import inspect import warnings from itertools import repeat from typing import Callable, List, Optional, Union import torch from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer from diffusers.image_processor import VaeImageProcessor from diffusers.models import AutoencoderKL, UNet2DConditionModel from diffusers.models.attention_processor import AttnProcessor, Attention from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker from diffusers.schedulers import DDIMScheduler from scheduling_dpmsolver_multistep_inject import DPMSolverMultistepSchedulerInject # from diffusers.utils import logging, randn_tensor from diffusers.utils import logging from diffusers.utils.torch_utils import randn_tensor from diffusers.pipelines.pipeline_utils import DiffusionPipeline from diffusers.pipelines.semantic_stable_diffusion import SemanticStableDiffusionPipelineOutput import numpy as np from PIL import Image from tqdm import tqdm import torch.nn.functional as F import math from collections.abc import Iterable logger = logging.get_logger(__name__) # pylint: disable=invalid-name class AttentionStore(): @staticmethod def get_empty_store(): return {"down_cross": [], "mid_cross": [], "up_cross": [], "down_self": [], "mid_self": [], "up_self": []} def __call__(self, attn, is_cross: bool, place_in_unet: str, editing_prompts, PnP=False): # attn.shape = batch_size * head_size, seq_len query, seq_len_key if attn.shape[1] <= self.max_size: bs = 1 + int(PnP) + editing_prompts skip = 2 if PnP else 1 # skip PnP & unconditional attn = torch.stack(attn.split(self.batch_size)).permute(1, 0, 2, 3) source_batch_size = int(attn.shape[1] // bs) self.forward( attn[:, skip * source_batch_size:], is_cross, place_in_unet) def forward(self, attn, is_cross: bool, place_in_unet: str): key = f"{place_in_unet}_{'cross' if is_cross else 'self'}" self.step_store[key].append(attn) def between_steps(self, store_step=True): if store_step: if self.average: if len(self.attention_store) == 0: self.attention_store = self.step_store else: for key in self.attention_store: for i in range(len(self.attention_store[key])): self.attention_store[key][i] += self.step_store[key][i] else: if len(self.attention_store) == 0: self.attention_store = [self.step_store] else: self.attention_store.append(self.step_store) self.cur_step += 1 self.step_store = self.get_empty_store() def get_attention(self, step: int): if self.average: attention = {key: [item / self.cur_step for item in self.attention_store[key]] for key in self.attention_store} else: assert (step is not None) attention = self.attention_store[step] return attention def aggregate_attention(self, attention_maps, prompts, res: int, from_where: List[str], is_cross: bool, select: int ): out = [[] for x in range(self.batch_size)] num_pixels = res ** 2 for location in from_where: for bs_item in attention_maps[f"{location}_{'cross' if is_cross else 'self'}"]: for batch, item in enumerate(bs_item): if item.shape[1] == num_pixels: cross_maps = item.reshape(len(prompts), -1, res, res, item.shape[-1])[select] out[batch].append(cross_maps) out = torch.stack([torch.cat(x, dim=0) for x in out]) # average over heads out = out.sum(1) / out.shape[1] return out def __init__(self, average: bool, batch_size=1, max_resolution=16): self.step_store = self.get_empty_store() self.attention_store = [] self.cur_step = 0 self.average = average self.batch_size = batch_size self.max_size = max_resolution ** 2 class CrossAttnProcessor: def __init__(self, attention_store, place_in_unet, PnP, editing_prompts): self.attnstore = attention_store self.place_in_unet = place_in_unet self.editing_prompts = editing_prompts self.PnP = PnP def __call__( self, attn: Attention, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None, ): assert (not attn.residual_connection) assert (attn.spatial_norm is None) assert (attn.group_norm is None) assert (hidden_states.ndim != 4) assert (encoder_hidden_states is not None) # is cross batch_size, sequence_length, _ = ( hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape ) attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states elif attn.norm_cross: encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) key = attn.to_k(encoder_hidden_states) value = attn.to_v(encoder_hidden_states) query = attn.head_to_batch_dim(query) key = attn.head_to_batch_dim(key) value = attn.head_to_batch_dim(value) attention_probs = attn.get_attention_scores(query, key, attention_mask) self.attnstore(attention_probs, is_cross=True, place_in_unet=self.place_in_unet, editing_prompts=self.editing_prompts, PnP=self.PnP) hidden_states = torch.bmm(attention_probs, value) hidden_states = attn.batch_to_head_dim(hidden_states) # linear proj hidden_states = attn.to_out[0](hidden_states) # dropout hidden_states = attn.to_out[1](hidden_states) hidden_states = hidden_states / attn.rescale_output_factor return hidden_states # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionAttendAndExcitePipeline.GaussianSmoothing class GaussianSmoothing(): def __init__(self, device): kernel_size = [3, 3] sigma = [0.5, 0.5] # The gaussian kernel is the product of the gaussian function of each dimension. kernel = 1 meshgrids = torch.meshgrid([torch.arange(size, dtype=torch.float32) for size in kernel_size]) for size, std, mgrid in zip(kernel_size, sigma, meshgrids): mean = (size - 1) / 2 kernel *= 1 / (std * math.sqrt(2 * math.pi)) * torch.exp(-(((mgrid - mean) / (2 * std)) ** 2)) # Make sure sum of values in gaussian kernel equals 1. kernel = kernel / torch.sum(kernel) # Reshape to depthwise convolutional weight kernel = kernel.view(1, 1, *kernel.size()) kernel = kernel.repeat(1, *[1] * (kernel.dim() - 1)) self.weight = kernel.to(device) def __call__(self, input): """ Arguments: Apply gaussian filter to input. input (torch.Tensor): Input to apply gaussian filter on. Returns: filtered (torch.Tensor): Filtered output. """ return F.conv2d(input, weight=self.weight.to(input.dtype)) def load_512(image_path, size, left=0, right=0, top=0, bottom=0, device=None, dtype=None): def pre_process(im, size, left=0, right=0, top=0, bottom=0): if type(im) is str: image = np.array(Image.open(im).convert('RGB'))[:, :, :3] elif isinstance(im, Image.Image): image = np.array((im).convert('RGB'))[:, :, :3] else: image = im h, w, c = image.shape left = min(left, w - 1) right = min(right, w - left - 1) top = min(top, h - left - 1) bottom = min(bottom, h - top - 1) image = image[top:h - bottom, left:w - right] h, w, c = image.shape if h < w: offset = (w - h) // 2 image = image[:, offset:offset + h] elif w < h: offset = (h - w) // 2 image = image[offset:offset + w] image = np.array(Image.fromarray(image).resize((size, size))) image = torch.from_numpy(image).float().permute(2, 0, 1) return image tmps = [] if isinstance(image_path, list): for item in image_path: tmps.append(pre_process(item, size, left, right, top, bottom)) else: tmps.append(pre_process(image_path, size, left, right, top, bottom)) image = torch.stack(tmps) / 127.5 - 1 image = image.to(device=device, dtype=dtype) return image # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionAttendAndExcitePipeline.GaussianSmoothing def reset_dpm(scheduler): if isinstance(scheduler, DPMSolverMultistepSchedulerInject): scheduler.model_outputs = [ None, ] * scheduler.config.solver_order scheduler.lower_order_nums = 0 class SemanticStableDiffusionImg2ImgPipeline_DPMSolver(DiffusionPipeline): r""" Pipeline for text-to-image generation with latent editing. 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.) This model builds on the implementation of ['StableDiffusionPipeline'] 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 latens. Can be one of [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. safety_checker ([`Q16SafetyChecker`]): 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 ([`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: Union[DDIMScheduler,DPMSolverMultistepSchedulerInject], safety_checker: StableDiffusionSafetyChecker, feature_extractor: CLIPImageProcessor, requires_safety_checker: bool = True, ): super().__init__() if not isinstance(scheduler, DDIMScheduler) or not isinstance(scheduler, DPMSolverMultistepSchedulerInject): scheduler = DPMSolverMultistepSchedulerInject.from_config(scheduler.config, algorithm_type="sde-dpmsolver++", solver_order=2) logger.warning("This pipeline only supports DDIMScheduler and DPMSolverMultistepSchedulerInject. " "The scheduler has been changed to DPMSolverMultistepSchedulerInject.") 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) def progress_bar(self, iterable=None, total=None, verbose=True): if not hasattr(self, "_progress_bar_config"): self._progress_bar_config = {} elif not isinstance(self._progress_bar_config, dict): raise ValueError( f"`self._progress_bar_config` should be of type `dict`, but is {type(self._progress_bar_config)}." ) if not verbose: return iterable elif iterable is not None: return tqdm(iterable, **self._progress_bar_config) elif total is not None: return tqdm(total=total, **self._progress_bar_config) else: raise ValueError("Either `total` or `iterable` has to be defined.") # 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 # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs def prepare_extra_step_kwargs(self, 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 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}." ) # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, latents): #shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) #if latents.shape != shape: # raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}") 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 prepare_unet(self, attention_store, PnP: bool = False): attn_procs = {} for name in self.unet.attn_processors.keys(): if name.startswith("mid_block"): place_in_unet = "mid" elif name.startswith("up_blocks"): place_in_unet = "up" elif name.startswith("down_blocks"): place_in_unet = "down" else: continue if "attn2" in name and place_in_unet != 'mid': attn_procs[name] = CrossAttnProcessor( attention_store=attention_store, place_in_unet=place_in_unet, PnP=PnP, editing_prompts=self.enabled_editing_prompts) else: attn_procs[name] = AttnProcessor() self.unet.set_attn_processor(attn_procs) @torch.no_grad() def __call__( self, eta: Optional[float] = 1.0, negative_prompt: Optional[Union[str, List[str]]] = None, output_type: Optional[str] = "pil", return_dict: bool = True, callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: int = 1, editing_prompt: Optional[Union[str, List[str]]] = None, editing_prompt_embeddings: Optional[torch.Tensor] = None, reverse_editing_direction: Optional[Union[bool, List[bool]]] = False, edit_guidance_scale: Optional[Union[float, List[float]]] = 5, edit_warmup_steps: Optional[Union[int, List[int]]] = 0, edit_cooldown_steps: Optional[Union[int, List[int]]] = None, edit_threshold: Optional[Union[float, List[float]]] = 0.9, user_mask: Optional[torch.FloatTensor] = None, edit_weights: Optional[List[float]] = None, sem_guidance: Optional[List[torch.Tensor]] = None, verbose=True, use_cross_attn_mask: bool = False, # Attention store (just for visualization purposes) attn_store_steps: Optional[List[int]] = [], store_averaged_over_steps: bool = True, use_intersect_mask: bool = False, init_latents = None, zs = None, ): r""" Function invoked when calling the pipeline for generation. Args: prompt (`str` or `List[str]`): The prompt or prompts to guide the image generation. 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. 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`, *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`. 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. editing_prompt (`str` or `List[str]`, *optional*): The prompt or prompts to use for Semantic guidance. Semantic guidance is disabled by setting `editing_prompt = None`. Guidance direction of prompt should be specified via `reverse_editing_direction`. editing_prompt_embeddings (`torch.Tensor>`, *optional*): Pre-computed embeddings to use for semantic guidance. Guidance direction of embedding should be specified via `reverse_editing_direction`. reverse_editing_direction (`bool` or `List[bool]`, *optional*, defaults to `False`): Whether the corresponding prompt in `editing_prompt` should be increased or decreased. edit_guidance_scale (`float` or `List[float]`, *optional*, defaults to 5): Guidance scale for semantic guidance. If provided as list values should correspond to `editing_prompt`. `edit_guidance_scale` is defined as `s_e` of equation 6 of [SEGA Paper](https://arxiv.org/pdf/2301.12247.pdf). edit_warmup_steps (`float` or `List[float]`, *optional*, defaults to 10): Number of diffusion steps (for each prompt) for which semantic guidance will not be applied. Momentum will still be calculated for those steps and applied once all warmup periods are over. `edit_warmup_steps` is defined as `delta` (δ) of [SEGA Paper](https://arxiv.org/pdf/2301.12247.pdf). edit_cooldown_steps (`float` or `List[float]`, *optional*, defaults to `None`): Number of diffusion steps (for each prompt) after which semantic guidance will no longer be applied. edit_threshold (`float` or `List[float]`, *optional*, defaults to 0.9): Threshold of semantic guidance. edit_momentum_scale (`float`, *optional*, defaults to 0.1): Scale of the momentum to be added to the semantic guidance at each diffusion step. If set to 0.0 momentum will be disabled. Momentum is already built up during warmup, i.e. for diffusion steps smaller than `sld_warmup_steps`. Momentum will only be added to latent guidance once all warmup periods are finished. `edit_momentum_scale` is defined as `s_m` of equation 7 of [SEGA Paper](https://arxiv.org/pdf/2301.12247.pdf). edit_mom_beta (`float`, *optional*, defaults to 0.4): Defines how semantic guidance momentum builds up. `edit_mom_beta` indicates how much of the previous momentum will be kept. Momentum is already built up during warmup, i.e. for diffusion steps smaller than `edit_warmup_steps`. `edit_mom_beta` is defined as `beta_m` (β) of equation 8 of [SEGA Paper](https://arxiv.org/pdf/2301.12247.pdf). edit_weights (`List[float]`, *optional*, defaults to `None`): Indicates how much each individual concept should influence the overall guidance. If no weights are provided all concepts are applied equally. `edit_mom_beta` is defined as `g_i` of equation 9 of [SEGA Paper](https://arxiv.org/pdf/2301.12247.pdf). sem_guidance (`List[torch.Tensor]`, *optional*): List of pre-generated guidance vectors to be applied at generation. Length of the list has to correspond to `num_inference_steps`. Returns: [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] or `tuple`: [`~pipelines.semantic_stable_diffusion.SemanticStableDiffusionPipelineOutput`] 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`. """ eta = 1.0 num_images_per_prompt = 1 # latents = self.init_latents latents = init_latents use_ddpm = True # zs = self.zs reset_dpm(self.scheduler) if use_intersect_mask: use_cross_attn_mask = True if use_cross_attn_mask: self.smoothing = GaussianSmoothing(self.device) org_prompt = "" # 2. Define call parameters batch_size = self.batch_size if editing_prompt: enable_edit_guidance = True if isinstance(editing_prompt, str): editing_prompt = [editing_prompt] self.enabled_editing_prompts = len(editing_prompt) elif editing_prompt_embeddings is not None: enable_edit_guidance = True self.enabled_editing_prompts = editing_prompt_embeddings.shape[0] else: self.enabled_editing_prompts = 0 enable_edit_guidance = False if enable_edit_guidance: # get safety text embeddings if editing_prompt_embeddings is None: edit_concepts_input = self.tokenizer( [x for item in editing_prompt for x in repeat(item, batch_size)], padding="max_length", max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt", return_length=True ) num_edit_tokens = edit_concepts_input.length - 2 # not counting startoftext and endoftext edit_concepts_input_ids = edit_concepts_input.input_ids untruncated_ids = self.tokenizer( [x for item in editing_prompt for x in repeat(item, batch_size)], padding="longest", return_tensors="pt").input_ids if untruncated_ids.shape[-1] >= edit_concepts_input_ids.shape[-1] and not torch.equal( edit_concepts_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}" ) edit_concepts = self.text_encoder(edit_concepts_input_ids.to(self.device))[0] else: edit_concepts = editing_prompt_embeddings.to(self.device).repeat(batch_size, 1, 1) # duplicate text embeddings for each generation per prompt, using mps friendly method bs_embed_edit, seq_len_edit, _ = edit_concepts.shape edit_concepts = edit_concepts.repeat(1, num_images_per_prompt, 1) edit_concepts = edit_concepts.view(bs_embed_edit * num_images_per_prompt, seq_len_edit, -1) # 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: List[str] if negative_prompt is None: uncond_tokens = [""] 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" has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" " the batch size of `prompt`." ) else: uncond_tokens = negative_prompt max_length = self.tokenizer.model_max_length 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, using mps friendly method seq_len = uncond_embeddings.shape[1] uncond_embeddings = uncond_embeddings.repeat(batch_size, num_images_per_prompt, 1) uncond_embeddings = uncond_embeddings.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 if enable_edit_guidance: text_embeddings = torch.cat([uncond_embeddings, edit_concepts]) self.text_cross_attention_maps = \ ([editing_prompt] if isinstance(editing_prompt, str) else editing_prompt) else: text_embeddings = torch.cat([uncond_embeddings]) # 4. Prepare timesteps #self.scheduler.set_timesteps(num_inference_steps, device=self.device) timesteps = self.scheduler.timesteps if use_ddpm: t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs.shape[0]:])} timesteps = timesteps[-zs.shape[0]:] if use_cross_attn_mask: self.attention_store = AttentionStore(average=store_averaged_over_steps, batch_size=batch_size) self.prepare_unet(self.attention_store, PnP=False) # 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, None, None, text_embeddings.dtype, self.device, latents, ) # 6. Prepare extra step kwargs. extra_step_kwargs = self.prepare_extra_step_kwargs(eta) self.uncond_estimates = None self.edit_estimates = None self.sem_guidance = None self.activation_mask = None for i, t in enumerate(self.progress_bar(timesteps, verbose=verbose)): # expand the latents if we are doing classifier free guidance if enable_edit_guidance: latent_model_input = torch.cat([latents] * (1 + self.enabled_editing_prompts)) else: latent_model_input = latents latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) text_embed_input = text_embeddings # predict the noise residual noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input).sample noise_pred_out = noise_pred.chunk(1 + self.enabled_editing_prompts) # [b,4, 64, 64] noise_pred_uncond = noise_pred_out[0] noise_pred_edit_concepts = noise_pred_out[1:] # default text guidance noise_guidance = torch.zeros_like(noise_pred_uncond) if self.uncond_estimates is None: self.uncond_estimates = torch.zeros((len(timesteps), *noise_pred_uncond.shape)) self.uncond_estimates[i] = noise_pred_uncond.detach().cpu() if sem_guidance is not None and len(sem_guidance) > i: edit_guidance = sem_guidance[i].to(self.device) noise_guidance = noise_guidance + edit_guidance elif enable_edit_guidance: if self.activation_mask is None: self.activation_mask = torch.zeros( (len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape) ) if self.edit_estimates is None and enable_edit_guidance: self.edit_estimates = torch.zeros( (len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape) ) if self.sem_guidance is None: self.sem_guidance = torch.zeros((len(timesteps), *noise_pred_uncond.shape)) concept_weights = torch.zeros( (len(noise_pred_edit_concepts), noise_guidance.shape[0]), device=self.device, dtype=noise_guidance.dtype, ) noise_guidance_edit = torch.zeros( (len(noise_pred_edit_concepts), *noise_guidance.shape), device=self.device, dtype=noise_guidance.dtype, ) warmup_inds = [] # noise_guidance_edit = torch.zeros_like(noise_guidance) for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts): self.edit_estimates[i, c] = noise_pred_edit_concept if isinstance(edit_warmup_steps, list): edit_warmup_steps_c = edit_warmup_steps[c] else: edit_warmup_steps_c = edit_warmup_steps if i >= edit_warmup_steps_c: warmup_inds.append(c) else: continue if isinstance(edit_guidance_scale, list): edit_guidance_scale_c = edit_guidance_scale[c] else: edit_guidance_scale_c = edit_guidance_scale if isinstance(edit_threshold, list): edit_threshold_c = edit_threshold[c] else: edit_threshold_c = edit_threshold if isinstance(reverse_editing_direction, list): reverse_editing_direction_c = reverse_editing_direction[c] else: reverse_editing_direction_c = reverse_editing_direction if edit_weights: edit_weight_c = edit_weights[c] else: edit_weight_c = 1.0 if isinstance(edit_cooldown_steps, list): edit_cooldown_steps_c = edit_cooldown_steps[c] elif edit_cooldown_steps is None: edit_cooldown_steps_c = i + 1 else: edit_cooldown_steps_c = edit_cooldown_steps if i >= edit_cooldown_steps_c: noise_guidance_edit[c, :, :, :, :] = torch.zeros_like(noise_pred_edit_concept) continue noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond # tmp_weights = (noise_pred_text - noise_pred_edit_concept).sum(dim=(1, 2, 3)) tmp_weights = (noise_guidance - noise_pred_edit_concept).sum(dim=(1, 2, 3)) tmp_weights = torch.full_like(tmp_weights, edit_weight_c) # * (1 / enabled_editing_prompts) if reverse_editing_direction_c: noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1 concept_weights[c, :] = tmp_weights noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c if user_mask is not None: noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask if use_cross_attn_mask: out = self.attention_store.aggregate_attention( attention_maps=self.attention_store.step_store, prompts=self.text_cross_attention_maps, res=16, from_where=["up", "down"], is_cross=True, select=self.text_cross_attention_maps.index(editing_prompt[c]), ) attn_map = out[:, :, :, 1:1 + num_edit_tokens[c]] # 0 -> startoftext # average over all tokens assert (attn_map.shape[3] == num_edit_tokens[c]) attn_map = torch.sum(attn_map, dim=3) # gaussian_smoothing attn_map = F.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect") attn_map = self.smoothing(attn_map).squeeze(1) # create binary mask if attn_map.dtype == torch.float32: tmp = torch.quantile(attn_map.flatten(start_dim=1), edit_threshold_c, dim=1) else: tmp = torch.quantile(attn_map.flatten(start_dim=1).to(torch.float32), edit_threshold_c, dim=1).to(attn_map.dtype) attn_mask = torch.where(attn_map >= tmp.unsqueeze(1).unsqueeze(1).repeat(1,16,16), 1.0, 0.0) # resolution must match latent space dimension attn_mask = F.interpolate( attn_mask.unsqueeze(1), noise_guidance_edit_tmp.shape[-2:] # 64,64 ).repeat(1, 4, 1, 1) self.activation_mask[i, c] = attn_mask.detach().cpu() if not use_intersect_mask: noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask if use_intersect_mask: noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp) noise_guidance_edit_tmp_quantile = torch.sum(noise_guidance_edit_tmp_quantile, dim=1, keepdim=True) noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1) # torch.quantile function expects float32 if noise_guidance_edit_tmp_quantile.dtype == torch.float32: tmp = torch.quantile( noise_guidance_edit_tmp_quantile.flatten(start_dim=2), edit_threshold_c, dim=2, keepdim=False, ) else: tmp = torch.quantile( noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32), edit_threshold_c, dim=2, keepdim=False, ).to(noise_guidance_edit_tmp_quantile.dtype) intersect_mask = torch.where( noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], torch.ones_like(noise_guidance_edit_tmp), torch.zeros_like(noise_guidance_edit_tmp), ) * attn_mask self.activation_mask[i, c] = intersect_mask.detach().cpu() noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask elif not use_cross_attn_mask: # calculate quantile noise_guidance_edit_tmp_quantile = torch.abs(noise_guidance_edit_tmp) noise_guidance_edit_tmp_quantile = torch.sum(noise_guidance_edit_tmp_quantile, dim=1, keepdim=True) noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.repeat(1, 4, 1, 1) # torch.quantile function expects float32 if noise_guidance_edit_tmp_quantile.dtype == torch.float32: tmp = torch.quantile( noise_guidance_edit_tmp_quantile.flatten(start_dim=2), edit_threshold_c, dim=2, keepdim=False, ) else: tmp = torch.quantile( noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(torch.float32), edit_threshold_c, dim=2, keepdim=False, ).to(noise_guidance_edit_tmp_quantile.dtype) self.activation_mask[i, c] = torch.where( noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], torch.ones_like(noise_guidance_edit_tmp), torch.zeros_like(noise_guidance_edit_tmp), ).detach().cpu() noise_guidance_edit_tmp = torch.where( noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None], noise_guidance_edit_tmp, torch.zeros_like(noise_guidance_edit_tmp), ) noise_guidance_edit[c, :, :, :, :] = noise_guidance_edit_tmp warmup_inds = torch.tensor(warmup_inds).to(self.device) concept_weights = torch.index_select(concept_weights, 0, warmup_inds) concept_weights = torch.where( concept_weights < 0, torch.zeros_like(concept_weights), concept_weights ) concept_weights = torch.nan_to_num(concept_weights) noise_guidance_edit = torch.einsum("cb,cbijk->bijk", concept_weights, noise_guidance_edit) noise_guidance = noise_guidance + noise_guidance_edit self.sem_guidance[i] = noise_guidance_edit.detach().cpu() noise_pred = noise_pred_uncond + noise_guidance # compute the previous noisy sample x_t -> x_t-1 if use_ddpm: idx = t_to_idx[int(t)] latents = self.scheduler.step(noise_pred, t, latents, variance_noise=zs[idx], **extra_step_kwargs).prev_sample else: # if not use_ddpm: latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample # step callback if use_cross_attn_mask: store_step = i in attn_store_steps if store_step: print(f"storing attention for step {i}") self.attention_store.between_steps(store_step) # call the callback, if provided if callback is not None and i % callback_steps == 0: callback(i, t, latents) # 8. Post-processing if not output_type == "latent": image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] image, has_nsfw_concept = self.run_safety_checker(image, self.device, text_embeddings.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 SemanticStableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) def encode_text(self, prompts): text_inputs = self.tokenizer( prompts, padding="max_length", max_length=self.tokenizer.model_max_length, return_tensors="pt", ) text_input_ids = 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] text_embeddings = self.text_encoder(text_input_ids.to(self.device))[0] return text_embeddings @torch.no_grad() def invert(self, image_path: str, source_prompt: str = "", source_guidance_scale=3.5, num_inversion_steps: int = 30, skip: int = 15, eta: float = 1.0, generator: Optional[torch.Generator] = None, verbose=True, ): """ Inverts a real image according to Algorihm 1 in https://arxiv.org/pdf/2304.06140.pdf, based on the code in https://github.com/inbarhub/DDPM_inversion returns: zs - noise maps xts - intermediate inverted latents """ # self.eta = eta # assert (self.eta > 0) skip = skip/100 train_steps = self.scheduler.config.num_train_timesteps timesteps = torch.from_numpy( np.linspace(train_steps - skip * train_steps - 1, 1, num_inversion_steps).astype(np.int64)).to(self.device) self.num_inversion_steps = timesteps.shape[0] self.scheduler.num_inference_steps = timesteps.shape[0] self.scheduler.timesteps = timesteps self.unet.set_attn_processor(AttnProcessor()) # 1. get embeddings uncond_embedding = self.encode_text("") # 2. encode image x0 = self.encode_image(image_path, dtype=uncond_embedding.dtype) self.batch_size = x0.shape[0] if not source_prompt == "": text_embeddings = self.encode_text(source_prompt).repeat((self.batch_size, 1, 1)) uncond_embedding = uncond_embedding.repeat((self.batch_size, 1, 1)) # autoencoder reconstruction # image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False)[0] # image_rec = self.image_processor.postprocess(image_rec, output_type="pil") # 3. find zs and xts variance_noise_shape = ( self.num_inversion_steps, self.batch_size, self.unet.config.in_channels, self.unet.sample_size, self.unet.sample_size) # intermediate latents t_to_idx = {int(v): k for k, v in enumerate(timesteps)} xts = torch.zeros(size=variance_noise_shape, device=self.device, dtype=uncond_embedding.dtype) for t in reversed(timesteps): idx = self.num_inversion_steps-t_to_idx[int(t)] - 1 noise = randn_tensor(shape=x0.shape, generator=generator, device=self.device, dtype=x0.dtype) xts[idx] = self.scheduler.add_noise(x0, noise, t) xts = torch.cat([x0.unsqueeze(0), xts], dim=0) reset_dpm(self.scheduler) # noise maps zs = torch.zeros(size=variance_noise_shape, device=self.device, dtype=uncond_embedding.dtype) for t in self.progress_bar(timesteps, verbose=verbose): idx = self.num_inversion_steps-t_to_idx[int(t)]-1 # 1. predict noise residual xt = xts[idx+1] noise_pred = self.unet(xt, timestep=t, encoder_hidden_states=uncond_embedding).sample if not source_prompt == "": noise_pred_cond = self.unet(xt, timestep=t, encoder_hidden_states=text_embeddings).sample noise_pred = noise_pred + source_guidance_scale * (noise_pred_cond - noise_pred) xtm1 = xts[idx] z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, eta) zs[idx] = z # correction to avoid error accumulation xts[idx] = xtm1_corrected # TODO: I don't think that the noise map for the last step should be discarded ?! # if not zs is None: # zs[-1] = torch.zeros_like(zs[-1]) # self.init_latents = xts[-1].expand(self.batch_size, -1, -1, -1) zs = zs.flip(0) # self.zs = zs return zs, xts @torch.no_grad() def encode_image(self, image_path, dtype=None): image = load_512(image_path, size=self.unet.sample_size * self.vae_scale_factor, device=self.device, dtype=dtype) x0 = self.vae.encode(image).latent_dist.mode() x0 = self.vae.config.scaling_factor * x0 return x0 def compute_noise_ddim(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 # modifed so that updated xtm1 is returned as well (to avoid error accumulation) mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction noise = (prev_latents - mu_xt) / (variance ** (0.5) * eta) return noise, mu_xt + (eta * variance ** 0.5) * noise # Copied from pipelines.StableDiffusion.CycleDiffusionPipeline.compute_noise def compute_noise_sde_dpm_pp_2nd(scheduler, prev_latents, latents, timestep, noise_pred, eta): def first_order_update(model_output, timestep, prev_timestep, sample): lambda_t, lambda_s = scheduler.lambda_t[prev_timestep], scheduler.lambda_t[timestep] alpha_t, alpha_s = scheduler.alpha_t[prev_timestep], scheduler.alpha_t[timestep] sigma_t, sigma_s = scheduler.sigma_t[prev_timestep], scheduler.sigma_t[timestep] h = lambda_t - lambda_s mu_xt = ( (sigma_t / sigma_s * torch.exp(-h)) * sample + (alpha_t * (1 - torch.exp(-2.0 * h))) * model_output ) sigma = sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) noise = (prev_latents - mu_xt) / sigma prev_sample = mu_xt + sigma * noise return noise, prev_sample def second_order_update(model_output_list, timestep_list, prev_timestep, sample): t, s0, s1 = prev_timestep, timestep_list[-1], timestep_list[-2] m0, m1 = model_output_list[-1], model_output_list[-2] lambda_t, lambda_s0, lambda_s1 = scheduler.lambda_t[t], scheduler.lambda_t[s0], scheduler.lambda_t[s1] alpha_t, alpha_s0 = scheduler.alpha_t[t], scheduler.alpha_t[s0] sigma_t, sigma_s0 = scheduler.sigma_t[t], scheduler.sigma_t[s0] h, h_0 = lambda_t - lambda_s0, lambda_s0 - lambda_s1 r0 = h_0 / h D0, D1 = m0, (1.0 / r0) * (m0 - m1) mu_xt = ( (sigma_t / sigma_s0 * torch.exp(-h)) * sample + (alpha_t * (1 - torch.exp(-2.0 * h))) * D0 + 0.5 * (alpha_t * (1 - torch.exp(-2.0 * h))) * D1 ) sigma = sigma_t * torch.sqrt(1.0 - torch.exp(-2 * h)) noise = (prev_latents - mu_xt) / sigma prev_sample = mu_xt + sigma * noise return noise, prev_sample step_index = (scheduler.timesteps == timestep).nonzero() if len(step_index) == 0: step_index = len(scheduler.timesteps) - 1 else: step_index = step_index.item() prev_timestep = 0 if step_index == len(scheduler.timesteps) - 1 else scheduler.timesteps[step_index + 1] model_output = scheduler.convert_model_output(noise_pred, timestep, latents) for i in range(scheduler.config.solver_order - 1): scheduler.model_outputs[i] = scheduler.model_outputs[i + 1] scheduler.model_outputs[-1] = model_output if scheduler.lower_order_nums < 1: noise, prev_sample = first_order_update(model_output, timestep, prev_timestep, latents) else: timestep_list = [scheduler.timesteps[step_index - 1], timestep] noise, prev_sample = second_order_update(scheduler.model_outputs, timestep_list, prev_timestep, latents) if scheduler.lower_order_nums < scheduler.config.solver_order: scheduler.lower_order_nums += 1 return noise, prev_sample def compute_noise(scheduler, *args): if isinstance(scheduler, DDIMScheduler): return compute_noise_ddim(scheduler, *args) elif isinstance(scheduler, DPMSolverMultistepSchedulerInject) and scheduler.config.algorithm_type == 'sde-dpmsolver++'\ and scheduler.config.solver_order == 2: return compute_noise_sde_dpm_pp_2nd(scheduler, *args) else: raise NotImplementedError