# Copyright 2023 MultiDiffusion Authors and The HuggingFace 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 copy import inspect import warnings from typing import Any, Callable, Dict, List, Optional, Union import torch from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer from ...image_processor import VaeImageProcessor from ...loaders import LoraLoaderMixin, TextualInversionLoaderMixin from ...models import AutoencoderKL, UNet2DConditionModel from ...schedulers import DDIMScheduler from ...utils import logging, randn_tensor, replace_example_docstring from ..pipeline_utils import DiffusionPipeline from . import StableDiffusionPipelineOutput from .safety_checker import StableDiffusionSafetyChecker logger = logging.get_logger(__name__) # pylint: disable=invalid-name EXAMPLE_DOC_STRING = """ Examples: ```py >>> import torch >>> from diffusers import StableDiffusionPanoramaPipeline, DDIMScheduler >>> model_ckpt = "stabilityai/stable-diffusion-2-base" >>> scheduler = DDIMScheduler.from_pretrained(model_ckpt, subfolder="scheduler") >>> pipe = StableDiffusionPanoramaPipeline.from_pretrained( ... model_ckpt, scheduler=scheduler, torch_dtype=torch.float16 ... ) >>> pipe = pipe.to("cuda") >>> prompt = "a photo of the dolomites" >>> image = pipe(prompt).images[0] ``` """ class StableDiffusionPanoramaPipeline(DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin): r""" Pipeline for text-to-image generation using "MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation". 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.). To generate panorama-like images, be sure to pass the `width` parameter accordingly when using the pipeline. Our recommendation for the `width` value is 2048. This is the default value of the `width` parameter for this pipeline. 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. The original work on Multi Diffsion used the [`DDIMScheduler`]. 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 ([`CLIPImageProcessor`]): Model that extracts features from generated images to be used as inputs for the `safety_checker`. """ _optional_components = ["safety_checker", "feature_extractor"] def __init__( self, vae: AutoencoderKL, text_encoder: CLIPTextModel, tokenizer: CLIPTokenizer, unet: UNet2DConditionModel, scheduler: DDIMScheduler, safety_checker: StableDiffusionSafetyChecker, feature_extractor: CLIPImageProcessor, requires_safety_checker: bool = True, ): super().__init__() if safety_checker is None and requires_safety_checker: logger.warning( 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 ." ) if safety_checker is not None and feature_extractor is None: raise ValueError( "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." ) self.register_modules( vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, safety_checker=safety_checker, feature_extractor=feature_extractor, ) self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) self.register_to_config(requires_safety_checker=requires_safety_checker) # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing def enable_vae_slicing(self): r""" Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. """ self.vae.enable_slicing() # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing def disable_vae_slicing(self): r""" Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to computing decoding in one step. """ self.vae.disable_slicing() # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt def _encode_prompt( self, prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt=None, prompt_embeds: Optional[torch.FloatTensor] = None, negative_prompt_embeds: Optional[torch.FloatTensor] = None, lora_scale: Optional[float] = None, ): r""" Encodes the prompt into text encoder hidden states. Args: prompt (`str` or `List[str]`, *optional*): prompt to be encoded 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`). 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. lora_scale (`float`, *optional*): A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. """ # 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, LoraLoaderMixin): self._lora_scale = lora_scale 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] if prompt_embeds is None: # textual inversion: procecss multi-vector tokens if necessary if isinstance(self, TextualInversionLoaderMixin): prompt = self.maybe_convert_prompt(prompt, self.tokenizer) text_inputs = self.tokenizer( prompt, padding="max_length", max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt", ) text_input_ids = text_inputs.input_ids untruncated_ids = self.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 = self.tokenizer.batch_decode( untruncated_ids[:, self.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" {self.tokenizer.model_max_length} tokens: {removed_text}" ) if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: attention_mask = text_inputs.attention_mask.to(device) else: attention_mask = None prompt_embeds = self.text_encoder( text_input_ids.to(device), attention_mask=attention_mask, ) prompt_embeds = prompt_embeds[0] prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.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) # get unconditional embeddings for classifier free guidance if do_classifier_free_guidance and negative_prompt_embeds is None: uncond_tokens: List[str] if negative_prompt is None: uncond_tokens = [""] * batch_size elif 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 isinstance(negative_prompt, str): uncond_tokens = [negative_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 # textual inversion: procecss multi-vector tokens if necessary if isinstance(self, TextualInversionLoaderMixin): uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) max_length = prompt_embeds.shape[1] uncond_input = self.tokenizer( uncond_tokens, padding="max_length", max_length=max_length, truncation=True, return_tensors="pt", ) if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: attention_mask = uncond_input.attention_mask.to(device) else: attention_mask = None negative_prompt_embeds = self.text_encoder( uncond_input.input_ids.to(device), attention_mask=attention_mask, ) negative_prompt_embeds = negative_prompt_embeds[0] if do_classifier_free_guidance: # duplicate unconditional embeddings for each generation per prompt, using mps friendly method seq_len = negative_prompt_embeds.shape[1] negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder.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) # 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 prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) return prompt_embeds # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker def run_safety_checker(self, image, device, dtype): if self.safety_checker is None: has_nsfw_concept = None else: if torch.is_tensor(image): feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") else: feature_extractor_input = self.image_processor.numpy_to_pil(image) safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) image, has_nsfw_concept = self.safety_checker( images=image, clip_input=safety_checker_input.pixel_values.to(dtype) ) return image, has_nsfw_concept # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents def decode_latents(self, latents): warnings.warn( "The decode_latents method is deprecated and will be removed in a future version. Please" " use VaeImageProcessor instead", FutureWarning, ) latents = 1 / self.vae.config.scaling_factor * latents image = self.vae.decode(latents, return_dict=False)[0] image = (image / 2 + 0.5).clamp(0, 1) # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 image = image.cpu().permute(0, 2, 3, 1).float().numpy() return image def decode_latents_with_padding(self, latents, padding=8): # Add padding to latents for circular inference # padding is the number of latents to add on each side # it would slightly increase the memory usage, but remove the boundary artifacts latents = 1 / self.vae.config.scaling_factor * latents latents_left = latents[..., :padding] latents_right = latents[..., -padding:] latents = torch.cat((latents_right, latents, latents_left), axis=-1) image = self.vae.decode(latents, return_dict=False)[0] padding_pix = self.vae_scale_factor * padding image = image[..., padding_pix:-padding_pix] return image # 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 # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs def check_inputs( self, prompt, height, width, callback_steps, negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None, ): if height % 8 != 0 or width % 8 != 0: raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") 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)}." ) 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 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)}") 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." ) 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}." ) # 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, height // self.vae_scale_factor, 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 def get_views(self, panorama_height, panorama_width, window_size=64, stride=8, circular_padding=False): # Here, we define the mappings F_i (see Eq. 7 in the MultiDiffusion paper https://arxiv.org/abs/2302.08113) # if panorama's height/width < window_size, num_blocks of height/width should return 1 panorama_height /= 8 panorama_width /= 8 num_blocks_height = (panorama_height - window_size) // stride + 1 if panorama_height > window_size else 1 if circular_padding: num_blocks_width = panorama_width // stride if panorama_width > window_size else 1 else: num_blocks_width = (panorama_width - window_size) // stride + 1 if panorama_width > window_size else 1 total_num_blocks = int(num_blocks_height * num_blocks_width) views = [] for i in range(total_num_blocks): h_start = int((i // num_blocks_width) * stride) h_end = h_start + window_size w_start = int((i % num_blocks_width) * stride) w_end = w_start + window_size views.append((h_start, h_end, w_start, w_end)) return views @torch.no_grad() @replace_example_docstring(EXAMPLE_DOC_STRING) def __call__( self, prompt: Union[str, List[str]] = None, height: Optional[int] = 512, width: Optional[int] = 2048, num_inference_steps: int = 50, guidance_scale: float = 7.5, view_batch_size: int = 1, 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: Optional[int] = 1, cross_attention_kwargs: Optional[Dict[str, Any]] = None, circular_padding: 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. height (`int`, *optional*, defaults to 512: The height in pixels of the generated image. width (`int`, *optional*, defaults to 2048): The width in pixels of the generated image. The width is kept to a high number because the pipeline is supposed to be used for generating panorama-like images. 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. view_batch_size (`int`, *optional*, defaults to 1): The batch size to denoise splited views. For some GPUs with high performance, higher view batch size can speedup the generation and increase the VRAM usage. 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). circular_padding (`bool`, *optional*, defaults to `False`): If set to True, circular padding is applied to ensure there are no stitching artifacts. Circular padding allows the model to seamlessly generate a transition from the rightmost part of the image to the leftmost part, maintaining consistency in a 360-degree sense. 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 = height or self.unet.config.sample_size * self.vae_scale_factor width = width or self.unet.config.sample_size * self.vae_scale_factor # 1. Check inputs. Raise error if not correct self.check_inputs( prompt, 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 text_encoder_lora_scale = ( cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None ) 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, lora_scale=text_encoder_lora_scale, ) # 4. Prepare timesteps self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps = self.scheduler.timesteps # 5. 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, ) # 6. Define panorama grid and initialize views for synthesis. # prepare batch grid views = self.get_views(height, width, circular_padding=circular_padding) views_batch = [views[i : i + view_batch_size] for i in range(0, len(views), view_batch_size)] views_scheduler_status = [copy.deepcopy(self.scheduler.__dict__)] * len(views_batch) count = torch.zeros_like(latents) value = torch.zeros_like(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 # Each denoising step also includes refinement of the latents with respect to the # views. 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): count.zero_() value.zero_() # generate views # Here, we iterate through different spatial crops of the latents and denoise them. These # denoised (latent) crops are then averaged to produce the final latent # for the current timestep via MultiDiffusion. Please see Sec. 4.1 in the # MultiDiffusion paper for more details: https://arxiv.org/abs/2302.08113 # Batch views denoise for j, batch_view in enumerate(views_batch): vb_size = len(batch_view) # get the latents corresponding to the current view coordinates if circular_padding: latents_for_view = [] for h_start, h_end, w_start, w_end in batch_view: if w_end > latents.shape[3]: # Add circular horizontal padding latent_view = torch.cat( ( latents[:, :, h_start:h_end, w_start:], latents[:, :, h_start:h_end, : w_end - latents.shape[3]], ), axis=-1, ) else: latent_view = latents[:, :, h_start:h_end, w_start:w_end] latents_for_view.append(latent_view) latents_for_view = torch.cat(latents_for_view) else: latents_for_view = torch.cat( [ latents[:, :, h_start:h_end, w_start:w_end] for h_start, h_end, w_start, w_end in batch_view ] ) # rematch block's scheduler status self.scheduler.__dict__.update(views_scheduler_status[j]) # expand the latents if we are doing classifier free guidance latent_model_input = ( latents_for_view.repeat_interleave(2, dim=0) if do_classifier_free_guidance else latents_for_view ) latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) # repeat prompt_embeds for batch prompt_embeds_input = torch.cat([prompt_embeds] * vb_size) # predict the noise residual noise_pred = self.unet( latent_model_input, t, encoder_hidden_states=prompt_embeds_input, cross_attention_kwargs=cross_attention_kwargs, ).sample # perform guidance if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred[::2], noise_pred[1::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_denoised_batch = self.scheduler.step( noise_pred, t, latents_for_view, **extra_step_kwargs ).prev_sample # save views scheduler status after sample views_scheduler_status[j] = copy.deepcopy(self.scheduler.__dict__) # extract value from batch for latents_view_denoised, (h_start, h_end, w_start, w_end) in zip( latents_denoised_batch.chunk(vb_size), batch_view ): if circular_padding and w_end > latents.shape[3]: # Case for circular padding value[:, :, h_start:h_end, w_start:] += latents_view_denoised[ :, :, h_start:h_end, : latents.shape[3] - w_start ] value[:, :, h_start:h_end, : w_end - latents.shape[3]] += latents_view_denoised[ :, :, h_start:h_end, latents.shape[3] - w_start : ] count[:, :, h_start:h_end, w_start:] += 1 count[:, :, h_start:h_end, : w_end - latents.shape[3]] += 1 else: value[:, :, h_start:h_end, w_start:w_end] += latents_view_denoised count[:, :, h_start:h_end, w_start:w_end] += 1 # take the MultiDiffusion step. Eq. 5 in MultiDiffusion paper: https://arxiv.org/abs/2302.08113 latents = torch.where(count > 0, value / count, value) # 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 not output_type == "latent": if circular_padding: image = self.decode_latents_with_padding(latents) else: image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) else: image = latents has_nsfw_concept = None if has_nsfw_concept is None: do_denormalize = [True] * image.shape[0] else: do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) if not return_dict: return (image, has_nsfw_concept) return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)