# -*- coding: utf-8 -*- # Copyright (c) XiMing Xing. All rights reserved. # Author: XiMing Xing # Description: import PIL from PIL import Image from typing import Callable, List, Optional, Union, Tuple, AnyStr import numpy as np import torch import torch.nn.functional as F from torch.cuda.amp import custom_bwd, custom_fwd from torchvision import transforms from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput from diffusers.pipelines.stable_diffusion import StableDiffusionPipeline from pytorch_svgrender.token2attn.attn_control import AttentionStore from pytorch_svgrender.token2attn.ptp_utils import text_under_image, view_images class Token2AttnMixinASDSPipeline(StableDiffusionPipeline): r""" Pipeline for text-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/runwayml/stable-diffusion-v1-5) for details. feature_extractor ([`CLIPFeatureExtractor`]): Model that extracts features from generated images to be used as inputs for the `safety_checker`. """ _optional_components = ["safety_checker", "feature_extractor"] @torch.no_grad() def __call__( self, prompt: Union[str, List[str]], height: Optional[int] = None, width: Optional[int] = None, controller: AttentionStore = None, # feed attention_store as control of ptp num_inference_steps: int = 50, guidance_scale: float = 7.5, negative_prompt: Optional[Union[str, List[str]]] = None, num_images_per_prompt: Optional[int] = 1, eta: float = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, latents: Optional[torch.FloatTensor] = None, output_type: Optional[str] = "pil", return_dict: bool = True, callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: Optional[int] = 1, ): 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. 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`. """ self.register_attention_control(controller) # add attention controller # 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) # 2. Define call parameters batch_size = 1 if isinstance(prompt, str) else len(prompt) 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_embeddings = self._encode_prompt( prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt ) # 4. Prepare timesteps self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps = self.scheduler.timesteps # 5. Prepare latent variables try: num_channels_latents = self.unet.config.in_channels except Exception or Warning: num_channels_latents = self.unet.in_channels latents = self.prepare_latents( batch_size * num_images_per_prompt, num_channels_latents, height, width, text_embeddings.dtype, device, generator, latents, ) # 6. Prepare extra step kwargs. inherit TODO: Logic should ideally just be moved out of the pipeline extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) # 7. Denoising loop num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order with self.progress_bar(total=num_inference_steps) as progress_bar: for i, t in enumerate(timesteps): # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) # predict the noise residual noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample # perform guidance if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample # step callback latents = controller.step_callback(latents) # 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) # image = self.decode_latents(latents) # 8. Post-processing # 9. Run safety checker 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, 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] # 10. Convert to output_type 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) def encode_(self, images): images = (2 * images - 1).clamp(-1.0, 1.0) # images: [B, 3, H, W] # encode images latents = self.vae.encode(images).latent_dist.sample() latents = self.vae.config.scaling_factor * latents # scale the initial noise by the standard deviation required by the scheduler latents = latents * self.scheduler.init_noise_sigma return latents @staticmethod def S_aug(sketch: torch.Tensor, crop_size: int = 512, augments: str = "affine_contrast"): # init augmentations augment_list = [] if "affine" in augments: augment_list.append( transforms.RandomPerspective(fill=0, p=1.0, distortion_scale=0.5) ) augment_list.append( transforms.RandomResizedCrop(crop_size, scale=(0.8, 0.8), ratio=(1.0, 1.0)) ) if "contrast" in augments: # 2: increases the sharpness by a factor of 2. augment_list.append( transforms.RandomAdjustSharpness(sharpness_factor=2) ) augment_compose = transforms.Compose(augment_list) return augment_compose(sketch) def score_distillation_sampling(self, pred_rgb: torch.Tensor, crop_size: int, augments: str, prompt: Union[List, str], negative_prompt: Union[List, str] = None, guidance_scale: float = 100, as_latent: bool = False, grad_scale: float = 1, t_range: Union[List[float], Tuple[float]] = (0.02, 0.98)): num_train_timesteps = self.scheduler.config.num_train_timesteps min_step = int(num_train_timesteps * t_range[0]) max_step = int(num_train_timesteps * t_range[1]) alphas = self.scheduler.alphas_cumprod.to(self.device) # for convenience # sketch augmentation pred_rgb_a = self.S_aug(pred_rgb, crop_size, augments) # interp to crop_size x crop_size to be fed into vae. if as_latent: latents = F.interpolate(pred_rgb_a, (64, 64), mode='bilinear', align_corners=False) * 2 - 1 else: # encode image into latents with vae, requires grad! latents = self.encode_(pred_rgb_a) # Encode input prompt num_images_per_prompt = 1 # the number of images to generate per prompt do_classifier_free_guidance = guidance_scale > 1.0 text_embeddings = self._encode_prompt( prompt, self.device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt=negative_prompt, ) # timestep ~ U(0.02, 0.98) to avoid very high/low noise level t = torch.randint(min_step, max_step + 1, [1], dtype=torch.long, device=self.device) # predict the noise residual with unet, stop gradient with torch.no_grad(): # add noise noise = torch.randn_like(latents) latents_noisy = self.scheduler.add_noise(latents, noise, t) # pred noise latent_model_input = torch.cat([latents_noisy] * 2) if do_classifier_free_guidance else latents_noisy noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings).sample # perform guidance (high scale from paper!) if do_classifier_free_guidance: noise_pred_uncond, noise_pred_pos = noise_pred.chunk(2) noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_pos - noise_pred_uncond) # w(t), sigma_t^2 w = (1 - alphas[t]) grad = grad_scale * w * (noise_pred - noise) grad = torch.nan_to_num(grad) # since we omitted an item in grad, we need to use the custom function to specify the gradient loss = SpecifyGradient.apply(latents, grad) return loss, grad.mean() def register_attention_control(self, controller): attn_procs = {} cross_att_count = 0 for name in self.unet.attn_processors.keys(): cross_attention_dim = None if name.endswith("attn1.processor") else self.unet.config.cross_attention_dim if name.startswith("mid_block"): hidden_size = self.unet.config.block_out_channels[-1] place_in_unet = "mid" elif name.startswith("up_blocks"): block_id = int(name[len("up_blocks.")]) hidden_size = list(reversed(self.unet.config.block_out_channels))[block_id] place_in_unet = "up" elif name.startswith("down_blocks"): block_id = int(name[len("down_blocks.")]) hidden_size = self.unet.config.block_out_channels[block_id] place_in_unet = "down" else: continue cross_att_count += 1 attn_procs[name] = P2PCrossAttnProcessor( controller=controller, place_in_unet=place_in_unet ) self.unet.set_attn_processor(attn_procs) controller.num_att_layers = cross_att_count @staticmethod def aggregate_attention(prompts, attention_store: AttentionStore, res: int, from_where: List[str], is_cross: bool, select: int): if isinstance(prompts, str): prompts = [prompts] assert isinstance(prompts, list) out = [] attention_maps = attention_store.get_average_attention() num_pixels = res ** 2 for location in from_where: for item in attention_maps[f"{location}_{'cross' if is_cross else 'self'}"]: if item.shape[1] == num_pixels: cross_maps = item.reshape(len(prompts), -1, res, res, item.shape[-1])[select] out.append(cross_maps) out = torch.cat(out, dim=0) out = out.sum(0) / out.shape[0] return out.cpu() def get_cross_attention(self, prompts, attention_store: AttentionStore, res: int, from_where: List[str], select: int = 0, save_path=None): tokens = self.tokenizer.encode(prompts[select]) decoder = self.tokenizer.decode # shape: [res ** 2, res ** 2, seq_len] attention_maps = self.aggregate_attention(prompts, attention_store, res, from_where, True, select) images = [] for i in range(len(tokens)): image = attention_maps[:, :, i] image = 255 * image / image.max() image = image.unsqueeze(-1).expand(*image.shape, 3) image = image.numpy().astype(np.uint8) image = np.array(Image.fromarray(image).resize((256, 256))) image = text_under_image(image, decoder(int(tokens[i]))) images.append(image) image_array = np.stack(images, axis=0) view_images(image_array, save_image=True, fp=save_path) return attention_maps, tokens def get_self_attention_comp(self, prompts, attention_store: AttentionStore, res: int, from_where: List[str], img_size: int = 224, max_com=10, select: int = 0, save_path: AnyStr = None): attention_maps = self.aggregate_attention(prompts, attention_store, res, from_where, False, select) attention_maps = attention_maps.numpy().reshape((res ** 2, res ** 2)) # shape: [res ** 2, res ** 2] u, s, vh = np.linalg.svd(attention_maps - np.mean(attention_maps, axis=1, keepdims=True)) print(f"self-attention_maps: {attention_maps.shape}, " f"u: {u.shape}, " f"s: {s.shape}, " f"vh: {vh.shape}") images = [] vh_returns = [] for i in range(max_com): image = vh[i].reshape(res, res) image = (image - image.min()) / (image.max() - image.min()) image = 255 * image ret_ = Image.fromarray(image).resize((img_size, img_size), resample=PIL.Image.Resampling.BILINEAR) vh_returns.append(np.array(ret_)) image = np.repeat(np.expand_dims(image, axis=2), 3, axis=2).astype(np.uint8) image = Image.fromarray(image).resize((256, 256)) image = np.array(image) images.append(image) image_array = np.stack(images, axis=0) view_images(image_array, num_rows=max_com // 10, offset_ratio=0, save_image=True, fp=save_path / "self-attn-vh.png") return attention_maps, (u, s, vh), np.stack(vh_returns, axis=0) class P2PCrossAttnProcessor: def __init__(self, controller, place_in_unet): super().__init__() self.controller = controller self.place_in_unet = place_in_unet def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None): batch_size, sequence_length, _ = hidden_states.shape attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size=batch_size) query = attn.to_q(hidden_states) is_cross = encoder_hidden_states is not None encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else 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) # one line change self.controller(attention_probs, is_cross, self.place_in_unet) 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) return hidden_states class SpecifyGradient(torch.autograd.Function): @staticmethod @custom_fwd def forward(ctx, input_tensor, gt_grad): ctx.save_for_backward(gt_grad) # we return a dummy value 1, which will be scaled by amp's scaler so we get the scale in backward. return torch.ones([1], device=input_tensor.device, dtype=input_tensor.dtype) @staticmethod @custom_bwd def backward(ctx, grad_scale): gt_grad, = ctx.saved_tensors gt_grad = gt_grad * grad_scale return gt_grad, None