# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved. import gc import logging import math import os import random import sys import types from contextlib import contextmanager from functools import partial from mmgp import offload import torch import torch.nn as nn import torch.cuda.amp as amp import torch.distributed as dist from tqdm import tqdm from PIL import Image import torchvision.transforms.functional as TF import torch.nn.functional as F from .distributed.fsdp import shard_model from .modules.model import WanModel from .modules.t5 import T5EncoderModel from .modules.vae import WanVAE from .utils.fm_solvers import (FlowDPMSolverMultistepScheduler, get_sampling_sigmas, retrieve_timesteps) from .utils.fm_solvers_unipc import FlowUniPCMultistepScheduler from wan.modules.posemb_layers import get_rotary_pos_embed from .utils.vace_preprocessor import VaceVideoProcessor from wan.utils.basic_flowmatch import FlowMatchScheduler def optimized_scale(positive_flat, negative_flat): # Calculate dot production dot_product = torch.sum(positive_flat * negative_flat, dim=1, keepdim=True) # Squared norm of uncondition squared_norm = torch.sum(negative_flat ** 2, dim=1, keepdim=True) + 1e-8 # st_star = v_cond^T * v_uncond / ||v_uncond||^2 st_star = dot_product / squared_norm return st_star class WanT2V: def __init__( self, config, checkpoint_dir, rank=0, model_filename = None, text_encoder_filename = None, quantizeTransformer = False, dtype = torch.bfloat16, VAE_dtype = torch.float32, mixed_precision_transformer = False ): self.device = torch.device(f"cuda") self.config = config self.rank = rank self.dtype = dtype self.num_train_timesteps = config.num_train_timesteps self.param_dtype = config.param_dtype self.text_encoder = T5EncoderModel( text_len=config.text_len, dtype=config.t5_dtype, device=torch.device('cpu'), checkpoint_path=text_encoder_filename, tokenizer_path=os.path.join(checkpoint_dir, config.t5_tokenizer), shard_fn= None) self.vae_stride = config.vae_stride self.patch_size = config.patch_size self.vae = WanVAE( vae_pth=os.path.join(checkpoint_dir, config.vae_checkpoint), dtype= VAE_dtype, device=self.device) logging.info(f"Creating WanModel from {model_filename[-1]}") from mmgp import offload # model_filename = "c:/temp/vace1.3/diffusion_pytorch_model.safetensors" # model_filename = "vace14B_quanto_bf16_int8.safetensors" # model_filename = "c:/temp/movii/diffusion_pytorch_model-00001-of-00007.safetensors" # config_filename= "c:/temp/movii/config.json" self.model = offload.fast_load_transformers_model(model_filename, modelClass=WanModel,do_quantize= quantizeTransformer, writable_tensors= False) # , forcedConfigPath= config_filename) # offload.load_model_data(self.model, "e:/vace.safetensors") # offload.load_model_data(self.model, "c:/temp/Phantom-Wan-1.3B.pth") # self.model.to(torch.bfloat16) # self.model.cpu() self.model.lock_layers_dtypes(torch.float32 if mixed_precision_transformer else dtype) # dtype = torch.bfloat16 offload.change_dtype(self.model, dtype, True) # offload.save_model(self.model, "wan2.1_moviigen_14B_mbf16.safetensors", config_file_path=config_filename) # offload.save_model(self.model, "wan2.1_moviigen_14B_quanto_fp16_int8.safetensors", do_quantize= True, config_file_path=config_filename) self.model.eval().requires_grad_(False) self.sample_neg_prompt = config.sample_neg_prompt if "Vace" in model_filename[-1]: self.vid_proc = VaceVideoProcessor(downsample=tuple([x * y for x, y in zip(config.vae_stride, self.patch_size)]), min_area=480*832, max_area=480*832, min_fps=config.sample_fps, max_fps=config.sample_fps, zero_start=True, seq_len=32760, keep_last=True) self.adapt_vace_model() def vace_encode_frames(self, frames, ref_images, masks=None, tile_size = 0, overlapped_latents = None): if ref_images is None: ref_images = [None] * len(frames) else: assert len(frames) == len(ref_images) if masks is None: latents = self.vae.encode(frames, tile_size = tile_size) else: inactive = [i * (1 - m) + 0 * m for i, m in zip(frames, masks)] reactive = [i * m + 0 * (1 - m) for i, m in zip(frames, masks)] inactive = self.vae.encode(inactive, tile_size = tile_size) self.toto = inactive[0].clone() if overlapped_latents != None : # inactive[0][:, 0:1] = self.vae.encode([frames[0][:, 0:1]], tile_size = tile_size)[0] # redundant inactive[0][:, 1:overlapped_latents.shape[1] + 1] = overlapped_latents reactive = self.vae.encode(reactive, tile_size = tile_size) latents = [torch.cat((u, c), dim=0) for u, c in zip(inactive, reactive)] cat_latents = [] for latent, refs in zip(latents, ref_images): if refs is not None: if masks is None: ref_latent = self.vae.encode(refs, tile_size = tile_size) else: ref_latent = self.vae.encode(refs, tile_size = tile_size) ref_latent = [torch.cat((u, torch.zeros_like(u)), dim=0) for u in ref_latent] assert all([x.shape[1] == 1 for x in ref_latent]) latent = torch.cat([*ref_latent, latent], dim=1) cat_latents.append(latent) return cat_latents def vace_encode_masks(self, masks, ref_images=None): if ref_images is None: ref_images = [None] * len(masks) else: assert len(masks) == len(ref_images) result_masks = [] for mask, refs in zip(masks, ref_images): c, depth, height, width = mask.shape new_depth = int((depth + 3) // self.vae_stride[0]) height = 2 * (int(height) // (self.vae_stride[1] * 2)) width = 2 * (int(width) // (self.vae_stride[2] * 2)) # reshape mask = mask[0, :, :, :] mask = mask.view( depth, height, self.vae_stride[1], width, self.vae_stride[1] ) # depth, height, 8, width, 8 mask = mask.permute(2, 4, 0, 1, 3) # 8, 8, depth, height, width mask = mask.reshape( self.vae_stride[1] * self.vae_stride[2], depth, height, width ) # 8*8, depth, height, width # interpolation mask = F.interpolate(mask.unsqueeze(0), size=(new_depth, height, width), mode='nearest-exact').squeeze(0) if refs is not None: length = len(refs) mask_pad = torch.zeros_like(mask[:, :length, :, :]) mask = torch.cat((mask_pad, mask), dim=1) result_masks.append(mask) return result_masks def vace_latent(self, z, m): return [torch.cat([zz, mm], dim=0) for zz, mm in zip(z, m)] def prepare_source(self, src_video, src_mask, src_ref_images, total_frames, image_size, device, original_video = False, keep_frames= [], start_frame = 0, fit_into_canvas = True, pre_src_video = None): image_sizes = [] trim_video = len(keep_frames) canvas_height, canvas_width = image_size for i, (sub_src_video, sub_src_mask, sub_pre_src_video) in enumerate(zip(src_video, src_mask,pre_src_video)): prepend_count = 0 if sub_pre_src_video == None else sub_pre_src_video.shape[1] num_frames = total_frames - prepend_count if sub_src_mask is not None and sub_src_video is not None: src_video[i], src_mask[i], _, _, _ = self.vid_proc.load_video_pair(sub_src_video, sub_src_mask, max_frames= num_frames, trim_video = trim_video - prepend_count, start_frame = start_frame, canvas_height = canvas_height, canvas_width = canvas_width, fit_into_canvas = fit_into_canvas) # src_video is [-1, 1] (at this function output), 0 = inpainting area (in fact 127 in [0, 255]) # src_mask is [-1, 1] (at this function output), 0 = preserve original video (in fact 127 in [0, 255]) and 1 = Inpainting (in fact 255 in [0, 255]) src_video[i] = src_video[i].to(device) src_mask[i] = src_mask[i].to(device) if prepend_count > 0: src_video[i] = torch.cat( [sub_pre_src_video, src_video[i]], dim=1) src_mask[i] = torch.cat( [torch.full_like(sub_pre_src_video, -1.0), src_mask[i]] ,1) src_video_shape = src_video[i].shape if src_video_shape[1] != total_frames: src_video[i] = torch.cat( [src_video[i], src_video[i].new_zeros(src_video_shape[0], total_frames -src_video_shape[1], *src_video_shape[-2:])], dim=1) src_mask[i] = torch.cat( [src_mask[i], src_mask[i].new_ones(src_video_shape[0], total_frames -src_video_shape[1], *src_video_shape[-2:])], dim=1) src_mask[i] = torch.clamp((src_mask[i][:1, :, :, :] + 1) / 2, min=0, max=1) image_sizes.append(src_video[i].shape[2:]) elif sub_src_video is None: if prepend_count > 0: src_video[i] = torch.cat( [sub_pre_src_video, torch.zeros((3, num_frames, image_size[0], image_size[1]), device=device)], dim=1) src_mask[i] = torch.cat( [torch.zeros_like(sub_pre_src_video), torch.ones((3, num_frames, image_size[0], image_size[1]), device=device)] ,1) else: src_video[i] = torch.zeros((3, num_frames, image_size[0], image_size[1]), device=device) src_mask[i] = torch.ones_like(src_video[i], device=device) image_sizes.append(image_size) else: src_video[i], _, _, _ = self.vid_proc.load_video(sub_src_video, max_frames= num_frames, trim_video = trim_video - prepend_count, start_frame = start_frame, canvas_height = canvas_height, canvas_width = canvas_width, fit_into_canvas = fit_into_canvas) src_video[i] = src_video[i].to(device) src_mask[i] = torch.zeros_like(src_video[i], device=device) if original_video else torch.ones_like(src_video[i], device=device) if prepend_count > 0: src_video[i] = torch.cat( [sub_pre_src_video, src_video[i]], dim=1) src_mask[i] = torch.cat( [torch.zeros_like(sub_pre_src_video), src_mask[i]] ,1) src_video_shape = src_video[i].shape if src_video_shape[1] != total_frames: src_video[i] = torch.cat( [src_video[i], src_video[i].new_zeros(src_video_shape[0], total_frames -src_video_shape[1], *src_video_shape[-2:])], dim=1) src_mask[i] = torch.cat( [src_mask[i], src_mask[i].new_ones(src_video_shape[0], total_frames -src_video_shape[1], *src_video_shape[-2:])], dim=1) image_sizes.append(src_video[i].shape[2:]) for k, keep in enumerate(keep_frames): if not keep: src_video[i][:, k:k+1] = 0 src_mask[i][:, k:k+1] = 1 for i, ref_images in enumerate(src_ref_images): if ref_images is not None: image_size = image_sizes[i] for j, ref_img in enumerate(ref_images): if ref_img is not None: ref_img = TF.to_tensor(ref_img).sub_(0.5).div_(0.5).unsqueeze(1) if ref_img.shape[-2:] != image_size: canvas_height, canvas_width = image_size ref_height, ref_width = ref_img.shape[-2:] white_canvas = torch.ones((3, 1, canvas_height, canvas_width), device=device) # [-1, 1] scale = min(canvas_height / ref_height, canvas_width / ref_width) new_height = int(ref_height * scale) new_width = int(ref_width * scale) resized_image = F.interpolate(ref_img.squeeze(1).unsqueeze(0), size=(new_height, new_width), mode='bilinear', align_corners=False).squeeze(0).unsqueeze(1) top = (canvas_height - new_height) // 2 left = (canvas_width - new_width) // 2 white_canvas[:, :, top:top + new_height, left:left + new_width] = resized_image ref_img = white_canvas src_ref_images[i][j] = ref_img.to(device) return src_video, src_mask, src_ref_images def decode_latent(self, zs, ref_images=None, tile_size= 0 ): if ref_images is None: ref_images = [None] * len(zs) else: assert len(zs) == len(ref_images) trimed_zs = [] for z, refs in zip(zs, ref_images): if refs is not None: z = z[:, len(refs):, :, :] trimed_zs.append(z) return self.vae.decode(trimed_zs, tile_size= tile_size) def get_vae_latents(self, ref_images, device, tile_size= 0): ref_vae_latents = [] for ref_image in ref_images: ref_image = TF.to_tensor(ref_image).sub_(0.5).div_(0.5).to(self.device) img_vae_latent = self.vae.encode([ref_image.unsqueeze(1)], tile_size= tile_size) ref_vae_latents.append(img_vae_latent[0]) return torch.cat(ref_vae_latents, dim=1) def generate(self, input_prompt, input_frames= None, input_masks = None, input_ref_images = None, input_video=None, target_camera=None, context_scale=1.0, width = 1280, height = 720, fit_into_canvas = True, frame_num=81, shift=5.0, sample_solver='unipc', sampling_steps=50, guide_scale=5.0, n_prompt="", seed=-1, offload_model=True, callback = None, enable_RIFLEx = None, VAE_tile_size = 0, joint_pass = False, slg_layers = None, slg_start = 0.0, slg_end = 1.0, cfg_star_switch = True, cfg_zero_step = 5, overlapped_latents = None, return_latent_slice = None, overlap_noise = 0, model_filename = None, **bbargs ): r""" Generates video frames from text prompt using diffusion process. Args: input_prompt (`str`): Text prompt for content generation size (tupele[`int`], *optional*, defaults to (1280,720)): Controls video resolution, (width,height). frame_num (`int`, *optional*, defaults to 81): How many frames to sample from a video. The number should be 4n+1 shift (`float`, *optional*, defaults to 5.0): Noise schedule shift parameter. Affects temporal dynamics sample_solver (`str`, *optional*, defaults to 'unipc'): Solver used to sample the video. sampling_steps (`int`, *optional*, defaults to 40): Number of diffusion sampling steps. Higher values improve quality but slow generation guide_scale (`float`, *optional*, defaults 5.0): Classifier-free guidance scale. Controls prompt adherence vs. creativity n_prompt (`str`, *optional*, defaults to ""): Negative prompt for content exclusion. If not given, use `config.sample_neg_prompt` seed (`int`, *optional*, defaults to -1): Random seed for noise generation. If -1, use random seed. offload_model (`bool`, *optional*, defaults to True): If True, offloads models to CPU during generation to save VRAM Returns: torch.Tensor: Generated video frames tensor. Dimensions: (C, N H, W) where: - C: Color channels (3 for RGB) - N: Number of frames (81) - H: Frame height (from size) - W: Frame width from size) """ # preprocess vace = "Vace" in model_filename if n_prompt == "": n_prompt = self.sample_neg_prompt seed = seed if seed >= 0 else random.randint(0, sys.maxsize) seed_g = torch.Generator(device=self.device) seed_g.manual_seed(seed) if self._interrupt: return None context = self.text_encoder([input_prompt], self.device)[0] context_null = self.text_encoder([n_prompt], self.device)[0] context = context.to(self.dtype) context_null = context_null.to(self.dtype) input_ref_images_neg = None phantom = False if target_camera != None: width = input_video.shape[2] height = input_video.shape[1] input_video = input_video.to(dtype=self.dtype , device=self.device) input_video = input_video.permute(3, 0, 1, 2).div_(127.5).sub_(1.) source_latents = self.vae.encode([input_video])[0] #.to(dtype=self.dtype, device=self.device) del input_video # Process target camera (recammaster) from wan.utils.cammmaster_tools import get_camera_embedding cam_emb = get_camera_embedding(target_camera) cam_emb = cam_emb.to(dtype=self.dtype, device=self.device) if vace : # vace context encode input_frames = [u.to(self.device) for u in input_frames] input_ref_images = [ None if u == None else [v.to(self.device) for v in u] for u in input_ref_images] input_masks = [u.to(self.device) for u in input_masks] previous_latents = None # if overlapped_latents != None: # input_ref_images = [u[-1:] for u in input_ref_images] z0 = self.vace_encode_frames(input_frames, input_ref_images, masks=input_masks, tile_size = VAE_tile_size, overlapped_latents = overlapped_latents ) m0 = self.vace_encode_masks(input_masks, input_ref_images) z = self.vace_latent(z0, m0) target_shape = list(z0[0].shape) target_shape[0] = int(target_shape[0] / 2) else: if input_ref_images != None: # Phantom Ref images phantom = True input_ref_images = self.get_vae_latents(input_ref_images, self.device) input_ref_images_neg = torch.zeros_like(input_ref_images) F = frame_num target_shape = (self.vae.model.z_dim, (F - 1) // self.vae_stride[0] + 1 + (input_ref_images.shape[1] if input_ref_images != None else 0), height // self.vae_stride[1], width // self.vae_stride[2]) seq_len = math.ceil((target_shape[2] * target_shape[3]) / (self.patch_size[1] * self.patch_size[2]) * target_shape[1]) if self._interrupt: return None noise = [ torch.randn( *target_shape, dtype=torch.float32, device=self.device, generator=seed_g) ] # evaluation mode if False: sample_scheduler = FlowMatchScheduler(num_inference_steps=sampling_steps, shift=shift, sigma_min=0, extra_one_step=True) timesteps = torch.tensor([1000, 934, 862, 756, 603, 410, 250, 140, 74, 0])[:sampling_steps].to(self.device) sample_scheduler.timesteps =timesteps elif sample_solver == 'unipc': sample_scheduler = FlowUniPCMultistepScheduler( num_train_timesteps=self.num_train_timesteps, shift=1, use_dynamic_shifting=False) sample_scheduler.set_timesteps( sampling_steps, device=self.device, shift=shift) timesteps = sample_scheduler.timesteps elif sample_solver == 'dpm++': sample_scheduler = FlowDPMSolverMultistepScheduler( num_train_timesteps=self.num_train_timesteps, shift=1, use_dynamic_shifting=False) sampling_sigmas = get_sampling_sigmas(sampling_steps, shift) timesteps, _ = retrieve_timesteps( sample_scheduler, device=self.device, sigmas=sampling_sigmas) else: raise NotImplementedError("Unsupported solver.") # sample videos latents = noise[0] del noise batch_size = 1 if target_camera != None: shape = list(latents.shape[1:]) shape[0] *= 2 freqs = get_rotary_pos_embed(shape, enable_RIFLEx= False) else: freqs = get_rotary_pos_embed(latents.shape[1:], enable_RIFLEx= enable_RIFLEx) kwargs = {'freqs': freqs, 'pipeline': self, 'callback': callback} if target_camera != None: kwargs.update({'cam_emb': cam_emb}) if vace: ref_images_count = len(input_ref_images[0]) if input_ref_images != None and input_ref_images[0] != None else 0 kwargs.update({'vace_context' : z, 'vace_context_scale' : context_scale}) if overlapped_latents != None: overlapped_latents_size = overlapped_latents.shape[1] + 1 z_reactive = [ zz[0:16, 0:overlapped_latents_size + ref_images_count].clone() for zz in z] if self.model.enable_teacache: x_count = 3 if phantom else 2 self.model.previous_residual = [None] * x_count self.model.compute_teacache_threshold(self.model.teacache_start_step, timesteps, self.model.teacache_multiplier) if callback != None: callback(-1, None, True) for i, t in enumerate(tqdm(timesteps)): if overlapped_latents != None: # overlap_noise_factor = overlap_noise *(i/(len(timesteps)-1)) / 1000 overlap_noise_factor = overlap_noise / 1000 latent_noise_factor = t / 1000 for zz, zz_r, ll in zip(z, z_reactive, [latents]): pass zz[0:16, ref_images_count:overlapped_latents_size + ref_images_count] = zz_r[:, ref_images_count:] * (1.0 - overlap_noise_factor) + torch.randn_like(zz_r[:, ref_images_count:] ) * overlap_noise_factor ll[:, 0:overlapped_latents_size + ref_images_count] = zz_r * (1.0 - latent_noise_factor) + torch.randn_like(zz_r ) * latent_noise_factor if target_camera != None: latent_model_input = torch.cat([latents, source_latents], dim=1) else: latent_model_input = latents kwargs["slg_layers"] = slg_layers if int(slg_start * sampling_steps) <= i < int(slg_end * sampling_steps) else None timestep = [t] offload.set_step_no_for_lora(self.model, i) timestep = torch.stack(timestep) kwargs["current_step"] = i kwargs["t"] = timestep if guide_scale == 1: noise_pred = self.model( [latent_model_input], x_id = 0, context = [context], **kwargs)[0] if self._interrupt: return None elif joint_pass: if phantom: pos_it, pos_i, neg = self.model( [ torch.cat([latent_model_input[:,:-input_ref_images.shape[1]], input_ref_images], dim=1) ] * 2 + [ torch.cat([latent_model_input[:,:-input_ref_images_neg.shape[1]], input_ref_images_neg], dim=1)], context = [context, context_null, context_null], **kwargs) else: noise_pred_cond, noise_pred_uncond = self.model( [latent_model_input, latent_model_input], context = [context, context_null], **kwargs) if self._interrupt: return None else: if phantom: pos_it = self.model( [ torch.cat([latent_model_input[:,:-input_ref_images.shape[1]], input_ref_images], dim=1) ], x_id = 0, context = [context], **kwargs )[0] if self._interrupt: return None pos_i = self.model( [ torch.cat([latent_model_input[:,:-input_ref_images.shape[1]], input_ref_images], dim=1) ], x_id = 1, context = [context_null],**kwargs )[0] if self._interrupt: return None neg = self.model( [ torch.cat([latent_model_input[:,:-input_ref_images_neg.shape[1]], input_ref_images_neg], dim=1) ], x_id = 2, context = [context_null], **kwargs )[0] if self._interrupt: return None else: noise_pred_cond = self.model( [latent_model_input], x_id = 0, context = [context], **kwargs)[0] if self._interrupt: return None noise_pred_uncond = self.model( [latent_model_input], x_id = 1, context = [context_null], **kwargs)[0] if self._interrupt: return None # del latent_model_input # CFG Zero *. Thanks to https://github.com/WeichenFan/CFG-Zero-star/ if guide_scale == 1: pass elif phantom: guide_scale_img= 5.0 guide_scale_text= guide_scale #7.5 noise_pred = neg + guide_scale_img * (pos_i - neg) + guide_scale_text * (pos_it - pos_i) else: noise_pred_text = noise_pred_cond if cfg_star_switch: positive_flat = noise_pred_text.view(batch_size, -1) negative_flat = noise_pred_uncond.view(batch_size, -1) alpha = optimized_scale(positive_flat,negative_flat) alpha = alpha.view(batch_size, 1, 1, 1) if (i <= cfg_zero_step): noise_pred = noise_pred_text*0. # it would be faster not to compute noise_pred... else: noise_pred_uncond *= alpha noise_pred = noise_pred_uncond + guide_scale * (noise_pred_text - noise_pred_uncond) noise_pred_uncond, noise_pred_cond, noise_pred_text, pos_it, pos_i, neg = None, None, None, None, None, None scheduler_kwargs = {} if isinstance(sample_scheduler, FlowMatchScheduler) else {"generator": seed_g} temp_x0 = sample_scheduler.step( noise_pred[:, :target_shape[1]].unsqueeze(0), t, latents.unsqueeze(0), # return_dict=False, **scheduler_kwargs)[0] latents = temp_x0.squeeze(0) del temp_x0 if callback is not None: callback(i, latents, False) x0 = [latents] if return_latent_slice != None: if overlapped_latents != None: # latents [:, 1:] = self.toto for zz, zz_r, ll in zip(z, z_reactive, [latents]): ll[:, 0:overlapped_latents_size + ref_images_count] = zz_r latent_slice = latents[:, return_latent_slice].clone() if input_frames == None: if phantom: # phantom post processing x0 = [x0_[:,:-input_ref_images.shape[1]] for x0_ in x0] videos = self.vae.decode(x0, VAE_tile_size) else: # vace post processing videos = self.decode_latent(x0, input_ref_images, VAE_tile_size) if return_latent_slice != None: return { "x" : videos[0], "latent_slice" : latent_slice } return videos[0] def adapt_vace_model(self): model = self.model modules_dict= { k: m for k, m in model.named_modules()} for model_layer, vace_layer in model.vace_layers_mapping.items(): module = modules_dict[f"vace_blocks.{vace_layer}"] target = modules_dict[f"blocks.{model_layer}"] setattr(target, "vace", module ) delattr(model, "vace_blocks")