Spaces:
Runtime error
Runtime error
File size: 12,677 Bytes
68aadbb 87e3aec 32fdae0 0a3b392 99731b2 87e3aec 32fdae0 87e3aec 68aadbb 3fc6280 39b1e18 4e5195b 39b1e18 4e5195b 39b1e18 4e5195b 39b1e18 4e5195b 3fc6280 32fdae0 4e5195b 68aadbb 1b86f61 4e5195b 87e3aec 4e5195b 68aadbb 4e5195b 87e3aec 4e5195b 68aadbb 4e5195b 68aadbb 4e5195b 68aadbb 87e3aec 4e5195b 87e3aec 4e5195b 87e3aec 4e5195b 87e3aec d7d00c9 68aadbb d7d00c9 4e5195b 4588985 4e5195b 4588985 dd28623 d7d00c9 68aadbb d7d00c9 68aadbb 4e5195b 68aadbb 1b86f61 4e5195b 68aadbb 4e5195b 68aadbb 3fc6280 4e5195b 68aadbb 4e5195b 68aadbb 4e5195b d7d00c9 4e5195b 39b1e18 d7d00c9 c06160a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 |
import gradio as gr
import torch
from diffusers import StableDiffusionPipeline, DDIMScheduler
from utils import video_to_frames, add_dict_to_yaml_file, save_video, seed_everything
# from diffusers.utils import export_to_video
from tokenflow_pnp import TokenFlow
from preprocess_utils import *
from tokenflow_utils import *
# load sd model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# model_id = "stabilityai/stable-diffusion-2-1-base"
# inv_pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16).to(device)
# inv_pipe.scheduler = DDIMScheduler.from_pretrained(model_id, subfolder="scheduler")
def randomize_seed_fn():
seed = random.randint(0, np.iinfo(np.int32).max)
return seed
def reset_do_inversion():
return True
# def get_example():
# case = [
# [
# 'examples/wolf.mp4',
# ],
# [
# 'examples/woman-running.mp4',
# ],
# ]
# return case
def prep(config):
# timesteps to save
if config["sd_version"] == '2.1':
model_key = "stabilityai/stable-diffusion-2-1-base"
elif config["sd_version"] == '2.0':
model_key = "stabilityai/stable-diffusion-2-base"
elif config["sd_version"] == '1.5' or config["sd_version"] == 'ControlNet':
model_key = "runwayml/stable-diffusion-v1-5"
elif config["sd_version"] == 'depth':
model_key = "stabilityai/stable-diffusion-2-depth"
toy_scheduler = DDIMScheduler.from_pretrained(model_key, subfolder="scheduler")
toy_scheduler.set_timesteps(config["save_steps"])
timesteps_to_save, num_inference_steps = get_timesteps(toy_scheduler, num_inference_steps=config["save_steps"],
strength=1.0,
device=device)
# seed_everything(config["seed"])
if not config["frames"]: # original non demo setting
save_path = os.path.join(config["save_dir"],
f'sd_{config["sd_version"]}',
Path(config["data_path"]).stem,
f'steps_{config["steps"]}',
f'nframes_{config["n_frames"]}')
os.makedirs(os.path.join(save_path, f'latents'), exist_ok=True)
add_dict_to_yaml_file(os.path.join(config["save_dir"], 'inversion_prompts.yaml'), Path(config["data_path"]).stem, config["inversion_prompt"])
# save inversion prompt in a txt file
with open(os.path.join(save_path, 'inversion_prompt.txt'), 'w') as f:
f.write(config["inversion_prompt"])
else:
save_path = None
model = Preprocess(device, config)
print(type(model.config["batch_size"]))
frames, latents, total_inverted_latents, rgb_reconstruction = model.extract_latents(
num_steps=model.config["steps"],
save_path=save_path,
batch_size=model.config["batch_size"],
timesteps_to_save=timesteps_to_save,
inversion_prompt=model.config["inversion_prompt"],
)
return frames, latents, total_inverted_latents, rgb_reconstruction
def preprocess_and_invert(input_video,
frames,
latents,
inverted_latents,
seed,
randomize_seed,
do_inversion,
# save_dir: str = "latents",
steps: int = 500,
batch_size: int = 8,
n_frames: int = 40,
inversion_prompt:str = '',
):
sd_version = "2.1"
height = 512
weidth: int = 512
save_steps = 50
if do_inversion or randomize_seed:
preprocess_config = {}
preprocess_config['H'] = height
preprocess_config['W'] = weidth
preprocess_config['save_dir'] = 'latents'
preprocess_config['sd_version'] = sd_version
preprocess_config['steps'] = steps
preprocess_config['batch_size'] = batch_size
preprocess_config['save_steps'] = save_steps
preprocess_config['n_frames'] = n_frames
preprocess_config['seed'] = seed
preprocess_config['inversion_prompt'] = inversion_prompt
preprocess_config['frames'] = video_to_frames(input_video)
preprocess_config['data_path'] = input_video.split(".")[0]
if randomize_seed:
seed = randomize_seed_fn()
seed_everything(seed)
frames, latents, total_inverted_latents, rgb_reconstruction = prep(preprocess_config)
frames = gr.State(value=frames)
latents = gr.State(value=latents)
inverted_latents = gr.State(value=total_inverted_latents)
do_inversion = False
return frames, latents, inverted_latents, do_inversion
def edit_with_pnp(input_video,
frames,
latents,
inverted_latents,
seed,
randomize_seed,
do_inversion,
steps,
prompt: str = "a marble sculpture of a woman running, Venus de Milo",
# negative_prompt: str = "ugly, blurry, low res, unrealistic, unaesthetic",
pnp_attn_t: float = 0.5,
pnp_f_t: float = 0.8,
batch_size: int = 8, #needs to be the same as for preprocess
n_frames: int = 40,#needs to be the same as for preprocess
n_timesteps: int = 50,
gudiance_scale: float = 7.5,
inversion_prompt: str = ""#needs to be the same as for preprocess
):
config = {}
config["sd_version"] = "2.1"
config["device"] = device
config["n_timesteps"] = n_timesteps
config["n_frames"] = n_frames
config["batch_size"] = batch_size
config["guidance_scale"] = gudiance_scale
config["prompt"] = prompt
config["negative_prompt"] = "ugly, blurry, low res, unrealistic, unaesthetic",
config["pnp_attn_t"] = pnp_attn_t
config["pnp_f_t"] = pnp_f_t
config["pnp_inversion_prompt"] = inversion_prompt
if do_inversion:
frames, latents, inverted_latents, do_inversion = preprocess_and_invert(
input_video,
frames,
latents,
inverted_latents,
seed,
randomize_seed,
do_inversion,
steps,
batch_size,
n_frames,
inversion_prompt)
do_inversion = False
if randomize_seed:
seed = randomize_seed_fn()
seed_everything(seed)
editor = TokenFlow(config=config, frames=frames.value, inverted_latents=inverted_latents.value)
edited_frames = editor.edit_video()
save_video(edited_frames, 'tokenflow_PnP_fps_30.mp4', fps=30)
# path = export_to_video(edited_frames)
return 'tokenflow_PnP_fps_30.mp4', frames, latents, inverted_latents, do_inversion
########
# demo #
########
intro = """
<h1 style="font-weight: 1400; text-align: center; margin-bottom: 7px;">
TokenFlow
</h1>
"""
with gr.Blocks(css="style.css") as demo:
gr.HTML(intro)
frames = gr.State()
inverted_latents = gr.State()
latents = gr.State()
do_inversion = gr.State(value=True)
with gr.Row():
input_video = gr.Video(label="Input Video", interactive=True, elem_id="input_video")
output_video = gr.Video(label="Edited Video", interactive=False, elem_id="output_video")
input_video.style(height=365, width=365)
output_video.style(height=365, width=365)
with gr.Row():
prompt = gr.Textbox(
label="Describe your edited video",
max_lines=1, value=""
)
# with gr.Group(visible=False) as share_btn_container:
# with gr.Group(elem_id="share-btn-container"):
# community_icon = gr.HTML(community_icon_html, visible=True)
# loading_icon = gr.HTML(loading_icon_html, visible=False)
# share_button = gr.Button("Share to community", elem_id="share-btn", visible=True)
# with gr.Row():
# inversion_progress = gr.Textbox(visible=False, label="Inversion progress")
with gr.Row():
run_button = gr.Button("Edit your video!", visible=True)
with gr.Accordion("Advanced Options", open=False):
with gr.Tabs() as tabs:
with gr.TabItem('General options', id=2):
with gr.Row():
with gr.Column(min_width=100):
seed = gr.Number(value=0, precision=0, label="Seed", interactive=True)
randomize_seed = gr.Checkbox(label='Randomize seed', value=False)
gudiance_scale = gr.Slider(label='Guidance Scale', minimum=1, maximum=30,
value=7.5, step=0.5, interactive=True)
steps = gr.Slider(label='Inversion steps', minimum=100, maximum=500,
value=500, step=1, interactive=True)
n_timesteps = gr.Slider(label='Diffusion steps', minimum=25, maximum=100,
value=50, step=1, interactive=True)
with gr.Column(min_width=100):
inversion_prompt = gr.Textbox(lines=1, label="Inversion prompt", interactive=True, placeholder="")
batch_size = gr.Slider(label='Batch size', minimum=1, maximum=10,
value=8, step=1, interactive=True)
n_frames = gr.Slider(label='Num frames', minimum=20, maximum=200,
value=40, step=1, interactive=True)
pnp_attn_t = gr.Slider(label='pnp attention threshold', minimum=0, maximum=1,
value=0.5, step=0.5, interactive=True)
pnp_f_t = gr.Slider(label='pnp feature threshold', minimum=0, maximum=1,
value=0.8, step=0.05, interactive=True)
input_video.change(
fn = reset_do_inversion,
outputs = [do_inversion],
queue = False)
input_video.upload(
fn = reset_do_inversion,
outputs = [do_inversion],
queue = False).then(fn = preprocess_and_invert,
inputs = [input_video,
frames,
latents,
inverted_latents,
seed,
randomize_seed,
do_inversion,
steps,
batch_size,
n_frames,
inversion_prompt
],
outputs = [frames,
latents,
inverted_latents,
do_inversion
])
run_button.click(fn = edit_with_pnp,
inputs = [input_video,
frames,
latents,
inverted_latents,
seed,
randomize_seed,
do_inversion,
steps,
prompt,
pnp_attn_t,
pnp_f_t,
batch_size,
n_frames,
n_timesteps,
gudiance_scale,
inversion_prompt ],
outputs = [output_video, frames, latents, inverted_latents, do_inversion]
)
# gr.Examples(
# examples=get_example(),
# label='Examples',
# inputs=[input_vid],
# outputs=[input_vid]
# )
demo.queue()
demo.launch() |