init demo

app.py

@@ -0,0 +1,332 @@
+import sys
+sys.path.append('./')
+
+
+import os 
+import cv2
+import torch
+import random
+import numpy as np
+from PIL import Image
+from diffusers import ControlNetModel, StableDiffusionXLControlNetPipeline
+
+import spaces
+import gradio as gr
+from huggingface_hub import hf_hub_download
+
+from ip_adapter import IPAdapterXL
+
+hf_hub_download(repo_id="h94/IP-Adapter", filename="sdxl_models/image_encoder", local_dir="./sdxl_models")
+hf_hub_download(repo_id="h94/IP-Adapter", filename="sdxl_models/ip-adapter_sdxl.bin", local_dir="./sdxl_models")
+
+# global variable
+MAX_SEED = np.iinfo(np.int32).max
+device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.float16 if str(device).__contains__("cuda") else torch.float32
+
+# initialization
+base_model_path = "stabilityai/stable-diffusion-xl-base-1.0"
+image_encoder_path = "sdxl_models/image_encoder"
+ip_ckpt = "sdxl_models/ip-adapter_sdxl.bin"
+
+controlnet_path = "diffusers/controlnet-canny-sdxl-1.0"
+controlnet = ControlNetModel.from_pretrained(controlnet_path, use_safetensors=False, torch_dtype=torch.float16).to(device)
+
+# load SDXL pipeline
+pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
+    base_model_path,
+    controlnet=controlnet,
+    torch_dtype=torch.float16,
+    add_watermarker=False,
+)
+
+# load ip-adapter
+# target_blocks=["block"] for original IP-Adapter
+# target_blocks=["up_blocks.0.attentions.1"] for style blocks only
+# target_blocks = ["up_blocks.0.attentions.1", "down_blocks.2.attentions.1"] # for style+layout blocks
+ip_model = IPAdapterXL(pipe, image_encoder_path, ip_ckpt, device, target_blocks=["up_blocks.0.attentions.1"])
+
+def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    return seed
+
+def resize_img(
+    input_image,
+    max_side=1280,
+    min_side=1024,
+    size=None,
+    pad_to_max_side=False,
+    mode=Image.BILINEAR,
+    base_pixel_number=64,
+):
+    w, h = input_image.size
+    if size is not None:
+        w_resize_new, h_resize_new = size
+    else:
+        ratio = min_side / min(h, w)
+        w, h = round(ratio * w), round(ratio * h)
+        ratio = max_side / max(h, w)
+        input_image = input_image.resize([round(ratio * w), round(ratio * h)], mode)
+        w_resize_new = (round(ratio * w) // base_pixel_number) * base_pixel_number
+        h_resize_new = (round(ratio * h) // base_pixel_number) * base_pixel_number
+    input_image = input_image.resize([w_resize_new, h_resize_new], mode)
+
+    if pad_to_max_side:
+        res = np.ones([max_side, max_side, 3], dtype=np.uint8) * 255
+        offset_x = (max_side - w_resize_new) // 2
+        offset_y = (max_side - h_resize_new) // 2
+        res[
+            offset_y : offset_y + h_resize_new, offset_x : offset_x + w_resize_new
+        ] = np.array(input_image)
+        input_image = Image.fromarray(res)
+    return input_image
+
+def get_example():
+    case = [
+        [
+            "./assets/0.jpg",
+            None,
+            "a cat, masterpiece, best quality, high quality",
+            1.0,
+            0.0
+        ],
+        [
+            "./assets/1.jpg",
+            None,
+            "a cat, masterpiece, best quality, high quality",
+            1.0,
+            0.0
+        ],
+        [
+            "./assets/2.jpg",
+            None,
+            "a cat, masterpiece, best quality, high quality",
+            1.0,
+            0.0
+        ],
+        [
+            "./assets/3.jpg",
+            None,
+            "a cat, masterpiece, best quality, high quality",
+            1.0,
+            0.0
+        ],
+        [
+            "./assets/2.jpg",
+            "./assets/yann-lecun.jpg",
+            "a man, masterpiece, best quality, high quality",
+            1.0,
+            0.6
+        ],
+    ]
+    return case
+
+def run_for_examples(style_image, source_image, prompt, scale, control_scale):
+
+    return create_image(
+        image_pil=style_image,
+        input_image=source_image,
+        prompt=prompt,
+        n_prompt="text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
+        scale=scale,
+        control_scale=control_scale,
+        guidance_scale=5,
+        num_samples=1,
+        num_inference_steps=30,
+        seed=42,
+        target="Load only style blocks",
+        neg_content_prompt="",
+        neg_content_scale=0,
+    )
+
+@spaces.GPU(enable_queue=True)
+def create_image(image_pil,
+                 input_image,
+                 prompt,
+                 n_prompt,
+                 scale, 
+                 control_scale, 
+                 guidance_scale,
+                 num_samples,
+                 num_inference_steps,
+                 seed,
+                 target="Load only style blocks",
+                 neg_content_prompt=None,
+                 neg_content_scale=0):
+
+    if target =="Load original IP-Adapter":
+        # target_blocks=["blocks"] for original IP-Adapter
+        ip_model = IPAdapterXL(pipe, image_encoder_path, ip_ckpt, device, target_blocks=["blocks"])
+    elif target=="Load only style blocks":
+        # target_blocks=["up_blocks.0.attentions.1"] for style blocks only
+        ip_model = IPAdapterXL(pipe, image_encoder_path, ip_ckpt, device, target_blocks=["up_blocks.0.attentions.1"])
+    elif target == "Load style+layout block":
+        # target_blocks = ["up_blocks.0.attentions.1", "down_blocks.2.attentions.1"] # for style+layout blocks
+        ip_model = IPAdapterXL(pipe, image_encoder_path, ip_ckpt, device, target_blocks=["up_blocks.0.attentions.1", "down_blocks.2.attentions.1"])
+    
+    if input_image is not None:
+        input_image = resize_img(input_image, max_side=1024)
+        cv_input_image = pil_to_cv2(input_image)
+        detected_map = cv2.Canny(cv_input_image, 50, 200)
+        canny_map = Image.fromarray(cv2.cvtColor(detected_map, cv2.COLOR_BGR2RGB))
+    else:
+        canny_map = Image.new('RGB', (1024, 1024), color=(255, 255, 255))
+        control_scale = 0
+
+    if float(control_scale) == 0:
+        canny_map = canny_map.resize((1024,1024))
+    
+    if len(neg_content_prompt) > 0 and neg_content_scale != 0:
+        images = ip_model.generate(pil_image=image_pil,
+                                prompt=prompt,
+                                negative_prompt=n_prompt,
+                                scale=scale,
+                                guidance_scale=guidance_scale,
+                                num_samples=num_samples,
+                                num_inference_steps=num_inference_steps, 
+                                seed=seed,
+                                image=canny_map,
+                                controlnet_conditioning_scale=float(control_scale),
+                                neg_content_prompt=neg_content_prompt,
+                                neg_content_scale=neg_content_scale
+                                )
+    else:
+        images = ip_model.generate(pil_image=image_pil,
+                                prompt=prompt,
+                                negative_prompt=n_prompt,
+                                scale=scale,
+                                guidance_scale=guidance_scale,
+                                num_samples=num_samples,
+                                num_inference_steps=num_inference_steps, 
+                                seed=seed,
+                                image=canny_map,
+                                controlnet_conditioning_scale=float(control_scale),
+                                )
+    return images
+
+def pil_to_cv2(image_pil):
+    image_np = np.array(image_pil)
+    image_cv2 = cv2.cvtColor(image_np, cv2.COLOR_RGB2BGR)
+    return image_cv2
+
+# Description
+title = r"""
+<h1 align="center">InstantStyle: Free Lunch towards Style-Preserving in Text-to-Image Generation</h1>
+"""
+
+description = r"""
+<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/InstantStyle/InstantStyle' target='_blank'><b>InstantStyle: Free Lunch towards Style-Preserving in Text-to-Image Generation</b></a>.<br>
+
+How to use:<br>
+1. Upload a style image.
+2. Set stylization mode, only use style block by default.
+2. Enter a text prompt, as done in normal text-to-image models.
+3. Click the <b>Submit</b> button to begin customization.
+4. Share your stylized photo with your friends and enjoy! 😊
+
+
+Advanced usage:<br>
+1. Click advanced options.
+2. Upload another source image for image-based stylization using ControlNet.
+3. Enter negative content prompt to avoid content leakage.
+"""
+
+article = r"""
+---
+📝 **Citation**
+<br>
+If our work is helpful for your research or applications, please cite us via:
+```bibtex
+@article{wang2024instantstyle,
+  title={InstantStyle: Free Lunch towards Style-Preserving in Text-to-Image Generation},
+  author={Wang, Haofan and Wang, Qixun and Bai, Xu and Qin, Zekui and Chen, Anthony},
+  journal={arXiv preprint arXiv:2404.02733},
+  year={2024}
+}
+```
+📧 **Contact**
+<br>
+If you have any questions, please feel free to open an issue or directly reach us out at <b>[email protected]</b>.
+"""
+
+block = gr.Blocks(css="footer {visibility: hidden}").queue(max_size=10, api_open=False)
+with block:
+    
+    # description
+    gr.Markdown(title)
+    gr.Markdown(description)
+    
+    with gr.Tabs():
+        with gr.Row():
+            with gr.Column():
+                
+                with gr.Row():
+                    with gr.Column():
+                        image_pil = gr.Image(label="Style Image", type='pil')
+                
+                target = gr.Radio(["Load only style blocks", "Load style+layout block", "Load original IP-Adapter"], 
+                                  value="Load only style blocks",
+                                  label="Style mode")
+                
+                prompt = gr.Textbox(label="Prompt",
+                                    value="a cat, masterpiece, best quality, high quality")
+                
+                scale = gr.Slider(minimum=0,maximum=2.0, step=0.01,value=1.0, label="Scale")
+                
+                with gr.Accordion(open=False, label="Advanced Options"):
+                    
+                    with gr.Column():
+                        src_image_pil = gr.Image(label="Source Image (optional)", type='pil')
+                    control_scale = gr.Slider(minimum=0,maximum=1.0, step=0.01,value=0.5, label="Controlnet conditioning scale")
+                    
+                    n_prompt = gr.Textbox(label="Neg Prompt", value="text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry")
+                    
+                    neg_content_prompt = gr.Textbox(label="Neg Content Prompt", value="")
+                    neg_content_scale = gr.Slider(minimum=0, maximum=1.0, step=0.01,value=0.5, label="Neg Content Scale")
+
+                    guidance_scale = gr.Slider(minimum=1,maximum=15.0, step=0.01,value=5.0, label="guidance scale")
+                    num_samples= gr.Slider(minimum=1,maximum=4.0, step=1.0,value=1.0, label="num samples")
+                    num_inference_steps = gr.Slider(minimum=5,maximum=50.0, step=1.0,value=20, label="num inference steps")
+                    seed = gr.Slider(minimum=-1000000,maximum=1000000,value=1, step=1, label="Seed Value")
+                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+                    
+                generate_button = gr.Button("Generate Image")
+                
+            with gr.Column():
+                generated_image = gr.Gallery(label="Generated Image")
+
+        generate_button.click(
+            fn=randomize_seed_fn,
+            inputs=[seed, randomize_seed],
+            outputs=seed,
+            queue=False,
+            api_name=False,
+        ).then(
+            fn=create_image,
+            inputs=[image_pil,
+                    src_image_pil,
+                    prompt,
+                    n_prompt,
+                    scale, 
+                    control_scale, 
+                    guidance_scale,
+                    num_samples,
+                    num_inference_steps,
+                    seed,
+                    target,
+                    neg_content_prompt,
+                    neg_content_scale], 
+            outputs=[generated_image])
+    
+    gr.Examples(
+        examples=get_example(),
+        inputs=[image_pil, src_image_pil, prompt, scale, control_scale],
+        fn=run_for_examples,
+        outputs=[generated_image],
+        cache_examples=True,
+    )
+    
+    gr.Markdown(article)
+
+block.launch(server_name="10.4.200.46")

ip_adapter/__init__.py

@@ -0,0 +1,9 @@
+from .ip_adapter import IPAdapter, IPAdapterPlus, IPAdapterPlusXL, IPAdapterXL, IPAdapterFull
+
+__all__ = [
+    "IPAdapter",
+    "IPAdapterPlus",
+    "IPAdapterPlusXL",
+    "IPAdapterXL",
+    "IPAdapterFull",
+]

ip_adapter/attention_processor.py

@@ -0,0 +1,562 @@
+# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class AttnProcessor(nn.Module):
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+        self.skip = skip
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        if not self.skip:
+            # for ip-adapter
+            ip_key = self.to_k_ip(ip_hidden_states)
+            ip_value = self.to_v_ip(ip_hidden_states)
+
+            ip_key = attn.head_to_batch_dim(ip_key)
+            ip_value = attn.head_to_batch_dim(ip_value)
+
+            ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+            self.attn_map = ip_attention_probs
+            ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+            ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+            hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+        self.skip = skip
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if not self.skip:
+            # for ip-adapter
+            ip_key = self.to_k_ip(ip_hidden_states)
+            ip_value = self.to_v_ip(ip_hidden_states)
+
+            ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            # the output of sdp = (batch, num_heads, seq_len, head_dim)
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            ip_hidden_states = F.scaled_dot_product_attention(
+                query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+            with torch.no_grad():
+                self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
+                #print(self.attn_map.shape)
+
+            ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+            ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+            hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+## for controlnet
+class CNAttnProcessor:
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(self, num_tokens=4):
+        self.num_tokens = num_tokens
+
+    def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class CNAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self, num_tokens=4):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.num_tokens = num_tokens
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states

ip_adapter/ip_adapter.py

@@ -0,0 +1,461 @@
+import os
+from typing import List
+
+import torch
+from diffusers import StableDiffusionPipeline
+from diffusers.pipelines.controlnet import MultiControlNetModel
+from PIL import Image
+from safetensors import safe_open
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+
+from .utils import is_torch2_available, get_generator
+
+if is_torch2_available():
+    from .attention_processor import (
+        AttnProcessor2_0 as AttnProcessor,
+    )
+    from .attention_processor import (
+        CNAttnProcessor2_0 as CNAttnProcessor,
+    )
+    from .attention_processor import (
+        IPAttnProcessor2_0 as IPAttnProcessor,
+    )
+else:
+    from .attention_processor import AttnProcessor, CNAttnProcessor, IPAttnProcessor
+from .resampler import Resampler
+
+
+class ImageProjModel(torch.nn.Module):
+    """Projection Model"""
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
+        super().__init__()
+
+        self.generator = None
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class MLPProjModel(torch.nn.Module):
+    """SD model with image prompt"""
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024):
+        super().__init__()
+        
+        self.proj = torch.nn.Sequential(
+            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
+            torch.nn.LayerNorm(cross_attention_dim)
+        )
+        
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class IPAdapter:
+    def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device, num_tokens=4, target_blocks=["block"]):
+        self.device = device
+        self.image_encoder_path = image_encoder_path
+        self.ip_ckpt = ip_ckpt
+        self.num_tokens = num_tokens
+        self.target_blocks = target_blocks
+
+        self.pipe = sd_pipe.to(self.device)
+        self.set_ip_adapter()
+
+        # load image encoder
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
+            self.device, dtype=torch.float16
+        )
+        self.clip_image_processor = CLIPImageProcessor()
+        # image proj model
+        self.image_proj_model = self.init_proj()
+
+        self.load_ip_adapter()
+
+    
+    def init_proj(self):
+        image_proj_model = ImageProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.projection_dim,
+            clip_extra_context_tokens=self.num_tokens,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    def set_ip_adapter(self):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor()
+            else:
+                selected = False
+                for block_name in self.target_blocks:
+                    if block_name in name:
+                        selected = True
+                        break
+                if selected:
+                    attn_procs[name] = IPAttnProcessor(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=self.num_tokens,
+                    ).to(self.device, dtype=torch.float16)
+                else:
+                    attn_procs[name] = IPAttnProcessor(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=self.num_tokens,
+                        skip=True
+                    ).to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        if hasattr(self.pipe, "controlnet"):
+            if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                for controlnet in self.pipe.controlnet.nets:
+                    controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
+            else:
+                self.pipe.controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
+
+    def load_ip_adapter(self):
+        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
+            state_dict = {"image_proj": {}, "ip_adapter": {}}
+            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    if key.startswith("image_proj."):
+                        state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                    elif key.startswith("ip_adapter."):
+                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+        else:
+            state_dict = torch.load(self.ip_ckpt, map_location="cpu")
+        self.image_proj_model.load_state_dict(state_dict["image_proj"])
+        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
+        ip_layers.load_state_dict(state_dict["ip_adapter"], strict=False)
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None, content_prompt_embeds=None):
+        if pil_image is not None:
+            if isinstance(pil_image, Image.Image):
+                pil_image = [pil_image]
+            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+        else:
+            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+        
+        if content_prompt_embeds is not None:
+            clip_image_embeds = clip_image_embeds - content_prompt_embeds
+
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds))
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def set_scale(self, scale):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, IPAttnProcessor):
+                attn_processor.scale = scale
+
+    def generate(
+        self,
+        pil_image=None,
+        clip_image_embeds=None,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        num_samples=4,
+        seed=None,
+        guidance_scale=7.5,
+        num_inference_steps=30,
+        neg_content_emb=None,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        if pil_image is not None:
+            num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
+        else:
+            num_prompts = clip_image_embeds.size(0)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
+            pil_image=pil_image, clip_image_embeds=clip_image_embeds, content_prompt_embeds=neg_content_emb
+        )
+        bs_embed, seq_len, _ = image_prompt_embeds.shape
+        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+
+        with torch.inference_mode():
+            prompt_embeds_, negative_prompt_embeds_ = self.pipe.encode_prompt(
+                prompt,
+                device=self.device,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds_, image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds_, uncond_image_prompt_embeds], dim=1)
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class IPAdapterXL(IPAdapter):
+    """SDXL"""
+
+    def generate(
+        self,
+        pil_image,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        num_samples=4,
+        seed=None,
+        num_inference_steps=30,
+        neg_content_emb=None,
+        neg_content_prompt=None,
+        neg_content_scale=1.0,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+        
+        if neg_content_emb is None:
+            if neg_content_prompt is not None:
+                with torch.inference_mode():
+                    (
+                        prompt_embeds_, # torch.Size([1, 77, 2048])
+                        negative_prompt_embeds_,
+                        pooled_prompt_embeds_, # torch.Size([1, 1280])
+                        negative_pooled_prompt_embeds_,
+                    ) = self.pipe.encode_prompt(
+                        neg_content_prompt,
+                        num_images_per_prompt=num_samples,
+                        do_classifier_free_guidance=True,
+                        negative_prompt=negative_prompt,
+                    )
+                    pooled_prompt_embeds_ *= neg_content_scale
+            else:
+                pooled_prompt_embeds_ = neg_content_emb
+        else:
+            pooled_prompt_embeds_ = None
+
+        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image, content_prompt_embeds=pooled_prompt_embeds_)
+        bs_embed, seq_len, _ = image_prompt_embeds.shape
+        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
+
+        self.generator = get_generator(seed, self.device)
+        
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            generator=self.generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class IPAdapterPlus(IPAdapter):
+    """IP-Adapter with fine-grained features"""
+
+    def init_proj(self):
+        image_proj_model = Resampler(
+            dim=self.pipe.unet.config.cross_attention_dim,
+            depth=4,
+            dim_head=64,
+            heads=12,
+            num_queries=self.num_tokens,
+            embedding_dim=self.image_encoder.config.hidden_size,
+            output_dim=self.pipe.unet.config.cross_attention_dim,
+            ff_mult=4,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        clip_image = clip_image.to(self.device, dtype=torch.float16)
+        clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = self.image_encoder(
+            torch.zeros_like(clip_image), output_hidden_states=True
+        ).hidden_states[-2]
+        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+class IPAdapterFull(IPAdapterPlus):
+    """IP-Adapter with full features"""
+
+    def init_proj(self):
+        image_proj_model = MLPProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.hidden_size,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+
+class IPAdapterPlusXL(IPAdapter):
+    """SDXL"""
+
+    def init_proj(self):
+        image_proj_model = Resampler(
+            dim=1280,
+            depth=4,
+            dim_head=64,
+            heads=20,
+            num_queries=self.num_tokens,
+            embedding_dim=self.image_encoder.config.hidden_size,
+            output_dim=self.pipe.unet.config.cross_attention_dim,
+            ff_mult=4,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        clip_image = clip_image.to(self.device, dtype=torch.float16)
+        clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = self.image_encoder(
+            torch.zeros_like(clip_image), output_hidden_states=True
+        ).hidden_states[-2]
+        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def generate(
+        self,
+        pil_image,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        num_samples=4,
+        seed=None,
+        num_inference_steps=30,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image)
+        bs_embed, seq_len, _ = image_prompt_embeds.shape
+        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images

ip_adapter/resampler.py

@@ -0,0 +1,158 @@
+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py
+
+import math
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+        max_seq_len: int = 257,  # CLIP tokens + CLS token
+        apply_pos_emb: bool = False,
+        num_latents_mean_pooled: int = 0,  # number of latents derived from mean pooled representation of the sequence
+    ):
+        super().__init__()
+        self.pos_emb = nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.to_latents_from_mean_pooled_seq = (
+            nn.Sequential(
+                nn.LayerNorm(dim),
+                nn.Linear(dim, dim * num_latents_mean_pooled),
+                Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
+            )
+            if num_latents_mean_pooled > 0
+            else None
+        )
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        if self.pos_emb is not None:
+            n, device = x.shape[1], x.device
+            pos_emb = self.pos_emb(torch.arange(n, device=device))
+            x = x + pos_emb
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        if self.to_latents_from_mean_pooled_seq:
+            meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool))
+            meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
+            latents = torch.cat((meanpooled_latents, latents), dim=-2)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+def masked_mean(t, *, dim, mask=None):
+    if mask is None:
+        return t.mean(dim=dim)
+
+    denom = mask.sum(dim=dim, keepdim=True)
+    mask = rearrange(mask, "b n -> b n 1")
+    masked_t = t.masked_fill(~mask, 0.0)
+
+    return masked_t.sum(dim=dim) / denom.clamp(min=1e-5)

ip_adapter/utils.py

@@ -0,0 +1,93 @@
+import torch
+import torch.nn.functional as F
+import numpy as np
+from PIL import Image
+
+attn_maps = {}
+def hook_fn(name):
+    def forward_hook(module, input, output):
+        if hasattr(module.processor, "attn_map"):
+            attn_maps[name] = module.processor.attn_map
+            del module.processor.attn_map
+
+    return forward_hook
+
+def register_cross_attention_hook(unet):
+    for name, module in unet.named_modules():
+        if name.split('.')[-1].startswith('attn2'):
+            module.register_forward_hook(hook_fn(name))
+
+    return unet
+
+def upscale(attn_map, target_size):
+    attn_map = torch.mean(attn_map, dim=0)
+    attn_map = attn_map.permute(1,0)
+    temp_size = None
+
+    for i in range(0,5):
+        scale = 2 ** i
+        if ( target_size[0] // scale ) * ( target_size[1] // scale) == attn_map.shape[1]*64:
+            temp_size = (target_size[0]//(scale*8), target_size[1]//(scale*8))
+            break
+
+    assert temp_size is not None, "temp_size cannot is None"
+
+    attn_map = attn_map.view(attn_map.shape[0], *temp_size)
+
+    attn_map = F.interpolate(
+        attn_map.unsqueeze(0).to(dtype=torch.float32),
+        size=target_size,
+        mode='bilinear',
+        align_corners=False
+    )[0]
+
+    attn_map = torch.softmax(attn_map, dim=0)
+    return attn_map
+def get_net_attn_map(image_size, batch_size=2, instance_or_negative=False, detach=True):
+
+    idx = 0 if instance_or_negative else 1
+    net_attn_maps = []
+
+    for name, attn_map in attn_maps.items():
+        attn_map = attn_map.cpu() if detach else attn_map
+        attn_map = torch.chunk(attn_map, batch_size)[idx].squeeze()
+        attn_map = upscale(attn_map, image_size) 
+        net_attn_maps.append(attn_map) 
+
+    net_attn_maps = torch.mean(torch.stack(net_attn_maps,dim=0),dim=0)
+
+    return net_attn_maps
+
+def attnmaps2images(net_attn_maps):
+
+    #total_attn_scores = 0
+    images = []
+
+    for attn_map in net_attn_maps:
+        attn_map = attn_map.cpu().numpy()
+        #total_attn_scores += attn_map.mean().item()
+
+        normalized_attn_map = (attn_map - np.min(attn_map)) / (np.max(attn_map) - np.min(attn_map)) * 255
+        normalized_attn_map = normalized_attn_map.astype(np.uint8)
+        #print("norm: ", normalized_attn_map.shape)
+        image = Image.fromarray(normalized_attn_map)
+
+        #image = fix_save_attn_map(attn_map)
+        images.append(image)
+
+    #print(total_attn_scores)
+    return images
+def is_torch2_available():
+    return hasattr(F, "scaled_dot_product_attention")
+
+def get_generator(seed, device):
+
+    if seed is not None:
+        if isinstance(seed, list):
+            generator = [torch.Generator(device).manual_seed(seed_item) for seed_item in seed]
+        else:
+            generator = torch.Generator(device).manual_seed(seed)
+    else:
+        generator = None
+
+    return generator

requirements.txt

@@ -0,0 +1,16 @@
+diffusers>=0.25.1
+torch>=2.0.0
+torchvision>=0.15.1
+transformers>=4.37.1
+accelerate
+safetensors
+einops
+spaces>=0.19.4
+omegaconf
+peft
+huggingface-hub==0.20.2
+opencv-python
+gradio
+controlnet_aux
+gdown
+peft