Spaces:
Running
on
L40S
Running
on
L40S
Create gradio_app.py
Browse files- gradio_app.py +137 -0
gradio_app.py
ADDED
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import os
|
3 |
+
import time
|
4 |
+
import argparse
|
5 |
+
from diffueraser.diffueraser import DiffuEraser
|
6 |
+
from propainter.inference import Propainter, get_device
|
7 |
+
|
8 |
+
# Download Weights
|
9 |
+
from huggingface_hub import snapshot_download
|
10 |
+
|
11 |
+
# List of subdirectories to create inside "checkpoints"
|
12 |
+
subfolders = [
|
13 |
+
"diffuEraser",
|
14 |
+
"stable-diffusion-v1-5",
|
15 |
+
"PCM_Weights",
|
16 |
+
"propainter",
|
17 |
+
"sd-vae-ft-mse"
|
18 |
+
]
|
19 |
+
# Create each subdirectory
|
20 |
+
for subfolder in subfolders:
|
21 |
+
os.makedirs(os.path.join("weigths", subfolder), exist_ok=True)
|
22 |
+
|
23 |
+
snapshot_download(
|
24 |
+
repo_id = "lixiaowen/diffuEraser",
|
25 |
+
local_dir = "./weights/diffuEraser"
|
26 |
+
)
|
27 |
+
|
28 |
+
snapshot_download(
|
29 |
+
repo_id = "stable-diffusion-v1-5/stable-diffusion-v1-5",
|
30 |
+
local_dir = "./weights/stable-diffusion-v1-5"
|
31 |
+
)
|
32 |
+
|
33 |
+
snapshot_download(
|
34 |
+
repo_id = "wangfuyun/PCM_Weights",
|
35 |
+
local_dir = "./weights/PCM_Weights"
|
36 |
+
)
|
37 |
+
|
38 |
+
snapshot_download(
|
39 |
+
repo_id = "camenduru/ProPainter",
|
40 |
+
local_dir = "./weights/propainter"
|
41 |
+
)
|
42 |
+
|
43 |
+
snapshot_download(
|
44 |
+
repo_id = "stabilityai/sd-vae-ft-mse",
|
45 |
+
local_dir = "./weights/sd-vae-ft-mse"
|
46 |
+
)
|
47 |
+
|
48 |
+
# βββββββββββββββββββββ
|
49 |
+
|
50 |
+
def infer(input_video, input_mask):
|
51 |
+
|
52 |
+
video_length = 10 # The maximum length of output video
|
53 |
+
mask_dilation_iter = 8 # Adjust it to change the degree of mask expansion
|
54 |
+
max_img_size = 960 # The maximum length of output width and height
|
55 |
+
save_path = "results" # Path to the output
|
56 |
+
|
57 |
+
ref_stride = 10
|
58 |
+
neighbor_length = 10
|
59 |
+
subvideo_length = 50
|
60 |
+
|
61 |
+
base_model_path = "weights/stable-diffusion-v1-5"
|
62 |
+
vae_path = "weights/sd-vae-ft-mse"
|
63 |
+
diffueraser_path = "weights/diffuEraser"
|
64 |
+
propainter_model_dir = "weights/propainter"
|
65 |
+
|
66 |
+
if not os.path.exists(save_path):
|
67 |
+
os.makedirs(save_path)
|
68 |
+
priori_path = os.path.join(save_path, "priori.mp4")
|
69 |
+
output_path = os.path.join(save_path, "diffueraser_result.mp4")
|
70 |
+
|
71 |
+
## model initialization
|
72 |
+
device = get_device()
|
73 |
+
# PCM params
|
74 |
+
ckpt = "2-Step"
|
75 |
+
video_inpainting_sd = DiffuEraser(device, base_model_path, vae_path, diffueraser_path, ckpt=ckpt)
|
76 |
+
propainter = Propainter(propainter_model_dir, device=device)
|
77 |
+
|
78 |
+
start_time = time.time()
|
79 |
+
|
80 |
+
## priori
|
81 |
+
propainter.forward(input_video, input_mask, priori_path, video_length=video_length,
|
82 |
+
ref_stride=ref_stride, neighbor_length=neighbor_length, subvideo_length = subvideo_length,
|
83 |
+
mask_dilation = mask_dilation_iter)
|
84 |
+
|
85 |
+
## diffueraser
|
86 |
+
guidance_scale = None # The default value is 0.
|
87 |
+
video_inpainting_sd.forward(input_video, input_mask, priori_path, output_path,
|
88 |
+
max_img_size = max_img_size, video_length=video_length, mask_dilation_iter=mask_dilation_iter,
|
89 |
+
guidance_scale=guidance_scale)
|
90 |
+
|
91 |
+
end_time = time.time()
|
92 |
+
inference_time = end_time - start_time
|
93 |
+
print(f"DiffuEraser inference time: {inference_time:.4f} s")
|
94 |
+
|
95 |
+
torch.cuda.empty_cache()
|
96 |
+
|
97 |
+
return output_path, priori_path
|
98 |
+
|
99 |
+
with gr.Blocks() as demo:
|
100 |
+
|
101 |
+
with gr.Column():
|
102 |
+
gr.Markdown("# DiffuEraser")
|
103 |
+
|
104 |
+
with gr.Row():
|
105 |
+
|
106 |
+
with gr.Column():
|
107 |
+
|
108 |
+
input_video = gr.Video(label="Input Video (MP4 ONLY")
|
109 |
+
input_mask = gr.Video(label="Input Mask Video (MP4 ONLY")
|
110 |
+
submit_btn = gr.Button("Submit")
|
111 |
+
|
112 |
+
gr.Examples(
|
113 |
+
examples = [
|
114 |
+
[".examples/example1/video.mp4", "./examples/example1/mask.mp4"],
|
115 |
+
[".examples/example2/video.mp4", "./examples/example2/mask.mp4"],
|
116 |
+
[".examples/example3/video.mp4", "./examples/example3/mask.mp4"],
|
117 |
+
],
|
118 |
+
inputs = [input_video, input_mask]
|
119 |
+
)
|
120 |
+
|
121 |
+
with gr.Column():
|
122 |
+
|
123 |
+
video_result = gr.Video(label="Result")
|
124 |
+
priori_result = gr.Video(label="ProPainter pass")
|
125 |
+
|
126 |
+
|
127 |
+
submit_btn.click(
|
128 |
+
fn = infer,
|
129 |
+
inputs = [input_video, input_mask],
|
130 |
+
outputs = [video_result, priori_result]
|
131 |
+
)
|
132 |
+
|
133 |
+
demo.queue().launch(show_api=False, show_error=True)
|
134 |
+
|
135 |
+
|
136 |
+
|
137 |
+
|