Spaces:
Running
Running
File size: 14,970 Bytes
94bafa8 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 |
import gradio as gr
import os
import torch
import argparse
import torchvision
from diffusers.schedulers import (DDIMScheduler, DDPMScheduler, PNDMScheduler,
EulerDiscreteScheduler, DPMSolverMultistepScheduler,
HeunDiscreteScheduler, EulerAncestralDiscreteScheduler,
DEISMultistepScheduler, KDPM2AncestralDiscreteScheduler)
from diffusers.schedulers.scheduling_dpmsolver_singlestep import DPMSolverSinglestepScheduler
from diffusers.models import AutoencoderKL, AutoencoderKLTemporalDecoder
from omegaconf import OmegaConf
from transformers import T5EncoderModel, T5Tokenizer
import os, sys
sys.path.append(os.path.split(sys.path[0])[0])
from sample.pipeline_latte import LattePipeline
from models import get_models
# import imageio
from torchvision.utils import save_image
import spaces
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, default="./configs/t2x/t2v_sample.yaml")
args = parser.parse_args()
args = OmegaConf.load(args.config)
torch.set_grad_enabled(False)
device = "cuda" if torch.cuda.is_available() else "cpu"
transformer_model = get_models(args).to(device, dtype=torch.float16)
# state_dict = find_model(args.ckpt)
# msg, unexp = transformer_model.load_state_dict(state_dict, strict=False)
if args.enable_vae_temporal_decoder:
vae = AutoencoderKLTemporalDecoder.from_pretrained(args.pretrained_model_path, subfolder="vae_temporal_decoder", torch_dtype=torch.float16).to(device)
else:
vae = AutoencoderKL.from_pretrained(args.pretrained_model_path, subfolder="vae", torch_dtype=torch.float16).to(device)
tokenizer = T5Tokenizer.from_pretrained(args.pretrained_model_path, subfolder="tokenizer")
text_encoder = T5EncoderModel.from_pretrained(args.pretrained_model_path, subfolder="text_encoder", torch_dtype=torch.float16).to(device)
# set eval mode
transformer_model.eval()
vae.eval()
text_encoder.eval()
@spaces.GPU
def gen_video(text_input, sample_method, scfg_scale, seed, height, width, video_length, diffusion_step):
torch.manual_seed(seed)
if sample_method == 'DDIM':
scheduler = DDIMScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type,
clip_sample=False)
elif sample_method == 'EulerDiscrete':
scheduler = EulerDiscreteScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type)
elif sample_method == 'DDPM':
scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type,
clip_sample=False)
elif sample_method == 'DPMSolverMultistep':
scheduler = DPMSolverMultistepScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type)
elif sample_method == 'DPMSolverSinglestep':
scheduler = DPMSolverSinglestepScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type)
elif sample_method == 'PNDM':
scheduler = PNDMScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type)
elif sample_method == 'HeunDiscrete':
scheduler = HeunDiscreteScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type)
elif sample_method == 'EulerAncestralDiscrete':
scheduler = EulerAncestralDiscreteScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type)
elif sample_method == 'DEISMultistep':
scheduler = DEISMultistepScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type)
elif sample_method == 'KDPM2AncestralDiscrete':
scheduler = KDPM2AncestralDiscreteScheduler.from_pretrained(args.pretrained_model_path,
subfolder="scheduler",
beta_start=args.beta_start,
beta_end=args.beta_end,
beta_schedule=args.beta_schedule,
variance_type=args.variance_type)
videogen_pipeline = LattePipeline(vae=vae,
text_encoder=text_encoder,
tokenizer=tokenizer,
scheduler=scheduler,
transformer=transformer_model).to(device)
# videogen_pipeline.enable_xformers_memory_efficient_attention()
videos = videogen_pipeline(text_input,
video_length=video_length,
height=height,
width=width,
num_inference_steps=diffusion_step,
guidance_scale=scfg_scale,
enable_temporal_attentions=args.enable_temporal_attentions,
num_images_per_prompt=1,
mask_feature=True,
enable_vae_temporal_decoder=args.enable_vae_temporal_decoder
).video
save_path = args.save_img_path + 'temp' + '.mp4'
torchvision.io.write_video(save_path, videos[0], fps=8)
return save_path
if not os.path.exists(args.save_img_path):
os.makedirs(args.save_img_path)
intro = """
<div style="display: flex;align-items: center;justify-content: center">
<h1 style="display: inline-block;margin-left: 10px;margin-top: 6px;font-weight: 500">Latte: Latent Diffusion Transformer for Video Generation</h1>
</div>
"""
with gr.Blocks() as demo:
# gr.HTML(intro)
# with gr.Accordion("README", open=False):
# gr.HTML(
# """
# <p style="font-size: 0.95rem;margin: 0rem;line-height: 1.2em;margin-top:1em;display: inline-block">
# <a href="https://maxin-cn.github.io/latte_project/" target="_blank">project page</a> | <a href="https://arxiv.org/abs/2401.03048" target="_blank">paper</a>
# </p>
# We will continue update Latte.
# """
# )
gr.Markdown("<font color=red size=10><center>Latte: Latent Diffusion Transformer for Video Generation</center></font>")
gr.Markdown(
"""<div style="display: flex;align-items: center;justify-content: center">
<h2 style="display: inline-block;margin-left: 10px;margin-top: 6px;font-weight: 500">Latte supports both T2I and T2V, and will be continuously updated, so stay tuned!</h2></div>
"""
)
gr.Markdown(
"""<div style="display: flex;align-items: center;justify-content: center">
[<a href="https://arxiv.org/abs/2401.03048">Arxiv Report</a>] | [<a href="https://maxin-cn.github.io/latte_project/">Project Page</a>] | [<a href="https://github.com/Vchitect/Latte">Github</a>]</div>
"""
)
with gr.Row():
with gr.Column(visible=True) as input_raws:
with gr.Row():
with gr.Column(scale=1.0):
# text_input = gr.Textbox(show_label=True, interactive=True, label="Text prompt").style(container=False)
text_input = gr.Textbox(show_label=True, interactive=True, label="Prompt")
# with gr.Row():
# with gr.Column(scale=0.5):
# image_input = gr.Image(show_label=True, interactive=True, label="Reference image").style(container=False)
# with gr.Column(scale=0.5):
# preframe_input = gr.Image(show_label=True, interactive=True, label="First frame").style(container=False)
with gr.Row():
with gr.Column(scale=0.5):
sample_method = gr.Dropdown(choices=["DDIM", "EulerDiscrete", "PNDM"], label="Sample Method", value="DDIM")
# with gr.Row():
# with gr.Column(scale=1.0):
# video_length = gr.Slider(
# minimum=1,
# maximum=24,
# value=1,
# step=1,
# interactive=True,
# label="Video Length (1 for T2I and 16 for T2V)",
# )
with gr.Column(scale=0.5):
video_length = gr.Dropdown(choices=[1, 16], label="Video Length (1 for T2I and 16 for T2V)", value=16)
with gr.Row():
with gr.Column(scale=1.0):
scfg_scale = gr.Slider(
minimum=1,
maximum=50,
value=7.5,
step=0.1,
interactive=True,
label="Guidence Scale",
)
with gr.Row():
with gr.Column(scale=1.0):
seed = gr.Slider(
minimum=1,
maximum=2147483647,
value=100,
step=1,
interactive=True,
label="Seed",
)
with gr.Row():
with gr.Column(scale=0.5):
height = gr.Slider(
minimum=256,
maximum=768,
value=512,
step=16,
interactive=False,
label="Height",
)
# with gr.Row():
with gr.Column(scale=0.5):
width = gr.Slider(
minimum=256,
maximum=768,
value=512,
step=16,
interactive=False,
label="Width",
)
with gr.Row():
with gr.Column(scale=1.0):
diffusion_step = gr.Slider(
minimum=20,
maximum=250,
value=50,
step=1,
interactive=True,
label="Sampling Step",
)
with gr.Column(scale=0.6, visible=True) as video_upload:
# with gr.Column(visible=True) as video_upload:
output = gr.Video(interactive=False, include_audio=True, elem_id="输出的视频") #.style(height=360)
# with gr.Column(elem_id="image", scale=0.5) as img_part:
# with gr.Tab("Video", elem_id='video_tab'):
# with gr.Tab("Image", elem_id='image_tab'):
# up_image = gr.Image(type="pil", interactive=True, elem_id="image_upload").style(height=360)
# upload_button = gr.Button(value="Upload & Start Chat", interactive=True, variant="primary")
# clear = gr.Button("Restart")
with gr.Row():
with gr.Column(scale=1.0, min_width=0):
run = gr.Button("💭Run")
# with gr.Column(scale=0.5, min_width=0):
# clear = gr.Button("🔄Clear️")
run.click(gen_video, [text_input, sample_method, scfg_scale, seed, height, width, video_length, diffusion_step], [output])
demo.launch(debug=False, share=True)
# demo.launch(server_name="0.0.0.0", server_port=10034, enable_queue=True)
|