File size: 9,855 Bytes
690f890 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 |
# -*- coding: utf-8 -*-
# Copyright (c) Alibaba, Inc. and its affiliates.
import argparse
import os
import random
import time
import torch
import numpy as np
from models.ltx.ltx_vace import LTXVace
from annotators.utils import save_one_video, save_one_image, get_annotator
MAX_HEIGHT = 720
MAX_WIDTH = 1280
MAX_NUM_FRAMES = 257
def get_total_gpu_memory():
if torch.cuda.is_available():
total_memory = torch.cuda.get_device_properties(0).total_memory / (1024**3)
return total_memory
return None
def seed_everething(seed: int):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(seed)
def get_parser():
parser = argparse.ArgumentParser(
description="Load models from separate directories and run the pipeline."
)
# Directories
parser.add_argument(
"--ckpt_path",
type=str,
default='models/VACE-LTX-Video-0.9/ltx-video-2b-v0.9.safetensors',
help="Path to a safetensors file that contains all model parts.",
)
parser.add_argument(
"--text_encoder_path",
type=str,
default='models/VACE-LTX-Video-0.9',
help="Path to a safetensors file that contains all model parts.",
)
parser.add_argument(
"--src_video",
type=str,
default=None,
help="The file of the source video. Default None.")
parser.add_argument(
"--src_mask",
type=str,
default=None,
help="The file of the source mask. Default None.")
parser.add_argument(
"--src_ref_images",
type=str,
default=None,
help="The file list of the source reference images. Separated by ','. Default None.")
parser.add_argument(
"--save_dir",
type=str,
default=None,
help="Path to the folder to save output video, if None will save in results/ directory.",
)
parser.add_argument("--seed", type=int, default="42")
# Pipeline parameters
parser.add_argument(
"--num_inference_steps", type=int, default=40, help="Number of inference steps"
)
parser.add_argument(
"--num_images_per_prompt",
type=int,
default=1,
help="Number of images per prompt",
)
parser.add_argument(
"--context_scale",
type=float,
default=1.0,
help="Context scale for the pipeline",
)
parser.add_argument(
"--guidance_scale",
type=float,
default=3,
help="Guidance scale for the pipeline",
)
parser.add_argument(
"--stg_scale",
type=float,
default=1,
help="Spatiotemporal guidance scale for the pipeline. 0 to disable STG.",
)
parser.add_argument(
"--stg_rescale",
type=float,
default=0.7,
help="Spatiotemporal guidance rescaling scale for the pipeline. 1 to disable rescale.",
)
parser.add_argument(
"--stg_mode",
type=str,
default="stg_a",
help="Spatiotemporal guidance mode for the pipeline. Can be either stg_a or stg_r.",
)
parser.add_argument(
"--stg_skip_layers",
type=str,
default="19",
help="Attention layers to skip for spatiotemporal guidance. Comma separated list of integers.",
)
parser.add_argument(
"--image_cond_noise_scale",
type=float,
default=0.15,
help="Amount of noise to add to the conditioned image",
)
parser.add_argument(
"--height",
type=int,
default=512,
help="The height of the output video only if src_video is empty.",
)
parser.add_argument(
"--width",
type=int,
default=768,
help="The width of the output video only if src_video is empty.",
)
parser.add_argument(
"--num_frames",
type=int,
default=97,
help="The frames of the output video only if src_video is empty.",
)
parser.add_argument(
"--frame_rate", type=int, default=25, help="Frame rate for the output video"
)
parser.add_argument(
"--precision",
choices=["bfloat16", "mixed_precision"],
default="bfloat16",
help="Sets the precision for the transformer and tokenizer. Default is bfloat16. If 'mixed_precision' is enabled, it moves to mixed-precision.",
)
# VAE noise augmentation
parser.add_argument(
"--decode_timestep",
type=float,
default=0.05,
help="Timestep for decoding noise",
)
parser.add_argument(
"--decode_noise_scale",
type=float,
default=0.025,
help="Noise level for decoding noise",
)
# Prompts
parser.add_argument(
"--prompt",
type=str,
required=True,
help="Text prompt to guide generation",
)
parser.add_argument(
"--negative_prompt",
type=str,
default="worst quality, inconsistent motion, blurry, jittery, distorted",
help="Negative prompt for undesired features",
)
parser.add_argument(
"--offload_to_cpu",
action="store_true",
help="Offloading unnecessary computations to CPU.",
)
parser.add_argument(
"--use_prompt_extend",
default='plain',
choices=['plain', 'ltx_en', 'ltx_en_ds'],
help="Whether to use prompt extend."
)
return parser
def main(args):
args = argparse.Namespace(**args) if isinstance(args, dict) else args
print(f"Running generation with arguments: {args}")
seed_everething(args.seed)
offload_to_cpu = False if not args.offload_to_cpu else get_total_gpu_memory() < 30
assert os.path.exists(args.ckpt_path) and os.path.exists(args.text_encoder_path)
ltx_vace = LTXVace(ckpt_path=args.ckpt_path,
text_encoder_path=args.text_encoder_path,
precision=args.precision,
stg_skip_layers=args.stg_skip_layers,
stg_mode=args.stg_mode,
offload_to_cpu=offload_to_cpu)
src_ref_images = args.src_ref_images.split(',') if args.src_ref_images is not None else []
if args.use_prompt_extend and args.use_prompt_extend != 'plain':
prompt = get_annotator(config_type='prompt', config_task=args.use_prompt_extend, return_dict=False).forward(args.prompt)
print(f"Prompt extended from '{args.prompt}' to '{prompt}'")
else:
prompt = args.prompt
output = ltx_vace.generate(src_video=args.src_video,
src_mask=args.src_mask,
src_ref_images=src_ref_images,
prompt=prompt,
negative_prompt=args.negative_prompt,
seed=args.seed,
num_inference_steps=args.num_inference_steps,
num_images_per_prompt=args.num_images_per_prompt,
context_scale=args.context_scale,
guidance_scale=args.guidance_scale,
stg_scale=args.stg_scale,
stg_rescale=args.stg_rescale,
frame_rate=args.frame_rate,
image_cond_noise_scale=args.image_cond_noise_scale,
decode_timestep=args.decode_timestep,
decode_noise_scale=args.decode_noise_scale,
output_height=args.height,
output_width=args.width,
num_frames=args.num_frames)
if args.save_dir is None:
save_dir = os.path.join('results', 'vace_ltxv', time.strftime('%Y-%m-%d-%H-%M-%S', time.localtime(time.time())))
else:
save_dir = args.save_dir
if not os.path.exists(save_dir):
os.makedirs(save_dir)
frame_rate = output['info']['frame_rate']
ret_data = {}
if output['out_video'] is not None:
save_path = os.path.join(save_dir, 'out_video.mp4')
out_video = (torch.clamp(output['out_video'] / 2 + 0.5, min=0.0, max=1.0).permute(1, 2, 3, 0) * 255).cpu().numpy().astype(np.uint8)
save_one_video(save_path, out_video, fps=frame_rate)
print(f"Save out_video to {save_path}")
ret_data['out_video'] = save_path
if output['src_video'] is not None:
save_path = os.path.join(save_dir, 'src_video.mp4')
src_video = (torch.clamp(output['src_video'] / 2 + 0.5, min=0.0, max=1.0).permute(1, 2, 3, 0) * 255).cpu().numpy().astype(np.uint8)
save_one_video(save_path, src_video, fps=frame_rate)
print(f"Save src_video to {save_path}")
ret_data['src_video'] = save_path
if output['src_mask'] is not None:
save_path = os.path.join(save_dir, 'src_mask.mp4')
src_mask = (torch.clamp(output['src_mask'], min=0.0, max=1.0).permute(1, 2, 3, 0) * 255).cpu().numpy().astype(np.uint8)
save_one_video(save_path, src_mask, fps=frame_rate)
print(f"Save src_mask to {save_path}")
ret_data['src_mask'] = save_path
if output['src_ref_images'] is not None:
for i, ref_img in enumerate(output['src_ref_images']): # [C, F=1, H, W]
save_path = os.path.join(save_dir, f'src_ref_image_{i}.png')
ref_img = (torch.clamp(ref_img.squeeze(1), min=0.0, max=1.0).permute(1, 2, 0) * 255).cpu().numpy().astype(np.uint8)
save_one_image(save_path, ref_img, use_type='pil')
print(f"Save src_ref_image_{i} to {save_path}")
ret_data[f'src_ref_image_{i}'] = save_path
return ret_data
if __name__ == "__main__":
args = get_parser().parse_args()
main(args) |