import inspect from typing import Callable, List, Optional, Union import numpy as np import torch import PIL from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer from ...configuration_utils import FrozenDict from ...models import AutoencoderKL, UNet2DConditionModel from ...pipeline_utils import DiffusionPipeline from ...schedulers import DDIMScheduler from ...utils import deprecate, logging from . import StableDiffusionPipelineOutput from .safety_checker import StableDiffusionSafetyChecker logger = logging.get_logger(__name__) # pylint: disable=invalid-name def preprocess(image): w, h = image.size w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 image = image.resize((w, h), resample=PIL.Image.LANCZOS) image = np.array(image).astype(np.float32) / 255.0 image = image[None].transpose(0, 3, 1, 2) image = torch.from_numpy(image) return 2.0 * image - 1.0 def posterior_sample(scheduler, latents, timestep, clean_latents, eta): # 1. get previous step value (=t-1) prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps if prev_timestep <= 0: return clean_latents # 2. compute alphas, betas alpha_prod_t = scheduler.alphas_cumprod[timestep] alpha_prod_t_prev = ( scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod ) variance = scheduler._get_variance(timestep, prev_timestep) std_dev_t = eta * variance ** (0.5) # direction pointing to x_t e_t = (latents - alpha_prod_t ** (0.5) * clean_latents) / (1 - alpha_prod_t) ** (0.5) dir_xt = (1.0 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * e_t noise = std_dev_t * torch.randn(clean_latents.shape, dtype=clean_latents.dtype, device=clean_latents.device) prev_latents = alpha_prod_t_prev ** (0.5) * clean_latents + dir_xt + noise return prev_latents def compute_noise(scheduler, prev_latents, latents, timestep, noise_pred, eta): # 1. get previous step value (=t-1) prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps # 2. compute alphas, betas alpha_prod_t = scheduler.alphas_cumprod[timestep] alpha_prod_t_prev = ( scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod ) beta_prod_t = 1 - alpha_prod_t # 3. compute predicted original sample from predicted noise also called # "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5) # 4. Clip "predicted x_0" if scheduler.config.clip_sample: pred_original_sample = torch.clamp(pred_original_sample, -1, 1) # 5. compute variance: "sigma_t(η)" -> see formula (16) # σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1) variance = scheduler._get_variance(timestep, prev_timestep) std_dev_t = eta * variance ** (0.5) # 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * noise_pred noise = (prev_latents - (alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction)) / ( variance ** (0.5) * eta ) return noise class CycleDiffusionPipeline(DiffusionPipeline): r""" Pipeline for text-guided image to image generation using Stable Diffusion. This model inherits from [`DiffusionPipeline`]. 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. 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/CompVis/stable-diffusion-v1-4) for details. feature_extractor ([`CLIPFeatureExtractor`]): Model that extracts features from generated images to be used as inputs for the `safety_checker`. """ def __init__( self, vae: AutoencoderKL, text_encoder: CLIPTextModel, tokenizer: CLIPTokenizer, unet: UNet2DConditionModel, scheduler: DDIMScheduler, safety_checker: StableDiffusionSafetyChecker, feature_extractor: CLIPFeatureExtractor, ): super().__init__() if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: deprecation_message = ( f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " "to update the config accordingly as leaving `steps_offset` might led to incorrect results" " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" " file" ) deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) new_config = dict(scheduler.config) new_config["steps_offset"] = 1 scheduler._internal_dict = FrozenDict(new_config) if safety_checker is None: logger.warn( f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" " results in services or applications open to the public. Both the diffusers team and Hugging Face" " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" " it only for use-cases that involve analyzing network behavior or auditing its results. For more" " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." ) self.register_modules( vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, safety_checker=safety_checker, feature_extractor=feature_extractor, ) def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): r""" Enable sliced attention computation. When this option is enabled, the attention module will split the input tensor in slices, to compute attention in several steps. This is useful to save some memory in exchange for a small speed decrease. Args: slice_size (`str` or `int`, *optional*, defaults to `"auto"`): When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` must be a multiple of `slice_size`. """ if slice_size == "auto": # half the attention head size is usually a good trade-off between # speed and memory slice_size = self.unet.config.attention_head_dim // 2 self.unet.set_attention_slice(slice_size) def disable_attention_slicing(self): r""" Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go back to computing attention in one step. """ # set slice_size = `None` to disable `set_attention_slice` self.enable_attention_slicing(None) @torch.no_grad() def __call__( self, prompt: Union[str, List[str]], source_prompt: Union[str, List[str]], init_image: Union[torch.FloatTensor, PIL.Image.Image], strength: float = 0.8, num_inference_steps: Optional[int] = 50, guidance_scale: Optional[float] = 7.5, source_guidance_scale: Optional[float] = 1, num_images_per_prompt: Optional[int] = 1, eta: Optional[float] = 0.1, generator: Optional[torch.Generator] = None, output_type: Optional[str] = "pil", return_dict: bool = True, callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: Optional[int] = 1, **kwargs, ): r""" Function invoked when calling the pipeline for generation. Args: prompt (`str` or `List[str]`): The prompt or prompts to guide the image generation. init_image (`torch.FloatTensor` or `PIL.Image.Image`): `Image`, or tensor representing an image batch, that will be used as the starting point for the process. strength (`float`, *optional*, defaults to 0.8): Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1. `init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will be maximum and the denoising process will run for the full number of iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores `init_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. This parameter will be modulated by `strength`. 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. source_guidance_scale (`float`, *optional*, defaults to 1): Guidance scale for the source prompt. This is useful to control the amount of influence the source prompt for encoding. num_images_per_prompt (`int`, *optional*, defaults to 1): The number of images to generate per prompt. eta (`float`, *optional*, defaults to 0.1): 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`, *optional*): A [torch generator](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make generation deterministic. 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. 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`. """ if isinstance(prompt, str): batch_size = 1 elif isinstance(prompt, list): batch_size = len(prompt) else: raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") if batch_size != 1: raise ValueError( "At the moment only `batch_size=1` is supported for prompts, but you seem to have passed multiple" f" prompts: {prompt}. Please make sure to pass only a single prompt." ) if strength < 0 or strength > 1: raise ValueError(f"The value of strength should in [0.0, 1.0] but is {strength}") if (callback_steps is None) or ( 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)}." ) # set timesteps self.scheduler.set_timesteps(num_inference_steps) if isinstance(init_image, PIL.Image.Image): init_image = preprocess(init_image) # get prompt text embeddings text_inputs = self.tokenizer( prompt, padding="max_length", max_length=self.tokenizer.model_max_length, return_tensors="pt", ) source_text_inputs = self.tokenizer( source_prompt, padding="max_length", max_length=self.tokenizer.model_max_length, return_tensors="pt", ) text_input_ids = text_inputs.input_ids source_text_input_ids = source_text_inputs.input_ids if text_input_ids.shape[-1] > self.tokenizer.model_max_length: removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length :]) logger.warning( "The following part of your input was truncated because CLIP can only handle sequences up to" f" {self.tokenizer.model_max_length} tokens: {removed_text}" ) text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length] if source_text_input_ids.shape[-1] > self.tokenizer.model_max_length: removed_text = self.tokenizer.batch_decode(source_text_input_ids[:, self.tokenizer.model_max_length :]) logger.warning( "The following part of your input was truncated because CLIP can only handle sequences up to" f" {self.tokenizer.model_max_length} tokens: {removed_text}" ) source_text_input_ids = source_text_input_ids[:, : self.tokenizer.model_max_length] text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0] source_text_embeddings = self.text_encoder(source_text_input_ids.to(self.device))[0] # duplicate text embeddings for each generation per prompt text_embeddings = text_embeddings.repeat_interleave(num_images_per_prompt, dim=0) source_text_embeddings = source_text_embeddings.repeat_interleave(num_images_per_prompt, dim=0) # 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. # get unconditional embeddings for classifier free guidance uncond_tokens = [""] max_length = text_input_ids.shape[-1] uncond_input = self.tokenizer( uncond_tokens, padding="max_length", max_length=max_length, truncation=True, return_tensors="pt", ) uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0] # duplicate unconditional embeddings for each generation per prompt uncond_embeddings = uncond_embeddings.repeat_interleave(batch_size * num_images_per_prompt, dim=0) # For classifier free guidance, we need to do two forward passes. # Here we concatenate the unconditional and text embeddings into a single batch # to avoid doing two forward passes text_embeddings = torch.cat([uncond_embeddings, text_embeddings]) source_uncond_tokens = [""] max_length = source_text_input_ids.shape[-1] source_uncond_input = self.tokenizer( source_uncond_tokens, padding="max_length", max_length=max_length, truncation=True, return_tensors="pt", ) source_uncond_embeddings = self.text_encoder(source_uncond_input.input_ids.to(self.device))[0] # duplicate unconditional embeddings for each generation per prompt source_uncond_embeddings = source_uncond_embeddings.repeat_interleave( batch_size * num_images_per_prompt, dim=0 ) # For classifier free guidance, we need to do two forward passes. # Here we concatenate the unconditional and text embeddings into a single batch # to avoid doing two forward passes source_text_embeddings = torch.cat([source_uncond_embeddings, source_text_embeddings]) # encode the init image into latents and scale the latents latents_dtype = text_embeddings.dtype init_image = init_image.to(device=self.device, dtype=latents_dtype) init_latent_dist = self.vae.encode(init_image).latent_dist init_latents = init_latent_dist.sample(generator=generator) init_latents = 0.18215 * init_latents if isinstance(prompt, str): prompt = [prompt] if len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] == 0: # expand init_latents for batch_size deprecation_message = ( f"You have passed {len(prompt)} text prompts (`prompt`), but only {init_latents.shape[0]} initial" " images (`init_image`). Initial images are now duplicating to match the number of text prompts. Note" " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" " your script to pass as many init images as text prompts to suppress this warning." ) deprecate("len(prompt) != len(init_image)", "1.0.0", deprecation_message, standard_warn=False) additional_image_per_prompt = len(prompt) // init_latents.shape[0] init_latents = torch.cat([init_latents] * additional_image_per_prompt * num_images_per_prompt, dim=0) elif len(prompt) > init_latents.shape[0] and len(prompt) % init_latents.shape[0] != 0: raise ValueError( f"Cannot duplicate `init_image` of batch size {init_latents.shape[0]} to {len(prompt)} text prompts." ) else: init_latents = torch.cat([init_latents] * num_images_per_prompt, dim=0) # get the original timestep using init_timestep offset = self.scheduler.config.get("steps_offset", 0) init_timestep = int(num_inference_steps * strength) + offset init_timestep = min(init_timestep, num_inference_steps) timesteps = self.scheduler.timesteps[-init_timestep] timesteps = torch.tensor([timesteps] * batch_size * num_images_per_prompt, device=self.device) # add noise to latents using the timesteps noise = torch.randn(init_latents.shape, generator=generator, device=self.device, dtype=latents_dtype) clean_latents = init_latents init_latents = self.scheduler.add_noise(init_latents, noise, timesteps) # 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 not (accepts_eta and (0 < eta <= 1)): raise ValueError( "Currently, only the DDIM scheduler is supported. Please make sure that `pipeline.scheduler` is of" f" type {DDIMScheduler.__class__} and not {self.scheduler.__class__}." ) extra_step_kwargs["eta"] = eta latents = init_latents source_latents = init_latents t_start = max(num_inference_steps - init_timestep + offset, 0) # Some schedulers like PNDM have timesteps as arrays # It's more optimized to move all timesteps to correct device beforehand timesteps = self.scheduler.timesteps[t_start:].to(self.device) for i, t in enumerate(self.progress_bar(timesteps)): # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) source_latent_model_input = torch.cat([source_latents] * 2) latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) source_latent_model_input = self.scheduler.scale_model_input(source_latent_model_input, t) # predict the noise residual concat_latent_model_input = torch.stack( [ source_latent_model_input[0], latent_model_input[0], source_latent_model_input[1], latent_model_input[1], ], dim=0, ) concat_text_embeddings = torch.stack( [ source_text_embeddings[0], text_embeddings[0], source_text_embeddings[1], text_embeddings[1], ], dim=0, ) concat_noise_pred = self.unet( concat_latent_model_input, t, encoder_hidden_states=concat_text_embeddings ).sample # perform guidance ( source_noise_pred_uncond, noise_pred_uncond, source_noise_pred_text, noise_pred_text, ) = concat_noise_pred.chunk(4, dim=0) noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) source_noise_pred = source_noise_pred_uncond + source_guidance_scale * ( source_noise_pred_text - source_noise_pred_uncond ) # Sample source_latents from the posterior distribution. prev_source_latents = posterior_sample( self.scheduler, source_latents, t, clean_latents, **extra_step_kwargs ) # Compute noise. noise = compute_noise( self.scheduler, prev_source_latents, source_latents, t, source_noise_pred, **extra_step_kwargs ) source_latents = prev_source_latents # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step( noise_pred, t, latents, variance_noise=noise, **extra_step_kwargs ).prev_sample # call the callback, if provided if callback is not None and i % callback_steps == 0: callback(i, t, latents) latents = 1 / 0.18215 * latents image = self.vae.decode(latents).sample image = (image / 2 + 0.5).clamp(0, 1) image = image.cpu().permute(0, 2, 3, 1).numpy() if self.safety_checker is not None: safety_checker_input = self.feature_extractor(self.numpy_to_pil(image), return_tensors="pt").to( self.device ) image, has_nsfw_concept = self.safety_checker( images=image, clip_input=safety_checker_input.pixel_values.to(text_embeddings.dtype) ) else: has_nsfw_concept = None if output_type == "pil": image = self.numpy_to_pil(image) if not return_dict: return (image, has_nsfw_concept) return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)