# Copyright 2024 The InstantX Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import inspect from typing import Any, Callable, Dict, List, Optional, Tuple, Union import numpy as np import PIL.Image import torch import torch.nn.functional as F from transformers import ( CLIPImageProcessor, CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer, CLIPVisionModelWithProjection, ) from diffusers.utils.import_utils import is_invisible_watermark_available from diffusers.image_processor import PipelineImageInput, VaeImageProcessor from diffusers.loaders import ( FromSingleFileMixin, IPAdapterMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin, ) from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel from diffusers.models.attention_processor import ( AttnProcessor2_0, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor, XFormersAttnProcessor, ) from diffusers.models.lora import adjust_lora_scale_text_encoder from diffusers.schedulers import KarrasDiffusionSchedulers, LCMScheduler from diffusers.utils import ( USE_PEFT_BACKEND, deprecate, logging, replace_example_docstring, scale_lora_layers, unscale_lora_layers, convert_unet_state_dict_to_peft ) from diffusers.utils.torch_utils import is_compiled_module, is_torch_version, randn_tensor from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin from diffusers.pipelines.stable_diffusion_xl.pipeline_output import StableDiffusionXLPipelineOutput if is_invisible_watermark_available(): from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker from peft import LoraConfig, set_peft_model_state_dict from module.aggregator import Aggregator logger = logging.get_logger(__name__) # pylint: disable=invalid-name EXAMPLE_DOC_STRING = """ Examples: ```py >>> # !pip install diffusers pillow transformers accelerate >>> import torch >>> from PIL import Image >>> from diffusers import DDPMScheduler >>> from schedulers.lcm_single_step_scheduler import LCMSingleStepScheduler >>> from module.ip_adapter.utils import load_adapter_to_pipe >>> from pipelines.sdxl_instantir import InstantIRPipeline >>> # download models under ./models >>> dcp_adapter = f'./models/adapter.pt' >>> previewer_lora_path = f'./models' >>> instantir_path = f'./models/aggregator.pt' >>> # load pretrained models >>> pipe = InstantIRPipeline.from_pretrained( ... "stabilityai/stable-diffusion-xl-base-1.0", controlnet=controlnet, vae=vae, torch_dtype=torch.float16 ... ) >>> # load adapter >>> load_adapter_to_pipe( ... pipe, ... dcp_adapter, ... image_encoder_or_path = 'facebook/dinov2-large', ... ) >>> # load previewer lora >>> pipe.prepare_previewers(previewer_lora_path) >>> pipe.scheduler = DDPMScheduler.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', subfolder="scheduler") >>> lcm_scheduler = LCMSingleStepScheduler.from_config(pipe.scheduler.config) >>> # load aggregator weights >>> pretrained_state_dict = torch.load(instantir_path) >>> pipe.aggregator.load_state_dict(pretrained_state_dict) >>> # send to GPU and fp16 >>> pipe.to(device="cuda", dtype=torch.float16) >>> pipe.aggregator.to(device="cuda", dtype=torch.float16) >>> pipe.enable_model_cpu_offload() >>> # load a broken image >>> low_quality_image = Image.open('path/to/your-image').convert("RGB") >>> # restoration >>> image = pipe( ... image=low_quality_image, ... previewer_scheduler=lcm_scheduler, ... ).images[0] ``` """ LCM_LORA_MODULES = [ "to_q", "to_k", "to_v", "to_out.0", "proj_in", "proj_out", "ff.net.0.proj", "ff.net.2", "conv1", "conv2", "conv_shortcut", "downsamplers.0.conv", "upsamplers.0.conv", "time_emb_proj", ] PREVIEWER_LORA_MODULES = [ "to_q", "to_kv", "0.to_out", "attn1.to_k", "attn1.to_v", "to_k_ip", "to_v_ip", "ln_k_ip.linear", "ln_v_ip.linear", "to_out.0", "proj_in", "proj_out", "ff.net.0.proj", "ff.net.2", "conv1", "conv2", "conv_shortcut", "downsamplers.0.conv", "upsamplers.0.conv", "time_emb_proj", ] def remove_attn2(model): def recursive_find_module(name, module): if not "up_blocks" in name and not "down_blocks" in name and not "mid_block" in name: return elif "resnets" in name: return if hasattr(module, "attn2"): setattr(module, "attn2", None) setattr(module, "norm2", None) return for sub_name, sub_module in module.named_children(): recursive_find_module(f"{name}.{sub_name}", sub_module) for name, module in model.named_children(): recursive_find_module(name, module) # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): """ Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 """ std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) # rescale the results from guidance (fixes overexposure) noise_pred_rescaled = noise_cfg * (std_text / std_cfg) # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg return noise_cfg # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps def retrieve_timesteps( scheduler, num_inference_steps: Optional[int] = None, device: Optional[Union[str, torch.device]] = None, timesteps: Optional[List[int]] = None, **kwargs, ): """ Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. Args: scheduler (`SchedulerMixin`): The scheduler to get timesteps from. num_inference_steps (`int`): The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps` must be `None`. device (`str` or `torch.device`, *optional*): The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. timesteps (`List[int]`, *optional*): Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` must be `None`. Returns: `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the second element is the number of inference steps. """ if timesteps is not None: accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) if not accepts_timesteps: raise ValueError( f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" f" timestep schedules. Please check whether you are using the correct scheduler." ) scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) timesteps = scheduler.timesteps num_inference_steps = len(timesteps) else: scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) timesteps = scheduler.timesteps return timesteps, num_inference_steps class InstantIRPipeline( DiffusionPipeline, StableDiffusionMixin, TextualInversionLoaderMixin, StableDiffusionXLLoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin, ): r""" Pipeline for text-to-image generation using Stable Diffusion XL with ControlNet guidance. This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular device, etc.). The pipeline also inherits the following loading methods: - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings - [`~loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] for loading LoRA weights - [`~loaders.StableDiffusionXLLoraLoaderMixin.save_lora_weights`] for saving LoRA weights - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters Args: vae ([`AutoencoderKL`]): Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. text_encoder ([`~transformers.CLIPTextModel`]): Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]): Second frozen text-encoder ([laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)). tokenizer ([`~transformers.CLIPTokenizer`]): A `CLIPTokenizer` to tokenize text. tokenizer_2 ([`~transformers.CLIPTokenizer`]): A `CLIPTokenizer` to tokenize text. unet ([`UNet2DConditionModel`]): A `UNet2DConditionModel` to denoise the encoded image latents. controlnet ([`ControlNetModel`] or `List[ControlNetModel]`): Provides additional conditioning to the `unet` during the denoising process. If you set multiple ControlNets as a list, the outputs from each ControlNet are added together to create one combined additional conditioning. scheduler ([`SchedulerMixin`]): A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`): Whether the negative prompt embeddings should always be set to 0. Also see the config of `stabilityai/stable-diffusion-xl-base-1-0`. add_watermarker (`bool`, *optional*): Whether to use the [invisible_watermark](https://github.com/ShieldMnt/invisible-watermark/) library to watermark output images. If not defined, it defaults to `True` if the package is installed; otherwise no watermarker is used. """ # leave controlnet out on purpose because it iterates with unet model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->unet->vae" _optional_components = [ "tokenizer", "tokenizer_2", "text_encoder", "text_encoder_2", "feature_extractor", "image_encoder", ] _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] def __init__( self, vae: AutoencoderKL, text_encoder: CLIPTextModel, text_encoder_2: CLIPTextModelWithProjection, tokenizer: CLIPTokenizer, tokenizer_2: CLIPTokenizer, unet: UNet2DConditionModel, scheduler: KarrasDiffusionSchedulers, aggregator: Aggregator = None, force_zeros_for_empty_prompt: bool = True, add_watermarker: Optional[bool] = None, feature_extractor: CLIPImageProcessor = None, image_encoder: CLIPVisionModelWithProjection = None, ): super().__init__() if aggregator is None: aggregator = Aggregator.from_unet(unet) remove_attn2(aggregator) self.register_modules( vae=vae, text_encoder=text_encoder, text_encoder_2=text_encoder_2, tokenizer=tokenizer, tokenizer_2=tokenizer_2, unet=unet, aggregator=aggregator, scheduler=scheduler, feature_extractor=feature_extractor, image_encoder=image_encoder, ) self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True) self.control_image_processor = VaeImageProcessor( vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=True ) add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() if add_watermarker: self.watermark = StableDiffusionXLWatermarker() else: self.watermark = None self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) def prepare_previewers(self, previewer_lora_path: str, use_lcm=False): if use_lcm: lora_state_dict, alpha_dict = self.lora_state_dict( previewer_lora_path, ) else: lora_state_dict, alpha_dict = self.lora_state_dict( previewer_lora_path, weight_name="previewer_lora_weights.bin" ) unet_state_dict = { f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.") } unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict) lora_state_dict = dict() for k, v in unet_state_dict.items(): if "ip" in k: k = k.replace("attn2", "attn2.processor") lora_state_dict[k] = v else: lora_state_dict[k] = v if alpha_dict: lora_alpha = next(iter(alpha_dict.values())) else: lora_alpha = 1 logger.info(f"use lora alpha {lora_alpha}") lora_config = LoraConfig( r=64, target_modules=LCM_LORA_MODULES if use_lcm else PREVIEWER_LORA_MODULES, lora_alpha=lora_alpha, lora_dropout=0.0, ) adapter_name = "lcm" if use_lcm else "previewer" self.unet.add_adapter(lora_config, adapter_name) incompatible_keys = set_peft_model_state_dict(self.unet, lora_state_dict, adapter_name=adapter_name) if incompatible_keys is not None: # check only for unexpected keys unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None) missing_keys = getattr(incompatible_keys, "missing_keys", None) if unexpected_keys: raise ValueError( f"Loading adapter weights from state_dict led to unexpected keys not found in the model: " f" {unexpected_keys}. " ) self.unet.disable_adapters() return lora_alpha # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.encode_prompt def encode_prompt( self, prompt: str, prompt_2: Optional[str] = None, device: Optional[torch.device] = None, num_images_per_prompt: int = 1, do_classifier_free_guidance: bool = True, negative_prompt: Optional[str] = None, negative_prompt_2: Optional[str] = None, prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.FloatTensor] = None, pooled_prompt_embeds: Optional[torch.FloatTensor] = None, negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, lora_scale: Optional[float] = None, clip_skip: Optional[int] = None, ): r""" Encodes the prompt into text encoder hidden states. Args: prompt (`str` or `List[str]`, *optional*): prompt to be encoded prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is used in both text-encoders device: (`torch.device`): torch device num_images_per_prompt (`int`): number of images that should be generated per prompt do_classifier_free_guidance (`bool`): whether to use classifier free guidance or not negative_prompt (`str` or `List[str]`, *optional*): The prompt or prompts not to guide the image generation. If not defined, one has to pass `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`). negative_prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. negative_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input argument. pooled_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, pooled text embeddings will be generated from `prompt` input argument. negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` input argument. lora_scale (`float`, *optional*): A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. clip_skip (`int`, *optional*): Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that the output of the pre-final layer will be used for computing the prompt embeddings. """ device = device or self._execution_device # set lora scale so that monkey patched LoRA # function of text encoder can correctly access it if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin): self._lora_scale = lora_scale # dynamically adjust the LoRA scale if self.text_encoder is not None: if not USE_PEFT_BACKEND: adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) else: scale_lora_layers(self.text_encoder, lora_scale) if self.text_encoder_2 is not None: if not USE_PEFT_BACKEND: adjust_lora_scale_text_encoder(self.text_encoder_2, lora_scale) else: scale_lora_layers(self.text_encoder_2, lora_scale) prompt = [prompt] if isinstance(prompt, str) else prompt if prompt is not None: batch_size = len(prompt) else: batch_size = prompt_embeds.shape[0] # Define tokenizers and text encoders tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] text_encoders = ( [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] ) if prompt_embeds is None: prompt_2 = prompt_2 or prompt prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2 # textual inversion: process multi-vector tokens if necessary prompt_embeds_list = [] prompts = [prompt, prompt_2] for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): if isinstance(self, TextualInversionLoaderMixin): prompt = self.maybe_convert_prompt(prompt, tokenizer) text_inputs = tokenizer( prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt", ) text_input_ids = text_inputs.input_ids untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( text_input_ids, untruncated_ids ): removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) logger.warning( "The following part of your input was truncated because CLIP can only handle sequences up to" f" {tokenizer.model_max_length} tokens: {removed_text}" ) prompt_embeds = text_encoder(text_input_ids.to(device), output_hidden_states=True) # We are only ALWAYS interested in the pooled output of the final text encoder pooled_prompt_embeds = prompt_embeds[0] if clip_skip is None: prompt_embeds = prompt_embeds.hidden_states[-2] else: # "2" because SDXL always indexes from the penultimate layer. prompt_embeds = prompt_embeds.hidden_states[-(clip_skip + 2)] prompt_embeds_list.append(prompt_embeds) prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) # get unconditional embeddings for classifier free guidance zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: negative_prompt_embeds = torch.zeros_like(prompt_embeds) negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) elif do_classifier_free_guidance and negative_prompt_embeds is None: negative_prompt = negative_prompt or "" negative_prompt_2 = negative_prompt_2 or negative_prompt # normalize str to list negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt negative_prompt_2 = ( batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2 ) uncond_tokens: List[str] if prompt is not None and type(prompt) is not type(negative_prompt): raise TypeError( f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" f" {type(prompt)}." ) elif batch_size != len(negative_prompt): raise ValueError( f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" " the batch size of `prompt`." ) else: uncond_tokens = [negative_prompt, negative_prompt_2] negative_prompt_embeds_list = [] for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): if isinstance(self, TextualInversionLoaderMixin): negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) max_length = prompt_embeds.shape[1] uncond_input = tokenizer( negative_prompt, padding="max_length", max_length=max_length, truncation=True, return_tensors="pt", ) negative_prompt_embeds = text_encoder( uncond_input.input_ids.to(device), output_hidden_states=True, ) # We are only ALWAYS interested in the pooled output of the final text encoder negative_pooled_prompt_embeds = negative_prompt_embeds[0] negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] negative_prompt_embeds_list.append(negative_prompt_embeds) negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) if self.text_encoder_2 is not None: prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) else: prompt_embeds = prompt_embeds.to(dtype=self.unet.dtype, device=device) bs_embed, seq_len, _ = prompt_embeds.shape # duplicate text embeddings for each generation per prompt, using mps friendly method prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) if do_classifier_free_guidance: # duplicate unconditional embeddings for each generation per prompt, using mps friendly method seq_len = negative_prompt_embeds.shape[1] if self.text_encoder_2 is not None: negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) else: negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.unet.dtype, device=device) negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( bs_embed * num_images_per_prompt, -1 ) if do_classifier_free_guidance: negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( bs_embed * num_images_per_prompt, -1 ) if self.text_encoder is not None: if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: # Retrieve the original scale by scaling back the LoRA layers unscale_lora_layers(self.text_encoder, lora_scale) if self.text_encoder_2 is not None: if isinstance(self, StableDiffusionXLLoraLoaderMixin) and USE_PEFT_BACKEND: # Retrieve the original scale by scaling back the LoRA layers unscale_lora_layers(self.text_encoder_2, lora_scale) return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): dtype = next(self.image_encoder.parameters()).dtype if not isinstance(image, torch.Tensor): image = self.feature_extractor(image, return_tensors="pt").pixel_values image = image.to(device=device, dtype=dtype) if output_hidden_states: image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) uncond_image_enc_hidden_states = self.image_encoder( torch.zeros_like(image), output_hidden_states=True ).hidden_states[-2] uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( num_images_per_prompt, dim=0 ) return image_enc_hidden_states, uncond_image_enc_hidden_states else: if isinstance(self.image_encoder, CLIPVisionModelWithProjection): # CLIP image encoder. image_embeds = self.image_encoder(image).image_embeds image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) uncond_image_embeds = torch.zeros_like(image_embeds) else: # DINO image encoder. image_embeds = self.image_encoder(image).last_hidden_state image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) uncond_image_embeds = self.image_encoder( torch.zeros_like(image) ).last_hidden_state uncond_image_embeds = uncond_image_embeds.repeat_interleave( num_images_per_prompt, dim=0 ) return image_embeds, uncond_image_embeds # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds def prepare_ip_adapter_image_embeds( self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance ): if ip_adapter_image_embeds is None: if not isinstance(ip_adapter_image, list): ip_adapter_image = [ip_adapter_image] if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): if isinstance(ip_adapter_image[0], list): raise ValueError( f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." ) else: logger.warning( f"Got {len(ip_adapter_image)} images for {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." " By default, these images will be sent to each IP-Adapter. If this is not your use-case, please specify `ip_adapter_image` as a list of image-list, with" f" length equals to the number of IP-Adapters." ) ip_adapter_image = [ip_adapter_image] * len(self.unet.encoder_hid_proj.image_projection_layers) image_embeds = [] for single_ip_adapter_image, image_proj_layer in zip( ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers ): output_hidden_state = isinstance(self.image_encoder, CLIPVisionModelWithProjection) and not isinstance(image_proj_layer, ImageProjection) single_image_embeds, single_negative_image_embeds = self.encode_image( single_ip_adapter_image, device, 1, output_hidden_state ) single_image_embeds = torch.stack([single_image_embeds] * (num_images_per_prompt//single_image_embeds.shape[0]), dim=0) single_negative_image_embeds = torch.stack( [single_negative_image_embeds] * (num_images_per_prompt//single_negative_image_embeds.shape[0]), dim=0 ) if do_classifier_free_guidance: single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds]) single_image_embeds = single_image_embeds.to(device) image_embeds.append(single_image_embeds) else: repeat_dims = [1] image_embeds = [] for single_image_embeds in ip_adapter_image_embeds: if do_classifier_free_guidance: single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2) single_image_embeds = single_image_embeds.repeat( num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:])) ) single_negative_image_embeds = single_negative_image_embeds.repeat( num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:])) ) single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds]) else: single_image_embeds = single_image_embeds.repeat( num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:])) ) image_embeds.append(single_image_embeds) return image_embeds # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs def prepare_extra_step_kwargs(self, generator, eta): # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 # and should be between [0, 1] accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) extra_step_kwargs = {} if accepts_eta: extra_step_kwargs["eta"] = eta # check if the scheduler accepts generator accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) if accepts_generator: extra_step_kwargs["generator"] = generator return extra_step_kwargs def check_inputs( self, prompt, prompt_2, image, callback_steps, negative_prompt=None, negative_prompt_2=None, prompt_embeds=None, negative_prompt_embeds=None, pooled_prompt_embeds=None, ip_adapter_image=None, ip_adapter_image_embeds=None, negative_pooled_prompt_embeds=None, controlnet_conditioning_scale=1.0, control_guidance_start=0.0, control_guidance_end=1.0, callback_on_step_end_tensor_inputs=None, ): if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): raise ValueError( f"`callback_steps` has to be a positive integer but is {callback_steps} of type" f" {type(callback_steps)}." ) if callback_on_step_end_tensor_inputs is not None and not all( k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs ): raise ValueError( f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" ) if prompt is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" " only forward one of the two." ) elif prompt_2 is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" " only forward one of the two." ) elif prompt is None and prompt_embeds is None: raise ValueError( "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." ) elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") if negative_prompt is not None and negative_prompt_embeds is not None: raise ValueError( f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" f" {negative_prompt_embeds}. Please make sure to only forward one of the two." ) elif negative_prompt_2 is not None and negative_prompt_embeds is not None: raise ValueError( f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" f" {negative_prompt_embeds}. Please make sure to only forward one of the two." ) if prompt_embeds is not None and negative_prompt_embeds is not None: if prompt_embeds.shape != negative_prompt_embeds.shape: raise ValueError( "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" f" {negative_prompt_embeds.shape}." ) if prompt_embeds is not None and pooled_prompt_embeds is None: raise ValueError( "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." ) if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: raise ValueError( "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." ) # Check `image` is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( self.aggregator, torch._dynamo.eval_frame.OptimizedModule ) if ( isinstance(self.aggregator, Aggregator) or is_compiled and isinstance(self.aggregator._orig_mod, Aggregator) ): self.check_image(image, prompt, prompt_embeds) else: assert False if control_guidance_start >= control_guidance_end: raise ValueError( f"control guidance start: {control_guidance_start} cannot be larger or equal to control guidance end: {control_guidance_end}." ) if control_guidance_start < 0.0: raise ValueError(f"control guidance start: {control_guidance_start} can't be smaller than 0.") if control_guidance_end > 1.0: raise ValueError(f"control guidance end: {control_guidance_end} can't be larger than 1.0.") if ip_adapter_image is not None and ip_adapter_image_embeds is not None: raise ValueError( "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined." ) if ip_adapter_image_embeds is not None: if not isinstance(ip_adapter_image_embeds, list): raise ValueError( f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}" ) elif ip_adapter_image_embeds[0].ndim not in [3, 4]: raise ValueError( f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D" ) # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image def check_image(self, image, prompt, prompt_embeds): image_is_pil = isinstance(image, PIL.Image.Image) image_is_tensor = isinstance(image, torch.Tensor) image_is_np = isinstance(image, np.ndarray) image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) if ( not image_is_pil and not image_is_tensor and not image_is_np and not image_is_pil_list and not image_is_tensor_list and not image_is_np_list ): raise TypeError( f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" ) if image_is_pil: image_batch_size = 1 else: image_batch_size = len(image) if prompt is not None and isinstance(prompt, str): prompt_batch_size = 1 elif prompt is not None and isinstance(prompt, list): prompt_batch_size = len(prompt) elif prompt_embeds is not None: prompt_batch_size = prompt_embeds.shape[0] if image_batch_size != 1 and image_batch_size != prompt_batch_size: raise ValueError( f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" ) # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.prepare_image def prepare_image( self, image, width, height, batch_size, num_images_per_prompt, device, dtype, do_classifier_free_guidance=False, ): image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) image_batch_size = image.shape[0] if image_batch_size == 1: repeat_by = batch_size else: # image batch size is the same as prompt batch size repeat_by = num_images_per_prompt image = image.repeat_interleave(repeat_by, dim=0) image = image.to(device=device, dtype=dtype) return image @torch.no_grad() def init_latents(self, latents, generator, timestep): print(f"device={self.vae.device}, generator={generator}") noise = torch.randn(latents.shape, generator=generator, device=self.vae.device, dtype=self.vae.dtype, layout=torch.strided) bsz = latents.shape[0] print(f"init latent at {timestep}") timestep = torch.tensor([timestep]*bsz, device=self.vae.device) # Note that the latents will be scaled aleady by scheduler.add_noise latents = self.scheduler.add_noise(latents, noise, timestep) return latents # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): shape = ( batch_size, num_channels_latents, int(height) // self.vae_scale_factor, int(width) // self.vae_scale_factor, ) if isinstance(generator, list) and len(generator) != batch_size: raise ValueError( f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" f" size of {batch_size}. Make sure the batch size matches the length of the generators." ) if latents is None: latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) else: latents = latents.to(device) # scale the initial noise by the standard deviation required by the scheduler latents = latents * self.scheduler.init_noise_sigma return latents # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline._get_add_time_ids def _get_add_time_ids( self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None ): add_time_ids = list(original_size + crops_coords_top_left + target_size) passed_add_embed_dim = ( self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim ) expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features if expected_add_embed_dim != passed_add_embed_dim: raise ValueError( f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." ) add_time_ids = torch.tensor([add_time_ids], dtype=dtype) return add_time_ids # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae def upcast_vae(self): dtype = self.vae.dtype self.vae.to(dtype=torch.float32) use_torch_2_0_or_xformers = isinstance( self.vae.decoder.mid_block.attentions[0].processor, ( AttnProcessor2_0, XFormersAttnProcessor, LoRAXFormersAttnProcessor, LoRAAttnProcessor2_0, ), ) # if xformers or torch_2_0 is used attention block does not need # to be in float32 which can save lots of memory if use_torch_2_0_or_xformers: self.vae.post_quant_conv.to(dtype) self.vae.decoder.conv_in.to(dtype) self.vae.decoder.mid_block.to(dtype) # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding def get_guidance_scale_embedding( self, w: torch.Tensor, embedding_dim: int = 512, dtype: torch.dtype = torch.float32 ) -> torch.FloatTensor: """ See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 Args: w (`torch.Tensor`): Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. embedding_dim (`int`, *optional*, defaults to 512): Dimension of the embeddings to generate. dtype (`torch.dtype`, *optional*, defaults to `torch.float32`): Data type of the generated embeddings. Returns: `torch.FloatTensor`: Embedding vectors with shape `(len(w), embedding_dim)`. """ assert len(w.shape) == 1 w = w * 1000.0 half_dim = embedding_dim // 2 emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) emb = w.to(dtype)[:, None] * emb[None, :] emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) if embedding_dim % 2 == 1: # zero pad emb = torch.nn.functional.pad(emb, (0, 1)) assert emb.shape == (w.shape[0], embedding_dim) return emb @property def guidance_scale(self): return self._guidance_scale @property def guidance_rescale(self): return self._guidance_rescale @property def clip_skip(self): return self._clip_skip # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` # corresponds to doing no classifier free guidance. @property def do_classifier_free_guidance(self): return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None @property def cross_attention_kwargs(self): return self._cross_attention_kwargs @property def denoising_end(self): return self._denoising_end @property def num_timesteps(self): return self._num_timesteps @torch.no_grad() @replace_example_docstring(EXAMPLE_DOC_STRING) def __call__( self, prompt: Union[str, List[str]] = None, prompt_2: Optional[Union[str, List[str]]] = None, image: PipelineImageInput = None, height: Optional[int] = None, width: Optional[int] = None, num_inference_steps: int = 30, timesteps: List[int] = None, denoising_end: Optional[float] = None, guidance_scale: float = 7.0, negative_prompt: Optional[Union[str, List[str]]] = None, negative_prompt_2: Optional[Union[str, List[str]]] = None, num_images_per_prompt: Optional[int] = 1, eta: float = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.FloatTensor] = None, pooled_prompt_embeds: Optional[torch.FloatTensor] = None, negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, ip_adapter_image: Optional[PipelineImageInput] = None, ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None, output_type: Optional[str] = "pil", return_dict: bool = True, save_preview_row: bool = False, init_latents_with_lq: bool = True, multistep_restore: bool = False, adastep_restore: bool = False, cross_attention_kwargs: Optional[Dict[str, Any]] = None, guidance_rescale: float = 0.0, controlnet_conditioning_scale: float = 1.0, control_guidance_start: float = 0.0, control_guidance_end: float = 1.0, preview_start: float = 0.0, preview_end: float = 1.0, original_size: Tuple[int, int] = None, crops_coords_top_left: Tuple[int, int] = (0, 0), target_size: Tuple[int, int] = None, negative_original_size: Optional[Tuple[int, int]] = None, negative_crops_coords_top_left: Tuple[int, int] = (0, 0), negative_target_size: Optional[Tuple[int, int]] = None, clip_skip: Optional[int] = None, callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, callback_on_step_end_tensor_inputs: List[str] = ["latents"], previewer_scheduler: KarrasDiffusionSchedulers = None, reference_latents: Optional[torch.FloatTensor] = None, **kwargs, ): r""" The call function to the pipeline for generation. Args: prompt (`str` or `List[str]`, *optional*): The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is used in both text-encoders. image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[np.ndarray]`,: `List[List[torch.FloatTensor]]`, `List[List[np.ndarray]]` or `List[List[PIL.Image.Image]]`): The ControlNet input condition to provide guidance to the `unet` for generation. If the type is specified as `torch.FloatTensor`, it is passed to ControlNet as is. `PIL.Image.Image` can also be accepted as an image. The dimensions of the output image defaults to `image`'s dimensions. If height and/or width are passed, `image` is resized accordingly. If multiple ControlNets are specified in `init`, images must be passed as a list such that each element of the list can be correctly batched for input to a single ControlNet. height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): The height in pixels of the generated image. Anything below 512 pixels won't work well for [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) and checkpoints that are not specifically fine-tuned on low resolutions. width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): The width in pixels of the generated image. Anything below 512 pixels won't work well for [stabilityai/stable-diffusion-xl-base-1.0](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0) and checkpoints that are not specifically fine-tuned on low resolutions. num_inference_steps (`int`, *optional*, defaults to 50): The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. timesteps (`List[int]`, *optional*): Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed will be used. Must be in descending order. denoising_end (`float`, *optional*): When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be completed before it is intentionally prematurely terminated. As a result, the returned sample will still retain a substantial amount of noise as determined by the discrete timesteps selected by the scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) guidance_scale (`float`, *optional*, defaults to 5.0): A higher guidance scale value encourages the model to generate images closely linked to the text `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. negative_prompt (`str` or `List[str]`, *optional*): The prompt or prompts to guide what to not include in image generation. If not defined, you need to pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). negative_prompt_2 (`str` or `List[str]`, *optional*): The prompt or prompts to guide what to not include in image generation. This is sent to `tokenizer_2` and `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders. num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. eta (`float`, *optional*, defaults to 0.0): Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. generator (`torch.Generator` or `List[torch.Generator]`, *optional*): A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation deterministic. latents (`torch.FloatTensor`, *optional*): Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor is generated by sampling using the supplied random `generator`. prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, text embeddings are generated from the `prompt` input argument. negative_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. pooled_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, pooled text embeddings are generated from `prompt` input argument. negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, pooled `negative_prompt_embeds` are generated from `negative_prompt` input argument. ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*): Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not provided, embeddings are computed from the `ip_adapter_image` input argument. output_type (`str`, *optional*, defaults to `"pil"`): The output format of the generated image. Choose between `PIL.Image` or `np.array`. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a plain tuple. cross_attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0): The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set the corresponding scale as a list. control_guidance_start (`float` or `List[float]`, *optional*, defaults to 0.0): The percentage of total steps at which the ControlNet starts applying. control_guidance_end (`float` or `List[float]`, *optional*, defaults to 1.0): The percentage of total steps at which the ControlNet stops applying. original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. `original_size` defaults to `(height, width)` if not specified. Part of SDXL's micro-conditioning as explained in section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): For most cases, `target_size` should be set to the desired height and width of the generated image. If not specified it will default to `(height, width)`. Part of SDXL's micro-conditioning as explained in section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). negative_original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): To negatively condition the generation process based on a specific image resolution. Part of SDXL's micro-conditioning as explained in section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. negative_crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): To negatively condition the generation process based on a specific crop coordinates. Part of SDXL's micro-conditioning as explained in section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. negative_target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): To negatively condition the generation process based on a target image resolution. It should be as same as the `target_size` for most cases. Part of SDXL's micro-conditioning as explained in section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). For more information, refer to this issue thread: https://github.com/huggingface/diffusers/issues/4208. clip_skip (`int`, *optional*): Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that the output of the pre-final layer will be used for computing the prompt embeddings. callback_on_step_end (`Callable`, *optional*): A function that calls at the end of each denoising steps during the inference. The function is called with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by `callback_on_step_end_tensor_inputs`. callback_on_step_end_tensor_inputs (`List`, *optional*): The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the `._callback_tensor_inputs` attribute of your pipeline class. Examples: Returns: [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, otherwise a `tuple` is returned containing the output images. """ callback = kwargs.pop("callback", None) callback_steps = kwargs.pop("callback_steps", None) if callback is not None: deprecate( "callback", "1.0.0", "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", ) if callback_steps is not None: deprecate( "callback_steps", "1.0.0", "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", ) aggregator = self.aggregator._orig_mod if is_compiled_module(self.aggregator) else self.aggregator if not isinstance(ip_adapter_image, list): ip_adapter_image = [ip_adapter_image] if ip_adapter_image is not None else [image] # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, prompt_2, image, callback_steps, negative_prompt, negative_prompt_2, prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, ip_adapter_image, ip_adapter_image_embeds, negative_pooled_prompt_embeds, controlnet_conditioning_scale, control_guidance_start, control_guidance_end, callback_on_step_end_tensor_inputs, ) self._guidance_scale = guidance_scale self._guidance_rescale = guidance_rescale self._clip_skip = clip_skip self._cross_attention_kwargs = cross_attention_kwargs self._denoising_end = denoising_end # 2. Define call parameters if prompt is not None and isinstance(prompt, str): if not isinstance(image, PIL.Image.Image): batch_size = len(image) else: batch_size = 1 prompt = [prompt] * batch_size elif prompt is not None and isinstance(prompt, list): batch_size = len(prompt) assert batch_size == len(image) or (isinstance(image, PIL.Image.Image) or len(image) == 1) else: batch_size = prompt_embeds.shape[0] assert batch_size == len(image) or (isinstance(image, PIL.Image.Image) or len(image) == 1) device = self._execution_device # 3.1 Encode input prompt text_encoder_lora_scale = ( self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None ) ( prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds, ) = self.encode_prompt( prompt=prompt, prompt_2=prompt_2, device=device, num_images_per_prompt=num_images_per_prompt, do_classifier_free_guidance=self.do_classifier_free_guidance, negative_prompt=negative_prompt, negative_prompt_2=negative_prompt_2, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds, negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, lora_scale=text_encoder_lora_scale, clip_skip=self.clip_skip, ) # 3.2 Encode ip_adapter_image if ip_adapter_image is not None or ip_adapter_image_embeds is not None: image_embeds = self.prepare_ip_adapter_image_embeds( ip_adapter_image, ip_adapter_image_embeds, device, batch_size * num_images_per_prompt, self.do_classifier_free_guidance, ) # 4. Prepare image image = self.prepare_image( image=image, width=width, height=height, batch_size=batch_size * num_images_per_prompt, num_images_per_prompt=num_images_per_prompt, device=device, dtype=aggregator.dtype, do_classifier_free_guidance=self.do_classifier_free_guidance, ) height, width = image.shape[-2:] if image.shape[1] != 4: needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast if needs_upcasting: image = image.float() self.vae.to(dtype=torch.float32) image = self.vae.encode(image).latent_dist.sample() image = image * self.vae.config.scaling_factor if needs_upcasting: self.vae.to(dtype=torch.float16) image = image.to(dtype=torch.float16) else: height = int(height * self.vae_scale_factor) width = int(width * self.vae_scale_factor) # 5. Prepare timesteps timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) # 6. Prepare latent variables if init_latents_with_lq: latents = self.init_latents(image, generator, timesteps[0]) else: num_channels_latents = self.unet.config.in_channels latents = self.prepare_latents( batch_size * num_images_per_prompt, num_channels_latents, height, width, prompt_embeds.dtype, device, generator, latents, ) # 6.5 Optionally get Guidance Scale Embedding timestep_cond = None if self.unet.config.time_cond_proj_dim is not None: guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) timestep_cond = self.get_guidance_scale_embedding( guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim ).to(device=device, dtype=latents.dtype) # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) # 7.1 Create tensor stating which controlnets to keep controlnet_keep = [] previewing = [] for i in range(len(timesteps)): keeps = 1.0 - float(i / len(timesteps) < control_guidance_start or (i + 1) / len(timesteps) > control_guidance_end) controlnet_keep.append(keeps) use_preview = 1.0 - float(i / len(timesteps) < preview_start or (i + 1) / len(timesteps) > preview_end) previewing.append(use_preview) if isinstance(controlnet_conditioning_scale, list): assert len(controlnet_conditioning_scale) == len(timesteps), f"{len(controlnet_conditioning_scale)} controlnet scales do not match number of sampling steps {len(timesteps)}" else: controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet_keep) # 7.2 Prepare added time ids & embeddings original_size = original_size or (height, width) target_size = target_size or (height, width) add_text_embeds = pooled_prompt_embeds if self.text_encoder_2 is None: text_encoder_projection_dim = int(pooled_prompt_embeds.shape[-1]) else: text_encoder_projection_dim = self.text_encoder_2.config.projection_dim add_time_ids = self._get_add_time_ids( original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype, text_encoder_projection_dim=text_encoder_projection_dim, ) if negative_original_size is not None and negative_target_size is not None: negative_add_time_ids = self._get_add_time_ids( negative_original_size, negative_crops_coords_top_left, negative_target_size, dtype=prompt_embeds.dtype, text_encoder_projection_dim=text_encoder_projection_dim, ) else: negative_add_time_ids = add_time_ids if self.do_classifier_free_guidance: prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) add_time_ids = torch.cat([negative_add_time_ids, add_time_ids], dim=0) image = torch.cat([image] * 2, dim=0) prompt_embeds = prompt_embeds.to(device) add_text_embeds = add_text_embeds.to(device) add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) # 8. Denoising loop num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) # 8.1 Apply denoising_end if ( self.denoising_end is not None and isinstance(self.denoising_end, float) and self.denoising_end > 0 and self.denoising_end < 1 ): discrete_timestep_cutoff = int( round( self.scheduler.config.num_train_timesteps - (self.denoising_end * self.scheduler.config.num_train_timesteps) ) ) num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) timesteps = timesteps[:num_inference_steps] is_unet_compiled = is_compiled_module(self.unet) is_aggregator_compiled = is_compiled_module(self.aggregator) is_torch_higher_equal_2_1 = is_torch_version(">=", "2.1") previewer_mean = torch.zeros_like(latents) unet_mean = torch.zeros_like(latents) preview_factor = torch.ones( (latents.shape[0], *((1,) * (len(latents.shape) - 1))), dtype=latents.dtype, device=latents.device ) self._num_timesteps = len(timesteps) preview_row = [] with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): # Relevant thread: # https://dev-discuss.pytorch.org/t/cudagraphs-in-pytorch-2-0/1428 if (is_unet_compiled and is_aggregator_compiled) and is_torch_higher_equal_2_1: torch._inductor.cudagraph_mark_step_begin() # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) prev_t = t unet_model_input = latent_model_input added_cond_kwargs = { "text_embeds": add_text_embeds, "time_ids": add_time_ids, "image_embeds": image_embeds } aggregator_added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} # prepare time_embeds in advance as adapter input cross_attention_t_emb = self.unet.get_time_embed(sample=latent_model_input, timestep=t) cross_attention_emb = self.unet.time_embedding(cross_attention_t_emb, timestep_cond) cross_attention_aug_emb = None cross_attention_aug_emb = self.unet.get_aug_embed( emb=cross_attention_emb, encoder_hidden_states=prompt_embeds, added_cond_kwargs=added_cond_kwargs ) cross_attention_emb = cross_attention_emb + cross_attention_aug_emb if cross_attention_aug_emb is not None else cross_attention_emb if self.unet.time_embed_act is not None: cross_attention_emb = self.unet.time_embed_act(cross_attention_emb) current_cross_attention_kwargs = {"temb": cross_attention_emb} if cross_attention_kwargs is not None: for k,v in cross_attention_kwargs.items(): current_cross_attention_kwargs[k] = v self._cross_attention_kwargs = current_cross_attention_kwargs # adaptive restoration factors adaRes_scale = preview_factor.to(latent_model_input.dtype).clamp(0.0, controlnet_conditioning_scale[i]) cond_scale = adaRes_scale * controlnet_keep[i] cond_scale = torch.cat([cond_scale] * 2) if self.do_classifier_free_guidance else cond_scale if (cond_scale>0.1).sum().item() > 0: if previewing[i] > 0: # preview with LCM self.unet.enable_adapters() preview_noise = self.unet( latent_model_input, t, encoder_hidden_states=prompt_embeds, timestep_cond=timestep_cond, cross_attention_kwargs=self.cross_attention_kwargs, added_cond_kwargs=added_cond_kwargs, return_dict=False, )[0] preview_latent = previewer_scheduler.step( preview_noise, t.to(dtype=torch.int64), # torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents, latent_model_input, # scaled latents here for compatibility return_dict=False )[0] self.unet.disable_adapters() if self.do_classifier_free_guidance: preview_row.append(preview_latent.chunk(2)[1].to('cpu')) else: preview_row.append(preview_latent.to('cpu')) # Prepare 2nd order step. if multistep_restore and i+1 < len(timesteps): noise_preview = preview_noise.chunk(2)[1] if self.do_classifier_free_guidance else preview_noise first_step = self.scheduler.step( noise_preview, t, latents, **extra_step_kwargs, return_dict=True, step_forward=False ) prev_t = timesteps[i + 1] unet_model_input = torch.cat([first_step.prev_sample] * 2) if self.do_classifier_free_guidance else first_step.prev_sample unet_model_input = self.scheduler.scale_model_input(unet_model_input, prev_t, heun_step=True) elif reference_latents is not None: preview_latent = torch.cat([reference_latents] * 2) if self.do_classifier_free_guidance else reference_latents else: preview_latent = image # Add fresh noise # preview_noise = torch.randn_like(preview_latent) # preview_latent = self.scheduler.add_noise(preview_latent, preview_noise, t) preview_latent=preview_latent.to(dtype=next(aggregator.parameters()).dtype) # Aggregator inference down_block_res_samples, mid_block_res_sample = aggregator( image, prev_t, encoder_hidden_states=prompt_embeds, controlnet_cond=preview_latent, # conditioning_scale=cond_scale, added_cond_kwargs=aggregator_added_cond_kwargs, return_dict=False, ) # aggregator features scaling down_block_res_samples = [sample*cond_scale for sample in down_block_res_samples] mid_block_res_sample = mid_block_res_sample*cond_scale # predict the noise residual noise_pred = self.unet( unet_model_input, prev_t, encoder_hidden_states=prompt_embeds, timestep_cond=timestep_cond, cross_attention_kwargs=self.cross_attention_kwargs, down_block_additional_residuals=down_block_res_samples, mid_block_additional_residual=mid_block_res_sample, added_cond_kwargs=added_cond_kwargs, return_dict=False, )[0] # perform guidance if self.do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) # compute the previous noisy sample x_t -> x_t-1 latents_dtype = latents.dtype unet_step = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=True) latents = unet_step.prev_sample # Update adaRes factors unet_pred_latent = unet_step.pred_original_sample # Adaptive restoration. if adastep_restore: pred_x0_l2 = ((preview_latent[latents.shape[0]:].float()-unet_pred_latent.float())).pow(2).sum(dim=(1,2,3)) previewer_l2 = ((preview_latent[latents.shape[0]:].float()-previewer_mean.float())).pow(2).sum(dim=(1,2,3)) # unet_l2 = ((unet_pred_latent.float()-unet_mean.float())).pow(2).sum(dim=(1,2,3)).sqrt() # l2_error = (((preview_latent[latents.shape[0]:]-previewer_mean) - (unet_pred_latent-unet_mean))).pow(2).mean(dim=(1,2,3)) # preview_error = torch.nn.functional.cosine_similarity(preview_latent[latents.shape[0]:].reshape(latents.shape[0], -1), unet_pred_latent.reshape(latents.shape[0],-1)) previewer_mean = preview_latent[latents.shape[0]:] unet_mean = unet_pred_latent preview_factor = (pred_x0_l2 / previewer_l2).reshape(-1, 1, 1, 1) if latents.dtype != latents_dtype: if torch.backends.mps.is_available(): # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 latents = latents.to(latents_dtype) if callback_on_step_end is not None: callback_kwargs = {} for k in callback_on_step_end_tensor_inputs: callback_kwargs[k] = locals()[k] callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) latents = callback_outputs.pop("latents", latents) prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) # call the callback, if provided if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): progress_bar.update() if callback is not None and i % callback_steps == 0: step_idx = i // getattr(self.scheduler, "order", 1) callback(step_idx, t, latents) if not output_type == "latent": # make sure the VAE is in float32 mode, as it overflows in float16 needs_upcasting = self.vae.dtype == torch.float16 and self.vae.config.force_upcast if needs_upcasting: self.upcast_vae() latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) # unscale/denormalize the latents # denormalize with the mean and std if available and not None has_latents_mean = hasattr(self.vae.config, "latents_mean") and self.vae.config.latents_mean is not None has_latents_std = hasattr(self.vae.config, "latents_std") and self.vae.config.latents_std is not None if has_latents_mean and has_latents_std: latents_mean = ( torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(latents.device, latents.dtype) ) latents_std = ( torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(latents.device, latents.dtype) ) latents = latents * latents_std / self.vae.config.scaling_factor + latents_mean else: latents = latents / self.vae.config.scaling_factor image = self.vae.decode(latents, return_dict=False)[0] # cast back to fp16 if needed if needs_upcasting: self.vae.to(dtype=torch.float16) else: image = latents if not output_type == "latent": # apply watermark if available if self.watermark is not None: image = self.watermark.apply_watermark(image) image = self.image_processor.postprocess(image, output_type=output_type) if save_preview_row: preview_image_row = [] if needs_upcasting: self.upcast_vae() for preview_latents in preview_row: preview_latents = preview_latents.to(device=self.device, dtype=next(iter(self.vae.post_quant_conv.parameters())).dtype) if has_latents_mean and has_latents_std: latents_mean = ( torch.tensor(self.vae.config.latents_mean).view(1, 4, 1, 1).to(preview_latents.device, preview_latents.dtype) ) latents_std = ( torch.tensor(self.vae.config.latents_std).view(1, 4, 1, 1).to(preview_latents.device, preview_latents.dtype) ) preview_latents = preview_latents * latents_std / self.vae.config.scaling_factor + latents_mean else: preview_latents = preview_latents / self.vae.config.scaling_factor preview_image = self.vae.decode(preview_latents, return_dict=False)[0] preview_image = self.image_processor.postprocess(preview_image, output_type=output_type) preview_image_row.append(preview_image) # cast back to fp16 if needed if needs_upcasting: self.vae.to(dtype=torch.float16) # Offload all models self.maybe_free_model_hooks() if not return_dict: if save_preview_row: return (image, preview_image_row) return (image,) return StableDiffusionXLPipelineOutput(images=image)