqa.
+
+ Args:
+ from_pretrained (Union[str, os.PathLike], optional): The *model id*
+ of a pretrained model or a path to a *directory* containing
+ model weights and config. Defaults to None.
+ from_config (Union[str, os.PathLike], optional): The *model id*
+ of a pretrained model or a path to a *directory* containing
+ model weights and config. Defaults to None.
+
+ *args, **kwargs: If `from_pretrained` is passed, *args and **kwargs
+ will be passed to `from_pretrained` function. Otherwise, *args
+ and **kwargs will be used to initialize the model by
+ `self._module_cls(*args, **kwargs)`.
+ """
+
+ def __init__(
+ self,
+ from_pretrained: Optional[Union[str, os.PathLike]] = None,
+ from_config: Optional[Union[str, os.PathLike]] = None,
+ *args,
+ **kwargs,
+ ):
+ # transformers = try_import("transformers")
+ module_cls = transformers.CLIPTokenizer
+
+ assert not (from_pretrained and from_config), (
+ "'from_pretrained' and 'from_config' should not be passed " "at the same time."
+ )
+
+ if from_config:
+ print_log(
+ "Tokenizers from Huggingface transformers do not support "
+ "'from_config'. Will call 'from_pretrained' instead "
+ "with the same argument.",
+ "current",
+ WARNING,
+ )
+ from_pretrained = from_config
+
+ if from_pretrained:
+ self.wrapped = module_cls.from_pretrained(from_pretrained, *args, **kwargs)
+ else:
+ self.wrapper = module_cls(*args, **kwargs)
+
+ self._from_pretrained = from_pretrained
+ self.token_map = {}
+
+ def __getattr__(self, name: str) -> Any:
+ if name == "wrapped":
+ return super().__getattr__("wrapped")
+
+ try:
+ return getattr(self.wrapped, name)
+ except AttributeError:
+ try:
+ return super().__getattr__(name)
+ except AttributeError:
+ raise AttributeError(
+ "'name' cannot be found in both "
+ f"'{self.__class__.__name__}' and "
+ f"'{self.__class__.__name__}.tokenizer'."
+ )
+
+ def try_adding_tokens(self, tokens: Union[str, List[str]], *args, **kwargs):
+ """Attempt to add tokens to the tokenizer.
+
+ Args:
+ tokens (Union[str, List[str]]): The tokens to be added.
+ """
+ num_added_tokens = self.wrapped.add_tokens(tokens, *args, **kwargs)
+ assert num_added_tokens != 0, (
+ f"The tokenizer already contains the token {tokens}. Please pass "
+ "a different `placeholder_token` that is not already in the "
+ "tokenizer."
+ )
+
+ def get_token_info(self, token: str) -> dict:
+ """Get the information of a token, including its start and end index in
+ the current tokenizer.
+
+ Args:
+ token (str): The token to be queried.
+
+ Returns:
+ dict: The information of the token, including its start and end
+ index in current tokenizer.
+ """
+ token_ids = self.__call__(token).input_ids
+ start, end = token_ids[1], token_ids[-2] + 1
+ return {"name": token, "start": start, "end": end}
+
+ def add_placeholder_token(self, placeholder_token: str, *args, num_vec_per_token: int = 1, **kwargs):
+ """Add placeholder tokens to the tokenizer.
+
+ Args:
+ placeholder_token (str): The placeholder token to be added.
+ num_vec_per_token (int, optional): The number of vectors of
+ the added placeholder token.
+ *args, **kwargs: The arguments for `self.wrapped.add_tokens`.
+ """
+ output = []
+ if num_vec_per_token == 1:
+ self.try_adding_tokens(placeholder_token, *args, **kwargs)
+ output.append(placeholder_token)
+ else:
+ output = []
+ for i in range(num_vec_per_token):
+ ith_token = placeholder_token + f"_{i}"
+ self.try_adding_tokens(ith_token, *args, **kwargs)
+ output.append(ith_token)
+
+ for token in self.token_map:
+ if token in placeholder_token:
+ raise ValueError(
+ f"The tokenizer already has placeholder token {token} "
+ f"that can get confused with {placeholder_token} "
+ "keep placeholder tokens independent"
+ )
+ self.token_map[placeholder_token] = output
+
+ def replace_placeholder_tokens_in_text(
+ self, text: Union[str, List[str]], vector_shuffle: bool = False, prop_tokens_to_load: float = 1.0
+ ) -> Union[str, List[str]]:
+ """Replace the keywords in text with placeholder tokens. This function
+ will be called in `self.__call__` and `self.encode`.
+
+ Args:
+ text (Union[str, List[str]]): The text to be processed.
+ vector_shuffle (bool, optional): Whether to shuffle the vectors.
+ Defaults to False.
+ prop_tokens_to_load (float, optional): The proportion of tokens to
+ be loaded. If 1.0, all tokens will be loaded. Defaults to 1.0.
+
+ Returns:
+ Union[str, List[str]]: The processed text.
+ """
+ if isinstance(text, list):
+ output = []
+ for i in range(len(text)):
+ output.append(self.replace_placeholder_tokens_in_text(text[i], vector_shuffle=vector_shuffle))
+ return output
+
+ for placeholder_token in self.token_map:
+ if placeholder_token in text:
+ tokens = self.token_map[placeholder_token]
+ tokens = tokens[: 1 + int(len(tokens) * prop_tokens_to_load)]
+ if vector_shuffle:
+ tokens = copy.copy(tokens)
+ random.shuffle(tokens)
+ text = text.replace(placeholder_token, " ".join(tokens))
+ return text
+
+ def replace_text_with_placeholder_tokens(self, text: Union[str, List[str]]) -> Union[str, List[str]]:
+ """Replace the placeholder tokens in text with the original keywords.
+ This function will be called in `self.decode`.
+
+ Args:
+ text (Union[str, List[str]]): The text to be processed.
+
+ Returns:
+ Union[str, List[str]]: The processed text.
+ """
+ if isinstance(text, list):
+ output = []
+ for i in range(len(text)):
+ output.append(self.replace_text_with_placeholder_tokens(text[i]))
+ return output
+
+ for placeholder_token, tokens in self.token_map.items():
+ merged_tokens = " ".join(tokens)
+ if merged_tokens in text:
+ text = text.replace(merged_tokens, placeholder_token)
+ return text
+
+ def __call__(
+ self,
+ text: Union[str, List[str]],
+ *args,
+ vector_shuffle: bool = False,
+ prop_tokens_to_load: float = 1.0,
+ **kwargs,
+ ):
+ """The call function of the wrapper.
+
+ Args:
+ text (Union[str, List[str]]): The text to be tokenized.
+ vector_shuffle (bool, optional): Whether to shuffle the vectors.
+ Defaults to False.
+ prop_tokens_to_load (float, optional): The proportion of tokens to
+ be loaded. If 1.0, all tokens will be loaded. Defaults to 1.0
+ *args, **kwargs: The arguments for `self.wrapped.__call__`.
+ """
+ replaced_text = self.replace_placeholder_tokens_in_text(
+ text, vector_shuffle=vector_shuffle, prop_tokens_to_load=prop_tokens_to_load
+ )
+
+ return self.wrapped.__call__(replaced_text, *args, **kwargs)
+
+ def encode(self, text: Union[str, List[str]], *args, **kwargs):
+ """Encode the passed text to token index.
+
+ Args:
+ text (Union[str, List[str]]): The text to be encode.
+ *args, **kwargs: The arguments for `self.wrapped.__call__`.
+ """
+ replaced_text = self.replace_placeholder_tokens_in_text(text)
+ return self.wrapped(replaced_text, *args, **kwargs)
+
+ def decode(self, token_ids, return_raw: bool = False, *args, **kwargs) -> Union[str, List[str]]:
+ """Decode the token index to text.
+
+ Args:
+ token_ids: The token index to be decoded.
+ return_raw: Whether keep the placeholder token in the text.
+ Defaults to False.
+ *args, **kwargs: The arguments for `self.wrapped.decode`.
+
+ Returns:
+ Union[str, List[str]]: The decoded text.
+ """
+ text = self.wrapped.decode(token_ids, *args, **kwargs)
+ if return_raw:
+ return text
+ replaced_text = self.replace_text_with_placeholder_tokens(text)
+ return replaced_text
+
+ def __repr__(self):
+ """The representation of the wrapper."""
+ s = super().__repr__()
+ prefix = f"Wrapped Module Class: {self._module_cls}\n"
+ prefix += f"Wrapped Module Name: {self._module_name}\n"
+ if self._from_pretrained:
+ prefix += f"From Pretrained: {self._from_pretrained}\n"
+ s = prefix + s
+ return s
+
+
+class EmbeddingLayerWithFixes(nn.Module):
+ """The revised embedding layer to support external embeddings. This design
+ of this class is inspired by https://github.com/AUTOMATIC1111/stable-
+ diffusion-webui/blob/22bcc7be428c94e9408f589966c2040187245d81/modules/sd_hi
+ jack.py#L224 # noqa.
+
+ Args:
+ wrapped (nn.Emebdding): The embedding layer to be wrapped.
+ external_embeddings (Union[dict, List[dict]], optional): The external
+ embeddings added to this layer. Defaults to None.
+ """
+
+ def __init__(self, wrapped: nn.Embedding, external_embeddings: Optional[Union[dict, List[dict]]] = None):
+ super().__init__()
+ self.wrapped = wrapped
+ self.num_embeddings = wrapped.weight.shape[0]
+
+ self.external_embeddings = []
+ if external_embeddings:
+ self.add_embeddings(external_embeddings)
+
+ self.trainable_embeddings = nn.ParameterDict()
+
+ @property
+ def weight(self):
+ """Get the weight of wrapped embedding layer."""
+ return self.wrapped.weight
+
+ def check_duplicate_names(self, embeddings: List[dict]):
+ """Check whether duplicate names exist in list of 'external
+ embeddings'.
+
+ Args:
+ embeddings (List[dict]): A list of embedding to be check.
+ """
+ names = [emb["name"] for emb in embeddings]
+ assert len(names) == len(set(names)), (
+ "Found duplicated names in 'external_embeddings'. Name list: " f"'{names}'"
+ )
+
+ def check_ids_overlap(self, embeddings):
+ """Check whether overlap exist in token ids of 'external_embeddings'.
+
+ Args:
+ embeddings (List[dict]): A list of embedding to be check.
+ """
+ ids_range = [[emb["start"], emb["end"], emb["name"]] for emb in embeddings]
+ ids_range.sort() # sort by 'start'
+ # check if 'end' has overlapping
+ for idx in range(len(ids_range) - 1):
+ name1, name2 = ids_range[idx][-1], ids_range[idx + 1][-1]
+ assert ids_range[idx][1] <= ids_range[idx + 1][0], (
+ f"Found ids overlapping between embeddings '{name1}' " f"and '{name2}'."
+ )
+
+ def add_embeddings(self, embeddings: Optional[Union[dict, List[dict]]]):
+ """Add external embeddings to this layer.
+
+ Use case:
+
+ >>> 1. Add token to tokenizer and get the token id.
+ >>> tokenizer = TokenizerWrapper('openai/clip-vit-base-patch32')
+ >>> # 'how much' in kiswahili
+ >>> tokenizer.add_placeholder_tokens('ngapi', num_vec_per_token=4)
+ >>>
+ >>> 2. Add external embeddings to the model.
+ >>> new_embedding = {
+ >>> 'name': 'ngapi', # 'how much' in kiswahili
+ >>> 'embedding': torch.ones(1, 15) * 4,
+ >>> 'start': tokenizer.get_token_info('kwaheri')['start'],
+ >>> 'end': tokenizer.get_token_info('kwaheri')['end'],
+ >>> 'trainable': False # if True, will registry as a parameter
+ >>> }
+ >>> embedding_layer = nn.Embedding(10, 15)
+ >>> embedding_layer_wrapper = EmbeddingLayerWithFixes(embedding_layer)
+ >>> embedding_layer_wrapper.add_embeddings(new_embedding)
+ >>>
+ >>> 3. Forward tokenizer and embedding layer!
+ >>> input_text = ['hello, ngapi!', 'hello my friend, ngapi?']
+ >>> input_ids = tokenizer(
+ >>> input_text, padding='max_length', truncation=True,
+ >>> return_tensors='pt')['input_ids']
+ >>> out_feat = embedding_layer_wrapper(input_ids)
+ >>>
+ >>> 4. Let's validate the result!
+ >>> assert (out_feat[0, 3: 7] == 2.3).all()
+ >>> assert (out_feat[2, 5: 9] == 2.3).all()
+
+ Args:
+ embeddings (Union[dict, list[dict]]): The external embeddings to
+ be added. Each dict must contain the following 4 fields: 'name'
+ (the name of this embedding), 'embedding' (the embedding
+ tensor), 'start' (the start token id of this embedding), 'end'
+ (the end token id of this embedding). For example:
+ `{name: NAME, start: START, end: END, embedding: torch.Tensor}`
+ """
+ if isinstance(embeddings, dict):
+ embeddings = [embeddings]
+
+ self.external_embeddings += embeddings
+ self.check_duplicate_names(self.external_embeddings)
+ self.check_ids_overlap(self.external_embeddings)
+
+ # set for trainable
+ added_trainable_emb_info = []
+ for embedding in embeddings:
+ trainable = embedding.get("trainable", False)
+ if trainable:
+ name = embedding["name"]
+ embedding["embedding"] = torch.nn.Parameter(embedding["embedding"])
+ self.trainable_embeddings[name] = embedding["embedding"]
+ added_trainable_emb_info.append(name)
+
+ added_emb_info = [emb["name"] for emb in embeddings]
+ added_emb_info = ", ".join(added_emb_info)
+ print_log(f"Successfully add external embeddings: {added_emb_info}.", "current")
+
+ if added_trainable_emb_info:
+ added_trainable_emb_info = ", ".join(added_trainable_emb_info)
+ print_log("Successfully add trainable external embeddings: " f"{added_trainable_emb_info}", "current")
+
+ def replace_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+ """Replace external input ids to 0.
+
+ Args:
+ input_ids (torch.Tensor): The input ids to be replaced.
+
+ Returns:
+ torch.Tensor: The replaced input ids.
+ """
+ input_ids_fwd = input_ids.clone()
+ input_ids_fwd[input_ids_fwd >= self.num_embeddings] = 0
+ return input_ids_fwd
+
+ def replace_embeddings(
+ self, input_ids: torch.Tensor, embedding: torch.Tensor, external_embedding: dict
+ ) -> torch.Tensor:
+ """Replace external embedding to the embedding layer. Noted that, in
+ this function we use `torch.cat` to avoid inplace modification.
+
+ Args:
+ input_ids (torch.Tensor): The original token ids. Shape like
+ [LENGTH, ].
+ embedding (torch.Tensor): The embedding of token ids after
+ `replace_input_ids` function.
+ external_embedding (dict): The external embedding to be replaced.
+
+ Returns:
+ torch.Tensor: The replaced embedding.
+ """
+ new_embedding = []
+
+ name = external_embedding["name"]
+ start = external_embedding["start"]
+ end = external_embedding["end"]
+ target_ids_to_replace = [i for i in range(start, end)]
+ ext_emb = external_embedding["embedding"]
+
+ # do not need to replace
+ if not (input_ids == start).any():
+ return embedding
+
+ # start replace
+ s_idx, e_idx = 0, 0
+ while e_idx < len(input_ids):
+ if input_ids[e_idx] == start:
+ if e_idx != 0:
+ # add embedding do not need to replace
+ new_embedding.append(embedding[s_idx:e_idx])
+
+ # check if the next embedding need to replace is valid
+ actually_ids_to_replace = [int(i) for i in input_ids[e_idx : e_idx + end - start]]
+ assert actually_ids_to_replace == target_ids_to_replace, (
+ f"Invalid 'input_ids' in position: {s_idx} to {e_idx}. "
+ f"Expect '{target_ids_to_replace}' for embedding "
+ f"'{name}' but found '{actually_ids_to_replace}'."
+ )
+
+ new_embedding.append(ext_emb)
+
+ s_idx = e_idx + end - start
+ e_idx = s_idx + 1
+ else:
+ e_idx += 1
+
+ if e_idx == len(input_ids):
+ new_embedding.append(embedding[s_idx:e_idx])
+
+ return torch.cat(new_embedding, dim=0)
+
+ def forward(self, input_ids: torch.Tensor, external_embeddings: Optional[List[dict]] = None):
+ """The forward function.
+
+ Args:
+ input_ids (torch.Tensor): The token ids shape like [bz, LENGTH] or
+ [LENGTH, ].
+ external_embeddings (Optional[List[dict]]): The external
+ embeddings. If not passed, only `self.external_embeddings`
+ will be used. Defaults to None.
+
+ input_ids: shape like [bz, LENGTH] or [LENGTH].
+ """
+ assert input_ids.ndim in [1, 2]
+ if input_ids.ndim == 1:
+ input_ids = input_ids.unsqueeze(0)
+
+ if external_embeddings is None and not self.external_embeddings:
+ return self.wrapped(input_ids)
+
+ input_ids_fwd = self.replace_input_ids(input_ids)
+ inputs_embeds = self.wrapped(input_ids_fwd)
+
+ vecs = []
+
+ if external_embeddings is None:
+ external_embeddings = []
+ elif isinstance(external_embeddings, dict):
+ external_embeddings = [external_embeddings]
+ embeddings = self.external_embeddings + external_embeddings
+
+ for input_id, embedding in zip(input_ids, inputs_embeds):
+ new_embedding = embedding
+ for external_embedding in embeddings:
+ new_embedding = self.replace_embeddings(input_id, new_embedding, external_embedding)
+ vecs.append(new_embedding)
+
+ return torch.stack(vecs)
+
+
+def add_tokens(
+ tokenizer, text_encoder, placeholder_tokens: list, initialize_tokens: list = None, num_vectors_per_token: int = 1
+):
+ """Add token for training.
+
+ # TODO: support add tokens as dict, then we can load pretrained tokens.
+ """
+ if initialize_tokens is not None:
+ assert len(initialize_tokens) == len(
+ placeholder_tokens
+ ), "placeholder_token should be the same length as initialize_token"
+ for ii in range(len(placeholder_tokens)):
+ tokenizer.add_placeholder_token(placeholder_tokens[ii], num_vec_per_token=num_vectors_per_token)
+
+ # text_encoder.set_embedding_layer()
+ embedding_layer = text_encoder.text_model.embeddings.token_embedding
+ text_encoder.text_model.embeddings.token_embedding = EmbeddingLayerWithFixes(embedding_layer)
+ embedding_layer = text_encoder.text_model.embeddings.token_embedding
+
+ assert embedding_layer is not None, (
+ "Do not support get embedding layer for current text encoder. " "Please check your configuration."
+ )
+ initialize_embedding = []
+ if initialize_tokens is not None:
+ for ii in range(len(placeholder_tokens)):
+ init_id = tokenizer(initialize_tokens[ii]).input_ids[1]
+ temp_embedding = embedding_layer.weight[init_id]
+ initialize_embedding.append(temp_embedding[None, ...].repeat(num_vectors_per_token, 1))
+ else:
+ for ii in range(len(placeholder_tokens)):
+ init_id = tokenizer("a").input_ids[1]
+ temp_embedding = embedding_layer.weight[init_id]
+ len_emb = temp_embedding.shape[0]
+ init_weight = (torch.rand(num_vectors_per_token, len_emb) - 0.5) / 2.0
+ initialize_embedding.append(init_weight)
+
+ # initialize_embedding = torch.cat(initialize_embedding,dim=0)
+
+ token_info_all = []
+ for ii in range(len(placeholder_tokens)):
+ token_info = tokenizer.get_token_info(placeholder_tokens[ii])
+ token_info["embedding"] = initialize_embedding[ii]
+ token_info["trainable"] = True
+ token_info_all.append(token_info)
+ embedding_layer.add_embeddings(token_info_all)