# Copyright (c) 2024 Jaerin Lee # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. from transformers import Blip2Processor, Blip2ForConditionalGeneration from diffusers import DiffusionPipeline, LCMScheduler, EulerDiscreteScheduler, AutoencoderTiny from huggingface_hub import hf_hub_download import torch import torch.nn as nn import torch.nn.functional as F import torchvision.transforms as T from einops import rearrange from collections import deque from typing import Tuple, List, Literal, Optional, Union from PIL import Image from util import load_model, gaussian_lowpass, shift_to_mask_bbox_center from data import BackgroundObject, LayerObject, BackgroundState #, LayerState class StreamMultiDiffusion(nn.Module): def __init__( self, device: torch.device, dtype: torch.dtype = torch.float16, sd_version: Literal['1.5'] = '1.5', hf_key: Optional[str] = None, lora_key: Optional[str] = None, use_tiny_vae: bool = True, t_index_list: List[int] = [0, 4, 12, 25, 37], # [0, 5, 16, 18, 20, 37], Magic number. width: int = 512, height: int = 512, frame_buffer_size: int = 1, num_inference_steps: int = 50, guidance_scale: float = 1.2, delta: float = 1.0, cfg_type: Literal['none', 'full', 'self', 'initialize'] = 'none', seed: int = 2024, autoflush: bool = True, default_mask_std: float = 8.0, default_mask_strength: float = 1.0, default_prompt_strength: float = 0.95, bootstrap_steps: int = 1, bootstrap_mix_steps: float = 1.0, # bootstrap_leak_sensitivity: float = 0.2, preprocess_mask_cover_alpha: float = 0.3, # TODO prompt_queue_capacity: int = 256, mask_type: Literal['discrete', 'semi-continuous', 'continuous'] = 'continuous', use_xformers: bool = True, ) -> None: super().__init__() self.device = device self.dtype = dtype self.seed = seed self.sd_version = sd_version self.autoflush = autoflush self.default_mask_std = default_mask_std self.default_mask_strength = default_mask_strength self.default_prompt_strength = default_prompt_strength self.bootstrap_steps = ( bootstrap_steps > torch.arange(len(t_index_list))).to(dtype=self.dtype, device=self.device) self.bootstrap_mix_steps = bootstrap_mix_steps self.bootstrap_mix_ratios = ( bootstrap_mix_steps - torch.arange(len(t_index_list), dtype=self.dtype, device=self.device)).clip_(0, 1) # self.bootstrap_leak_sensitivity = bootstrap_leak_sensitivity self.preprocess_mask_cover_alpha = preprocess_mask_cover_alpha self.mask_type = mask_type ### State definition # [0. Start] -(prepare)-> [1. Initialized] # [1. Initialized] -(update_background)-> [2. Background Registered] (len(self.prompts)==0) # [2. Background Registered] -(update_layers)-> [3. Unflushed] (len(self.prompts)>0) # [3. Unflushed] -(flush)-> [4. Ready] # [4. Ready] -(any updates)-> [3. Unflushed] # [4. Ready] -(__call__)-> [4. Ready], continuously returns generated image. self.ready_checklist = { 'initialized': False, 'background_registered': False, 'layers_ready': False, 'flushed': False, } ### Session state update queue: for lazy update policy for streaming applications. self.update_buffer = { 'background': None, # Maintains a single instance of BackgroundObject 'layers': deque(maxlen=prompt_queue_capacity), # Maintains a queue of LayerObjects } print(f'[INFO] Loading Stable Diffusion...') get_scheduler = lambda pipe: LCMScheduler.from_config(pipe.scheduler.config) lora_weight_name = None if self.sd_version == '1.5': if hf_key is not None: print(f'[INFO] Using custom model key: {hf_key}') model_key = hf_key else: model_key = 'runwayml/stable-diffusion-v1-5' lora_key = 'latent-consistency/lcm-lora-sdv1-5' lora_weight_name = 'pytorch_lora_weights.safetensors' # elif self.sd_version == 'xl': # model_key = 'stabilityai/stable-diffusion-xl-base-1.0' # lora_key = 'latent-consistency/lcm-lora-sdxl' # lora_weight_name = 'pytorch_lora_weights.safetensors' else: raise ValueError(f'Stable Diffusion version {self.sd_version} not supported.') ### Internally stored "Session" states self.state = { 'background': BackgroundState(), # Maintains a single instance of BackgroundState # 'layers': LayerState(), # Maintains a single instance of LayerState 'model_key': model_key, # The Hugging Face model ID. } # Create model self.i2t_processor = Blip2Processor.from_pretrained('Salesforce/blip2-opt-2.7b') self.i2t_model = Blip2ForConditionalGeneration.from_pretrained('Salesforce/blip2-opt-2.7b') self.pipe = load_model(model_key, self.sd_version, self.device, self.dtype) self.pipe.load_lora_weights(lora_key, weight_name=lora_weight_name, adapter_name='lcm') self.pipe.fuse_lora( fuse_unet=True, fuse_text_encoder=True, lora_scale=1.0, safe_fusing=False, ) if use_xformers: self.pipe.enable_xformers_memory_efficient_attention() self.vae = ( AutoencoderTiny.from_pretrained('madebyollin/taesd').to(device=self.device, dtype=self.dtype) if use_tiny_vae else self.pipe.vae ) # self.tokenizer = self.pipe.tokenizer self.text_encoder = self.pipe.text_encoder self.unet = self.pipe.unet self.vae_scale_factor = self.pipe.vae_scale_factor self.scheduler = get_scheduler(self.pipe) self.scheduler.set_timesteps(num_inference_steps) self.generator = None # Lock the canvas size--changing the canvas size can be implemented by reloading the module. self.height = height self.width = width self.latent_height = int(height // self.pipe.vae_scale_factor) self.latent_width = int(width // self.pipe.vae_scale_factor) # For bootstrapping. self.white = self.encode_imgs(torch.ones(1, 3, height, width, dtype=self.dtype, device=self.device)) # StreamDiffusion setting. self.t_list = t_index_list assert len(self.t_list) > 1, 'Current version only supports diffusion models with multiple steps.' self.frame_bff_size = frame_buffer_size # f self.denoising_steps_num = len(self.t_list) # t=2 self.cfg_type = cfg_type self.num_inference_steps = num_inference_steps self.guidance_scale = 1.0 if self.cfg_type == 'none' else guidance_scale self.delta = delta self.batch_size = self.denoising_steps_num * frame_buffer_size # T = t*f if self.cfg_type == 'initialize': self.trt_unet_batch_size = (self.denoising_steps_num + 1) * self.frame_bff_size elif self.cfg_type == 'full': self.trt_unet_batch_size = 2 * self.denoising_steps_num * self.frame_bff_size else: self.trt_unet_batch_size = self.denoising_steps_num * frame_buffer_size print(f'[INFO] Model is loaded!') self.reset_seed(self.generator, seed) self.reset_latent() self.prepare() print(f'[INFO] Parameters prepared!') self.ready_checklist['initialized'] = True @property def background(self) -> BackgroundState: return self.state['background'] # @property # def layers(self) -> LayerState: # return self.state['layers'] @property def num_layers(self) -> int: return len(self.prompts) if hasattr(self, 'prompts') else 0 @property def is_ready_except_flush(self) -> bool: return all(v for k, v in self.ready_checklist.items() if k != 'flushed') @property def is_flush_needed(self) -> bool: return self.autoflush and not self.ready_checklist['flushed'] @property def is_ready(self) -> bool: return self.is_ready_except_flush and not self.is_flush_needed @property def is_dirty(self) -> bool: return not (self.update_buffer['background'] is None and len(self.update_buffer['layers']) == 0) @property def has_background(self) -> bool: return self.background.is_empty # @property # def has_layers(self) -> bool: # return len(self.layers) > 0 def __repr__(self) -> str: return ( f'{type(self).__name__}(\n\tbackground: {str(self.background)},\n\t' f'model_key: {self.state["model_key"]}\n)' # f'layers: {str(self.layers)},\n\tmodel_key: {self.state["model_key"]}\n)' ) def check_integrity(self, throw_error: bool = True) -> bool: p = len(self.prompts) flag = ( p != len(self.negative_prompts) or p != len(self.prompt_strengths) or p != len(self.masks) or p != len(self.mask_strengths) or p != len(self.mask_stds) or p != len(self.original_masks) ) if flag and throw_error: print( f'LayerState(\n\tlen(prompts): {p},\n\tlen(negative_prompts): {len(self.negative_prompts)},\n\t' f'len(prompt_strengths): {len(self.prompt_strengths)},\n\tlen(masks): {len(self.masks)},\n\t' f'len(mask_stds): {len(self.mask_stds)},\n\tlen(mask_strengths): {len(self.mask_strengths)},\n\t' f'len(original_masks): {len(self.original_masks)}\n)' ) raise ValueError('[ERROR] LayerState is corrupted!') return not flag def check_ready(self) -> bool: all_except_flushed = all(v for k, v in self.ready_checklist.items() if k != 'flushed') if all_except_flushed: if self.is_flush_needed: self.flush() return True print('[WARNING] MagicDraw module is not ready yet! Complete the checklist:') for k, v in self.ready_checklist.items(): prefix = ' [ v ] ' if v else ' [ x ] ' print(prefix + k.replace('_', ' ')) return False def reset_seed(self, generator: Optional[torch.Generator] = None, seed: Optional[int] = None) -> None: generator = torch.Generator(self.device) if generator is None else generator seed = self.seed if seed is None else seed self.generator = generator self.generator.manual_seed(seed) self.init_noise = torch.randn((self.batch_size, 4, self.latent_height, self.latent_width), generator=generator, device=self.device, dtype=self.dtype) self.stock_noise = torch.zeros_like(self.init_noise) self.ready_checklist['flushed'] = False def reset_latent(self) -> None: # initialize x_t_latent (it can be any random tensor) b = (self.denoising_steps_num - 1) * self.frame_bff_size self.x_t_latent_buffer = torch.zeros( (b, 4, self.latent_height, self.latent_width), dtype=self.dtype, device=self.device) def reset_state(self) -> None: # TODO Reset states for context switch between multiple users. pass def prepare(self) -> None: # make sub timesteps list based on the indices in the t_list list and the values in the timesteps list self.timesteps = self.scheduler.timesteps.to(self.device) self.sub_timesteps = [] for t in self.t_list: self.sub_timesteps.append(self.timesteps[t]) sub_timesteps_tensor = torch.tensor(self.sub_timesteps, dtype=torch.long, device=self.device) self.sub_timesteps_tensor = sub_timesteps_tensor.repeat_interleave(self.frame_bff_size, dim=0) c_skip_list = [] c_out_list = [] for timestep in self.sub_timesteps: c_skip, c_out = self.scheduler.get_scalings_for_boundary_condition_discrete(timestep) c_skip_list.append(c_skip) c_out_list.append(c_out) self.c_skip = torch.stack(c_skip_list).view(len(self.t_list), 1, 1, 1).to(dtype=self.dtype, device=self.device) self.c_out = torch.stack(c_out_list).view(len(self.t_list), 1, 1, 1).to(dtype=self.dtype, device=self.device) alpha_prod_t_sqrt_list = [] beta_prod_t_sqrt_list = [] for timestep in self.sub_timesteps: alpha_prod_t_sqrt = self.scheduler.alphas_cumprod[timestep].sqrt() beta_prod_t_sqrt = (1 - self.scheduler.alphas_cumprod[timestep]).sqrt() alpha_prod_t_sqrt_list.append(alpha_prod_t_sqrt) beta_prod_t_sqrt_list.append(beta_prod_t_sqrt) alpha_prod_t_sqrt = (torch.stack(alpha_prod_t_sqrt_list).view(len(self.t_list), 1, 1, 1) .to(dtype=self.dtype, device=self.device)) beta_prod_t_sqrt = (torch.stack(beta_prod_t_sqrt_list).view(len(self.t_list), 1, 1, 1) .to(dtype=self.dtype, device=self.device)) self.alpha_prod_t_sqrt = alpha_prod_t_sqrt.repeat_interleave(self.frame_bff_size, dim=0) self.beta_prod_t_sqrt = beta_prod_t_sqrt.repeat_interleave(self.frame_bff_size, dim=0) noise_lvs = ((1 - self.scheduler.alphas_cumprod.to(self.device)[self.sub_timesteps_tensor]) ** 0.5) self.noise_lvs = noise_lvs[None, :, None, None, None] self.next_noise_lvs = torch.cat([noise_lvs[1:], noise_lvs.new_zeros(1)])[None, :, None, None, None] @torch.no_grad() def get_text_prompts(self, image: Image.Image) -> str: r"""A convenient method to extract text prompt from an image. This is called if the user does not provide background prompt but only the background image. We use BLIP-2 to automatically generate prompts. Args: image (Image.Image): A PIL image. Returns: A single string of text prompt. """ question = 'Question: What are in the image? Answer:' inputs = self.i2t_processor(image, question, return_tensors='pt') out = self.i2t_model.generate(**inputs, max_new_tokens=77) prompt = self.i2t_processor.decode(out[0], skip_special_tokens=True).strip() return prompt @torch.no_grad() def encode_imgs( self, imgs: torch.Tensor, generator: Optional[torch.Generator] = None, add_noise: bool = False, ) -> torch.Tensor: r"""A wrapper function for VAE encoder of the latent diffusion model. Args: imgs (torch.Tensor): An image to get StableDiffusion latents. Expected shape: (B, 3, H, W). Expected pixel scale: [0, 1]. generator (Optional[torch.Generator]): Seed for KL-Autoencoder. add_noise (bool): Turn this on for a noisy latent. Returns: An image latent embedding with 1/8 size (depending on the auto- encoder. Shape: (B, 4, H//8, W//8). """ def _retrieve_latents( encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = 'sample', ): if hasattr(encoder_output, 'latent_dist') and sample_mode == 'sample': return encoder_output.latent_dist.sample(generator) elif hasattr(encoder_output, 'latent_dist') and sample_mode == 'argmax': return encoder_output.latent_dist.mode() elif hasattr(encoder_output, 'latents'): return encoder_output.latents else: raise AttributeError('[ERROR] Could not access latents of provided encoder_output') imgs = 2 * imgs - 1 latents = self.vae.config.scaling_factor * _retrieve_latents(self.vae.encode(imgs), generator=generator) if add_noise: latents = self.alpha_prod_t_sqrt[0] * latents + self.beta_prod_t_sqrt[0] * self.init_noise[0] return latents @torch.no_grad() def decode_latents(self, latents: torch.Tensor) -> torch.Tensor: r"""A wrapper function for VAE decoder of the latent diffusion model. Args: latents (torch.Tensor): An image latent to get associated images. Expected shape: (B, 4, H//8, W//8). Returns: An image latent embedding with 1/8 size (depending on the auto- encoder. Shape: (B, 3, H, W). """ latents = 1 / self.vae.config.scaling_factor * latents imgs = self.vae.decode(latents).sample imgs = (imgs / 2 + 0.5).clip_(0, 1) return imgs @torch.no_grad() def update_background( self, image: Optional[Image.Image] = None, prompt: Optional[str] = None, negative_prompt: Optional[str] = None, ) -> bool: flag_changed = False if image is not None: image_ = image.resize((self.width, self.height)) prompt = self.get_text_prompts(image_) if prompt is None else prompt negative_prompt = '' if negative_prompt is None else negative_prompt embed = self.pipe.encode_prompt( prompt=[prompt], device=self.device, num_images_per_prompt=1, do_classifier_free_guidance=(self.guidance_scale > 1.0), negative_prompt=[negative_prompt], ) # ((1, 77, 768): cond, (1, 77, 768): uncond) self.state['background'].image = image self.state['background'].latent = ( self.encode_imgs(T.ToTensor()(image_)[None].to(self.device, self.dtype)) ) # (1, 3, H, W) self.state['background'].prompt = prompt self.state['background'].negative_prompt = negative_prompt self.state['background'].embed = embed if self.bootstrap_steps[0] > 0: mix_ratio = self.bootstrap_mix_ratios[:, None, None, None] self.bootstrap_latent = mix_ratio * self.white + (1.0 - mix_ratio) * self.state['background'].latent self.ready_checklist['background_registered'] = True flag_changed = True else: if not self.ready_checklist['background_registered']: print('[WARNING] Register background image first! Request ignored.') return False if prompt is not None: self.background.prompt = prompt flag_changed = True if negative_prompt is not None: self.background.negative_prompt = negative_prompt flag_changed = True if flag_changed: self.background.embed = self.pipe.encode_prompt( prompt=[self.background.prompt], device=self.device, num_images_per_prompt=1, do_classifier_free_guidance=(self.guidance_scale > 1.0), negative_prompt=[self.background.negative_prompt], ) # ((1, 77, 768): cond, (1, 77, 768): uncond) self.ready_checklist['flushed'] = not flag_changed return flag_changed @torch.no_grad() def process_mask( self, masks: Optional[Union[torch.Tensor, Image.Image, List[Image.Image]]] = None, strength: Optional[Union[torch.Tensor, float]] = None, std: Optional[Union[torch.Tensor, float]] = None, ) -> Tuple[torch.Tensor]: r"""Fast preprocess of masks for region-based generation with fine- grained controls. Mask preprocessing is done in four steps: 1. Resizing: Resize the masks into the specified width and height by nearest neighbor interpolation. 2. (Optional) Ordering: Masks with higher indices are considered to cover the masks with smaller indices. Covered masks are decayed in its alpha value by the specified factor of `preprocess_mask_cover_alpha`. 3. Blurring: Gaussian blur is applied to the mask with the specified standard deviation (isotropic). This results in gradual increase of masked region as the timesteps evolve, naturally blending fore- ground and the predesignated background. Not strictly required if you want to produce images from scratch withoout background. 4. Quantization: Split the real-numbered masks of value between [0, 1] into predefined noise levels for each quantized scheduling step of the diffusion sampler. For example, if the diffusion model sampler has noise level of [0.9977, 0.9912, 0.9735, 0.8499, 0.5840], which is the default noise level of this module with schedule [0, 4, 12, 25, 37], the masks are split into binary masks whose values are greater than these levels. This results in tradual increase of mask region as the timesteps increase. Details are described in our paper at https://arxiv.org/pdf/2403.09055.pdf. On the Three Modes of `mask_type`: `self.mask_type` is predefined at the initialization stage of this pipeline. Three possible modes are available: 'discrete', 'semi- continuous', and 'continuous'. These define the mask quantization modes we use. Basically, this (subtly) controls the smoothness of foreground-background blending. Continuous modes produces nonbinary masks to further blend foreground and background latents by linear- ly interpolating between them. Semi-continuous masks only applies continuous mask at the last step of the LCM sampler. Due to the large step size of the LCM scheduler, we find that our continuous blending helps generating seamless inpainting and editing results. Args: masks (Union[torch.Tensor, Image.Image, List[Image.Image]]): Masks. strength (Optional[Union[torch.Tensor, float]]): Mask strength that overrides the default value. A globally multiplied factor to the mask at the initial stage of processing. Can be applied seperately for each mask. std (Optional[Union[torch.Tensor, float]]): Mask blurring Gaussian kernel's standard deviation. Overrides the default value. Can be applied seperately for each mask. Returns: A tuple of tensors. - masks: Preprocessed (ordered, blurred, and quantized) binary/non- binary masks (see the explanation on `mask_type` above) for region-based image synthesis. - strengths: Return mask strengths for caching. - std: Return mask blur standard deviations for caching. - original_masks: Return original masks for caching. """ if masks is None: kwargs = {'dtype': self.dtype, 'device': self.device} original_masks = torch.zeros((0, 1, self.latent_height, self.latent_width), dtype=self.dtype) masks = torch.zeros((0, self.batch_size, 1, self.latent_height, self.latent_width), **kwargs) strength = torch.zeros((0,), **kwargs) std = torch.zeros((0,), **kwargs) return masks, strength, std, original_masks if isinstance(masks, Image.Image): masks = [masks] if isinstance(masks, (tuple, list)): # Assumes white background for Image.Image; # inverted boolean masks with shape (1, 1, H, W) for torch.Tensor. masks = torch.cat([ # (T.ToTensor()(mask.resize((self.width, self.height), Image.NEAREST)) < 0.5)[None, :1] (1.0 - T.ToTensor()(mask.resize((self.width, self.height), Image.BILINEAR)))[None, :1] for mask in masks ], dim=0).float().clip_(0, 1) original_masks = masks masks = masks.float().to(self.device) # Background mask alpha is decayed by the specified factor where foreground masks covers it. if self.preprocess_mask_cover_alpha > 0: masks = torch.stack([ torch.where( masks[i + 1:].sum(dim=0) > 0, mask * self.preprocess_mask_cover_alpha, mask, ) if i < len(masks) - 1 else mask for i, mask in enumerate(masks) ], dim=0) if std is None: std = self.default_mask_std if isinstance(std, (int, float)): std = [std] * len(masks) if isinstance(std, (list, tuple)): std = torch.as_tensor(std, dtype=torch.float, device=self.device) # Mask preprocessing parameters are fetched from the default settings. if strength is None: strength = self.default_mask_strength if isinstance(strength, (int, float)): strength = [strength] * len(masks) if isinstance(strength, (list, tuple)): strength = torch.as_tensor(strength, dtype=torch.float, device=self.device) if (std > 0).any(): std = torch.where(std > 0, std, 1e-5) masks = gaussian_lowpass(masks, std) # NOTE: This `strength` aligns with `denoising strength`. However, with LCM, using strength < 0.96 # gives unpleasant results. masks = masks * strength[:, None, None, None] masks = masks.unsqueeze(1).repeat(1, self.noise_lvs.shape[1], 1, 1, 1) if self.mask_type == 'discrete': # Discrete mode. masks = masks > self.noise_lvs elif self.mask_type == 'semi-continuous': # Semi-continuous mode (continuous at the last step only). masks = torch.cat(( masks[:, :-1] > self.noise_lvs[:, :-1], ( (masks[:, -1:] - self.next_noise_lvs[:, -1:]) / (self.noise_lvs[:, -1:] - self.next_noise_lvs[:, -1:]) ).clip_(0, 1), ), dim=1) elif self.mask_type == 'continuous': # Continuous mode: Have the exact same `1` coverage with discrete mode, but the mask gradually # decreases continuously after the discrete mode boundary to become `0` at the # next lower threshold. masks = ((masks - self.next_noise_lvs) / (self.noise_lvs - self.next_noise_lvs)).clip_(0, 1) # NOTE: Post processing mask strength does not align with conventional 'denoising_strength'. However, # fine-grained mask alpha channel tuning is available with this form. # masks = masks * strength[None, :, None, None, None] masks = rearrange(masks.float(), 'p t () h w -> (p t) () h w') masks = F.interpolate(masks, size=(self.latent_height, self.latent_width), mode='nearest') masks = rearrange(masks.to(self.dtype), '(p t) () h w -> p t () h w', p=len(std)) return masks, strength, std, original_masks @torch.no_grad() def update_layers( self, prompts: Union[str, List[str]], negative_prompts: Optional[Union[str, List[str]]] = None, suffix: Optional[str] = None, #', background is ', prompt_strengths: Optional[Union[torch.Tensor, float, List[float]]] = None, masks: Optional[Union[torch.Tensor, Image.Image, List[Image.Image]]] = None, mask_strengths: Optional[Union[torch.Tensor, float, List[float]]] = None, mask_stds: Optional[Union[torch.Tensor, float, List[float]]] = None, ) -> None: if not self.ready_checklist['background_registered']: print('[WARNING] Register background image first! Request ignored.') return ### Register prompts if isinstance(prompts, str): prompts = [prompts] if negative_prompts is None: negative_prompts = '' if isinstance(negative_prompts, str): negative_prompts = [negative_prompts] fg_prompt = [p + suffix + self.background.prompt if suffix is not None else p for p in prompts] self.prompts = fg_prompt self.negative_prompts = negative_prompts p = self.num_layers e = self.pipe.encode_prompt( prompt=fg_prompt, device=self.device, num_images_per_prompt=1, do_classifier_free_guidance=(self.guidance_scale > 1.0), negative_prompt=negative_prompts, ) # (p, 77, 768) if prompt_strengths is None: prompt_strengths = self.default_prompt_strength if isinstance(prompt_strengths, (int, float)): prompt_strengths = [prompt_strengths] * p if isinstance(prompt_strengths, (list, tuple)): prompt_strengths = torch.as_tensor(prompt_strengths, dtype=self.dtype, device=self.device) self.prompt_strengths = prompt_strengths s = prompt_strengths[:, None, None] self.prompt_embeds = torch.lerp(self.background.embed[0], e[0], s).repeat(self.batch_size, 1, 1) # (T * p, 77, 768) if self.guidance_scale > 1.0 and self.cfg_type in ('initialize', 'full'): b = self.batch_size if self.cfg_type == 'full' else self.frame_bff_size uncond_prompt_embeds = torch.lerp(self.background.embed[1], e[1], s).repeat(b, 1, 1) # (T * p, 77, 768) self.prompt_embeds = torch.cat([uncond_prompt_embeds, self.prompt_embeds], dim=0) # (2 * T * p, 77, 768) self.sub_timesteps_tensor_ = self.sub_timesteps_tensor.repeat_interleave(p) # (T * p,) self.init_noise_ = self.init_noise.repeat_interleave(p, dim=0) # (T * p, 77, 768) self.stock_noise_ = self.stock_noise.repeat_interleave(p, dim=0) # (T * p, 77, 768) self.c_out_ = self.c_out.repeat_interleave(p, dim=0) # (T * p, 1, 1, 1) self.c_skip_ = self.c_skip.repeat_interleave(p, dim=0) # (T * p, 1, 1, 1) self.beta_prod_t_sqrt_ = self.beta_prod_t_sqrt.repeat_interleave(p, dim=0) # (T * p, 1, 1, 1) self.alpha_prod_t_sqrt_ = self.alpha_prod_t_sqrt.repeat_interleave(p, dim=0) # (T * p, 1, 1, 1) ### Register new masks if isinstance(masks, Image.Image): masks = [masks] n = len(masks) if masks is not None else 0 # Modificiation. masks, mask_strengths, mask_stds, original_masks = self.process_mask(masks, mask_strengths, mask_stds) self.counts = masks.sum(dim=0) # (T, 1, h, w) self.bg_mask = (1 - self.counts).clip_(0, 1) # (T, 1, h, w) self.masks = masks # (p, T, 1, h, w) self.mask_strengths = mask_strengths # (p,) self.mask_stds = mask_stds # (p,) self.original_masks = original_masks # (p, 1, h, w) if p > n: # Add more masks: counts and bg_masks are not changed, but only masks are changed. self.masks = torch.cat(( self.masks, torch.zeros( (p - n, self.batch_size, 1, self.latent_height, self.latent_width), dtype=self.dtype, device=self.device, ), ), dim=0) print(f'[WARNING] Detected more prompts ({p}) than masks ({n}). ' 'Automatically adds blank masks for the additional prompts.') elif p < n: # Warns user to add more prompts. print(f'[WARNING] Detected more masks ({n}) than prompts ({p}). ' 'Additional masks are ignored until more prompts are provided.') self.ready_checklist['layers_ready'] = True self.ready_checklist['flushed'] = False @torch.no_grad() def update_single_layer( self, idx: Optional[int] = None, prompt: Optional[str] = None, negative_prompt: Optional[str] = None, suffix: Optional[str] = None, #', background is ', prompt_strength: Optional[float] = None, mask: Optional[Union[torch.Tensor, Image.Image]] = None, mask_strength: Optional[float] = None, mask_std: Optional[float] = None, ) -> None: ### Possible input combinations and expected behaviors # The module will consider a layer, a pair of (prompt, mask), to be 'active' only if a prompt # is registered. A blank mask will be assigned if no mask is provided for the 'active' layer. # The layers should be in either of ('active', 'inactive') states. 'inactive' layers will not # receive any input unless equipped with prompt. 'active' layers receive any input and modify # their states accordingly. In the actual implementation, only the 'active' layers are stored # and can be accessed by the fields. Values len(self.prompts) = self.num_layers is the number # of 'active' layers. # If no background is registered. The layers should be all 'inactive'. if not self.ready_checklist['background_registered']: print('[WARNING] Register background image first! Request ignored.') return # The first layer create request should be carrying a prompt. If only mask is drawn without a # prompt, it just ignores the request--the user will update her request soon. if self.num_layers == 0: if prompt is not None: self.update_layers( prompts=prompt, negative_prompts=negative_prompt, suffix=suffix, prompt_strengths=prompt_strength, masks=mask, mask_strengths=mask_strength, mask_stds=mask_std, ) return # Invalid request indices -> considered as a layer add request. if idx is None or idx > self.num_layers or idx < 0: idx = self.num_layers # Two modes for the layer edits: 'append mode' and 'edit mode'. 'append mode' appends a new # layer at the end of the layers list. 'edit mode' modifies internal variables for the given # index. 'append mode' is defined by the request index and strictly requires a prompt input. is_appending = idx == self.num_layers if is_appending and prompt is None: print(f'[WARNING] Creating a new prompt at index ({idx}) but found no prompt. Request ignored.') return ### Register prompts # | prompt | neg_prompt | append mode (idx==len) | edit mode (idx 1.0 and self.cfg_type in ('initialize', 'full') # Synchonize the internal state. # We have asserted that prompt is not None if the mode is 'appending'. if prompt is not None: if suffix is not None: prompt = prompt + suffix + self.background.prompt if is_appending: self.prompts.append(prompt) else: self.prompts[idx] = prompt if negative_prompt is not None: if is_appending: self.negative_prompts.append(negative_prompt) else: self.negative_prompts[idx] = negative_prompt elif is_appending: # Make sure that negative prompts are well specified. self.negative_prompts.append('') if is_appending: if prompt_strength is None: prompt_strength = self.default_prompt_strength self.prompt_strengths = torch.cat(( self.prompt_strengths, torch.as_tensor([prompt_strength], dtype=self.dtype, device=self.device), ), dim=0) elif prompt_strength is not None: self.prompt_strengths[idx] = prompt_strength # Edit currently stored prompt embeddings. if is_double_cond: uncond_prompt_embed_, prompt_embed_ = torch.chunk(self.prompt_embeds, 2, dim=0) uncond_prompt_embed_ = rearrange(uncond_prompt_embed_, '(t p) c1 c2 -> t p c1 c2', p=p) prompt_embed_ = rearrange(prompt_embed_, '(t p) c1 c2 -> t p c1 c2', p=p) else: uncond_prompt_embed_ = None prompt_embed_ = rearrange(self.prompt_embeds, '(t p) c1 c2 -> t p c1 c2', p=p) e = self.pipe.encode_prompt( prompt=self.prompts[idx], device=self.device, num_images_per_prompt=1, do_classifier_free_guidance=(self.guidance_scale > 1.0), negative_prompt=self.negative_prompts[idx], ) # (1, 77, 768), (1, 77, 768) s = self.prompt_strengths[idx] t = prompt_embed_.shape[0] prompt_embed = torch.lerp(self.background.embed[0], e[0], s)[None].repeat(t, 1, 1, 1) # (1, 77, 768) if is_double_cond: uncond_prompt_embed = torch.lerp(self.background.embed[1], e[1], s)[None].repeat(t, 1, 1, 1) # (1, 77, 768) if is_appending: prompt_embed_ = torch.cat((prompt_embed_, prompt_embed), dim=1) if is_double_cond: uncond_prompt_embed_ = torch.cat((uncond_prompt_embed_, uncond_prompt_embed), dim=1) else: prompt_embed_[:, idx:(idx + 1)] = prompt_embed if is_double_cond: uncond_prompt_embed_[:, idx:(idx + 1)] = uncond_prompt_embed self.prompt_embeds = rearrange(prompt_embed_, 't p c1 c2 -> (t p) c1 c2') if is_double_cond: uncond_prompt_embeds = rearrange(uncond_prompt_embed_, 't p c1 c2 -> (t p) c1 c2') self.prompt_embeds = torch.cat([uncond_prompt_embeds, self.prompt_embeds], dim=0) # (2 * T * p, 77, 768) self.ready_checklist['flushed'] = False if is_appending: p = self.num_layers self.sub_timesteps_tensor_ = self.sub_timesteps_tensor.repeat_interleave(p) # (T * p,) self.init_noise_ = self.init_noise.repeat_interleave(p, dim=0) # (T * p, 77, 768) self.stock_noise_ = self.stock_noise.repeat_interleave(p, dim=0) # (T * p, 77, 768) self.c_out_ = self.c_out.repeat_interleave(p, dim=0) # (T * p, 1, 1, 1) self.c_skip_ = self.c_skip.repeat_interleave(p, dim=0) # (T * p, 1, 1, 1) self.beta_prod_t_sqrt_ = self.beta_prod_t_sqrt.repeat_interleave(p, dim=0) # (T * p, 1, 1, 1) self.alpha_prod_t_sqrt_ = self.alpha_prod_t_sqrt.repeat_interleave(p, dim=0) # (T * p, 1, 1, 1) ### Register new masks # | mask | std / str | append mode (idx==len) | edit mode (idx create mask. mask, strength, std, original_mask = self.process_mask(mask, mask_strength, mask_std) flag_nonzero_mask = True elif is_appending: # No given mask & append mode -> create white mask. mask = torch.zeros( (1, self.batch_size, 1, self.latent_height, self.latent_width), dtype=self.dtype, device=self.device, ) strength = torch.as_tensor([self.default_mask_strength], dtype=self.dtype, device=self.device) std = torch.as_tensor([self.default_mask_std], dtype=self.dtype, device=self.device) original_mask = torch.zeros((1, 1, self.latent_height, self.latent_width), dtype=self.dtype) elif mask_std is not None or mask_strength is not None: # No given mask & edit mode & given std / str -> replace existing mask with given std / str. if mask_std is None: mask_std = self.mask_stds[idx:(idx + 1)] if mask_strength is None: mask_strength = self.mask_strengths[idx:(idx + 1)] mask, strength, std, original_mask = self.process_mask( self.original_masks[idx:(idx + 1)], mask_strength, mask_std) flag_nonzero_mask = True else: # No given mask & no given std & edit mode -> Do nothing. return if is_appending: # Append mode. self.masks = torch.cat((self.masks, mask), dim=0) # (p, T, 1, h, w) self.mask_strengths = torch.cat((self.mask_strengths, strength), dim=0) # (p,) self.mask_stds = torch.cat((self.mask_stds, std), dim=0) # (p,) self.original_masks = torch.cat((self.original_masks, original_mask), dim=0) # (p, 1, h, w) if flag_nonzero_mask: self.counts = self.counts + mask[0] if hasattr(self, 'counts') else mask[0] # (T, 1, h, w) self.bg_mask = (1 - self.counts).clip_(0, 1) # (T, 1, h, w) else: # Edit mode. if flag_nonzero_mask: self.counts = self.counts - self.masks[idx] + mask[0] # (T, 1, h, w) self.bg_mask = (1 - self.counts).clip_(0, 1) # (T, 1, h, w) self.masks[idx:(idx + 1)] = mask # (p, T, 1, h, w) self.mask_strengths[idx:(idx + 1)] = strength # (p,) self.mask_stds[idx:(idx + 1)] = std # (p,) self.original_masks[idx:(idx + 1)] = original_mask # (p, 1, h, w) # if flag_nonzero_mask: # self.ready_checklist['flushed'] = False @torch.no_grad() def register_all( self, prompts: Union[str, List[str]], masks: Union[Image.Image, List[Image.Image]], background: Image.Image, background_prompt: Optional[str] = None, background_negative_prompt: str = '', negative_prompts: Union[str, List[str]] = '', suffix: Optional[str] = None, #', background is ', prompt_strengths: float = 1.0, mask_strengths: float = 1.0, mask_stds: Union[torch.Tensor, float] = 10.0, ) -> None: # The order of this registration should not be changed! self.update_background(background, background_prompt, background_negative_prompt) self.update_layers(prompts, negative_prompts, suffix, prompt_strengths, masks, mask_strengths, mask_stds) def update( self, background: Optional[Image.Image] = None, background_prompt: Optional[str] = None, background_negative_prompt: Optional[str] = None, idx: Optional[int] = None, prompt: Optional[str] = None, negative_prompt: Optional[str] = None, suffix: Optional[str] = None, prompt_strength: Optional[float] = None, mask: Optional[Union[torch.Tensor, Image.Image]] = None, mask_strength: Optional[float] = None, mask_std: Optional[float] = None, ) -> None: # For lazy update (to solve minor synchonization problem with gradio). bq = BackgroundObject( image=background, prompt=background_prompt, negative_prompt=background_negative_prompt, ) if not bq.is_empty: self.update_buffer['background'] = bq lq = LayerObject( idx=idx, prompt=prompt, negative_prompt=negative_prompt, suffix=suffix, prompt_strength=prompt_strength, mask=mask, mask_strength=mask_strength, mask_std=mask_std, ) if not lq.is_empty: limit = self.update_buffer['layers'].maxlen # Optimize the prompt queue: Overrride uncommitted layers with the same idx. new_q = deque(maxlen=limit) for _ in range(len(self.update_buffer['layers'])): # Check from the newest to the oldest. # Copy old requests only if the current query does not carry those requests. query = self.update_buffer['layers'].pop() overriden = lq.merge(query) if not overriden: new_q.appendleft(query) self.update_buffer['layers'] = new_q if len(self.update_buffer['layers']) == limit: print(f'[WARNING] Maximum prompt change query limit ({limit}) is reached. ' f'Current query {lq} will be ignored.') else: self.update_buffer['layers'].append(lq) @torch.no_grad() def commit(self) -> None: flag_changed = self.is_dirty bq = self.update_buffer['background'] lq = self.update_buffer['layers'] count_bq_req = int(bq is not None and not bq.is_empty) count_lq_req = len(lq) if flag_changed: print(f'[INFO] Requests found: {count_bq_req} background requests ' f'& {count_lq_req} layer requests:\n{str(bq)}, {", ".join([str(l) for l in lq])}') bq = self.update_buffer['background'] if bq is not None: self.update_background(**vars(bq)) self.update_buffer['background'] = None while len(lq) > 0: l = lq.popleft() self.update_single_layer(**vars(l)) if flag_changed: print(f'[INFO] Requests resolved: {count_bq_req} background requests ' f'& {count_lq_req} layer requests.') def scheduler_step_batch( self, model_pred_batch: torch.Tensor, x_t_latent_batch: torch.Tensor, idx: Optional[int] = None, ) -> torch.Tensor: r"""Denoise-only step for reverse diffusion scheduler. Args: model_pred_batch (torch.Tensor): Noise prediction results. x_t_latent_batch (torch.Tensor): Noisy latent. idx (Optional[int]): Instead of timesteps (in [0, 1000]-scale) use indices for the timesteps tensor (ranged in [0, len(timesteps)-1]). Specify only if a single-index, not stream-batched inference is what you want. Returns: A denoised tensor with the same size as latent. """ if idx is None: F_theta = (x_t_latent_batch - self.beta_prod_t_sqrt_ * model_pred_batch) / self.alpha_prod_t_sqrt_ denoised_batch = self.c_out_ * F_theta + self.c_skip_ * x_t_latent_batch else: F_theta = (x_t_latent_batch - self.beta_prod_t_sqrt[idx] * model_pred_batch) / self.alpha_prod_t_sqrt[idx] denoised_batch = self.c_out[idx] * F_theta + self.c_skip[idx] * x_t_latent_batch return denoised_batch def unet_step( self, x_t_latent: torch.Tensor, # (T, 4, h, w) idx: Optional[int] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: p = self.num_layers x_t_latent = x_t_latent.repeat_interleave(p, dim=0) # (T * p, 4, h, w) if self.bootstrap_steps[0] > 0: # Background bootstrapping. bootstrap_latent = self.scheduler.add_noise( self.bootstrap_latent, self.stock_noise, torch.tensor(self.sub_timesteps_tensor, device=self.device), ) x_t_latent = rearrange(x_t_latent, '(t p) c h w -> p t c h w', p=p) bootstrap_mask = ( self.masks * self.bootstrap_steps[None, :, None, None, None] + (1.0 - self.bootstrap_steps[None, :, None, None, None]) ) # (p, t, c, h, w) x_t_latent = (1.0 - bootstrap_mask) * bootstrap_latent[None] + bootstrap_mask * x_t_latent x_t_latent = rearrange(x_t_latent, 'p t c h w -> (t p) c h w') # Centering. x_t_latent = shift_to_mask_bbox_center(x_t_latent, rearrange(self.masks, 'p t c h w -> (t p) c h w'), reverse=True) t_list = self.sub_timesteps_tensor_ # (T * p,) if self.guidance_scale > 1.0 and self.cfg_type == 'initialize': x_t_latent_plus_uc = torch.concat([x_t_latent[:p], x_t_latent], dim=0) # (T * p + 1, 4, h, w) t_list = torch.concat([t_list[:p], t_list], dim=0) # (T * p + 1, 4, h, w) elif self.guidance_scale > 1.0 and self.cfg_type == 'full': x_t_latent_plus_uc = torch.concat([x_t_latent, x_t_latent], dim=0) # (2 * T * p, 4, h, w) t_list = torch.concat([t_list, t_list], dim=0) # (2 * T * p,) else: x_t_latent_plus_uc = x_t_latent # (T * p, 4, h, w) model_pred = self.unet( x_t_latent_plus_uc, # (B, 4, h, w) t_list, # (B,) encoder_hidden_states=self.prompt_embeds, # (B, 77, 768) return_dict=False, # TODO: Add SDXL Support. # added_cond_kwargs={'text_embeds': add_text_embeds, 'time_ids': add_time_ids}, )[0] # (B, 4, h, w) if self.bootstrap_steps[0] > 0: # Uncentering. bootstrap_mask = rearrange(self.masks, 'p t c h w -> (t p) c h w') if self.guidance_scale > 1.0 and self.cfg_type == 'initialize': bootstrap_mask_ = torch.concat([bootstrap_mask[:p], bootstrap_mask], dim=0) elif self.guidance_scale > 1.0 and self.cfg_type == 'full': bootstrap_mask_ = torch.concat([bootstrap_mask, bootstrap_mask], dim=0) else: bootstrap_mask_ = bootstrap_mask model_pred = shift_to_mask_bbox_center(model_pred, bootstrap_mask_) x_t_latent = shift_to_mask_bbox_center(x_t_latent, bootstrap_mask) # # Remove leakage (optional). # leak = (latent_ - bg_latent_).pow(2).mean(dim=1, keepdim=True) # leak_sigmoid = torch.sigmoid(leak / self.bootstrap_leak_sensitivity) * 2 - 1 # fg_mask_ = fg_mask_ * leak_sigmoid ### noise_pred_text, noise_pred_uncond: (T * p, 4, h, w) ### self.stock_noise, init_noise: (T, 4, h, w) if self.guidance_scale > 1.0 and self.cfg_type == 'initialize': noise_pred_text = model_pred[p:] self.stock_noise_ = torch.concat([model_pred[:p], self.stock_noise_[p:]], dim=0) elif self.guidance_scale > 1.0 and self.cfg_type == 'full': noise_pred_uncond, noise_pred_text = model_pred.chunk(2) else: noise_pred_text = model_pred if self.guidance_scale > 1.0 and self.cfg_type in ('self', 'initialize'): noise_pred_uncond = self.stock_noise_ * self.delta if self.guidance_scale > 1.0 and self.cfg_type != 'none': model_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) else: model_pred = noise_pred_text # compute the previous noisy sample x_t -> x_t-1 denoised_batch = self.scheduler_step_batch(model_pred, x_t_latent, idx) if self.cfg_type in ('self' , 'initialize'): scaled_noise = self.beta_prod_t_sqrt_ * self.stock_noise_ delta_x = self.scheduler_step_batch(model_pred, scaled_noise, idx) # Do mask edit. alpha_next = torch.concat([self.alpha_prod_t_sqrt_[p:], torch.ones_like(self.alpha_prod_t_sqrt_[:p])], dim=0) delta_x = alpha_next * delta_x beta_next = torch.concat([self.beta_prod_t_sqrt_[p:], torch.ones_like(self.beta_prod_t_sqrt_[:p])], dim=0) delta_x = delta_x / beta_next init_noise = torch.concat([self.init_noise_[p:], self.init_noise_[:p]], dim=0) self.stock_noise_ = init_noise + delta_x p2 = len(self.t_list) - 1 background = torch.concat([ self.scheduler.add_noise( self.background.latent.repeat(p2, 1, 1, 1), self.stock_noise[1:], torch.tensor(self.t_list[1:], device=self.device), ), self.background.latent, ], dim=0) denoised_batch = rearrange(denoised_batch, '(t p) c h w -> p t c h w', p=p) latent = (self.masks * denoised_batch).sum(dim=0) # (T, 4, h, w) latent = torch.where(self.counts > 0, latent / self.counts, latent) # latent = ( # (1 - self.bg_mask) * self.mask_strengths * latent # + ((1 - self.bg_mask) * (1.0 - self.mask_strengths) + self.bg_mask) * background # ) latent = (1 - self.bg_mask) * latent + self.bg_mask * background return latent @torch.no_grad() def __call__( self, no_decode: bool = False, ignore_check_ready: bool = False, ) -> Optional[Union[torch.Tensor, Image.Image]]: if not ignore_check_ready and not self.check_ready(): return if not ignore_check_ready and self.is_dirty: print("I'm so dirty now!") self.commit() self.flush() latent = torch.randn((1, self.unet.config.in_channels, self.latent_height, self.latent_width), dtype=self.dtype, device=self.device) # (1, 4, h, w) latent = torch.cat((latent, self.x_t_latent_buffer), dim=0) # (t, 4, h, w) self.stock_noise = torch.cat((self.init_noise[:1], self.stock_noise[:-1]), dim=0) # (t, 4, h, w) if self.cfg_type in ('self', 'initialize'): self.stock_noise_ = self.stock_noise.repeat_interleave(self.num_layers, dim=0) # (T * p, 77, 768) x_0_pred_batch = self.unet_step(latent) latent = x_0_pred_batch[-1:] self.x_t_latent_buffer = ( self.alpha_prod_t_sqrt[1:] * x_0_pred_batch[:-1] + self.beta_prod_t_sqrt[1:] * self.init_noise[1:] ) # For pipeline flushing. if no_decode: return latent imgs = self.decode_latents(latent.half()) # (1, 3, H, W) img = T.ToPILImage()(imgs[0].cpu()) return img def flush(self) -> None: for _ in self.t_list: self(True, True) self.ready_checklist['flushed'] = True