Spaces:
Running
on
Zero
Running
on
Zero
app
Browse files- .gitignore +0 -2
- app1_a.py +0 -386
- app1_bf.py +0 -388
- app1_bf2.py +0 -388
- app_bf.py +0 -391
.gitignore
CHANGED
@@ -1,5 +1,3 @@
|
|
1 |
-
app1.py
|
2 |
-
app2.py
|
3 |
demo_utils1.py
|
4 |
tmp
|
5 |
models
|
|
|
|
|
|
|
1 |
demo_utils1.py
|
2 |
tmp
|
3 |
models
|
app1_a.py
DELETED
@@ -1,386 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
import gradio as gr
|
3 |
-
import numpy as np
|
4 |
-
from enum import Enum
|
5 |
-
import db_examples
|
6 |
-
import cv2
|
7 |
-
|
8 |
-
|
9 |
-
from demo_utils1 import *
|
10 |
-
|
11 |
-
from misc_utils.train_utils import unit_test_create_model
|
12 |
-
from misc_utils.image_utils import save_tensor_to_gif, save_tensor_to_images
|
13 |
-
import os
|
14 |
-
from PIL import Image
|
15 |
-
import torch
|
16 |
-
import torchvision
|
17 |
-
from torchvision import transforms
|
18 |
-
from einops import rearrange
|
19 |
-
import imageio
|
20 |
-
import time
|
21 |
-
|
22 |
-
from torchvision.transforms import functional as F
|
23 |
-
from torch.hub import download_url_to_file
|
24 |
-
|
25 |
-
import os
|
26 |
-
|
27 |
-
# 推理设置
|
28 |
-
from pl_trainer.inference.inference import InferenceIP2PVideo
|
29 |
-
from tqdm import tqdm
|
30 |
-
|
31 |
-
|
32 |
-
# if not os.path.exists(filename):
|
33 |
-
# original_path = os.getcwd()
|
34 |
-
# base_path = './models'
|
35 |
-
# os.makedirs(base_path, exist_ok=True)
|
36 |
-
|
37 |
-
# # 直接在代码中写入 Token(注意安全风险)
|
38 |
-
# GIT_TOKEN = "955b8ea91095840b76fe38b90a088c200d4c813c"
|
39 |
-
# repo_url = f"https://YeFang:{GIT_TOKEN}@code.openxlab.org.cn/YeFang/RIV_models.git"
|
40 |
-
|
41 |
-
# try:
|
42 |
-
# if os.system(f'git clone {repo_url} {base_path}') != 0:
|
43 |
-
# raise RuntimeError("Git 克隆失败")
|
44 |
-
# os.chdir(base_path)
|
45 |
-
# if os.system('git lfs pull') != 0:
|
46 |
-
# raise RuntimeError("Git LFS 拉取失败")
|
47 |
-
# finally:
|
48 |
-
# os.chdir(original_path)
|
49 |
-
|
50 |
-
def tensor_to_pil_image(x):
|
51 |
-
"""
|
52 |
-
将 4D PyTorch 张量转换为 PIL 图像。
|
53 |
-
"""
|
54 |
-
x = x.float() # 确保张量类型为 float
|
55 |
-
grid_img = torchvision.utils.make_grid(x, nrow=4).permute(1, 2, 0).detach().cpu().numpy()
|
56 |
-
grid_img = (grid_img * 255).clip(0, 255).astype("uint8") # 将 [0, 1] 范围转换为 [0, 255]
|
57 |
-
return Image.fromarray(grid_img)
|
58 |
-
|
59 |
-
def frame_to_batch(x):
|
60 |
-
"""
|
61 |
-
将帧维度转换为批次维度。
|
62 |
-
"""
|
63 |
-
return rearrange(x, 'b f c h w -> (b f) c h w')
|
64 |
-
|
65 |
-
def clip_image(x, min=0., max=1.):
|
66 |
-
"""
|
67 |
-
将图像张量裁剪到指定的最小和最大值。
|
68 |
-
"""
|
69 |
-
return torch.clamp(x, min=min, max=max)
|
70 |
-
|
71 |
-
def unnormalize(x):
|
72 |
-
"""
|
73 |
-
将张量范围从 [-1, 1] 转换到 [0, 1]。
|
74 |
-
"""
|
75 |
-
return (x + 1) / 2
|
76 |
-
|
77 |
-
|
78 |
-
# 读取图像文件
|
79 |
-
def read_images_from_directory(directory, num_frames=16):
|
80 |
-
images = []
|
81 |
-
for i in range(num_frames):
|
82 |
-
img_path = os.path.join(directory, f'{i:04d}.png')
|
83 |
-
img = imageio.imread(img_path)
|
84 |
-
images.append(torch.tensor(img).permute(2, 0, 1)) # Convert to Tensor (C, H, W)
|
85 |
-
return images
|
86 |
-
|
87 |
-
def load_and_process_images(folder_path):
|
88 |
-
"""
|
89 |
-
读取文件夹中的所有图片,将它们转换为 [-1, 1] 范围的张量并返回一个 4D 张量。
|
90 |
-
"""
|
91 |
-
processed_images = []
|
92 |
-
transform = transforms.Compose([
|
93 |
-
transforms.ToTensor(),
|
94 |
-
transforms.Lambda(lambda x: x * 2 - 1) # 将 [0, 1] 转换为 [-1, 1]
|
95 |
-
])
|
96 |
-
for filename in sorted(os.listdir(folder_path)):
|
97 |
-
if filename.endswith(".png"):
|
98 |
-
img_path = os.path.join(folder_path, filename)
|
99 |
-
image = Image.open(img_path).convert("RGB")
|
100 |
-
processed_image = transform(image)
|
101 |
-
processed_images.append(processed_image)
|
102 |
-
return torch.stack(processed_images) # 返回 4D 张量
|
103 |
-
|
104 |
-
def load_and_process_video(video_path, num_frames=16, crop_size=512):
|
105 |
-
"""
|
106 |
-
读取视频文件中的前 num_frames 帧,将每一帧转换为 [-1, 1] 范围的张量,
|
107 |
-
并进行中心裁剪至 crop_size x crop_size,返回一个 4D 张量。
|
108 |
-
"""
|
109 |
-
processed_frames = []
|
110 |
-
transform = transforms.Compose([
|
111 |
-
transforms.CenterCrop(crop_size), # 中心裁剪
|
112 |
-
transforms.ToTensor(),
|
113 |
-
transforms.Lambda(lambda x: x * 2 - 1) # 将 [0, 1] 转换为 [-1, 1]
|
114 |
-
])
|
115 |
-
|
116 |
-
# 使用 OpenCV 读取视频
|
117 |
-
cap = cv2.VideoCapture(video_path)
|
118 |
-
|
119 |
-
if not cap.isOpened():
|
120 |
-
raise ValueError(f"无法打开视频文件: {video_path}")
|
121 |
-
|
122 |
-
frame_count = 0
|
123 |
-
|
124 |
-
while frame_count < num_frames:
|
125 |
-
ret, frame = cap.read()
|
126 |
-
if not ret:
|
127 |
-
break # 视频帧读取完毕或视频帧不足
|
128 |
-
|
129 |
-
# 转换为 RGB 格式
|
130 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
131 |
-
image = Image.fromarray(frame)
|
132 |
-
|
133 |
-
# 应用转换
|
134 |
-
processed_frame = transform(image)
|
135 |
-
processed_frames.append(processed_frame)
|
136 |
-
|
137 |
-
frame_count += 1
|
138 |
-
|
139 |
-
cap.release() # 释放视频资源
|
140 |
-
|
141 |
-
if len(processed_frames) < num_frames:
|
142 |
-
raise ValueError(f"视频帧不足 {num_frames} 帧,仅找到 {len(processed_frames)} 帧。")
|
143 |
-
|
144 |
-
return torch.stack(processed_frames) # 返回 4D 张量 (帧数, 通道数, 高度, 宽度)
|
145 |
-
|
146 |
-
|
147 |
-
def clear_cache(output_path):
|
148 |
-
if os.path.exists(output_path):
|
149 |
-
os.remove(output_path)
|
150 |
-
return None
|
151 |
-
|
152 |
-
|
153 |
-
#! 加载模型
|
154 |
-
# 配置路径和加载模型
|
155 |
-
config_path = 'configs/instruct_v2v_ic_gradio.yaml'
|
156 |
-
diffusion_model = unit_test_create_model(config_path)
|
157 |
-
diffusion_model = diffusion_model.to('cuda')
|
158 |
-
|
159 |
-
# 加载模型检查点
|
160 |
-
# ckpt_path = 'models/relvid_mm_sd15_fbc_unet.pth' #! change
|
161 |
-
# ckpt_path = 'tmp/pytorch_model.bin'
|
162 |
-
# 下载文件
|
163 |
-
|
164 |
-
os.makedirs('models', exist_ok=True)
|
165 |
-
model_path = "models/relvid_mm_sd15_fbc_unet.pth"
|
166 |
-
|
167 |
-
if not os.path.exists(model_path):
|
168 |
-
download_url_to_file(url='https://huggingface.co/aleafy/RelightVid/resolve/main/relvid_mm_sd15_fbc_unet.pth', dst=model_path)
|
169 |
-
|
170 |
-
|
171 |
-
ckpt = torch.load(model_path, map_location='cpu')
|
172 |
-
diffusion_model.load_state_dict(ckpt, strict=False)
|
173 |
-
|
174 |
-
|
175 |
-
# import pdb; pdb.set_trace()
|
176 |
-
|
177 |
-
# 更改全局临时目录
|
178 |
-
new_tmp_dir = "./demo/gradio_bg"
|
179 |
-
os.makedirs(new_tmp_dir, exist_ok=True)
|
180 |
-
|
181 |
-
# import pdb; pdb.set_trace()
|
182 |
-
|
183 |
-
def save_video_from_frames(image_pred, save_pth, fps=8):
|
184 |
-
"""
|
185 |
-
将 image_pred 中的帧保存为视频文件。
|
186 |
-
|
187 |
-
参数:
|
188 |
-
- image_pred: Tensor,形状为 (1, 16, 3, 512, 512)
|
189 |
-
- save_pth: 保存视频的路径,例如 "output_video.mp4"
|
190 |
-
- fps: 视频的帧率
|
191 |
-
"""
|
192 |
-
# 视频参数
|
193 |
-
num_frames = image_pred.shape[1]
|
194 |
-
frame_height, frame_width = 512, 512 # 目标尺寸
|
195 |
-
fourcc = cv2.VideoWriter_fourcc(*'mp4v') # 使用 mp4 编码格式
|
196 |
-
|
197 |
-
# 创建 VideoWriter 对象
|
198 |
-
out = cv2.VideoWriter(save_pth, fourcc, fps, (frame_width, frame_height))
|
199 |
-
|
200 |
-
for i in range(num_frames):
|
201 |
-
# 反归一化 + 转换为 0-255 范围
|
202 |
-
pred_frame = clip_image(unnormalize(image_pred[0][i].unsqueeze(0))) * 255
|
203 |
-
pred_frame_resized = pred_frame.squeeze(0).detach().cpu() # (3, 512, 512)
|
204 |
-
pred_frame_resized = pred_frame_resized.permute(1, 2, 0).numpy().astype("uint8") # (512, 512, 3)
|
205 |
-
|
206 |
-
# Resize 到 256x256
|
207 |
-
pred_frame_resized = cv2.resize(pred_frame_resized, (frame_width, frame_height))
|
208 |
-
|
209 |
-
# 将 RGB 转为 BGR(因为 OpenCV 使用 BGR 格式)
|
210 |
-
pred_frame_bgr = cv2.cvtColor(pred_frame_resized, cv2.COLOR_RGB2BGR)
|
211 |
-
|
212 |
-
# 写入帧到视频
|
213 |
-
out.write(pred_frame_bgr)
|
214 |
-
|
215 |
-
# 释放 VideoWriter 资源
|
216 |
-
out.release()
|
217 |
-
print(f"视频已保存至 {save_pth}")
|
218 |
-
|
219 |
-
|
220 |
-
inf_pipe = InferenceIP2PVideo(
|
221 |
-
diffusion_model.unet,
|
222 |
-
scheduler='ddpm',
|
223 |
-
num_ddim_steps=20
|
224 |
-
)
|
225 |
-
|
226 |
-
# 伪函数占位(生成空白视频)
|
227 |
-
def dummy_process(input_fg, input_bg):
|
228 |
-
# import pdb; pdb.set_trace()
|
229 |
-
|
230 |
-
diffusion_model.to(torch.float16)
|
231 |
-
fg_tensor = load_and_process_video(input_fg).cuda().unsqueeze(0).to(dtype=torch.float16)
|
232 |
-
bg_tensor = load_and_process_video(input_bg).cuda().unsqueeze(0).to(dtype=torch.float16) # (1, 16, 4, 64, 64)
|
233 |
-
|
234 |
-
cond_fg_tensor = diffusion_model.encode_image_to_latent(fg_tensor) # (1, 16, 4, 64, 64)
|
235 |
-
cond_bg_tensor = diffusion_model.encode_image_to_latent(bg_tensor)
|
236 |
-
cond_tensor = torch.cat((cond_fg_tensor, cond_bg_tensor), dim=2)
|
237 |
-
|
238 |
-
# 初始化潜变量
|
239 |
-
init_latent = torch.randn_like(cond_fg_tensor)
|
240 |
-
|
241 |
-
EDIT_PROMPT = 'change the background'
|
242 |
-
VIDEO_CFG = 1.2
|
243 |
-
TEXT_CFG = 7.5
|
244 |
-
text_cond = diffusion_model.encode_text([EDIT_PROMPT]) # (1, 77, 768)
|
245 |
-
text_uncond = diffusion_model.encode_text([''])
|
246 |
-
# to float16
|
247 |
-
print('------------to float 16----------------')
|
248 |
-
init_latent, text_cond, text_uncond, cond_tensor = (
|
249 |
-
init_latent.to(dtype=torch.float16),
|
250 |
-
text_cond.to(dtype=torch.float16),
|
251 |
-
text_uncond.to(dtype=torch.float16),
|
252 |
-
cond_tensor.to(dtype=torch.float16)
|
253 |
-
)
|
254 |
-
inf_pipe.unet.to(torch.float16)
|
255 |
-
latent_pred = inf_pipe(
|
256 |
-
latent=init_latent,
|
257 |
-
text_cond=text_cond,
|
258 |
-
text_uncond=text_uncond,
|
259 |
-
img_cond=cond_tensor,
|
260 |
-
text_cfg=TEXT_CFG,
|
261 |
-
img_cfg=VIDEO_CFG,
|
262 |
-
)['latent']
|
263 |
-
|
264 |
-
|
265 |
-
image_pred = diffusion_model.decode_latent_to_image(latent_pred) # (1,16,3,512,512)
|
266 |
-
output_path = os.path.join(new_tmp_dir, f"output_{int(time.time())}.mp4")
|
267 |
-
# clear_cache(output_path)
|
268 |
-
|
269 |
-
save_video_from_frames(image_pred, output_path)
|
270 |
-
# import pdb; pdb.set_trace()
|
271 |
-
# fps = 8
|
272 |
-
# frames = []
|
273 |
-
# for i in range(16):
|
274 |
-
# pred_frame = clip_image(unnormalize(image_pred[0][i].unsqueeze(0))) * 255
|
275 |
-
# pred_frame_resized = pred_frame.squeeze(0).detach().cpu() #(3,512,512)
|
276 |
-
# pred_frame_resized = pred_frame_resized.permute(1, 2, 0).detach().cpu().numpy().astype("uint8") #(512,512,3) np
|
277 |
-
# Image.fromarray(pred_frame_resized).save(save_pth)
|
278 |
-
|
279 |
-
# # 生成一个简单的黑色视频作为示例
|
280 |
-
# output_path = os.path.join(new_tmp_dir, "output.mp4")
|
281 |
-
# fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
282 |
-
# out = cv2.VideoWriter(output_path, fourcc, 20.0, (512, 512))
|
283 |
-
|
284 |
-
# for _ in range(60): # 生成 3 秒的视频(20fps)
|
285 |
-
# frame = np.zeros((512, 512, 3), dtype=np.uint8)
|
286 |
-
# out.write(frame)
|
287 |
-
# out.release()
|
288 |
-
torch.cuda.empty_cache()
|
289 |
-
|
290 |
-
return output_path
|
291 |
-
|
292 |
-
# 枚举类用于背景选择
|
293 |
-
class BGSource(Enum):
|
294 |
-
UPLOAD = "Use Background Video"
|
295 |
-
UPLOAD_FLIP = "Use Flipped Background Video"
|
296 |
-
UPLOAD_REVERSE = "Use Reversed Background Video"
|
297 |
-
|
298 |
-
|
299 |
-
# Quick prompts 示例
|
300 |
-
quick_prompts = [
|
301 |
-
'beautiful woman',
|
302 |
-
'handsome man',
|
303 |
-
'beautiful woman, cinematic lighting',
|
304 |
-
'handsome man, cinematic lighting',
|
305 |
-
'beautiful woman, natural lighting',
|
306 |
-
'handsome man, natural lighting',
|
307 |
-
'beautiful woman, neo punk lighting, cyberpunk',
|
308 |
-
'handsome man, neo punk lighting, cyberpunk',
|
309 |
-
]
|
310 |
-
quick_prompts = [[x] for x in quick_prompts]
|
311 |
-
|
312 |
-
# Gradio UI 结构
|
313 |
-
block = gr.Blocks().queue()
|
314 |
-
with block:
|
315 |
-
with gr.Row():
|
316 |
-
gr.Markdown("## IC-Light (Relighting with Foreground and Background Video Condition)")
|
317 |
-
|
318 |
-
with gr.Row():
|
319 |
-
with gr.Column():
|
320 |
-
with gr.Row():
|
321 |
-
input_fg = gr.Video(label="Foreground Video", height=370, width=370, visible=True)
|
322 |
-
input_bg = gr.Video(label="Background Video", height=370, width=370, visible=True)
|
323 |
-
|
324 |
-
prompt = gr.Textbox(label="Prompt")
|
325 |
-
bg_source = gr.Radio(choices=[e.value for e in BGSource],
|
326 |
-
value=BGSource.UPLOAD.value,
|
327 |
-
label="Background Source", type='value')
|
328 |
-
|
329 |
-
example_prompts = gr.Dataset(samples=quick_prompts, label='Prompt Quick List', components=[prompt])
|
330 |
-
bg_gallery = gr.Gallery(height=450, object_fit='contain', label='Background Quick List', value=db_examples.bg_samples, columns=5, allow_preview=False)
|
331 |
-
relight_button = gr.Button(value="Relight")
|
332 |
-
|
333 |
-
with gr.Group():
|
334 |
-
with gr.Row():
|
335 |
-
num_samples = gr.Slider(label="Videos", minimum=1, maximum=12, value=1, step=1)
|
336 |
-
seed = gr.Number(label="Seed", value=12345, precision=0)
|
337 |
-
with gr.Row():
|
338 |
-
video_width = gr.Slider(label="Video Width", minimum=256, maximum=1024, value=512, step=64)
|
339 |
-
video_height = gr.Slider(label="Video Height", minimum=256, maximum=1024, value=640, step=64)
|
340 |
-
|
341 |
-
with gr.Accordion("Advanced options", open=False):
|
342 |
-
steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=20, step=1)
|
343 |
-
cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=7.0, step=0.01)
|
344 |
-
highres_scale = gr.Slider(label="Highres Scale", minimum=1.0, maximum=3.0, value=1.5, step=0.01)
|
345 |
-
highres_denoise = gr.Slider(label="Highres Denoise", minimum=0.1, maximum=0.9, value=0.5, step=0.01)
|
346 |
-
a_prompt = gr.Textbox(label="Added Prompt", value='best quality')
|
347 |
-
n_prompt = gr.Textbox(label="Negative Prompt", value='lowres, bad anatomy, bad hands, cropped, worst quality')
|
348 |
-
normal_button = gr.Button(value="Compute Normal (4x Slower)")
|
349 |
-
|
350 |
-
with gr.Column():
|
351 |
-
result_video = gr.Video(label='Output Video', height=600, width=600, visible=True)
|
352 |
-
fg_gallery = gr.Gallery(width=600, object_fit='contain', label='Foreground Quick List', value=db_examples.bg_samples, columns=4, allow_preview=False)
|
353 |
-
|
354 |
-
# 输入列表
|
355 |
-
# ips = [input_fg, input_bg, prompt, video_width, video_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source]
|
356 |
-
ips = [input_fg, input_bg]
|
357 |
-
|
358 |
-
# 按钮绑定处理函数
|
359 |
-
# relight_button.click(fn=lambda: None, inputs=[], outputs=[result_video])
|
360 |
-
|
361 |
-
relight_button.click(fn=dummy_process, inputs=ips, outputs=[result_video])
|
362 |
-
|
363 |
-
normal_button.click(fn=dummy_process, inputs=ips, outputs=[result_video])
|
364 |
-
|
365 |
-
# 背景库选择
|
366 |
-
def bg_gallery_selected(gal, evt: gr.SelectData):
|
367 |
-
# import pdb; pdb.set_trace()
|
368 |
-
# img_path = gal[evt.index][0]
|
369 |
-
img_path = db_examples.bg_samples[evt.index]
|
370 |
-
video_path = img_path.replace('frames/0000.png', 'cropped_video.mp4')
|
371 |
-
return video_path
|
372 |
-
|
373 |
-
bg_gallery.select(bg_gallery_selected, inputs=bg_gallery, outputs=input_bg)
|
374 |
-
|
375 |
-
# 示例
|
376 |
-
# dummy_video_for_outputs = gr.Video(visible=False, label='Result')
|
377 |
-
gr.Examples(
|
378 |
-
fn=lambda *args: args[-1],
|
379 |
-
examples=db_examples.background_conditioned_examples,
|
380 |
-
inputs=[input_fg, input_bg, prompt, bg_source, video_width, video_height, seed, result_video],
|
381 |
-
outputs=[result_video],
|
382 |
-
run_on_click=True, examples_per_page=1024
|
383 |
-
)
|
384 |
-
|
385 |
-
# 启动 Gradio 应用
|
386 |
-
block.launch(server_name='0.0.0.0', server_port=10003, share=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app1_bf.py
DELETED
@@ -1,388 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
import gradio as gr
|
3 |
-
import numpy as np
|
4 |
-
from enum import Enum
|
5 |
-
import db_examples
|
6 |
-
import cv2
|
7 |
-
|
8 |
-
|
9 |
-
from demo_utils1 import *
|
10 |
-
|
11 |
-
from misc_utils.train_utils import unit_test_create_model
|
12 |
-
from misc_utils.image_utils import save_tensor_to_gif, save_tensor_to_images
|
13 |
-
import os
|
14 |
-
from PIL import Image
|
15 |
-
import torch
|
16 |
-
import torchvision
|
17 |
-
from torchvision import transforms
|
18 |
-
from einops import rearrange
|
19 |
-
import imageio
|
20 |
-
import time
|
21 |
-
|
22 |
-
from torchvision.transforms import functional as F
|
23 |
-
from torch.hub import download_url_to_file
|
24 |
-
|
25 |
-
import os
|
26 |
-
|
27 |
-
# 推理设置
|
28 |
-
from pl_trainer.inference.inference import InferenceIP2PVideo
|
29 |
-
from tqdm import tqdm
|
30 |
-
|
31 |
-
|
32 |
-
# if not os.path.exists(filename):
|
33 |
-
# original_path = os.getcwd()
|
34 |
-
# base_path = './models'
|
35 |
-
# os.makedirs(base_path, exist_ok=True)
|
36 |
-
|
37 |
-
# # 直接在代码中写入 Token(注意安全风险)
|
38 |
-
# GIT_TOKEN = "955b8ea91095840b76fe38b90a088c200d4c813c"
|
39 |
-
# repo_url = f"https://YeFang:{GIT_TOKEN}@code.openxlab.org.cn/YeFang/RIV_models.git"
|
40 |
-
|
41 |
-
# try:
|
42 |
-
# if os.system(f'git clone {repo_url} {base_path}') != 0:
|
43 |
-
# raise RuntimeError("Git 克隆失败")
|
44 |
-
# os.chdir(base_path)
|
45 |
-
# if os.system('git lfs pull') != 0:
|
46 |
-
# raise RuntimeError("Git LFS 拉取失败")
|
47 |
-
# finally:
|
48 |
-
# os.chdir(original_path)
|
49 |
-
|
50 |
-
def tensor_to_pil_image(x):
|
51 |
-
"""
|
52 |
-
将 4D PyTorch 张量转换为 PIL 图像。
|
53 |
-
"""
|
54 |
-
x = x.float() # 确保张量类型为 float
|
55 |
-
grid_img = torchvision.utils.make_grid(x, nrow=4).permute(1, 2, 0).detach().cpu().numpy()
|
56 |
-
grid_img = (grid_img * 255).clip(0, 255).astype("uint8") # 将 [0, 1] 范围转换为 [0, 255]
|
57 |
-
return Image.fromarray(grid_img)
|
58 |
-
|
59 |
-
def frame_to_batch(x):
|
60 |
-
"""
|
61 |
-
将帧维度转换为批次维度。
|
62 |
-
"""
|
63 |
-
return rearrange(x, 'b f c h w -> (b f) c h w')
|
64 |
-
|
65 |
-
def clip_image(x, min=0., max=1.):
|
66 |
-
"""
|
67 |
-
将图像张量裁剪到指定的最小和最大值。
|
68 |
-
"""
|
69 |
-
return torch.clamp(x, min=min, max=max)
|
70 |
-
|
71 |
-
def unnormalize(x):
|
72 |
-
"""
|
73 |
-
将张量范围从 [-1, 1] 转换到 [0, 1]。
|
74 |
-
"""
|
75 |
-
return (x + 1) / 2
|
76 |
-
|
77 |
-
|
78 |
-
# 读取图像文件
|
79 |
-
def read_images_from_directory(directory, num_frames=16):
|
80 |
-
images = []
|
81 |
-
for i in range(num_frames):
|
82 |
-
img_path = os.path.join(directory, f'{i:04d}.png')
|
83 |
-
img = imageio.imread(img_path)
|
84 |
-
images.append(torch.tensor(img).permute(2, 0, 1)) # Convert to Tensor (C, H, W)
|
85 |
-
return images
|
86 |
-
|
87 |
-
def load_and_process_images(folder_path):
|
88 |
-
"""
|
89 |
-
读取文件夹中的所有图片,将它们转换为 [-1, 1] 范围的张量并返回一个 4D 张量。
|
90 |
-
"""
|
91 |
-
processed_images = []
|
92 |
-
transform = transforms.Compose([
|
93 |
-
transforms.ToTensor(),
|
94 |
-
transforms.Lambda(lambda x: x * 2 - 1) # 将 [0, 1] 转换为 [-1, 1]
|
95 |
-
])
|
96 |
-
for filename in sorted(os.listdir(folder_path)):
|
97 |
-
if filename.endswith(".png"):
|
98 |
-
img_path = os.path.join(folder_path, filename)
|
99 |
-
image = Image.open(img_path).convert("RGB")
|
100 |
-
processed_image = transform(image)
|
101 |
-
processed_images.append(processed_image)
|
102 |
-
return torch.stack(processed_images) # 返回 4D 张量
|
103 |
-
|
104 |
-
def load_and_process_video(video_path, num_frames=16, crop_size=512):
|
105 |
-
"""
|
106 |
-
读取视频文件中的前 num_frames 帧,将每一帧转换为 [-1, 1] 范围的张量,
|
107 |
-
并进行中心裁剪至 crop_size x crop_size,返回一个 4D 张量。
|
108 |
-
"""
|
109 |
-
processed_frames = []
|
110 |
-
transform = transforms.Compose([
|
111 |
-
transforms.CenterCrop(crop_size), # 中心裁剪
|
112 |
-
transforms.ToTensor(),
|
113 |
-
transforms.Lambda(lambda x: x * 2 - 1) # 将 [0, 1] 转换为 [-1, 1]
|
114 |
-
])
|
115 |
-
|
116 |
-
# 使用 OpenCV 读取视频
|
117 |
-
cap = cv2.VideoCapture(video_path)
|
118 |
-
|
119 |
-
if not cap.isOpened():
|
120 |
-
raise ValueError(f"无法打开视频文件: {video_path}")
|
121 |
-
|
122 |
-
frame_count = 0
|
123 |
-
|
124 |
-
while frame_count < num_frames:
|
125 |
-
ret, frame = cap.read()
|
126 |
-
if not ret:
|
127 |
-
break # 视频帧读取完毕或视频帧不足
|
128 |
-
|
129 |
-
# 转换为 RGB 格式
|
130 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
131 |
-
image = Image.fromarray(frame)
|
132 |
-
|
133 |
-
# 应用转换
|
134 |
-
processed_frame = transform(image)
|
135 |
-
processed_frames.append(processed_frame)
|
136 |
-
|
137 |
-
frame_count += 1
|
138 |
-
|
139 |
-
cap.release() # 释放视频资源
|
140 |
-
|
141 |
-
if len(processed_frames) < num_frames:
|
142 |
-
raise ValueError(f"视频帧不足 {num_frames} 帧,仅找到 {len(processed_frames)} 帧。")
|
143 |
-
|
144 |
-
return torch.stack(processed_frames) # 返回 4D 张量 (帧数, 通道数, 高度, 宽度)
|
145 |
-
|
146 |
-
|
147 |
-
def clear_cache(output_path):
|
148 |
-
if os.path.exists(output_path):
|
149 |
-
os.remove(output_path)
|
150 |
-
return None
|
151 |
-
|
152 |
-
|
153 |
-
#! 加载模型
|
154 |
-
# 配置路径和加载模型
|
155 |
-
config_path = 'configs/instruct_v2v_ic_gradio.yaml'
|
156 |
-
diffusion_model = unit_test_create_model(config_path)
|
157 |
-
diffusion_model = diffusion_model.to('cuda')
|
158 |
-
|
159 |
-
# 加载模型检查点
|
160 |
-
# ckpt_path = 'models/relvid_mm_sd15_fbc_unet.pth' #! change
|
161 |
-
# ckpt_path = 'tmp/pytorch_model.bin'
|
162 |
-
# 下载文件
|
163 |
-
|
164 |
-
os.makedirs('models', exist_ok=True)
|
165 |
-
model_path = "models/relvid_mm_sd15_fbc_unet.pth"
|
166 |
-
|
167 |
-
if not os.path.exists(model_path):
|
168 |
-
download_url_to_file(url='https://huggingface.co/aleafy/RelightVid/resolve/main/relvid_mm_sd15_fbc_unet.pth', dst=model_path)
|
169 |
-
|
170 |
-
|
171 |
-
ckpt = torch.load(model_path, map_location='cpu')
|
172 |
-
diffusion_model.load_state_dict(ckpt, strict=False)
|
173 |
-
|
174 |
-
|
175 |
-
# import pdb; pdb.set_trace()
|
176 |
-
|
177 |
-
# 更改全局临时目录
|
178 |
-
new_tmp_dir = "./demo/gradio_bg"
|
179 |
-
os.makedirs(new_tmp_dir, exist_ok=True)
|
180 |
-
|
181 |
-
# import pdb; pdb.set_trace()
|
182 |
-
|
183 |
-
def save_video_from_frames(image_pred, save_pth, fps=8):
|
184 |
-
"""
|
185 |
-
将 image_pred 中的帧保存为视频文件。
|
186 |
-
|
187 |
-
参数:
|
188 |
-
- image_pred: Tensor,形状为 (1, 16, 3, 512, 512)
|
189 |
-
- save_pth: 保存视频的路径,例如 "output_video.mp4"
|
190 |
-
- fps: 视频的帧率
|
191 |
-
"""
|
192 |
-
# 视频参数
|
193 |
-
num_frames = image_pred.shape[1]
|
194 |
-
frame_height, frame_width = 512, 512 # 目标尺寸
|
195 |
-
fourcc = cv2.VideoWriter_fourcc(*'mp4v') # 使用 mp4 编码格式
|
196 |
-
|
197 |
-
# 创建 VideoWriter 对象
|
198 |
-
out = cv2.VideoWriter(save_pth, fourcc, fps, (frame_width, frame_height))
|
199 |
-
|
200 |
-
for i in range(num_frames):
|
201 |
-
# 反归一化 + 转换为 0-255 范围
|
202 |
-
pred_frame = clip_image(unnormalize(image_pred[0][i].unsqueeze(0))) * 255
|
203 |
-
pred_frame_resized = pred_frame.squeeze(0).detach().cpu() # (3, 512, 512)
|
204 |
-
pred_frame_resized = pred_frame_resized.permute(1, 2, 0).numpy().astype("uint8") # (512, 512, 3)
|
205 |
-
|
206 |
-
# Resize 到 256x256
|
207 |
-
pred_frame_resized = cv2.resize(pred_frame_resized, (frame_width, frame_height))
|
208 |
-
|
209 |
-
# 将 RGB 转为 BGR(因为 OpenCV 使用 BGR 格式)
|
210 |
-
pred_frame_bgr = cv2.cvtColor(pred_frame_resized, cv2.COLOR_RGB2BGR)
|
211 |
-
|
212 |
-
# 写入帧到视频
|
213 |
-
out.write(pred_frame_bgr)
|
214 |
-
|
215 |
-
# 释放 VideoWriter 资源
|
216 |
-
out.release()
|
217 |
-
print(f"视频已保存至 {save_pth}")
|
218 |
-
|
219 |
-
|
220 |
-
inf_pipe = InferenceIP2PVideo(
|
221 |
-
diffusion_model.unet,
|
222 |
-
scheduler='ddpm',
|
223 |
-
num_ddim_steps=20
|
224 |
-
)
|
225 |
-
|
226 |
-
# 伪函数占位(生成空白视频)
|
227 |
-
def dummy_process(input_fg, input_bg):
|
228 |
-
# import pdb; pdb.set_trace()
|
229 |
-
|
230 |
-
diffusion_model.to(torch.float16)
|
231 |
-
fg_tensor = load_and_process_video(input_fg).cuda().unsqueeze(0).to(dtype=torch.float16)
|
232 |
-
bg_tensor = load_and_process_video(input_bg).cuda().unsqueeze(0).to(dtype=torch.float16) # (1, 16, 4, 64, 64)
|
233 |
-
|
234 |
-
cond_fg_tensor = diffusion_model.encode_image_to_latent(fg_tensor) # (1, 16, 4, 64, 64)
|
235 |
-
cond_bg_tensor = diffusion_model.encode_image_to_latent(bg_tensor)
|
236 |
-
cond_tensor = torch.cat((cond_fg_tensor, cond_bg_tensor), dim=2)
|
237 |
-
|
238 |
-
# 初始化潜变量
|
239 |
-
init_latent = torch.randn_like(cond_fg_tensor)
|
240 |
-
|
241 |
-
EDIT_PROMPT = 'change the background'
|
242 |
-
VIDEO_CFG = 1.2
|
243 |
-
TEXT_CFG = 7.5
|
244 |
-
text_cond = diffusion_model.encode_text([EDIT_PROMPT]) # (1, 77, 768)
|
245 |
-
text_uncond = diffusion_model.encode_text([''])
|
246 |
-
# to float16
|
247 |
-
print('------------to float 16----------------')
|
248 |
-
init_latent, text_cond, text_uncond, cond_tensor = (
|
249 |
-
init_latent.to(dtype=torch.float16),
|
250 |
-
text_cond.to(dtype=torch.float16),
|
251 |
-
text_uncond.to(dtype=torch.float16),
|
252 |
-
cond_tensor.to(dtype=torch.float16)
|
253 |
-
)
|
254 |
-
inf_pipe.unet.to(torch.float16)
|
255 |
-
latent_pred = inf_pipe(
|
256 |
-
latent=init_latent,
|
257 |
-
text_cond=text_cond,
|
258 |
-
text_uncond=text_uncond,
|
259 |
-
img_cond=cond_tensor,
|
260 |
-
text_cfg=TEXT_CFG,
|
261 |
-
img_cfg=VIDEO_CFG,
|
262 |
-
)['latent']
|
263 |
-
|
264 |
-
|
265 |
-
image_pred = diffusion_model.decode_latent_to_image(latent_pred) # (1,16,3,512,512)
|
266 |
-
output_path = os.path.join(new_tmp_dir, f"output_{int(time.time())}.mp4")
|
267 |
-
# clear_cache(output_path)
|
268 |
-
|
269 |
-
save_video_from_frames(image_pred, output_path)
|
270 |
-
# import pdb; pdb.set_trace()
|
271 |
-
# fps = 8
|
272 |
-
# frames = []
|
273 |
-
# for i in range(16):
|
274 |
-
# pred_frame = clip_image(unnormalize(image_pred[0][i].unsqueeze(0))) * 255
|
275 |
-
# pred_frame_resized = pred_frame.squeeze(0).detach().cpu() #(3,512,512)
|
276 |
-
# pred_frame_resized = pred_frame_resized.permute(1, 2, 0).detach().cpu().numpy().astype("uint8") #(512,512,3) np
|
277 |
-
# Image.fromarray(pred_frame_resized).save(save_pth)
|
278 |
-
|
279 |
-
# # 生成一个简单的黑色视频作为示例
|
280 |
-
# output_path = os.path.join(new_tmp_dir, "output.mp4")
|
281 |
-
# fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
282 |
-
# out = cv2.VideoWriter(output_path, fourcc, 20.0, (512, 512))
|
283 |
-
|
284 |
-
# for _ in range(60): # 生成 3 秒的视频(20fps)
|
285 |
-
# frame = np.zeros((512, 512, 3), dtype=np.uint8)
|
286 |
-
# out.write(frame)
|
287 |
-
# out.release()
|
288 |
-
torch.cuda.empty_cache()
|
289 |
-
|
290 |
-
return output_path
|
291 |
-
|
292 |
-
# 枚举类用于背景选择
|
293 |
-
class BGSource(Enum):
|
294 |
-
UPLOAD = "Use Background Video"
|
295 |
-
UPLOAD_FLIP = "Use Flipped Background Video"
|
296 |
-
LEFT = "Left Light"
|
297 |
-
RIGHT = "Right Light"
|
298 |
-
TOP = "Top Light"
|
299 |
-
BOTTOM = "Bottom Light"
|
300 |
-
GREY = "Ambient"
|
301 |
-
|
302 |
-
# Quick prompts 示例
|
303 |
-
quick_prompts = [
|
304 |
-
'beautiful woman',
|
305 |
-
'handsome man',
|
306 |
-
'beautiful woman, cinematic lighting',
|
307 |
-
'handsome man, cinematic lighting',
|
308 |
-
'beautiful woman, natural lighting',
|
309 |
-
'handsome man, natural lighting',
|
310 |
-
'beautiful woman, neo punk lighting, cyberpunk',
|
311 |
-
'handsome man, neo punk lighting, cyberpunk',
|
312 |
-
]
|
313 |
-
quick_prompts = [[x] for x in quick_prompts]
|
314 |
-
|
315 |
-
# Gradio UI 结构
|
316 |
-
block = gr.Blocks().queue()
|
317 |
-
with block:
|
318 |
-
with gr.Row():
|
319 |
-
gr.Markdown("## IC-Light (Relighting with Foreground and Background Video Condition)")
|
320 |
-
|
321 |
-
with gr.Row():
|
322 |
-
with gr.Column():
|
323 |
-
with gr.Row():
|
324 |
-
input_fg = gr.Video(label="Foreground Video", height=370, width=370, visible=True)
|
325 |
-
input_bg = gr.Video(label="Background Video", height=370, width=370, visible=True)
|
326 |
-
|
327 |
-
prompt = gr.Textbox(label="Prompt")
|
328 |
-
bg_source = gr.Radio(choices=[e.value for e in BGSource],
|
329 |
-
value=BGSource.UPLOAD.value,
|
330 |
-
label="Background Source", type='value')
|
331 |
-
|
332 |
-
example_prompts = gr.Dataset(samples=quick_prompts, label='Prompt Quick List', components=[prompt])
|
333 |
-
bg_gallery = gr.Gallery(height=450, object_fit='contain', label='Background Quick List', value=db_examples.bg_samples, columns=5, allow_preview=False)
|
334 |
-
relight_button = gr.Button(value="Relight")
|
335 |
-
|
336 |
-
with gr.Group():
|
337 |
-
with gr.Row():
|
338 |
-
num_samples = gr.Slider(label="Videos", minimum=1, maximum=12, value=1, step=1)
|
339 |
-
seed = gr.Number(label="Seed", value=12345, precision=0)
|
340 |
-
with gr.Row():
|
341 |
-
video_width = gr.Slider(label="Video Width", minimum=256, maximum=1024, value=512, step=64)
|
342 |
-
video_height = gr.Slider(label="Video Height", minimum=256, maximum=1024, value=640, step=64)
|
343 |
-
|
344 |
-
with gr.Accordion("Advanced options", open=False):
|
345 |
-
steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=20, step=1)
|
346 |
-
cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=7.0, step=0.01)
|
347 |
-
highres_scale = gr.Slider(label="Highres Scale", minimum=1.0, maximum=3.0, value=1.5, step=0.01)
|
348 |
-
highres_denoise = gr.Slider(label="Highres Denoise", minimum=0.1, maximum=0.9, value=0.5, step=0.01)
|
349 |
-
a_prompt = gr.Textbox(label="Added Prompt", value='best quality')
|
350 |
-
n_prompt = gr.Textbox(label="Negative Prompt", value='lowres, bad anatomy, bad hands, cropped, worst quality')
|
351 |
-
normal_button = gr.Button(value="Compute Normal (4x Slower)")
|
352 |
-
|
353 |
-
with gr.Column():
|
354 |
-
result_video = gr.Video(label='Output Video', height=600, width=600, visible=True)
|
355 |
-
|
356 |
-
# 输入列表
|
357 |
-
# ips = [input_fg, input_bg, prompt, video_width, video_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source]
|
358 |
-
ips = [input_fg, input_bg]
|
359 |
-
|
360 |
-
# 按钮绑定处理函数
|
361 |
-
# relight_button.click(fn=lambda: None, inputs=[], outputs=[result_video])
|
362 |
-
|
363 |
-
relight_button.click(fn=dummy_process, inputs=ips, outputs=[result_video])
|
364 |
-
|
365 |
-
normal_button.click(fn=dummy_process, inputs=ips, outputs=[result_video])
|
366 |
-
|
367 |
-
# 背景库选择
|
368 |
-
def bg_gallery_selected(gal, evt: gr.SelectData):
|
369 |
-
# import pdb; pdb.set_trace()
|
370 |
-
# img_path = gal[evt.index][0]
|
371 |
-
img_path = db_examples.bg_samples[evt.index]
|
372 |
-
video_path = img_path.replace('frames/0000.png', 'cropped_video.mp4')
|
373 |
-
return video_path
|
374 |
-
|
375 |
-
bg_gallery.select(bg_gallery_selected, inputs=bg_gallery, outputs=input_bg)
|
376 |
-
|
377 |
-
# 示例
|
378 |
-
# dummy_video_for_outputs = gr.Video(visible=False, label='Result')
|
379 |
-
gr.Examples(
|
380 |
-
fn=lambda *args: args[-1],
|
381 |
-
examples=db_examples.background_conditioned_examples,
|
382 |
-
inputs=[input_fg, input_bg, prompt, bg_source, video_width, video_height, seed, result_video],
|
383 |
-
outputs=[result_video],
|
384 |
-
run_on_click=True, examples_per_page=1024
|
385 |
-
)
|
386 |
-
|
387 |
-
# 启动 Gradio 应用
|
388 |
-
block.launch(server_name='0.0.0.0', server_port=10002, share=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app1_bf2.py
DELETED
@@ -1,388 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
import gradio as gr
|
3 |
-
import numpy as np
|
4 |
-
from enum import Enum
|
5 |
-
import db_examples
|
6 |
-
import cv2
|
7 |
-
|
8 |
-
|
9 |
-
from demo_utils1 import *
|
10 |
-
|
11 |
-
from misc_utils.train_utils import unit_test_create_model
|
12 |
-
from misc_utils.image_utils import save_tensor_to_gif, save_tensor_to_images
|
13 |
-
import os
|
14 |
-
from PIL import Image
|
15 |
-
import torch
|
16 |
-
import torchvision
|
17 |
-
from torchvision import transforms
|
18 |
-
from einops import rearrange
|
19 |
-
import imageio
|
20 |
-
import time
|
21 |
-
|
22 |
-
from torchvision.transforms import functional as F
|
23 |
-
from torch.hub import download_url_to_file
|
24 |
-
|
25 |
-
import os
|
26 |
-
|
27 |
-
# 推理设置
|
28 |
-
from pl_trainer.inference.inference import InferenceIP2PVideo
|
29 |
-
from tqdm import tqdm
|
30 |
-
|
31 |
-
|
32 |
-
# if not os.path.exists(filename):
|
33 |
-
# original_path = os.getcwd()
|
34 |
-
# base_path = './models'
|
35 |
-
# os.makedirs(base_path, exist_ok=True)
|
36 |
-
|
37 |
-
# # 直接在代码中写入 Token(注意安全风险)
|
38 |
-
# GIT_TOKEN = "955b8ea91095840b76fe38b90a088c200d4c813c"
|
39 |
-
# repo_url = f"https://YeFang:{GIT_TOKEN}@code.openxlab.org.cn/YeFang/RIV_models.git"
|
40 |
-
|
41 |
-
# try:
|
42 |
-
# if os.system(f'git clone {repo_url} {base_path}') != 0:
|
43 |
-
# raise RuntimeError("Git 克隆失败")
|
44 |
-
# os.chdir(base_path)
|
45 |
-
# if os.system('git lfs pull') != 0:
|
46 |
-
# raise RuntimeError("Git LFS 拉取失败")
|
47 |
-
# finally:
|
48 |
-
# os.chdir(original_path)
|
49 |
-
|
50 |
-
def tensor_to_pil_image(x):
|
51 |
-
"""
|
52 |
-
将 4D PyTorch 张量转换为 PIL 图像。
|
53 |
-
"""
|
54 |
-
x = x.float() # 确保张量类型为 float
|
55 |
-
grid_img = torchvision.utils.make_grid(x, nrow=4).permute(1, 2, 0).detach().cpu().numpy()
|
56 |
-
grid_img = (grid_img * 255).clip(0, 255).astype("uint8") # 将 [0, 1] 范围转换为 [0, 255]
|
57 |
-
return Image.fromarray(grid_img)
|
58 |
-
|
59 |
-
def frame_to_batch(x):
|
60 |
-
"""
|
61 |
-
将帧维度转换为批次维度。
|
62 |
-
"""
|
63 |
-
return rearrange(x, 'b f c h w -> (b f) c h w')
|
64 |
-
|
65 |
-
def clip_image(x, min=0., max=1.):
|
66 |
-
"""
|
67 |
-
将图像张量裁剪到指定的最小和最大值。
|
68 |
-
"""
|
69 |
-
return torch.clamp(x, min=min, max=max)
|
70 |
-
|
71 |
-
def unnormalize(x):
|
72 |
-
"""
|
73 |
-
将张量范围从 [-1, 1] 转换到 [0, 1]。
|
74 |
-
"""
|
75 |
-
return (x + 1) / 2
|
76 |
-
|
77 |
-
|
78 |
-
# 读取图像文件
|
79 |
-
def read_images_from_directory(directory, num_frames=16):
|
80 |
-
images = []
|
81 |
-
for i in range(num_frames):
|
82 |
-
img_path = os.path.join(directory, f'{i:04d}.png')
|
83 |
-
img = imageio.imread(img_path)
|
84 |
-
images.append(torch.tensor(img).permute(2, 0, 1)) # Convert to Tensor (C, H, W)
|
85 |
-
return images
|
86 |
-
|
87 |
-
def load_and_process_images(folder_path):
|
88 |
-
"""
|
89 |
-
读取文件夹中的所有图片,将它们转换为 [-1, 1] 范围的张量并返回一个 4D 张量。
|
90 |
-
"""
|
91 |
-
processed_images = []
|
92 |
-
transform = transforms.Compose([
|
93 |
-
transforms.ToTensor(),
|
94 |
-
transforms.Lambda(lambda x: x * 2 - 1) # 将 [0, 1] 转换为 [-1, 1]
|
95 |
-
])
|
96 |
-
for filename in sorted(os.listdir(folder_path)):
|
97 |
-
if filename.endswith(".png"):
|
98 |
-
img_path = os.path.join(folder_path, filename)
|
99 |
-
image = Image.open(img_path).convert("RGB")
|
100 |
-
processed_image = transform(image)
|
101 |
-
processed_images.append(processed_image)
|
102 |
-
return torch.stack(processed_images) # 返回 4D 张量
|
103 |
-
|
104 |
-
def load_and_process_video(video_path, num_frames=16, crop_size=512):
|
105 |
-
"""
|
106 |
-
读取视频文件中的前 num_frames 帧,将每一帧转换为 [-1, 1] 范围的张量,
|
107 |
-
并进行中心裁剪至 crop_size x crop_size,返回一个 4D 张量。
|
108 |
-
"""
|
109 |
-
processed_frames = []
|
110 |
-
transform = transforms.Compose([
|
111 |
-
transforms.CenterCrop(crop_size), # 中心裁剪
|
112 |
-
transforms.ToTensor(),
|
113 |
-
transforms.Lambda(lambda x: x * 2 - 1) # 将 [0, 1] 转换为 [-1, 1]
|
114 |
-
])
|
115 |
-
|
116 |
-
# 使用 OpenCV 读取视频
|
117 |
-
cap = cv2.VideoCapture(video_path)
|
118 |
-
|
119 |
-
if not cap.isOpened():
|
120 |
-
raise ValueError(f"无法打开视频文件: {video_path}")
|
121 |
-
|
122 |
-
frame_count = 0
|
123 |
-
|
124 |
-
while frame_count < num_frames:
|
125 |
-
ret, frame = cap.read()
|
126 |
-
if not ret:
|
127 |
-
break # 视频帧读取完毕或视频帧不足
|
128 |
-
|
129 |
-
# 转换为 RGB 格式
|
130 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
131 |
-
image = Image.fromarray(frame)
|
132 |
-
|
133 |
-
# 应用转换
|
134 |
-
processed_frame = transform(image)
|
135 |
-
processed_frames.append(processed_frame)
|
136 |
-
|
137 |
-
frame_count += 1
|
138 |
-
|
139 |
-
cap.release() # 释放视频资源
|
140 |
-
|
141 |
-
if len(processed_frames) < num_frames:
|
142 |
-
raise ValueError(f"视频帧不足 {num_frames} 帧,仅找到 {len(processed_frames)} 帧。")
|
143 |
-
|
144 |
-
return torch.stack(processed_frames) # 返回 4D 张量 (帧数, 通道数, 高度, 宽度)
|
145 |
-
|
146 |
-
|
147 |
-
def clear_cache(output_path):
|
148 |
-
if os.path.exists(output_path):
|
149 |
-
os.remove(output_path)
|
150 |
-
return None
|
151 |
-
|
152 |
-
|
153 |
-
#! 加载模型
|
154 |
-
# 配置路径和加载模型
|
155 |
-
config_path = 'configs/instruct_v2v_ic_gradio.yaml'
|
156 |
-
diffusion_model = unit_test_create_model(config_path)
|
157 |
-
diffusion_model = diffusion_model.to('cuda')
|
158 |
-
|
159 |
-
# 加载模型检查点
|
160 |
-
# ckpt_path = 'models/relvid_mm_sd15_fbc_unet.pth' #! change
|
161 |
-
# ckpt_path = 'tmp/pytorch_model.bin'
|
162 |
-
# 下载文件
|
163 |
-
|
164 |
-
os.makedirs('models', exist_ok=True)
|
165 |
-
model_path = "models/relvid_mm_sd15_fbc_unet.pth"
|
166 |
-
|
167 |
-
if not os.path.exists(model_path):
|
168 |
-
download_url_to_file(url='https://huggingface.co/aleafy/RelightVid/resolve/main/relvid_mm_sd15_fbc_unet.pth', dst=model_path)
|
169 |
-
|
170 |
-
|
171 |
-
ckpt = torch.load(model_path, map_location='cpu')
|
172 |
-
diffusion_model.load_state_dict(ckpt, strict=False)
|
173 |
-
|
174 |
-
|
175 |
-
# import pdb; pdb.set_trace()
|
176 |
-
|
177 |
-
# 更改全局临时目录
|
178 |
-
new_tmp_dir = "./demo/gradio_bg"
|
179 |
-
os.makedirs(new_tmp_dir, exist_ok=True)
|
180 |
-
|
181 |
-
# import pdb; pdb.set_trace()
|
182 |
-
|
183 |
-
def save_video_from_frames(image_pred, save_pth, fps=8):
|
184 |
-
"""
|
185 |
-
将 image_pred 中的帧保存为视频文件。
|
186 |
-
|
187 |
-
参数:
|
188 |
-
- image_pred: Tensor,形状为 (1, 16, 3, 512, 512)
|
189 |
-
- save_pth: 保存视频的路径,例如 "output_video.mp4"
|
190 |
-
- fps: 视频的帧率
|
191 |
-
"""
|
192 |
-
# 视频参数
|
193 |
-
num_frames = image_pred.shape[1]
|
194 |
-
frame_height, frame_width = 512, 512 # 目标尺寸
|
195 |
-
fourcc = cv2.VideoWriter_fourcc(*'mp4v') # 使用 mp4 编码格式
|
196 |
-
|
197 |
-
# 创建 VideoWriter 对象
|
198 |
-
out = cv2.VideoWriter(save_pth, fourcc, fps, (frame_width, frame_height))
|
199 |
-
|
200 |
-
for i in range(num_frames):
|
201 |
-
# 反归一化 + 转换为 0-255 范围
|
202 |
-
pred_frame = clip_image(unnormalize(image_pred[0][i].unsqueeze(0))) * 255
|
203 |
-
pred_frame_resized = pred_frame.squeeze(0).detach().cpu() # (3, 512, 512)
|
204 |
-
pred_frame_resized = pred_frame_resized.permute(1, 2, 0).numpy().astype("uint8") # (512, 512, 3)
|
205 |
-
|
206 |
-
# Resize 到 256x256
|
207 |
-
pred_frame_resized = cv2.resize(pred_frame_resized, (frame_width, frame_height))
|
208 |
-
|
209 |
-
# 将 RGB 转为 BGR(因为 OpenCV 使用 BGR 格式)
|
210 |
-
pred_frame_bgr = cv2.cvtColor(pred_frame_resized, cv2.COLOR_RGB2BGR)
|
211 |
-
|
212 |
-
# 写入帧到视频
|
213 |
-
out.write(pred_frame_bgr)
|
214 |
-
|
215 |
-
# 释放 VideoWriter 资源
|
216 |
-
out.release()
|
217 |
-
print(f"视频已保存至 {save_pth}")
|
218 |
-
|
219 |
-
|
220 |
-
inf_pipe = InferenceIP2PVideo(
|
221 |
-
diffusion_model.unet,
|
222 |
-
scheduler='ddpm',
|
223 |
-
num_ddim_steps=20
|
224 |
-
)
|
225 |
-
|
226 |
-
# 伪函数占位(生成空白视频)
|
227 |
-
def dummy_process(input_fg, input_bg):
|
228 |
-
# import pdb; pdb.set_trace()
|
229 |
-
|
230 |
-
diffusion_model.to(torch.float16)
|
231 |
-
fg_tensor = load_and_process_video(input_fg).cuda().unsqueeze(0).to(dtype=torch.float16)
|
232 |
-
bg_tensor = load_and_process_video(input_bg).cuda().unsqueeze(0).to(dtype=torch.float16) # (1, 16, 4, 64, 64)
|
233 |
-
|
234 |
-
cond_fg_tensor = diffusion_model.encode_image_to_latent(fg_tensor) # (1, 16, 4, 64, 64)
|
235 |
-
cond_bg_tensor = diffusion_model.encode_image_to_latent(bg_tensor)
|
236 |
-
cond_tensor = torch.cat((cond_fg_tensor, cond_bg_tensor), dim=2)
|
237 |
-
|
238 |
-
# 初始化潜变量
|
239 |
-
init_latent = torch.randn_like(cond_fg_tensor)
|
240 |
-
|
241 |
-
EDIT_PROMPT = 'change the background'
|
242 |
-
VIDEO_CFG = 1.2
|
243 |
-
TEXT_CFG = 7.5
|
244 |
-
text_cond = diffusion_model.encode_text([EDIT_PROMPT]) # (1, 77, 768)
|
245 |
-
text_uncond = diffusion_model.encode_text([''])
|
246 |
-
# to float16
|
247 |
-
print('------------to float 16----------------')
|
248 |
-
init_latent, text_cond, text_uncond, cond_tensor = (
|
249 |
-
init_latent.to(dtype=torch.float16),
|
250 |
-
text_cond.to(dtype=torch.float16),
|
251 |
-
text_uncond.to(dtype=torch.float16),
|
252 |
-
cond_tensor.to(dtype=torch.float16)
|
253 |
-
)
|
254 |
-
inf_pipe.unet.to(torch.float16)
|
255 |
-
latent_pred = inf_pipe(
|
256 |
-
latent=init_latent,
|
257 |
-
text_cond=text_cond,
|
258 |
-
text_uncond=text_uncond,
|
259 |
-
img_cond=cond_tensor,
|
260 |
-
text_cfg=TEXT_CFG,
|
261 |
-
img_cfg=VIDEO_CFG,
|
262 |
-
)['latent']
|
263 |
-
|
264 |
-
|
265 |
-
image_pred = diffusion_model.decode_latent_to_image(latent_pred) # (1,16,3,512,512)
|
266 |
-
output_path = os.path.join(new_tmp_dir, f"output_{int(time.time())}.mp4")
|
267 |
-
# clear_cache(output_path)
|
268 |
-
|
269 |
-
save_video_from_frames(image_pred, output_path)
|
270 |
-
# import pdb; pdb.set_trace()
|
271 |
-
# fps = 8
|
272 |
-
# frames = []
|
273 |
-
# for i in range(16):
|
274 |
-
# pred_frame = clip_image(unnormalize(image_pred[0][i].unsqueeze(0))) * 255
|
275 |
-
# pred_frame_resized = pred_frame.squeeze(0).detach().cpu() #(3,512,512)
|
276 |
-
# pred_frame_resized = pred_frame_resized.permute(1, 2, 0).detach().cpu().numpy().astype("uint8") #(512,512,3) np
|
277 |
-
# Image.fromarray(pred_frame_resized).save(save_pth)
|
278 |
-
|
279 |
-
# # 生成一个简单的黑色视频作为示例
|
280 |
-
# output_path = os.path.join(new_tmp_dir, "output.mp4")
|
281 |
-
# fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
282 |
-
# out = cv2.VideoWriter(output_path, fourcc, 20.0, (512, 512))
|
283 |
-
|
284 |
-
# for _ in range(60): # 生成 3 秒的视频(20fps)
|
285 |
-
# frame = np.zeros((512, 512, 3), dtype=np.uint8)
|
286 |
-
# out.write(frame)
|
287 |
-
# out.release()
|
288 |
-
torch.cuda.empty_cache()
|
289 |
-
|
290 |
-
return output_path
|
291 |
-
|
292 |
-
# 枚举类用于背景选择
|
293 |
-
class BGSource(Enum):
|
294 |
-
UPLOAD = "Use Background Video"
|
295 |
-
UPLOAD_FLIP = "Use Flipped Background Video"
|
296 |
-
UPLOAD_REVERSE = "Use Reversed Background Video"
|
297 |
-
|
298 |
-
# Quick prompts 示例
|
299 |
-
quick_prompts = [
|
300 |
-
'beautiful woman',
|
301 |
-
'handsome man',
|
302 |
-
'beautiful woman, cinematic lighting',
|
303 |
-
'handsome man, cinematic lighting',
|
304 |
-
'beautiful woman, natural lighting',
|
305 |
-
'handsome man, natural lighting',
|
306 |
-
'beautiful woman, neo punk lighting, cyberpunk',
|
307 |
-
'handsome man, neo punk lighting, cyberpunk',
|
308 |
-
]
|
309 |
-
quick_prompts = [[x] for x in quick_prompts]
|
310 |
-
|
311 |
-
# Gradio UI 结构
|
312 |
-
block = gr.Blocks().queue()
|
313 |
-
with block:
|
314 |
-
with gr.Row():
|
315 |
-
gr.Markdown("## IC-Light (Relighting with Foreground and Background Video Condition)")
|
316 |
-
|
317 |
-
with gr.Row():
|
318 |
-
with gr.Column():
|
319 |
-
input_fg = gr.Video(label="Foreground Video", height=450, visible=True)
|
320 |
-
with gr.Column():
|
321 |
-
input_bg = gr.Video(label="Background Video", height=450, visible=True)
|
322 |
-
with gr.Column():
|
323 |
-
result_video = gr.Video(label='Output Video', height=450, visible=True)
|
324 |
-
|
325 |
-
with gr.Row():
|
326 |
-
with gr.Column():
|
327 |
-
prompt = gr.Textbox(label="Prompt")
|
328 |
-
bg_source = gr.Radio(choices=[e.value for e in BGSource],
|
329 |
-
value=BGSource.UPLOAD.value,
|
330 |
-
label="Background Source", type='value')
|
331 |
-
|
332 |
-
example_prompts = gr.Dataset(samples=quick_prompts, label='Prompt Quick List', components=[prompt])
|
333 |
-
bg_gallery = gr.Gallery(height=450, object_fit='contain', label='Background Quick List', value=db_examples.bg_samples, columns=5, allow_preview=False)
|
334 |
-
relight_button = gr.Button(value="Relight")
|
335 |
-
|
336 |
-
with gr.Group():
|
337 |
-
with gr.Row():
|
338 |
-
num_samples = gr.Slider(label="Videos", minimum=1, maximum=12, value=1, step=1)
|
339 |
-
seed = gr.Number(label="Seed", value=12345, precision=0)
|
340 |
-
with gr.Row():
|
341 |
-
video_width = gr.Slider(label="Video Width", minimum=256, maximum=1024, value=512, step=64)
|
342 |
-
video_height = gr.Slider(label="Video Height", minimum=256, maximum=1024, value=640, step=64)
|
343 |
-
|
344 |
-
|
345 |
-
with gr.Column():
|
346 |
-
with gr.Accordion("Advanced options", open=False):
|
347 |
-
steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=20, step=1)
|
348 |
-
cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=7.0, step=0.01)
|
349 |
-
highres_scale = gr.Slider(label="Highres Scale", minimum=1.0, maximum=3.0, value=1.5, step=0.01)
|
350 |
-
highres_denoise = gr.Slider(label="Highres Denoise", minimum=0.1, maximum=0.9, value=0.5, step=0.01)
|
351 |
-
a_prompt = gr.Textbox(label="Added Prompt", value='best quality')
|
352 |
-
n_prompt = gr.Textbox(label="Negative Prompt", value='lowres, bad anatomy, bad hands, cropped, worst quality')
|
353 |
-
normal_button = gr.Button(value="Compute Normal (4x Slower)")
|
354 |
-
|
355 |
-
|
356 |
-
# 输入列表
|
357 |
-
# ips = [input_fg, input_bg, prompt, video_width, video_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source]
|
358 |
-
ips = [input_fg, input_bg]
|
359 |
-
|
360 |
-
# 按钮绑定处理函数
|
361 |
-
# relight_button.click(fn=lambda: None, inputs=[], outputs=[result_video])
|
362 |
-
|
363 |
-
relight_button.click(fn=dummy_process, inputs=ips, outputs=[result_video])
|
364 |
-
|
365 |
-
normal_button.click(fn=dummy_process, inputs=ips, outputs=[result_video])
|
366 |
-
|
367 |
-
# 背景库选择
|
368 |
-
def bg_gallery_selected(gal, evt: gr.SelectData):
|
369 |
-
# import pdb; pdb.set_trace()
|
370 |
-
# img_path = gal[evt.index][0]
|
371 |
-
img_path = db_examples.bg_samples[evt.index]
|
372 |
-
video_path = img_path.replace('frames/0000.png', 'cropped_video.mp4')
|
373 |
-
return video_path
|
374 |
-
|
375 |
-
bg_gallery.select(bg_gallery_selected, inputs=bg_gallery, outputs=input_bg)
|
376 |
-
|
377 |
-
# 示例
|
378 |
-
# dummy_video_for_outputs = gr.Video(visible=False, label='Result')
|
379 |
-
gr.Examples(
|
380 |
-
fn=lambda *args: args[-1],
|
381 |
-
examples=db_examples.background_conditioned_examples,
|
382 |
-
inputs=[input_fg, input_bg, prompt, bg_source, video_width, video_height, seed, result_video],
|
383 |
-
outputs=[result_video],
|
384 |
-
run_on_click=True, examples_per_page=1024
|
385 |
-
)
|
386 |
-
|
387 |
-
# 启动 Gradio 应用
|
388 |
-
block.launch(server_name='0.0.0.0', server_port=10002, share=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
app_bf.py
DELETED
@@ -1,391 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
import gradio as gr
|
3 |
-
import numpy as np
|
4 |
-
from enum import Enum
|
5 |
-
import db_examples
|
6 |
-
import cv2
|
7 |
-
|
8 |
-
import spaces
|
9 |
-
|
10 |
-
from demo_utils1 import *
|
11 |
-
|
12 |
-
from misc_utils.train_utils import unit_test_create_model
|
13 |
-
from misc_utils.image_utils import save_tensor_to_gif, save_tensor_to_images
|
14 |
-
import os
|
15 |
-
from PIL import Image
|
16 |
-
import torch
|
17 |
-
import torchvision
|
18 |
-
from torchvision import transforms
|
19 |
-
from einops import rearrange
|
20 |
-
import imageio
|
21 |
-
import time
|
22 |
-
|
23 |
-
from torchvision.transforms import functional as F
|
24 |
-
from torch.hub import download_url_to_file
|
25 |
-
|
26 |
-
import os
|
27 |
-
|
28 |
-
# 推理设置
|
29 |
-
from pl_trainer.inference.inference import InferenceIP2PVideo
|
30 |
-
from tqdm import tqdm
|
31 |
-
|
32 |
-
|
33 |
-
# if not os.path.exists(filename):
|
34 |
-
# original_path = os.getcwd()
|
35 |
-
# base_path = './models'
|
36 |
-
# os.makedirs(base_path, exist_ok=True)
|
37 |
-
|
38 |
-
# # 直接在代码中写入 Token(注意安全风险)
|
39 |
-
# GIT_TOKEN = "955b8ea91095840b76fe38b90a088c200d4c813c"
|
40 |
-
# repo_url = f"https://YeFang:{GIT_TOKEN}@code.openxlab.org.cn/YeFang/RIV_models.git"
|
41 |
-
|
42 |
-
# try:
|
43 |
-
# if os.system(f'git clone {repo_url} {base_path}') != 0:
|
44 |
-
# raise RuntimeError("Git 克隆失败")
|
45 |
-
# os.chdir(base_path)
|
46 |
-
# if os.system('git lfs pull') != 0:
|
47 |
-
# raise RuntimeError("Git LFS 拉取失败")
|
48 |
-
# finally:
|
49 |
-
# os.chdir(original_path)
|
50 |
-
|
51 |
-
def tensor_to_pil_image(x):
|
52 |
-
"""
|
53 |
-
将 4D PyTorch 张量转换为 PIL 图像。
|
54 |
-
"""
|
55 |
-
x = x.float() # 确保张量类型为 float
|
56 |
-
grid_img = torchvision.utils.make_grid(x, nrow=4).permute(1, 2, 0).detach().cpu().numpy()
|
57 |
-
grid_img = (grid_img * 255).clip(0, 255).astype("uint8") # 将 [0, 1] 范围转换为 [0, 255]
|
58 |
-
return Image.fromarray(grid_img)
|
59 |
-
|
60 |
-
def frame_to_batch(x):
|
61 |
-
"""
|
62 |
-
将帧维度转换为批次维度。
|
63 |
-
"""
|
64 |
-
return rearrange(x, 'b f c h w -> (b f) c h w')
|
65 |
-
|
66 |
-
def clip_image(x, min=0., max=1.):
|
67 |
-
"""
|
68 |
-
将图像张量裁剪到指定的最小和最大值。
|
69 |
-
"""
|
70 |
-
return torch.clamp(x, min=min, max=max)
|
71 |
-
|
72 |
-
def unnormalize(x):
|
73 |
-
"""
|
74 |
-
将张量范围从 [-1, 1] 转换到 [0, 1]。
|
75 |
-
"""
|
76 |
-
return (x + 1) / 2
|
77 |
-
|
78 |
-
|
79 |
-
# 读取图像文件
|
80 |
-
def read_images_from_directory(directory, num_frames=16):
|
81 |
-
images = []
|
82 |
-
for i in range(num_frames):
|
83 |
-
img_path = os.path.join(directory, f'{i:04d}.png')
|
84 |
-
img = imageio.imread(img_path)
|
85 |
-
images.append(torch.tensor(img).permute(2, 0, 1)) # Convert to Tensor (C, H, W)
|
86 |
-
return images
|
87 |
-
|
88 |
-
def load_and_process_images(folder_path):
|
89 |
-
"""
|
90 |
-
读取文件夹中的所有图片,将它们转换为 [-1, 1] 范围的张量并返回一个 4D 张量。
|
91 |
-
"""
|
92 |
-
processed_images = []
|
93 |
-
transform = transforms.Compose([
|
94 |
-
transforms.ToTensor(),
|
95 |
-
transforms.Lambda(lambda x: x * 2 - 1) # 将 [0, 1] 转换为 [-1, 1]
|
96 |
-
])
|
97 |
-
for filename in sorted(os.listdir(folder_path)):
|
98 |
-
if filename.endswith(".png"):
|
99 |
-
img_path = os.path.join(folder_path, filename)
|
100 |
-
image = Image.open(img_path).convert("RGB")
|
101 |
-
processed_image = transform(image)
|
102 |
-
processed_images.append(processed_image)
|
103 |
-
return torch.stack(processed_images) # 返回 4D 张量
|
104 |
-
|
105 |
-
def load_and_process_video(video_path, num_frames=16, crop_size=512):
|
106 |
-
"""
|
107 |
-
读取视频文件中的前 num_frames 帧,将每一帧转换为 [-1, 1] 范围的张量,
|
108 |
-
并进行中心裁剪至 crop_size x crop_size,返回一个 4D 张量。
|
109 |
-
"""
|
110 |
-
processed_frames = []
|
111 |
-
transform = transforms.Compose([
|
112 |
-
transforms.CenterCrop(crop_size), # 中心裁剪
|
113 |
-
transforms.ToTensor(),
|
114 |
-
transforms.Lambda(lambda x: x * 2 - 1) # 将 [0, 1] 转换为 [-1, 1]
|
115 |
-
])
|
116 |
-
|
117 |
-
# 使用 OpenCV 读取视频
|
118 |
-
cap = cv2.VideoCapture(video_path)
|
119 |
-
|
120 |
-
if not cap.isOpened():
|
121 |
-
raise ValueError(f"无法打开视频文件: {video_path}")
|
122 |
-
|
123 |
-
frame_count = 0
|
124 |
-
|
125 |
-
while frame_count < num_frames:
|
126 |
-
ret, frame = cap.read()
|
127 |
-
if not ret:
|
128 |
-
break # 视频帧读取完毕或视频帧不足
|
129 |
-
|
130 |
-
# 转换为 RGB 格式
|
131 |
-
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
132 |
-
image = Image.fromarray(frame)
|
133 |
-
|
134 |
-
# 应用转换
|
135 |
-
processed_frame = transform(image)
|
136 |
-
processed_frames.append(processed_frame)
|
137 |
-
|
138 |
-
frame_count += 1
|
139 |
-
|
140 |
-
cap.release() # 释放视频资源
|
141 |
-
|
142 |
-
if len(processed_frames) < num_frames:
|
143 |
-
raise ValueError(f"视频帧不足 {num_frames} 帧,仅找到 {len(processed_frames)} 帧。")
|
144 |
-
|
145 |
-
return torch.stack(processed_frames) # 返回 4D 张量 (帧数, 通道数, 高度, 宽度)
|
146 |
-
|
147 |
-
|
148 |
-
def clear_cache(output_path):
|
149 |
-
if os.path.exists(output_path):
|
150 |
-
os.remove(output_path)
|
151 |
-
return None
|
152 |
-
|
153 |
-
|
154 |
-
#! 加载模型
|
155 |
-
# 配置路径和加载模型
|
156 |
-
config_path = 'configs/instruct_v2v_ic_gradio.yaml'
|
157 |
-
diffusion_model = unit_test_create_model(config_path)
|
158 |
-
diffusion_model = diffusion_model.to('cuda')
|
159 |
-
|
160 |
-
# 加载模型检查点
|
161 |
-
# ckpt_path = 'models/relvid_mm_sd15_fbc_unet.pth' #! change
|
162 |
-
# ckpt_path = 'tmp/pytorch_model.bin'
|
163 |
-
# 下载文件
|
164 |
-
|
165 |
-
os.makedirs('models', exist_ok=True)
|
166 |
-
model_path = "models/relvid_mm_sd15_fbc_unet.pth"
|
167 |
-
|
168 |
-
if not os.path.exists(model_path):
|
169 |
-
download_url_to_file(url='https://huggingface.co/aleafy/RelightVid/resolve/main/relvid_mm_sd15_fbc_unet.pth', dst=model_path)
|
170 |
-
|
171 |
-
|
172 |
-
ckpt = torch.load(model_path, map_location='cpu')
|
173 |
-
diffusion_model.load_state_dict(ckpt, strict=False)
|
174 |
-
|
175 |
-
|
176 |
-
# import pdb; pdb.set_trace()
|
177 |
-
|
178 |
-
# 更改全局临时目录
|
179 |
-
new_tmp_dir = "./demo/gradio_bg"
|
180 |
-
os.makedirs(new_tmp_dir, exist_ok=True)
|
181 |
-
|
182 |
-
# import pdb; pdb.set_trace()
|
183 |
-
|
184 |
-
def save_video_from_frames(image_pred, save_pth, fps=8):
|
185 |
-
"""
|
186 |
-
将 image_pred 中的帧保存为视频文件。
|
187 |
-
|
188 |
-
参数:
|
189 |
-
- image_pred: Tensor,形状为 (1, 16, 3, 512, 512)
|
190 |
-
- save_pth: 保存视频的路径,例如 "output_video.mp4"
|
191 |
-
- fps: 视频的帧率
|
192 |
-
"""
|
193 |
-
# 视频参数
|
194 |
-
num_frames = image_pred.shape[1]
|
195 |
-
frame_height, frame_width = 512, 512 # 目标尺寸
|
196 |
-
fourcc = cv2.VideoWriter_fourcc(*'mp4v') # 使用 mp4 编码格式
|
197 |
-
|
198 |
-
# 创建 VideoWriter 对象
|
199 |
-
out = cv2.VideoWriter(save_pth, fourcc, fps, (frame_width, frame_height))
|
200 |
-
|
201 |
-
for i in range(num_frames):
|
202 |
-
# 反归一化 + 转换为 0-255 范围
|
203 |
-
pred_frame = clip_image(unnormalize(image_pred[0][i].unsqueeze(0))) * 255
|
204 |
-
pred_frame_resized = pred_frame.squeeze(0).detach().cpu() # (3, 512, 512)
|
205 |
-
pred_frame_resized = pred_frame_resized.permute(1, 2, 0).numpy().astype("uint8") # (512, 512, 3)
|
206 |
-
|
207 |
-
# Resize 到 256x256
|
208 |
-
pred_frame_resized = cv2.resize(pred_frame_resized, (frame_width, frame_height))
|
209 |
-
|
210 |
-
# 将 RGB 转为 BGR(因为 OpenCV 使用 BGR 格式)
|
211 |
-
pred_frame_bgr = cv2.cvtColor(pred_frame_resized, cv2.COLOR_RGB2BGR)
|
212 |
-
|
213 |
-
# 写入帧到视频
|
214 |
-
out.write(pred_frame_bgr)
|
215 |
-
|
216 |
-
# 释放 VideoWriter 资源
|
217 |
-
out.release()
|
218 |
-
print(f"视频已保存至 {save_pth}")
|
219 |
-
|
220 |
-
|
221 |
-
inf_pipe = InferenceIP2PVideo(
|
222 |
-
diffusion_model.unet,
|
223 |
-
scheduler='ddpm',
|
224 |
-
num_ddim_steps=20
|
225 |
-
)
|
226 |
-
|
227 |
-
# 伪函数占位(生成空白视频)
|
228 |
-
@spaces.GPU
|
229 |
-
def dummy_process(input_fg, input_bg):
|
230 |
-
# import pdb; pdb.set_trace()
|
231 |
-
|
232 |
-
diffusion_model.to(torch.float16)
|
233 |
-
fg_tensor = load_and_process_video(input_fg).cuda().unsqueeze(0).to(dtype=torch.float16)
|
234 |
-
bg_tensor = load_and_process_video(input_bg).cuda().unsqueeze(0).to(dtype=torch.float16) # (1, 16, 4, 64, 64)
|
235 |
-
|
236 |
-
cond_fg_tensor = diffusion_model.encode_image_to_latent(fg_tensor) # (1, 16, 4, 64, 64)
|
237 |
-
cond_bg_tensor = diffusion_model.encode_image_to_latent(bg_tensor)
|
238 |
-
cond_tensor = torch.cat((cond_fg_tensor, cond_bg_tensor), dim=2)
|
239 |
-
|
240 |
-
# 初始化潜变量
|
241 |
-
init_latent = torch.randn_like(cond_fg_tensor)
|
242 |
-
|
243 |
-
EDIT_PROMPT = 'change the background'
|
244 |
-
VIDEO_CFG = 1.2
|
245 |
-
TEXT_CFG = 7.5
|
246 |
-
text_cond = diffusion_model.encode_text([EDIT_PROMPT]) # (1, 77, 768)
|
247 |
-
text_uncond = diffusion_model.encode_text([''])
|
248 |
-
# to float16
|
249 |
-
print('------------to float 16----------------')
|
250 |
-
init_latent, text_cond, text_uncond, cond_tensor = (
|
251 |
-
init_latent.to(dtype=torch.float16),
|
252 |
-
text_cond.to(dtype=torch.float16),
|
253 |
-
text_uncond.to(dtype=torch.float16),
|
254 |
-
cond_tensor.to(dtype=torch.float16)
|
255 |
-
)
|
256 |
-
inf_pipe.unet.to(torch.float16)
|
257 |
-
latent_pred = inf_pipe(
|
258 |
-
latent=init_latent,
|
259 |
-
text_cond=text_cond,
|
260 |
-
text_uncond=text_uncond,
|
261 |
-
img_cond=cond_tensor,
|
262 |
-
text_cfg=TEXT_CFG,
|
263 |
-
img_cfg=VIDEO_CFG,
|
264 |
-
)['latent']
|
265 |
-
|
266 |
-
|
267 |
-
image_pred = diffusion_model.decode_latent_to_image(latent_pred) # (1,16,3,512,512)
|
268 |
-
output_path = os.path.join(new_tmp_dir, f"output_{int(time.time())}.mp4")
|
269 |
-
# clear_cache(output_path)
|
270 |
-
|
271 |
-
save_video_from_frames(image_pred, output_path)
|
272 |
-
# import pdb; pdb.set_trace()
|
273 |
-
# fps = 8
|
274 |
-
# frames = []
|
275 |
-
# for i in range(16):
|
276 |
-
# pred_frame = clip_image(unnormalize(image_pred[0][i].unsqueeze(0))) * 255
|
277 |
-
# pred_frame_resized = pred_frame.squeeze(0).detach().cpu() #(3,512,512)
|
278 |
-
# pred_frame_resized = pred_frame_resized.permute(1, 2, 0).detach().cpu().numpy().astype("uint8") #(512,512,3) np
|
279 |
-
# Image.fromarray(pred_frame_resized).save(save_pth)
|
280 |
-
|
281 |
-
# # 生成一个简单的黑色视频作为示例
|
282 |
-
# output_path = os.path.join(new_tmp_dir, "output.mp4")
|
283 |
-
# fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
284 |
-
# out = cv2.VideoWriter(output_path, fourcc, 20.0, (512, 512))
|
285 |
-
|
286 |
-
# for _ in range(60): # 生成 3 秒的视频(20fps)
|
287 |
-
# frame = np.zeros((512, 512, 3), dtype=np.uint8)
|
288 |
-
# out.write(frame)
|
289 |
-
# out.release()
|
290 |
-
torch.cuda.empty_cache()
|
291 |
-
|
292 |
-
return output_path
|
293 |
-
|
294 |
-
# 枚举类用于背景选择
|
295 |
-
class BGSource(Enum):
|
296 |
-
UPLOAD = "Use Background Video"
|
297 |
-
UPLOAD_FLIP = "Use Flipped Background Video"
|
298 |
-
LEFT = "Left Light"
|
299 |
-
RIGHT = "Right Light"
|
300 |
-
TOP = "Top Light"
|
301 |
-
BOTTOM = "Bottom Light"
|
302 |
-
GREY = "Ambient"
|
303 |
-
|
304 |
-
# Quick prompts ��例
|
305 |
-
quick_prompts = [
|
306 |
-
'beautiful woman',
|
307 |
-
'handsome man',
|
308 |
-
'beautiful woman, cinematic lighting',
|
309 |
-
'handsome man, cinematic lighting',
|
310 |
-
'beautiful woman, natural lighting',
|
311 |
-
'handsome man, natural lighting',
|
312 |
-
'beautiful woman, neo punk lighting, cyberpunk',
|
313 |
-
'handsome man, neo punk lighting, cyberpunk',
|
314 |
-
]
|
315 |
-
quick_prompts = [[x] for x in quick_prompts]
|
316 |
-
|
317 |
-
# Gradio UI 结构
|
318 |
-
block = gr.Blocks().queue()
|
319 |
-
with block:
|
320 |
-
with gr.Row():
|
321 |
-
gr.Markdown("## IC-Light (Relighting with Foreground and Background Video Condition)")
|
322 |
-
|
323 |
-
with gr.Row():
|
324 |
-
with gr.Column():
|
325 |
-
with gr.Row():
|
326 |
-
input_fg = gr.Video(label="Foreground Video", height=370, width=370, visible=True)
|
327 |
-
input_bg = gr.Video(label="Background Video", height=370, width=370, visible=True)
|
328 |
-
|
329 |
-
prompt = gr.Textbox(label="Prompt")
|
330 |
-
bg_source = gr.Radio(choices=[e.value for e in BGSource],
|
331 |
-
value=BGSource.UPLOAD.value,
|
332 |
-
label="Background Source", type='value')
|
333 |
-
|
334 |
-
example_prompts = gr.Dataset(samples=quick_prompts, label='Prompt Quick List', components=[prompt])
|
335 |
-
bg_gallery = gr.Gallery(height=450, object_fit='contain', label='Background Quick List', value=db_examples.bg_samples, columns=5, allow_preview=False)
|
336 |
-
relight_button = gr.Button(value="Relight")
|
337 |
-
|
338 |
-
with gr.Group():
|
339 |
-
with gr.Row():
|
340 |
-
num_samples = gr.Slider(label="Videos", minimum=1, maximum=12, value=1, step=1)
|
341 |
-
seed = gr.Number(label="Seed", value=12345, precision=0)
|
342 |
-
with gr.Row():
|
343 |
-
video_width = gr.Slider(label="Video Width", minimum=256, maximum=1024, value=512, step=64)
|
344 |
-
video_height = gr.Slider(label="Video Height", minimum=256, maximum=1024, value=640, step=64)
|
345 |
-
|
346 |
-
with gr.Accordion("Advanced options", open=False):
|
347 |
-
steps = gr.Slider(label="Steps", minimum=1, maximum=100, value=20, step=1)
|
348 |
-
cfg = gr.Slider(label="CFG Scale", minimum=1.0, maximum=32.0, value=7.0, step=0.01)
|
349 |
-
highres_scale = gr.Slider(label="Highres Scale", minimum=1.0, maximum=3.0, value=1.5, step=0.01)
|
350 |
-
highres_denoise = gr.Slider(label="Highres Denoise", minimum=0.1, maximum=0.9, value=0.5, step=0.01)
|
351 |
-
a_prompt = gr.Textbox(label="Added Prompt", value='best quality')
|
352 |
-
n_prompt = gr.Textbox(label="Negative Prompt", value='lowres, bad anatomy, bad hands, cropped, worst quality')
|
353 |
-
normal_button = gr.Button(value="Compute Normal (4x Slower)")
|
354 |
-
|
355 |
-
with gr.Column():
|
356 |
-
result_video = gr.Video(label='Output Video', height=600, width=600, visible=True)
|
357 |
-
|
358 |
-
# 输入列表
|
359 |
-
# ips = [input_fg, input_bg, prompt, video_width, video_height, num_samples, seed, steps, a_prompt, n_prompt, cfg, highres_scale, highres_denoise, bg_source]
|
360 |
-
ips = [input_fg, input_bg]
|
361 |
-
|
362 |
-
# 按钮绑定处理函数
|
363 |
-
# relight_button.click(fn=lambda: None, inputs=[], outputs=[result_video])
|
364 |
-
|
365 |
-
relight_button.click(fn=dummy_process, inputs=ips, outputs=[result_video])
|
366 |
-
|
367 |
-
normal_button.click(fn=dummy_process, inputs=ips, outputs=[result_video])
|
368 |
-
|
369 |
-
# 背景库选择
|
370 |
-
def bg_gallery_selected(gal, evt: gr.SelectData):
|
371 |
-
# import pdb; pdb.set_trace()
|
372 |
-
# img_path = gal[evt.index][0]
|
373 |
-
img_path = db_examples.bg_samples[evt.index]
|
374 |
-
video_path = img_path.replace('frames/0000.png', 'cropped_video.mp4')
|
375 |
-
return video_path
|
376 |
-
|
377 |
-
bg_gallery.select(bg_gallery_selected, inputs=bg_gallery, outputs=input_bg)
|
378 |
-
|
379 |
-
# 示例
|
380 |
-
# dummy_video_for_outputs = gr.Video(visible=False, label='Result')
|
381 |
-
gr.Examples(
|
382 |
-
fn=lambda *args: args[-1],
|
383 |
-
examples=db_examples.background_conditioned_examples,
|
384 |
-
inputs=[input_fg, input_bg, prompt, bg_source, video_width, video_height, seed, result_video],
|
385 |
-
outputs=[result_video],
|
386 |
-
run_on_click=True, examples_per_page=1024
|
387 |
-
)
|
388 |
-
|
389 |
-
# 启动 Gradio 应用
|
390 |
-
# block.launch(server_name='0.0.0.0', server_port=10002, share=True)
|
391 |
-
block.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|