Spaces:
Running
on
Zero
Running
on
Zero
File size: 10,398 Bytes
703e263 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 |
from ..models import SDXLTextEncoder, SDXLTextEncoder2, SDXLUNet, SDXLVAEDecoder, SDXLVAEEncoder, SDXLIpAdapter, IpAdapterXLCLIPImageEmbedder, SDXLMotionModel
from ..models.kolors_text_encoder import ChatGLMModel
from ..models.model_manager import ModelManager
from ..controlnets import MultiControlNetManager, ControlNetUnit, ControlNetConfigUnit, Annotator
from ..prompters import SDXLPrompter, KolorsPrompter
from ..schedulers import EnhancedDDIMScheduler
from .sdxl_image import SDXLImagePipeline
from .dancer import lets_dance_xl
from typing import List
import torch
from tqdm import tqdm
class SDXLVideoPipeline(SDXLImagePipeline):
def __init__(self, device="cuda", torch_dtype=torch.float16, use_original_animatediff=True):
super().__init__(device=device, torch_dtype=torch_dtype)
self.scheduler = EnhancedDDIMScheduler(beta_schedule="linear" if use_original_animatediff else "scaled_linear")
self.prompter = SDXLPrompter()
# models
self.text_encoder: SDXLTextEncoder = None
self.text_encoder_2: SDXLTextEncoder2 = None
self.text_encoder_kolors: ChatGLMModel = None
self.unet: SDXLUNet = None
self.vae_decoder: SDXLVAEDecoder = None
self.vae_encoder: SDXLVAEEncoder = None
# self.controlnet: MultiControlNetManager = None (TODO)
self.ipadapter_image_encoder: IpAdapterXLCLIPImageEmbedder = None
self.ipadapter: SDXLIpAdapter = None
self.motion_modules: SDXLMotionModel = None
def fetch_models(self, model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[], prompt_refiner_classes=[]):
# Main models
self.text_encoder = model_manager.fetch_model("sdxl_text_encoder")
self.text_encoder_2 = model_manager.fetch_model("sdxl_text_encoder_2")
self.text_encoder_kolors = model_manager.fetch_model("kolors_text_encoder")
self.unet = model_manager.fetch_model("sdxl_unet")
self.vae_decoder = model_manager.fetch_model("sdxl_vae_decoder")
self.vae_encoder = model_manager.fetch_model("sdxl_vae_encoder")
self.prompter.fetch_models(self.text_encoder)
self.prompter.load_prompt_refiners(model_manager, prompt_refiner_classes)
# ControlNets (TODO)
# IP-Adapters
self.ipadapter = model_manager.fetch_model("sdxl_ipadapter")
self.ipadapter_image_encoder = model_manager.fetch_model("sdxl_ipadapter_clip_image_encoder")
# Motion Modules
self.motion_modules = model_manager.fetch_model("sdxl_motion_modules")
if self.motion_modules is None:
self.scheduler = EnhancedDDIMScheduler(beta_schedule="scaled_linear")
# Kolors
if self.text_encoder_kolors is not None:
print("Switch to Kolors. The prompter will be replaced.")
self.prompter = KolorsPrompter()
self.prompter.fetch_models(self.text_encoder_kolors)
# The schedulers of AniamteDiff and Kolors are incompatible. We align it with AniamteDiff.
if self.motion_modules is None:
self.scheduler = EnhancedDDIMScheduler(beta_end=0.014, num_train_timesteps=1100)
else:
self.prompter.fetch_models(self.text_encoder, self.text_encoder_2)
@staticmethod
def from_model_manager(model_manager: ModelManager, controlnet_config_units: List[ControlNetConfigUnit]=[], prompt_refiner_classes=[]):
pipe = SDXLVideoPipeline(
device=model_manager.device,
torch_dtype=model_manager.torch_dtype,
)
pipe.fetch_models(model_manager, controlnet_config_units, prompt_refiner_classes)
return pipe
def decode_video(self, latents, tiled=False, tile_size=64, tile_stride=32):
images = [
self.decode_image(latents[frame_id: frame_id+1], tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
for frame_id in range(latents.shape[0])
]
return images
def encode_video(self, processed_images, tiled=False, tile_size=64, tile_stride=32):
latents = []
for image in processed_images:
image = self.preprocess_image(image).to(device=self.device, dtype=self.torch_dtype)
latent = self.encode_image(image, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
latents.append(latent.cpu())
latents = torch.concat(latents, dim=0)
return latents
@torch.no_grad()
def __call__(
self,
prompt,
negative_prompt="",
cfg_scale=7.5,
clip_skip=1,
num_frames=None,
input_frames=None,
ipadapter_images=None,
ipadapter_scale=1.0,
ipadapter_use_instant_style=False,
controlnet_frames=None,
denoising_strength=1.0,
height=512,
width=512,
num_inference_steps=20,
animatediff_batch_size = 16,
animatediff_stride = 8,
unet_batch_size = 1,
controlnet_batch_size = 1,
cross_frame_attention = False,
smoother=None,
smoother_progress_ids=[],
tiled=False,
tile_size=64,
tile_stride=32,
progress_bar_cmd=tqdm,
progress_bar_st=None,
):
# Tiler parameters, batch size ...
tiler_kwargs = {"tiled": tiled, "tile_size": tile_size, "tile_stride": tile_stride}
# Prepare scheduler
self.scheduler.set_timesteps(num_inference_steps, denoising_strength)
# Prepare latent tensors
if self.motion_modules is None:
noise = torch.randn((1, 4, height//8, width//8), device="cpu", dtype=self.torch_dtype).repeat(num_frames, 1, 1, 1)
else:
noise = torch.randn((num_frames, 4, height//8, width//8), device="cpu", dtype=self.torch_dtype)
if input_frames is None or denoising_strength == 1.0:
latents = noise
else:
latents = self.encode_video(input_frames, **tiler_kwargs)
latents = self.scheduler.add_noise(latents, noise, timestep=self.scheduler.timesteps[0])
latents = latents.to(self.device) # will be deleted for supporting long videos
# Encode prompts
prompt_emb_posi = self.encode_prompt(prompt, clip_skip=clip_skip, positive=True)
prompt_emb_nega = self.encode_prompt(negative_prompt, clip_skip=clip_skip, positive=False)
# IP-Adapter
if ipadapter_images is not None:
if ipadapter_use_instant_style:
self.ipadapter.set_less_adapter()
else:
self.ipadapter.set_full_adapter()
ipadapter_image_encoding = self.ipadapter_image_encoder(ipadapter_images)
ipadapter_kwargs_list_posi = {"ipadapter_kwargs_list": self.ipadapter(ipadapter_image_encoding, scale=ipadapter_scale)}
ipadapter_kwargs_list_nega = {"ipadapter_kwargs_list": self.ipadapter(torch.zeros_like(ipadapter_image_encoding))}
else:
ipadapter_kwargs_list_posi, ipadapter_kwargs_list_nega = {"ipadapter_kwargs_list": {}}, {"ipadapter_kwargs_list": {}}
# Prepare ControlNets
if controlnet_frames is not None:
if isinstance(controlnet_frames[0], list):
controlnet_frames_ = []
for processor_id in range(len(controlnet_frames)):
controlnet_frames_.append(
torch.stack([
self.controlnet.process_image(controlnet_frame, processor_id=processor_id).to(self.torch_dtype)
for controlnet_frame in progress_bar_cmd(controlnet_frames[processor_id])
], dim=1)
)
controlnet_frames = torch.concat(controlnet_frames_, dim=0)
else:
controlnet_frames = torch.stack([
self.controlnet.process_image(controlnet_frame).to(self.torch_dtype)
for controlnet_frame in progress_bar_cmd(controlnet_frames)
], dim=1)
controlnet_kwargs = {"controlnet_frames": controlnet_frames}
else:
controlnet_kwargs = {"controlnet_frames": None}
# Prepare extra input
extra_input = self.prepare_extra_input(latents)
# Denoise
for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
timestep = timestep.unsqueeze(0).to(self.device)
# Classifier-free guidance
noise_pred_posi = lets_dance_xl(
self.unet, motion_modules=self.motion_modules, controlnet=None,
sample=latents, timestep=timestep,
**prompt_emb_posi, **controlnet_kwargs, **ipadapter_kwargs_list_posi, **extra_input, **tiler_kwargs,
device=self.device,
)
noise_pred_nega = lets_dance_xl(
self.unet, motion_modules=self.motion_modules, controlnet=None,
sample=latents, timestep=timestep,
**prompt_emb_nega, **controlnet_kwargs, **ipadapter_kwargs_list_nega, **extra_input, **tiler_kwargs,
device=self.device,
)
noise_pred = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
# DDIM and smoother
if smoother is not None and progress_id in smoother_progress_ids:
rendered_frames = self.scheduler.step(noise_pred, timestep, latents, to_final=True)
rendered_frames = self.decode_video(rendered_frames)
rendered_frames = smoother(rendered_frames, original_frames=input_frames)
target_latents = self.encode_video(rendered_frames)
noise_pred = self.scheduler.return_to_timestep(timestep, latents, target_latents)
latents = self.scheduler.step(noise_pred, timestep, latents)
# UI
if progress_bar_st is not None:
progress_bar_st.progress(progress_id / len(self.scheduler.timesteps))
# Decode image
output_frames = self.decode_video(latents, **tiler_kwargs)
# Post-process
if smoother is not None and (num_inference_steps in smoother_progress_ids or -1 in smoother_progress_ids):
output_frames = smoother(output_frames, original_frames=input_frames)
return output_frames
|