import inspect import warnings import math from typing import List, Dict, Any, Optional, Union, Callable import numpy as np import torch import torch.nn.functional as F import PIL from diffusers.configuration_utils import FrozenDict from diffusers.utils import is_accelerate_available from diffusers import ( AutoencoderKL, DiffusionPipeline, UNet2DConditionModel, DDIMScheduler, DPMSolverMultistepScheduler, EulerAncestralDiscreteScheduler, EulerDiscreteScheduler, LMSDiscreteScheduler, PNDMScheduler, ) from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput from diffusers.utils import logging, randn_tensor from transformers import CLIPTextModel, CLIPTokenizer from models.clip.util import unify_input_ids, get_extended_embeddings logger = logging.get_logger(__name__) # pylint: disable=invalid-name def preprocess(image): w, h = image.size w, h = map(lambda x: x - x % 8, (w, h)) # resize to integer multiple of 8 image = image.resize((w, h), resample=PIL.Image.LANCZOS) image = np.array(image).astype(np.float32) / 255.0 image = image[None].transpose(0, 3, 1, 2) image = torch.from_numpy(image) return 2.0 * image - 1.0 def gaussian_blur_2d(img, kernel_size, sigma): ksize_half = (kernel_size - 1) * 0.5 x = torch.linspace(-ksize_half, ksize_half, steps=kernel_size) pdf = torch.exp(-0.5 * (x / sigma).pow(2)) x_kernel = pdf / pdf.sum() x_kernel = x_kernel.to(device=img.device, dtype=img.dtype) kernel2d = torch.mm(x_kernel[:, None], x_kernel[None, :]) kernel2d = kernel2d.expand(img.shape[-3], 1, kernel2d.shape[0], kernel2d.shape[1]) padding = [kernel_size // 2, kernel_size // 2, kernel_size // 2, kernel_size // 2] img = F.pad(img, padding, mode="reflect") img = F.conv2d(img, kernel2d, groups=img.shape[-3]) return img class CrossAttnStoreProcessor: def __init__(self): self.attention_probs = None 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) query = attn.to_q(hidden_states) if encoder_hidden_states is None: encoder_hidden_states = hidden_states elif attn.cross_attention_norm: encoder_hidden_states = attn.norm_cross(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) self.attention_probs = attn.get_attention_scores(query, key, attention_mask) hidden_states = torch.bmm(self.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 VlpnStableDiffusion(DiffusionPipeline): def __init__( self, vae: AutoencoderKL, text_encoder: CLIPTextModel, tokenizer: CLIPTokenizer, unet: UNet2DConditionModel, scheduler: Union[ DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler, EulerDiscreteScheduler, EulerAncestralDiscreteScheduler, DPMSolverMultistepScheduler, ], ): super().__init__() if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: warnings.warn( f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " "to update the config accordingly as leaving `steps_offset` might led to incorrect results" " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" " file", DeprecationWarning, ) new_config = dict(scheduler.config) new_config["steps_offset"] = 1 scheduler._internal_dict = FrozenDict(new_config) self.register_modules( vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, ) self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"): r""" Enable sliced attention computation. When this option is enabled, the attention module will split the input tensor in slices, to compute attention in several steps. This is useful to save some memory in exchange for a small speed decrease. Args: slice_size (`str` or `int`, *optional*, defaults to `"auto"`): When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If a number is provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` must be a multiple of `slice_size`. """ if slice_size == "auto": if isinstance(self.unet.config.attention_head_dim, int): # half the attention head size is usually a good trade-off between # speed and memory slice_size = self.unet.config.attention_head_dim // 2 else: # if `attention_head_dim` is a list, take the smallest head size slice_size = min(self.unet.config.attention_head_dim) self.unet.set_attention_slice(slice_size) def disable_attention_slicing(self): r""" Disable sliced attention computation. If `enable_attention_slicing` was previously invoked, this method will go back to computing attention in one step. """ # set slice_size = `None` to disable `attention slicing` self.enable_attention_slicing(None) def enable_sequential_cpu_offload(self): r""" Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only when their specific submodule has its `forward` method called. """ if is_accelerate_available(): from accelerate import cpu_offload else: raise ImportError("Please install accelerate via `pip install accelerate`") device = torch.device("cuda") for cpu_offloaded_model in [self.unet, self.text_encoder, self.vae, self.safety_checker]: if cpu_offloaded_model is not None: cpu_offload(cpu_offloaded_model, device) 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() 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() @property def execution_device(self): r""" Returns the device on which the pipeline's models will be executed. After calling `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module hooks. """ if self.device != torch.device("meta") or not hasattr(self.unet, "_hf_hook"): return self.device for module in self.unet.modules(): if ( hasattr(module, "_hf_hook") and hasattr(module._hf_hook, "execution_device") and module._hf_hook.execution_device is not None ): return torch.device(module._hf_hook.execution_device) return self.device def check_inputs( self, prompt: Union[str, List[str], List[int], List[List[int]]], negative_prompt: Optional[Union[str, List[str], List[int], List[List[int]]]], width: int, height: int, strength: float, callback_steps: Optional[int] ): if isinstance(prompt, str) or (isinstance(prompt, list) and isinstance(prompt[0], int)): prompt = [prompt] if negative_prompt is None: negative_prompt = "" if isinstance(negative_prompt, str) or (isinstance(negative_prompt, list) and isinstance(negative_prompt[0], int)): negative_prompt = [negative_prompt] * len(prompt) if not isinstance(prompt, list): raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") if not isinstance(negative_prompt, list): raise ValueError(f"`negative_prompt` has to be of type `str` or `list` but is {type(negative_prompt)}") if len(negative_prompt) != len(prompt): raise ValueError( f"`prompt` and `negative_prompt` have to be the same length, but are {len(prompt)} and {len(negative_prompt)}") if strength < 0 or strength > 1: raise ValueError(f"`strength` should in [0.0, 1.0] but is {strength}") 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)}." ) return prompt, negative_prompt def encode_prompt( self, prompt: Union[List[str], List[List[int]]], negative_prompt: Union[List[str], List[List[int]]], num_images_per_prompt: int, do_classifier_free_guidance: bool, device ): if isinstance(prompt[0], str): text_input_ids = self.tokenizer(prompt, padding="do_not_pad").input_ids else: text_input_ids = prompt text_input_ids *= num_images_per_prompt if do_classifier_free_guidance: if isinstance(prompt[0], str): unconditional_input_ids = self.tokenizer(negative_prompt, padding="do_not_pad").input_ids else: unconditional_input_ids = negative_prompt unconditional_input_ids *= num_images_per_prompt text_input_ids = unconditional_input_ids + text_input_ids text_inputs = unify_input_ids(self.tokenizer, text_input_ids) text_input_ids = text_inputs.input_ids 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 = get_extended_embeddings(self.text_encoder, text_input_ids, attention_mask) return prompt_embeds def get_timesteps(self, latents_are_image, num_inference_steps, strength, device): if latents_are_image: # get the original timestep using init_timestep offset = self.scheduler.config.get("steps_offset", 0) init_timestep = int(num_inference_steps * strength) + offset init_timestep = min(init_timestep, num_inference_steps) t_start = max(num_inference_steps - init_timestep + offset, 0) timesteps = self.scheduler.timesteps[t_start:] else: timesteps = self.scheduler.timesteps timesteps = timesteps.to(device) return timesteps 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=device, dtype=dtype) # scale the initial noise by the standard deviation required by the scheduler latents = latents * self.scheduler.init_noise_sigma return latents def prepare_latents_from_image(self, init_image, timestep, batch_size, dtype, device, generator=None): init_image = init_image.to(device=device, dtype=dtype) init_latent_dist = self.vae.encode(init_image).latent_dist init_latents = init_latent_dist.sample(generator=generator) init_latents = 0.18215 * init_latents if batch_size > init_latents.shape[0]: raise ValueError( f"Cannot duplicate `init_image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." ) else: init_latents = torch.cat([init_latents] * batch_size, dim=0) # add noise to latents using the timesteps noise = torch.randn(init_latents.shape, generator=generator, device=device, dtype=dtype) # get latents init_latents = self.scheduler.add_noise(init_latents, noise, timestep) latents = init_latents return latents 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 def decode_latents(self, latents): latents = 1 / 0.18215 * latents image = self.vae.decode(latents).sample image = (image / 2 + 0.5).clamp(0, 1) # we always cast to float32 as this does not cause significant overhead and is compatible with bfloa16 image = image.cpu().permute(0, 2, 3, 1).float().numpy() return image @torch.no_grad() def __call__( self, prompt: Union[str, List[str], List[int], List[List[int]]], negative_prompt: Optional[Union[str, List[str], List[int], List[List[int]]]] = None, num_images_per_prompt: int = 1, strength: float = 0.8, height: Optional[int] = None, width: Optional[int] = None, num_inference_steps: int = 50, guidance_scale: float = 7.5, sag_scale: float = 0.75, eta: float = 0.0, generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, image: Optional[Union[torch.FloatTensor, PIL.Image.Image]] = None, output_type: str = "pil", return_dict: bool = True, callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, callback_steps: int = 1, cross_attention_kwargs: Optional[Dict[str, Any]] = None, ): r""" Function invoked when calling the pipeline for generation. Args: prompt (`str` or `List[str]`): The prompt or prompts to guide the image generation. strength (`float`, *optional*, defaults to 0.8): Conceptually, indicates how much to transform the reference `init_image`. Must be between 0 and 1. `init_image` will be used as a starting point, adding more noise to it the larger the `strength`. The number of denoising steps depends on the amount of noise initially added. When `strength` is 1, added noise will be maximum and the denoising process will run for the full number of iterations specified in `num_inference_steps`. A value of 1, therefore, essentially ignores `init_image`. height (`int`, *optional*, defaults to 768): The height in pixels of the generated image. width (`int`, *optional*, defaults to 768): 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. 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*): A [torch generator](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. cross_attention_kwargs (`dict`, *optional*): A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under `self.processor` in [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py). 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 prompt, negative_prompt = self.check_inputs(prompt, negative_prompt, width, height, strength, callback_steps) # 2. Define call parameters batch_size = len(prompt) device = self.execution_device do_classifier_free_guidance = guidance_scale > 1.0 do_self_attention_guidance = sag_scale > 0.0 latents_are_image = isinstance(image, PIL.Image.Image) # 3. Encode input prompt prompt_embeds = self.encode_prompt( prompt, negative_prompt, num_images_per_prompt, do_classifier_free_guidance, device ) # 4. Prepare timesteps self.scheduler.set_timesteps(num_inference_steps, device=device) timesteps = self.get_timesteps(latents_are_image, num_inference_steps, strength, device) # 5. Prepare latent variables num_channels_latents = self.unet.in_channels if latents_are_image: image = preprocess(image) latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) latents = self.prepare_latents_from_image( image, latent_timestep, batch_size * num_images_per_prompt, prompt_embeds.dtype, device, generator ) else: latents = self.prepare_latents( batch_size * num_images_per_prompt, num_channels_latents, height, width, prompt_embeds.dtype, device, generator, image, ) # 6. 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) # 7. Denoising loo if do_self_attention_guidance: store_processor = CrossAttnStoreProcessor() self.unet.mid_block.attentions[0].transformer_blocks[0].attn1.processor = store_processor 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=prompt_embeds, cross_attention_kwargs=cross_attention_kwargs, ).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) if do_self_attention_guidance: # classifier-free guidance produces two chunks of attention map # and we only use unconditional one according to equation (24) # in https://arxiv.org/pdf/2210.00939.pdf if do_classifier_free_guidance: # DDIM-like prediction of x0 pred_x0 = self.pred_x0(latents, noise_pred_uncond, t) # get the stored attention maps uncond_attn, cond_attn = store_processor.attention_probs.chunk(2) # self-attention-based degrading of latents degraded_latents = self.sag_masking( pred_x0, uncond_attn, t, self.pred_epsilon(latents, noise_pred_uncond, t) ) uncond_emb, _ = prompt_embeds.chunk(2) # forward and give guidance degraded_pred = self.unet(degraded_latents, t, encoder_hidden_states=uncond_emb).sample noise_pred += sag_scale * (noise_pred_uncond - degraded_pred) else: # DDIM-like prediction of x0 pred_x0 = self.pred_x0(latents, noise_pred, t) # get the stored attention maps cond_attn = store_processor.attention_probs # self-attention-based degrading of latents degraded_latents = self.sag_masking( pred_x0, cond_attn, t, self.pred_epsilon(latents, noise_pred, t) ) # forward and give guidance degraded_pred = self.unet(degraded_latents, t, encoder_hidden_states=prompt_embeds).sample noise_pred += sag_scale * (noise_pred - degraded_pred) # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample # call the callback, if provided if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): progress_bar.update() if callback is not None and i % callback_steps == 0: callback(i, t, latents) # 8. Post-processing image = self.decode_latents(latents) # 9. Run safety checker has_nsfw_concept = None # 10. Convert to PIL if output_type == "pil": image = self.numpy_to_pil(image) if not return_dict: return (image, has_nsfw_concept) return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) # Self-Attention-Guided (SAG) Stable Diffusion def sag_masking(self, original_latents, attn_map, t, eps): # Same masking process as in SAG paper: https://arxiv.org/pdf/2210.00939.pdf bh, hw1, hw2 = attn_map.shape b, latent_channel, latent_h, latent_w = original_latents.shape h = self.unet.attention_head_dim if isinstance(h, list): h = h[-1] map_size = math.isqrt(hw1) # Produce attention mask attn_map = attn_map.reshape(b, h, hw1, hw2) attn_mask = attn_map.mean(1, keepdim=False).sum(1, keepdim=False) > 1.0 attn_mask = ( attn_mask.reshape(b, map_size, map_size).unsqueeze(1).repeat(1, latent_channel, 1, 1).type(attn_map.dtype) ) attn_mask = torch.nn.functional.interpolate(attn_mask, (latent_h, latent_w)) # Blur according to the self-attention mask degraded_latents = gaussian_blur_2d(original_latents, kernel_size=9, sigma=1.0) degraded_latents = degraded_latents * attn_mask + original_latents * (1 - attn_mask) # Noise it again to match the noise level degraded_latents = self.scheduler.add_noise(degraded_latents, noise=eps, timesteps=t) return degraded_latents # Modified from diffusers.schedulers.scheduling_ddim.DDIMScheduler.step # Note: there are some schedulers that clip or do not return x_0 (PNDMScheduler, DDIMScheduler, etc.) def pred_x0(self, sample, model_output, timestep): alpha_prod_t = self.scheduler.alphas_cumprod[timestep] beta_prod_t = 1 - alpha_prod_t if self.scheduler.config.prediction_type == "epsilon": pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5) elif self.scheduler.config.prediction_type == "sample": pred_original_sample = model_output elif self.scheduler.config.prediction_type == "v_prediction": pred_original_sample = (alpha_prod_t**0.5) * sample - (beta_prod_t**0.5) * model_output # predict V model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample else: raise ValueError( f"prediction_type given as {self.scheduler.config.prediction_type} must be one of `epsilon`, `sample`," " or `v_prediction`" ) return pred_original_sample def pred_epsilon(self, sample, model_output, timestep): alpha_prod_t = self.scheduler.alphas_cumprod[timestep] beta_prod_t = 1 - alpha_prod_t if self.scheduler.config.prediction_type == "epsilon": pred_eps = model_output elif self.scheduler.config.prediction_type == "sample": pred_eps = (sample - (alpha_prod_t**0.5) * model_output) / (beta_prod_t**0.5) elif self.scheduler.config.prediction_type == "v_prediction": pred_eps = (beta_prod_t**0.5) * sample + (alpha_prod_t**0.5) * model_output else: raise ValueError( f"prediction_type given as {self.scheduler.config.prediction_type} must be one of `epsilon`, `sample`," " or `v_prediction`" ) return pred_eps