Spaces:
Sleeping
Sleeping
File size: 9,433 Bytes
b1531dc |
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 |
# TODO
import numpy as np
import argparse
import torch
from torchvision.utils import make_grid
import tempfile
import gradio as gr
from omegaconf import OmegaConf
from einops import rearrange
from scripts.pub.V3D_512 import (
sample_one,
get_batch,
get_unique_embedder_keys_from_conditioner,
load_model,
)
from sgm.util import default, instantiate_from_config
from safetensors.torch import load_file as load_safetensors
from PIL import Image
from kiui.op import recenter
from torchvision.transforms import ToTensor
from einops import rearrange, repeat
import rembg
import os
from glob import glob
from mediapy import write_video
from pathlib import Path
import spaces
@spaces.GPU
def do_sample(
image,
model,
clip_model,
ae_model,
device,
num_frames,
num_steps,
decoding_t,
border_ratio,
ignore_alpha,
rembg_session,
output_folder,
):
# if image.mode == "RGBA":
# image = image.convert("RGB")
image = Image.fromarray(image)
w, h = image.size
if border_ratio > 0:
if image.mode != "RGBA" or ignore_alpha:
image = image.convert("RGB")
image = np.asarray(image)
carved_image = rembg.remove(image, session=rembg_session) # [H, W, 4]
else:
image = np.asarray(image)
carved_image = image
mask = carved_image[..., -1] > 0
image = recenter(carved_image, mask, border_ratio=border_ratio)
image = image.astype(np.float32) / 255.0
if image.shape[-1] == 4:
image = image[..., :3] * image[..., 3:4] + (1 - image[..., 3:4])
image = Image.fromarray((image * 255).astype(np.uint8))
else:
print("Ignore border ratio")
image = image.resize((512, 512))
image = ToTensor()(image)
image = image * 2.0 - 1.0
image = image.unsqueeze(0).to(device)
H, W = image.shape[2:]
assert image.shape[1] == 3
F = 8
C = 4
shape = (num_frames, C, H // F, W // F)
value_dict = {}
value_dict["motion_bucket_id"] = 0
value_dict["fps_id"] = 0
value_dict["cond_aug"] = 0.05
value_dict["cond_frames_without_noise"] = clip_model(image)
value_dict["cond_frames"] = ae_model.encode(image)
value_dict["cond_frames"] += 0.05 * torch.randn_like(value_dict["cond_frames"])
value_dict["cond_aug"] = 0.05
with torch.no_grad():
with torch.autocast(device):
batch, batch_uc = get_batch(
get_unique_embedder_keys_from_conditioner(model.conditioner),
value_dict,
[1, num_frames],
T=num_frames,
device=device,
)
c, uc = model.conditioner.get_unconditional_conditioning(
batch,
batch_uc=batch_uc,
force_uc_zero_embeddings=[
"cond_frames",
"cond_frames_without_noise",
],
)
for k in ["crossattn", "concat"]:
uc[k] = repeat(uc[k], "b ... -> b t ...", t=num_frames)
uc[k] = rearrange(uc[k], "b t ... -> (b t) ...", t=num_frames)
c[k] = repeat(c[k], "b ... -> b t ...", t=num_frames)
c[k] = rearrange(c[k], "b t ... -> (b t) ...", t=num_frames)
randn = torch.randn(shape, device=device)
randn = randn.to(device)
additional_model_inputs = {}
additional_model_inputs["image_only_indicator"] = torch.zeros(
2, num_frames
).to(device)
additional_model_inputs["num_video_frames"] = batch["num_video_frames"]
def denoiser(input, sigma, c):
return model.denoiser(
model.model, input, sigma, c, **additional_model_inputs
)
samples_z = model.sampler(denoiser, randn, cond=c, uc=uc)
model.en_and_decode_n_samples_a_time = decoding_t
samples_x = model.decode_first_stage(samples_z)
samples = torch.clamp((samples_x + 1.0) / 2.0, min=0.0, max=1.0)
os.makedirs(output_folder, exist_ok=True)
base_count = len(glob(os.path.join(output_folder, "*.mp4")))
video_path = os.path.join(output_folder, f"{base_count:06d}.mp4")
frames = (
(rearrange(samples, "t c h w -> t h w c") * 255)
.cpu()
.numpy()
.astype(np.uint8)
)
write_video(video_path, frames, fps=6)
return video_path
def change_model_params(model, min_cfg, max_cfg):
model.sampler.guider.max_scale = max_cfg
model.sampler.guider.min_scale = min_cfg
@spaces.GPU
def launch(device="cuda", share=False):
model_config = "./scripts/pub/configs/V3D_512.yaml"
num_frames = OmegaConf.load(
model_config
).model.params.sampler_config.params.guider_config.params.num_frames
print("Detected num_frames:", num_frames)
# num_steps = default(num_steps, 25)
num_steps = 25
output_folder = "outputs/V3D_512"
sd = load_safetensors("./ckpts/svd_xt.safetensors")
clip_model_config = OmegaConf.load("./configs/embedder/clip_image.yaml")
clip_model = instantiate_from_config(clip_model_config).eval()
clip_sd = dict()
for k, v in sd.items():
if "conditioner.embedders.0" in k:
clip_sd[k.replace("conditioner.embedders.0.", "")] = v
clip_model.load_state_dict(clip_sd)
clip_model = clip_model.to(device)
ae_model_config = OmegaConf.load("./configs/ae/video.yaml")
ae_model = instantiate_from_config(ae_model_config).eval()
encoder_sd = dict()
for k, v in sd.items():
if "first_stage_model" in k:
encoder_sd[k.replace("first_stage_model.", "")] = v
ae_model.load_state_dict(encoder_sd)
ae_model = ae_model.to(device)
rembg_session = rembg.new_session()
model, _ = load_model(
model_config, device, num_frames, num_steps, min_cfg=3.5, max_cfg=3.5
)
with gr.Blocks(title="V3D", theme=gr.themes.Monochrome()) as demo:
with gr.Row(equal_height=True):
with gr.Column():
input_image = gr.Image(value=None, label="Input Image")
border_ratio_slider = gr.Slider(
value=0.3,
label="Border Ratio",
minimum=0.05,
maximum=0.5,
step=0.05,
)
decoding_t_slider = gr.Slider(
value=1,
label="Number of Decoding frames",
minimum=1,
maximum=num_frames,
step=1,
)
min_guidance_slider = gr.Slider(
value=3.5,
label="Min CFG Value",
minimum=0.05,
maximum=0.5,
step=0.05,
)
max_guidance_slider = gr.Slider(
value=3.5,
label="Max CFG Value",
minimum=0.05,
maximum=0.5,
step=0.05,
)
run_button = gr.Button(value="Run V3D")
with gr.Column():
output_video = gr.Video(value=None, label="Output Orbit Video")
@run_button.click(
inputs=[
input_image,
border_ratio_slider,
min_guidance_slider,
max_guidance_slider,
decoding_t_slider,
],
outputs=[output_video],
)
def _(image, border_ratio, min_guidance, max_guidance, decoding_t):
change_model_params(model, min_guidance, max_guidance)
return do_sample(
image,
model,
clip_model,
ae_model,
device,
num_frames,
num_steps,
int(decoding_t),
border_ratio,
False,
rembg_session,
output_folder,
)
# do_sample(
# np.asarray(Image.open("assets/baby_yoda.png")),
# model,
# clip_model,
# ae_model,
# device,
# num_frames,
# num_steps,
# 1,
# 0.3,
# False,
# rembg_session,
# output_folder,
# )
demo.launch(inbrowser=True, inline=False, share=share, show_error=True)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--device", type=str, default="cuda")
parser.add_argument("--share", action="store_true")
opt = parser.parse_args()
test = OmegaConf.load("./scripts/pub/configs/V3D_512.yaml")
print(test)
def download_if_need(path, url):
if Path(path).exists():
return
import wget
path = Path(path)
path.parent.mkdir(parents=True, exist_ok=True)
wget.download(url, out=str(path))
# download_if_need(
# "ckpts/svd_xt.safetensors",
# "https://huggingface.co/stabilityai/stable-video-diffusion-img2vid-xt/resolve/main/svd_xt.safetensors",
# )
# download_if_need(
# "ckpts/V3D_512.ckpt", "https://huggingface.co/heheyas/V3D/resolve/main/V3D.ckpt"
# )
launch(opt.device, opt.share)
|