import torch import PIL.Image import numpy as np from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_controlnet import * EXAMPLE_DOC_STRING = """ Examples: ```py >>> # !pip install opencv-python transformers accelerate >>> from diffusers import StableDiffusionControlNetInpaintPipeline, ControlNetModel, UniPCMultistepScheduler >>> from diffusers.utils import load_image >>> import numpy as np >>> import torch >>> import cv2 >>> from PIL import Image >>> # download an image >>> image = load_image( ... "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png" ... ) >>> image = np.array(image) >>> mask_image = load_image( ... "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png" ... ) >>> mask_image = np.array(mask_image) >>> # get canny image >>> canny_image = cv2.Canny(image, 100, 200) >>> canny_image = canny_image[:, :, None] >>> canny_image = np.concatenate([canny_image, canny_image, canny_image], axis=2) >>> canny_image = Image.fromarray(canny_image) >>> # load control net and stable diffusion v1-5 >>> controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16) >>> pipe = StableDiffusionControlNetInpaintPipeline.from_pretrained( ... "runwayml/stable-diffusion-inpainting", controlnet=controlnet, torch_dtype=torch.float16 ... ) >>> # speed up diffusion process with faster scheduler and memory optimization >>> pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config) >>> # remove following line if xformers is not installed >>> pipe.enable_xformers_memory_efficient_attention() >>> pipe.enable_model_cpu_offload() >>> # generate image >>> generator = torch.manual_seed(0) >>> image = pipe( ... "futuristic-looking doggo", ... num_inference_steps=20, ... generator=generator, ... image=image, ... control_image=canny_image, ... mask_image=mask_image ... ).images[0] ``` """ def prepare_mask_and_masked_image(image, mask): """ 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 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): 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 = 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) return mask, masked_image class StableDiffusionControlNetInpaintPipeline(StableDiffusionControlNetPipeline): r""" Pipeline for text-guided image inpainting using Stable Diffusion with ControlNet guidance. This model inherits from [`StableDiffusionControlNetPipeline`]. Check the superclass documentation for the generic methods the library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) Args: vae ([`AutoencoderKL`]): Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. text_encoder ([`CLIPTextModel`]): Frozen text-encoder. Stable Diffusion uses the text portion of [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. tokenizer (`CLIPTokenizer`): Tokenizer of class [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. controlnet ([`ControlNetModel`]): Provides additional conditioning to the unet during the denoising process 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`]. safety_checker ([`StableDiffusionSafetyChecker`]): Classification module that estimates whether generated images could be considered offensive or harmful. Please, refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for details. feature_extractor ([`CLIPFeatureExtractor`]): Model that extracts features from generated images to be used as inputs for the `safety_checker`. """ def prepare_mask_latents( self, mask, masked_image, batch_size, height, width, dtype, device, generator, do_classifier_free_guidance ): # resize the mask to latents shape as we concatenate the mask to the latents # we do that before converting to dtype to avoid breaking in case we're using cpu_offload # and half precision mask = torch.nn.functional.interpolate( mask, size=(height // self.vae_scale_factor, width // self.vae_scale_factor) ) mask = mask.to(device=device, dtype=dtype) masked_image = masked_image.to(device=device, dtype=dtype) # encode the mask image into latents space so we can concatenate it to the latents if isinstance(generator, list): masked_image_latents = [ self.vae.encode(masked_image[i : i + 1]).latent_dist.sample(generator=generator[i]) for i in range(batch_size) ] masked_image_latents = torch.cat(masked_image_latents, dim=0) else: masked_image_latents = self.vae.encode(masked_image).latent_dist.sample(generator=generator) masked_image_latents = self.vae.config.scaling_factor * masked_image_latents # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method if mask.shape[0] < batch_size: if not batch_size % mask.shape[0] == 0: raise ValueError( "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to" f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number" " of masks that you pass is divisible by the total requested batch size." ) mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1) if masked_image_latents.shape[0] < batch_size: if not batch_size % masked_image_latents.shape[0] == 0: raise ValueError( "The passed images and the required batch size don't match. Images are supposed to be duplicated" f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed." " Make sure the number of images that you pass is divisible by the total requested batch size." ) masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1) mask = torch.cat([mask] * 2) if do_classifier_free_guidance else mask masked_image_latents = ( torch.cat([masked_image_latents] * 2) if do_classifier_free_guidance else masked_image_latents ) # aligning device to prevent device errors when concating it with the latent model input masked_image_latents = masked_image_latents.to(device=device, dtype=dtype) return mask, masked_image_latents @torch.no_grad() @replace_example_docstring(EXAMPLE_DOC_STRING) def __call__( self, prompt: Union[str, List[str]] = None, image: Union[torch.FloatTensor, PIL.Image.Image] = None, control_image: Union[torch.FloatTensor, PIL.Image.Image, List[torch.FloatTensor], List[PIL.Image.Image]] = None, mask_image: Union[torch.FloatTensor, 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: float = 1.0, ): 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 (`PIL.Image.Image`): `Image`, or tensor representing an image batch which will be inpainted, *i.e.* parts of the image will be masked out with `mask_image` and repainted according to `prompt`. control_image (`torch.FloatTensor`, `PIL.Image.Image`, `List[torch.FloatTensor]` or `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 control image is automatically resized to fit the output image. mask_image (`PIL.Image.Image`): `Image`, or tensor representing an image batch, to mask `image`. White pixels in the mask will be repainted, while black pixels will be preserved. If `mask_image` is a PIL image, it will be converted to a single channel (luminance) before use. If it's a tensor, it should contain one color channel (L) instead of 3, so the expected shape would be `(B, H, W, 1)`. 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. 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 `AttnProcessor` 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`, *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. 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, control_image) # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, control_image, height, width, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds ) # 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 # 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 control_image = self.prepare_image( control_image, width, height, batch_size * num_images_per_prompt, num_images_per_prompt, device, self.controlnet.dtype, ) if do_classifier_free_guidance: control_image = torch.cat([control_image] * 2) # 5. Prepare timesteps self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps = self.scheduler.timesteps # 6. Prepare latent variables num_channels_latents = self.controlnet.in_channels latents = self.prepare_latents( batch_size * num_images_per_prompt, num_channels_latents, height, width, prompt_embeds.dtype, device, generator, latents, ) # EXTRA: prepare mask latents mask, masked_image = prepare_mask_and_masked_image(image, mask_image) mask, masked_image_latents = self.prepare_mask_latents( mask, masked_image, batch_size * num_images_per_prompt, height, width, prompt_embeds.dtype, device, generator, do_classifier_free_guidance, ) # 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) down_block_res_samples, mid_block_res_sample = self.controlnet( latent_model_input, t, encoder_hidden_states=prompt_embeds, controlnet_cond=control_image, return_dict=False, ) down_block_res_samples = [ down_block_res_sample * controlnet_conditioning_scale for down_block_res_sample in down_block_res_samples ] mid_block_res_sample *= controlnet_conditioning_scale # predict the noise residual latent_model_input = torch.cat([latent_model_input, mask, masked_image_latents], dim=1) 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, ).sample # 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).prev_sample # 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)