# Inspired by: https://github.com/haofanwang/ControlNet-for-Diffusers/ # From https://raw.githubusercontent.com/huggingface/diffusers/53377ef83c6446033f3ee506e3ef718db817b293/examples/community/stable_diffusion_controlnet_inpaint.py import inspect from typing import Any, Callable, Dict, List, Optional, Union, Tuple import numpy as np import PIL.Image import torch import torch.nn.functional as F from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer from diffusers import AutoencoderKL, ControlNetModel, DiffusionPipeline, \ UNet2DConditionModel, logging, StableDiffusionControlNetPipeline from diffusers.models.controlnet import ControlNetOutput from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput, StableDiffusionSafetyChecker from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_controlnet import MultiControlNetModel from diffusers.schedulers import KarrasDiffusionSchedulers from diffusers.utils import ( PIL_INTERPOLATION, is_accelerate_available, is_accelerate_version, is_compiled_module, randn_tensor, replace_example_docstring, ) from diffusers.loaders import LoraLoaderMixin logger = logging.get_logger(__name__) # pylint: disable=invalid-name EXAMPLE_DOC_STRING = """ Examples: ```py >>> import numpy as np >>> import torch >>> from PIL import Image >>> from stable_diffusion_controlnet_inpaint import StableDiffusionControlNetInpaintPipeline >>> from transformers import AutoImageProcessor, UperNetForSemanticSegmentation >>> from diffusers import ControlNetModel, UniPCMultistepScheduler >>> from diffusers.utils import load_image >>> def ade_palette(): return [[120, 120, 120], [180, 120, 120], [6, 230, 230], [80, 50, 50], [4, 200, 3], [120, 120, 80], [140, 140, 140], [204, 5, 255], [230, 230, 230], [4, 250, 7], [224, 5, 255], [235, 255, 7], [150, 5, 61], [120, 120, 70], [8, 255, 51], [255, 6, 82], [143, 255, 140], [204, 255, 4], [255, 51, 7], [204, 70, 3], [0, 102, 200], [61, 230, 250], [255, 6, 51], [11, 102, 255], [255, 7, 71], [255, 9, 224], [9, 7, 230], [220, 220, 220], [255, 9, 92], [112, 9, 255], [8, 255, 214], [7, 255, 224], [255, 184, 6], [10, 255, 71], [255, 41, 10], [7, 255, 255], [224, 255, 8], [102, 8, 255], [255, 61, 6], [255, 194, 7], [255, 122, 8], [0, 255, 20], [255, 8, 41], [255, 5, 153], [6, 51, 255], [235, 12, 255], [160, 150, 20], [0, 163, 255], [140, 140, 140], [250, 10, 15], [20, 255, 0], [31, 255, 0], [255, 31, 0], [255, 224, 0], [153, 255, 0], [0, 0, 255], [255, 71, 0], [0, 235, 255], [0, 173, 255], [31, 0, 255], [11, 200, 200], [255, 82, 0], [0, 255, 245], [0, 61, 255], [0, 255, 112], [0, 255, 133], [255, 0, 0], [255, 163, 0], [255, 102, 0], [194, 255, 0], [0, 143, 255], [51, 255, 0], [0, 82, 255], [0, 255, 41], [0, 255, 173], [10, 0, 255], [173, 255, 0], [0, 255, 153], [255, 92, 0], [255, 0, 255], [255, 0, 245], [255, 0, 102], [255, 173, 0], [255, 0, 20], [255, 184, 184], [0, 31, 255], [0, 255, 61], [0, 71, 255], [255, 0, 204], [0, 255, 194], [0, 255, 82], [0, 10, 255], [0, 112, 255], [51, 0, 255], [0, 194, 255], [0, 122, 255], [0, 255, 163], [255, 153, 0], [0, 255, 10], [255, 112, 0], [143, 255, 0], [82, 0, 255], [163, 255, 0], [255, 235, 0], [8, 184, 170], [133, 0, 255], [0, 255, 92], [184, 0, 255], [255, 0, 31], [0, 184, 255], [0, 214, 255], [255, 0, 112], [92, 255, 0], [0, 224, 255], [112, 224, 255], [70, 184, 160], [163, 0, 255], [153, 0, 255], [71, 255, 0], [255, 0, 163], [255, 204, 0], [255, 0, 143], [0, 255, 235], [133, 255, 0], [255, 0, 235], [245, 0, 255], [255, 0, 122], [255, 245, 0], [10, 190, 212], [214, 255, 0], [0, 204, 255], [20, 0, 255], [255, 255, 0], [0, 153, 255], [0, 41, 255], [0, 255, 204], [41, 0, 255], [41, 255, 0], [173, 0, 255], [0, 245, 255], [71, 0, 255], [122, 0, 255], [0, 255, 184], [0, 92, 255], [184, 255, 0], [0, 133, 255], [255, 214, 0], [25, 194, 194], [102, 255, 0], [92, 0, 255]] >>> image_processor = AutoImageProcessor.from_pretrained("openmmlab/upernet-convnext-small") >>> image_segmentor = UperNetForSemanticSegmentation.from_pretrained("openmmlab/upernet-convnext-small") >>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-seg", torch_dtype=torch.float16) >>> pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( "runwayml/stable-diffusion-inpainting", controlnet=controlnet, safety_checker=None, torch_dtype=torch.float16 ) >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) >>> pipe.enable_xformers_memory_efficient_attention() >>> pipe.enable_model_cpu_offload() >>> def image_to_seg(image): pixel_values = image_processor(image, return_tensors="pt").pixel_values with torch.no_grad(): outputs = image_segmentor(pixel_values) seg = image_processor.post_process_semantic_segmentation(outputs, target_sizes=[image.size[::-1]])[0] color_seg = np.zeros((seg.shape[0], seg.shape[1], 3), dtype=np.uint8) # height, width, 3 palette = np.array(ade_palette()) for label, color in enumerate(palette): color_seg[seg == label, :] = color color_seg = color_seg.astype(np.uint8) seg_image = Image.fromarray(color_seg) return seg_image >>> image = load_image( "https://github.com/CompVis/latent-diffusion/raw/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" ) >>> mask_image = load_image( "https://github.com/CompVis/latent-diffusion/raw/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" ) >>> controlnet_conditioning_image = image_to_seg(image) >>> image = pipe( "Face of a yellow cat, high resolution, sitting on a park bench", image, mask_image, controlnet_conditioning_image, num_inference_steps=20, ).images[0] >>> image.save("out.png") ``` """ def prepare_image(image): if isinstance(image, torch.Tensor): # Batch single image if image.ndim == 3: image = image.unsqueeze(0) image = image.to(dtype=torch.float32) else: # preprocess image if isinstance(image, (PIL.Image.Image, np.ndarray)): image = [image] if isinstance(image, list) and isinstance(image[0], PIL.Image.Image): image = [np.array(i.convert("RGB"))[None, :] for i in image] image = np.concatenate(image, axis=0) elif isinstance(image, list) and isinstance(image[0], np.ndarray): image = np.concatenate([i[None, :] for i in image], axis=0) image = image.transpose(0, 3, 1, 2) image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 return image # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_inpaint.prepare_mask_and_masked_image def prepare_mask_and_masked_image(image, mask, height, width, return_image=False): """ Prepares a pair (image, mask) to be consumed by the Stable Diffusion pipeline. This means that those inputs will be converted to ``torch.Tensor`` with shapes ``batch x channels x height x width`` where ``channels`` is ``3`` for the ``image`` and ``1`` for the ``mask``. The ``image`` will be converted to ``torch.float32`` and normalized to be in ``[-1, 1]``. The ``mask`` will be binarized (``mask > 0.5``) and cast to ``torch.float32`` too. Args: image (Union[np.array, PIL.Image, torch.Tensor]): The image to inpaint. It can be a ``PIL.Image``, or a ``height x width x 3`` ``np.array`` or a ``channels x height x width`` ``torch.Tensor`` or a ``batch x channels x height x width`` ``torch.Tensor``. mask (_type_): The mask to apply to the image, i.e. regions to inpaint. It can be a ``PIL.Image``, or a ``height x width`` ``np.array`` or a ``1 x height x width`` ``torch.Tensor`` or a ``batch x 1 x height x width`` ``torch.Tensor``. Raises: ValueError: ``torch.Tensor`` images should be in the ``[-1, 1]`` range. ValueError: ``torch.Tensor`` mask should be in the ``[0, 1]`` range. ValueError: ``mask`` and ``image`` should have the same spatial dimensions. TypeError: ``mask`` is a ``torch.Tensor`` but ``image`` is not (ot the other way around). Returns: tuple[torch.Tensor]: The pair (mask, masked_image) as ``torch.Tensor`` with 4 dimensions: ``batch x channels x height x width``. """ if image is None: raise ValueError("`image` input cannot be undefined.") if mask is None: raise ValueError("`mask_image` input cannot be undefined.") if isinstance(image, torch.Tensor): if not isinstance(mask, torch.Tensor): raise TypeError(f"`image` is a torch.Tensor but `mask` (type: {type(mask)} is not") # Batch single image if image.ndim == 3: assert image.shape[0] == 3, "Image outside a batch should be of shape (3, H, W)" image = image.unsqueeze(0) # Batch and add channel dim for single mask if mask.ndim == 2: mask = mask.unsqueeze(0).unsqueeze(0) # Batch single mask or add channel dim if mask.ndim == 3: # Single batched mask, no channel dim or single mask not batched but channel dim if mask.shape[0] == 1: mask = mask.unsqueeze(0) # Batched masks no channel dim else: mask = mask.unsqueeze(1) assert image.ndim == 4 and mask.ndim == 4, "Image and Mask must have 4 dimensions" assert image.shape[-2:] == mask.shape[-2:], "Image and Mask must have the same spatial dimensions" assert image.shape[0] == mask.shape[0], "Image and Mask must have the same batch size" # Check image is in [-1, 1] if image.min() < -1 or image.max() > 1: raise ValueError("Image should be in [-1, 1] range") # Check mask is in [0, 1] if mask.min() < 0 or mask.max() > 1: raise ValueError("Mask should be in [0, 1] range") # Binarize mask mask[mask < 0.5] = 0 mask[mask >= 0.5] = 1 # Image as float32 image = image.to(dtype=torch.float32) elif isinstance(mask, torch.Tensor): raise TypeError(f"`mask` is a torch.Tensor but `image` (type: {type(image)} is not") else: # preprocess image if isinstance(image, (PIL.Image.Image, np.ndarray)): image = [image] if isinstance(image, list) and isinstance(image[0], PIL.Image.Image): # resize all images w.r.t passed height an width image = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in image] image = [np.array(i.convert("RGB"))[None, :] for i in image] image = np.concatenate(image, axis=0) elif isinstance(image, list) and isinstance(image[0], np.ndarray): image = np.concatenate([i[None, :] for i in image], axis=0) image = image.transpose(0, 3, 1, 2) image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0 # preprocess mask if isinstance(mask, (PIL.Image.Image, np.ndarray)): mask = [mask] if isinstance(mask, list) and isinstance(mask[0], PIL.Image.Image): mask = [i.resize((width, height), resample=PIL.Image.LANCZOS) for i in mask] mask = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask], axis=0) mask = mask.astype(np.float32) / 255.0 elif isinstance(mask, list) and isinstance(mask[0], np.ndarray): mask = np.concatenate([m[None, None, :] for m in mask], axis=0) mask[mask < 0.5] = 0 mask[mask >= 0.5] = 1 mask = torch.from_numpy(mask) masked_image = image * (mask < 0.5) # n.b. ensure backwards compatibility as old function does not return image if return_image: return mask, masked_image, image return mask, masked_image def prepare_mask_image(mask_image): if isinstance(mask_image, torch.Tensor): if mask_image.ndim == 2: # Batch and add channel dim for single mask mask_image = mask_image.unsqueeze(0).unsqueeze(0) elif mask_image.ndim == 3 and mask_image.shape[0] == 1: # Single mask, the 0'th dimension is considered to be # the existing batch size of 1 mask_image = mask_image.unsqueeze(0) elif mask_image.ndim == 3 and mask_image.shape[0] != 1: # Batch of mask, the 0'th dimension is considered to be # the batching dimension mask_image = mask_image.unsqueeze(1) # Binarize mask mask_image[mask_image < 0.5] = 0 mask_image[mask_image >= 0.5] = 1 else: # preprocess mask if isinstance(mask_image, (PIL.Image.Image, np.ndarray)): mask_image = [mask_image] if isinstance(mask_image, list) and isinstance(mask_image[0], PIL.Image.Image): mask_image = np.concatenate([np.array(m.convert("L"))[None, None, :] for m in mask_image], axis=0) mask_image = mask_image.astype(np.float32) / 255.0 elif isinstance(mask_image, list) and isinstance(mask_image[0], np.ndarray): mask_image = np.concatenate([m[None, None, :] for m in mask_image], axis=0) mask_image[mask_image < 0.5] = 0 mask_image[mask_image >= 0.5] = 1 mask_image = torch.from_numpy(mask_image) return mask_image def prepare_controlnet_conditioning_image( controlnet_conditioning_image, width, height, batch_size, num_images_per_prompt, device, dtype, do_classifier_free_guidance, ): if not isinstance(controlnet_conditioning_image, torch.Tensor): if isinstance(controlnet_conditioning_image, PIL.Image.Image): controlnet_conditioning_image = [controlnet_conditioning_image] if isinstance(controlnet_conditioning_image[0], PIL.Image.Image): controlnet_conditioning_image = [ np.array(i.resize((width, height), resample=PIL_INTERPOLATION["lanczos"]))[None, :] for i in controlnet_conditioning_image ] controlnet_conditioning_image = np.concatenate(controlnet_conditioning_image, axis=0) controlnet_conditioning_image = np.array(controlnet_conditioning_image).astype(np.float32) / 255.0 controlnet_conditioning_image = controlnet_conditioning_image.transpose(0, 3, 1, 2) controlnet_conditioning_image = torch.from_numpy(controlnet_conditioning_image) elif isinstance(controlnet_conditioning_image[0], torch.Tensor): controlnet_conditioning_image = torch.cat(controlnet_conditioning_image, dim=0) image_batch_size = controlnet_conditioning_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 controlnet_conditioning_image = controlnet_conditioning_image.repeat_interleave(repeat_by, dim=0) controlnet_conditioning_image = controlnet_conditioning_image.to(device=device, dtype=dtype) if do_classifier_free_guidance: controlnet_conditioning_image = torch.cat([controlnet_conditioning_image] * 2) return controlnet_conditioning_image class StableDiffusionControlNetPipeline2(StableDiffusionControlNetPipeline): def __call__( self, prompt: Union[str, List[str]] = None, image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]] = None, height: Optional[int] = None, width: Optional[int] = None, num_inference_steps: int = 50, guidance_scale: float = 7.5, negative_prompt: 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, output_type: Optional[str] = "pil", return_dict: bool = True, callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: int = 1, cross_attention_kwargs: Optional[Dict[str, Any]] = None, controlnet_conditioning_scale: Union[float, List[float]] = 1.0, controlnet_conditioning_scale_map=None, guess_mode: bool = False, ): r""" Function invoked when calling the pipeline for generation. Args: prompt (`str` or `List[str]`, *optional*): The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. instead. image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, `List[List[torch.FloatTensor]]`, or `List[List[PIL.Image.Image]]`): The ControlNet input condition. ControlNet uses this input condition to generate guidance to Unet. 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 according to them. 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. width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): The width in pixels of the generated image. 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. guidance_scale (`float`, *optional*, defaults to 7.5): Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). `guidance_scale` is defined as `w` of equation 2. of [Imagen Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, usually at the expense of lower image quality. 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`). 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) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to [`schedulers.DDIMScheduler`], will be ignored for others. generator (`torch.Generator` or `List[torch.Generator]`, *optional*): One or a list of [torch generator(s)](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 will ge 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, *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. output_type (`str`, *optional*, defaults to `"pil"`): The output format of the generate image. Choose between [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.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. callback (`Callable`, *optional*): A function that will be called every `callback_steps` steps during inference. The function will be called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. callback_steps (`int`, *optional*, defaults to 1): The frequency at which the `callback` function will be called. If not specified, the callback will be called at every step. cross_attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.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. guess_mode (`bool`, *optional*, defaults to `False`): In this mode, the ControlNet encoder will try best to recognize the content of the input image even if you remove all prompts. The `guidance_scale` between 3.0 and 5.0 is recommended. Examples: Returns: [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When returning a tuple, the first element is a list with the generated images, and the second element is a list of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw) content, according to the `safety_checker`. """ # 0. Default height and width to unet height, width = self._default_height_width(height, width, image) # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, image, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds, controlnet_conditioning_scale, ) # 2. Define call parameters if prompt is not None and isinstance(prompt, str): batch_size = 1 elif prompt is not None and isinstance(prompt, list): batch_size = len(prompt) else: batch_size = prompt_embeds.shape[0] device = self._execution_device # 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. do_classifier_free_guidance = guidance_scale > 1.0 controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) if controlnet_conditioning_scale_map is not None: if isinstance(controlnet_conditioning_scale, list): controlnet_conditioning_scale = [scale * controlnet_conditioning_scale_map for scale in controlnet_conditioning_scale] else: controlnet_conditioning_scale = controlnet_conditioning_scale * controlnet_conditioning_scale_map global_pool_conditions = ( controlnet.config.global_pool_conditions if isinstance(controlnet, ControlNetModel) else controlnet.nets[0].config.global_pool_conditions ) guess_mode = guess_mode or global_pool_conditions # 3. Encode input prompt prompt_embeds = self._encode_prompt( prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, ) # 4. Prepare image if isinstance(controlnet, ControlNetModel): 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=controlnet.dtype, do_classifier_free_guidance=do_classifier_free_guidance, guess_mode=guess_mode, ) elif isinstance(controlnet, MultiControlNetModel): images = [] for image_ in 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=controlnet.dtype, do_classifier_free_guidance=do_classifier_free_guidance, guess_mode=guess_mode, ) images.append(image_) image = images else: assert False # 5. Prepare timesteps self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps = self.scheduler.timesteps # 6. Prepare latent variables 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, ) # 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) # 8. Denoising loop num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) # controlnet(s) inference if guess_mode and do_classifier_free_guidance: # Infer ControlNet only for the conditional batch. controlnet_latent_model_input = latents controlnet_prompt_embeds = prompt_embeds.chunk(2)[1] else: controlnet_latent_model_input = latent_model_input controlnet_prompt_embeds = prompt_embeds down_block_res_samples, mid_block_res_sample = self.controlnet( controlnet_latent_model_input, t, encoder_hidden_states=controlnet_prompt_embeds, controlnet_cond=image, conditioning_scale=controlnet_conditioning_scale, guess_mode=guess_mode, return_dict=False, ) if guess_mode and do_classifier_free_guidance: # Infered ControlNet only for the conditional batch. # To apply the output of ControlNet to both the unconditional and conditional batches, # add 0 to the unconditional batch to keep it unchanged. down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) # predict the noise residual noise_pred = self.unet( latent_model_input, t, encoder_hidden_states=prompt_embeds, cross_attention_kwargs=cross_attention_kwargs, down_block_additional_residuals=down_block_res_samples, mid_block_additional_residual=mid_block_res_sample, return_dict=False, )[0] # perform guidance if 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) # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] # 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: callback(i, t, latents) # If we do sequential model offloading, let's offload unet and controlnet # manually for max memory savings if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: self.unet.to("cpu") self.controlnet.to("cpu") torch.cuda.empty_cache() if output_type == "latent": image = latents has_nsfw_concept = None elif output_type == "pil": # 8. Post-processing image = self.decode_latents(latents) # 9. Run safety checker image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) # 10. Convert to PIL image = self.numpy_to_pil(image) else: # 8. Post-processing image = self.decode_latents(latents) # 9. Run safety checker image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) # Offload last model to CPU if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: self.final_offload_hook.offload() if not return_dict: return (image, has_nsfw_concept) return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) class ControlNetModel2(ControlNetModel): def forward( self, sample: torch.FloatTensor, timestep: Union[torch.Tensor, float, int], encoder_hidden_states: torch.Tensor, controlnet_cond: torch.FloatTensor, conditioning_scale: float = 1.0, class_labels: Optional[torch.Tensor] = None, timestep_cond: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, cross_attention_kwargs: Optional[Dict[str, Any]] = None, guess_mode: bool = False, return_dict: bool = True, ) -> Union[ControlNetOutput, Tuple]: # check channel order channel_order = self.config.controlnet_conditioning_channel_order if channel_order == "rgb": # in rgb order by default ... elif channel_order == "bgr": controlnet_cond = torch.flip(controlnet_cond, dims=[1]) else: raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}") # prepare attention_mask if attention_mask is not None: attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 attention_mask = attention_mask.unsqueeze(1) # 1. time timesteps = timestep if not torch.is_tensor(timesteps): # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can # This would be a good case for the `match` statement (Python 3.10+) is_mps = sample.device.type == "mps" if isinstance(timestep, float): dtype = torch.float32 if is_mps else torch.float64 else: dtype = torch.int32 if is_mps else torch.int64 timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) elif len(timesteps.shape) == 0: timesteps = timesteps[None].to(sample.device) # broadcast to batch dimension in a way that's compatible with ONNX/Core ML timesteps = timesteps.expand(sample.shape[0]) t_emb = self.time_proj(timesteps) # timesteps does not contain any weights and will always return f32 tensors # but time_embedding might actually be running in fp16. so we need to cast here. # there might be better ways to encapsulate this. t_emb = t_emb.to(dtype=sample.dtype) emb = self.time_embedding(t_emb, timestep_cond) if self.class_embedding is not None: if class_labels is None: raise ValueError("class_labels should be provided when num_class_embeds > 0") if self.config.class_embed_type == "timestep": class_labels = self.time_proj(class_labels) class_emb = self.class_embedding(class_labels).to(dtype=self.dtype) emb = emb + class_emb # 2. pre-process sample = self.conv_in(sample) controlnet_cond = self.controlnet_cond_embedding(controlnet_cond) sample = sample + controlnet_cond # 3. down down_block_res_samples = (sample,) for downsample_block in self.down_blocks: if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: sample, res_samples = downsample_block( hidden_states=sample, temb=emb, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask, cross_attention_kwargs=cross_attention_kwargs, ) else: sample, res_samples = downsample_block(hidden_states=sample, temb=emb) down_block_res_samples += res_samples # 4. mid if self.mid_block is not None: sample = self.mid_block( sample, emb, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask, cross_attention_kwargs=cross_attention_kwargs, ) # 5. Control net blocks controlnet_down_block_res_samples = () for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks): down_block_res_sample = controlnet_block(down_block_res_sample) controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,) down_block_res_samples = controlnet_down_block_res_samples mid_block_res_sample = self.controlnet_mid_block(sample) # 6. scaling if guess_mode and not self.config.global_pool_conditions: scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device) # 0.1 to 1.0 scales = scales * conditioning_scale down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)] mid_block_res_sample = mid_block_res_sample * scales[-1] # last one else: if isinstance(conditioning_scale, float): down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples] mid_block_res_sample = mid_block_res_sample * conditioning_scale else: assert isinstance(conditioning_scale, torch.Tensor) if len(conditioning_scale.shape) == 2: conditioning_scale = conditioning_scale[None, None] elif len(conditioning_scale.shape) == 3: conditioning_scale = conditioning_scale[None] down_block_res_samples = [ sample * F.interpolate(conditioning_scale, sample.shape[-2:], mode='bilinear', align_corners=True).type(sample.dtype) for sample in down_block_res_samples ] mid_block_res_sample = mid_block_res_sample * F.interpolate( conditioning_scale, mid_block_res_sample.shape[-2:], mode='bilinear', align_corners=True ).type(sample.dtype) if self.config.global_pool_conditions: down_block_res_samples = [ torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples ] mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True) if not return_dict: return (down_block_res_samples, mid_block_res_sample) return ControlNetOutput( down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample )