Spaces:
Running
on
L40S
Running
on
L40S
svfr (#1)
Browse files- (a4ce973fcbf35fcb2439184e4113aa6f190aa210)
- update (081a49f6f59d76700471269a023d40dd3658aa7b)
- update (f4e85a8429c3aed1881b7fd7cb6dff187c470d29)
Co-authored-by: minixiami <[email protected]>
- ORIGINAL_README.md +6 -0
- README.md +3 -1
- app.py +385 -40
- infer.py +29 -5
- src/dataset/dataset.py +68 -0
ORIGINAL_README.md
CHANGED
@@ -108,6 +108,8 @@ https://github.com/user-attachments/assets/efdac23c-0ba5-4dad-ab8c-48904af5dd89
|
|
108 |
|
109 |
## 🚀 Getting Started
|
110 |
|
|
|
|
|
111 |
## Setup
|
112 |
|
113 |
Use the following command to install a conda environment for SVFR from scratch:
|
@@ -200,6 +202,10 @@ The code of SVFR is released under the MIT License. There is no limitation for b
|
|
200 |
|
201 |
**The pretrained models we provided with this library are available for non-commercial research purposes only, including both auto-downloading models and manual-downloading models.**
|
202 |
|
|
|
|
|
|
|
|
|
203 |
|
204 |
## BibTex
|
205 |
```
|
|
|
108 |
|
109 |
## 🚀 Getting Started
|
110 |
|
111 |
+
> Note: It is recommended to use a GPU with 16GB or more VRAM.
|
112 |
+
|
113 |
## Setup
|
114 |
|
115 |
Use the following command to install a conda environment for SVFR from scratch:
|
|
|
202 |
|
203 |
**The pretrained models we provided with this library are available for non-commercial research purposes only, including both auto-downloading models and manual-downloading models.**
|
204 |
|
205 |
+
## Acknowledgments
|
206 |
+
|
207 |
+
This work is built on the architecture of [Sonic](https://github.com/jixiaozhong/Sonic).
|
208 |
+
|
209 |
|
210 |
## BibTex
|
211 |
```
|
README.md
CHANGED
@@ -10,4 +10,6 @@ pinned: false
|
|
10 |
short_description: Unified Framework for Generalized Video Face Restoration
|
11 |
---
|
12 |
|
13 |
-
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
|
|
|
10 |
short_description: Unified Framework for Generalized Video Face Restoration
|
11 |
---
|
12 |
|
13 |
+
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
14 |
+
|
15 |
+
|
app.py
CHANGED
@@ -1,3 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import torch
|
2 |
import sys
|
3 |
import os
|
@@ -8,6 +20,45 @@ import uuid
|
|
8 |
import gradio as gr
|
9 |
from glob import glob
|
10 |
from huggingface_hub import snapshot_download
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
# Download models
|
13 |
os.makedirs("models", exist_ok=True)
|
@@ -30,44 +81,326 @@ snapshot_download(
|
|
30 |
local_dir = "./models/stable-video-diffusion-img2vid-xt"
|
31 |
)
|
32 |
|
33 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
|
35 |
unique_id = str(uuid.uuid4())
|
36 |
output_dir = f"results_{unique_id}"
|
37 |
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
|
|
|
|
44 |
|
45 |
try:
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
)
|
57 |
|
58 |
# Search for the mp4 file in a subfolder of output_dir
|
59 |
-
output_video = glob(os.path.join(output_dir,"
|
60 |
-
|
|
|
61 |
|
62 |
if output_video:
|
63 |
output_video_path = output_video[0] # Get the first match
|
|
|
64 |
else:
|
65 |
output_video_path = None
|
|
|
66 |
|
67 |
-
print(output_video_path)
|
68 |
-
|
|
|
69 |
|
70 |
except subprocess.CalledProcessError as e:
|
|
|
71 |
raise gr.Error(f"Error during inference: {str(e)}")
|
72 |
|
73 |
css="""
|
@@ -91,38 +424,50 @@ with gr.Blocks(css=css) as demo:
|
|
91 |
<a href="https://arxiv.org/pdf/2501.01235">
|
92 |
<img src='https://img.shields.io/badge/ArXiv-Paper-red'>
|
93 |
</a>
|
94 |
-
<a href="https://huggingface.co/spaces/fffiloni/SVFR-demo?duplicate=true">
|
95 |
-
<img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/duplicate-this-space-sm.svg" alt="Duplicate this Space">
|
96 |
-
</a>
|
97 |
-
<a href="https://huggingface.co/fffiloni">
|
98 |
-
<img src="https://huggingface.co/datasets/huggingface/badges/resolve/main/follow-me-on-HF-sm-dark.svg" alt="Follow me on HF">
|
99 |
-
</a>
|
100 |
</div>
|
101 |
""")
|
102 |
with gr.Row():
|
103 |
with gr.Column():
|
104 |
input_seq = gr.Video(label="Video LQ")
|
105 |
-
task_name = gr.
|
106 |
label="Task",
|
107 |
-
choices=["BFR", "
|
108 |
-
value="BFR"
|
109 |
)
|
110 |
-
|
|
|
|
|
|
|
|
|
|
|
111 |
with gr.Column():
|
|
|
112 |
output_res = gr.Video(label="Restored")
|
113 |
gr.Examples(
|
114 |
examples = [
|
115 |
-
["./assert/lq/lq1.mp4", "BFR"],
|
116 |
-
["./assert/lq/lq2.mp4", "BFR
|
117 |
-
["./assert/lq/lq3.mp4", "
|
118 |
],
|
119 |
-
inputs = [input_seq, task_name]
|
120 |
)
|
121 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
122 |
submit_btn.click(
|
123 |
fn = infer,
|
124 |
-
inputs = [input_seq, task_name],
|
125 |
-
outputs = [output_res]
|
|
|
|
|
|
|
|
|
|
|
126 |
)
|
127 |
|
128 |
-
demo.queue().launch(show_api=False, show_error=True)
|
|
|
1 |
+
"""
|
2 |
+
This script is based on the original project by https://huggingface.co/fffiloni.
|
3 |
+
URL: https://huggingface.co/spaces/fffiloni/SVFR-demo/blob/main/app.py
|
4 |
+
|
5 |
+
Modifications made:
|
6 |
+
- Synced the infer code updates from GitHub repo.
|
7 |
+
- Added an inpainting option to enhance functionality.
|
8 |
+
|
9 |
+
Author of modifications: https://github.com/wangzhiyaoo
|
10 |
+
Date: 2025/01/15
|
11 |
+
"""
|
12 |
+
|
13 |
import torch
|
14 |
import sys
|
15 |
import os
|
|
|
20 |
import gradio as gr
|
21 |
from glob import glob
|
22 |
from huggingface_hub import snapshot_download
|
23 |
+
import random
|
24 |
+
|
25 |
+
import argparse
|
26 |
+
import warnings
|
27 |
+
import os
|
28 |
+
import numpy as np
|
29 |
+
import torch
|
30 |
+
import torch.utils.checkpoint
|
31 |
+
from PIL import Image
|
32 |
+
import random
|
33 |
+
|
34 |
+
from omegaconf import OmegaConf
|
35 |
+
from diffusers import AutoencoderKLTemporalDecoder
|
36 |
+
from diffusers.schedulers import EulerDiscreteScheduler
|
37 |
+
from transformers import CLIPVisionModelWithProjection
|
38 |
+
import torchvision.transforms as transforms
|
39 |
+
import torch.nn.functional as F
|
40 |
+
from src.models.svfr_adapter.unet_3d_svd_condition_ip import UNet3DConditionSVDModel
|
41 |
+
|
42 |
+
# pipeline
|
43 |
+
from src.pipelines.pipeline import LQ2VideoLongSVDPipeline
|
44 |
+
|
45 |
+
from src.utils.util import (
|
46 |
+
save_videos_grid,
|
47 |
+
seed_everything,
|
48 |
+
)
|
49 |
+
from torchvision.utils import save_image
|
50 |
+
|
51 |
+
from src.models.id_proj import IDProjConvModel
|
52 |
+
from src.models import model_insightface_360k
|
53 |
+
|
54 |
+
from src.dataset.face_align.align import AlignImage
|
55 |
+
|
56 |
+
warnings.filterwarnings("ignore")
|
57 |
+
|
58 |
+
import decord
|
59 |
+
import cv2
|
60 |
+
from src.dataset.dataset import get_affine_transform, mean_face_lm5p_256, get_union_bbox, process_bbox, crop_resize_img
|
61 |
+
|
62 |
|
63 |
# Download models
|
64 |
os.makedirs("models", exist_ok=True)
|
|
|
81 |
local_dir = "./models/stable-video-diffusion-img2vid-xt"
|
82 |
)
|
83 |
|
84 |
+
BASE_DIR = '.'
|
85 |
+
|
86 |
+
config = OmegaConf.load("./config/infer.yaml")
|
87 |
+
vae = AutoencoderKLTemporalDecoder.from_pretrained(
|
88 |
+
f"{BASE_DIR}/{config.pretrained_model_name_or_path}",
|
89 |
+
subfolder="vae",
|
90 |
+
variant="fp16")
|
91 |
+
|
92 |
+
val_noise_scheduler = EulerDiscreteScheduler.from_pretrained(
|
93 |
+
f"{BASE_DIR}/{config.pretrained_model_name_or_path}",
|
94 |
+
subfolder="scheduler")
|
95 |
+
|
96 |
+
image_encoder = CLIPVisionModelWithProjection.from_pretrained(
|
97 |
+
f"{BASE_DIR}/{config.pretrained_model_name_or_path}",
|
98 |
+
subfolder="image_encoder",
|
99 |
+
variant="fp16")
|
100 |
+
unet = UNet3DConditionSVDModel.from_pretrained(
|
101 |
+
f"{BASE_DIR}/{config.pretrained_model_name_or_path}",
|
102 |
+
subfolder="unet",
|
103 |
+
variant="fp16")
|
104 |
+
|
105 |
+
weight_dir = 'models/face_align'
|
106 |
+
det_path = os.path.join(BASE_DIR, weight_dir, 'yoloface_v5m.pt')
|
107 |
+
align_instance = AlignImage("cuda", det_path=det_path)
|
108 |
+
|
109 |
+
to_tensor = transforms.Compose([
|
110 |
+
transforms.ToTensor(),
|
111 |
+
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
|
112 |
+
])
|
113 |
+
|
114 |
+
import torch.nn as nn
|
115 |
+
class InflatedConv3d(nn.Conv2d):
|
116 |
+
def forward(self, x):
|
117 |
+
x = super().forward(x)
|
118 |
+
return x
|
119 |
+
# Add ref channel
|
120 |
+
old_weights = unet.conv_in.weight
|
121 |
+
old_bias = unet.conv_in.bias
|
122 |
+
new_conv1 = InflatedConv3d(
|
123 |
+
12,
|
124 |
+
old_weights.shape[0],
|
125 |
+
kernel_size=unet.conv_in.kernel_size,
|
126 |
+
stride=unet.conv_in.stride,
|
127 |
+
padding=unet.conv_in.padding,
|
128 |
+
bias=True if old_bias is not None else False,
|
129 |
+
)
|
130 |
+
param = torch.zeros((320, 4, 3, 3), requires_grad=True)
|
131 |
+
new_conv1.weight = torch.nn.Parameter(torch.cat((old_weights, param), dim=1))
|
132 |
+
if old_bias is not None:
|
133 |
+
new_conv1.bias = old_bias
|
134 |
+
unet.conv_in = new_conv1
|
135 |
+
unet.config["in_channels"] = 12
|
136 |
+
unet.config.in_channels = 12
|
137 |
+
|
138 |
+
|
139 |
+
id_linear = IDProjConvModel(in_channels=512, out_channels=1024).to(device='cuda')
|
140 |
+
|
141 |
+
# load pretrained weights
|
142 |
+
unet_checkpoint_path = os.path.join(BASE_DIR, config.unet_checkpoint_path)
|
143 |
+
unet.load_state_dict(
|
144 |
+
torch.load(unet_checkpoint_path, map_location="cpu"),
|
145 |
+
strict=True,
|
146 |
+
)
|
147 |
+
|
148 |
+
id_linear_checkpoint_path = os.path.join(BASE_DIR, config.id_linear_checkpoint_path)
|
149 |
+
id_linear.load_state_dict(
|
150 |
+
torch.load(id_linear_checkpoint_path, map_location="cpu"),
|
151 |
+
strict=True,
|
152 |
+
)
|
153 |
+
|
154 |
+
net_arcface = model_insightface_360k.getarcface(f'{BASE_DIR}/{config.net_arcface_checkpoint_path}').eval().to(device="cuda")
|
155 |
+
|
156 |
+
if config.weight_dtype == "fp16":
|
157 |
+
weight_dtype = torch.float16
|
158 |
+
elif config.weight_dtype == "fp32":
|
159 |
+
weight_dtype = torch.float32
|
160 |
+
elif config.weight_dtype == "bf16":
|
161 |
+
weight_dtype = torch.bfloat16
|
162 |
+
else:
|
163 |
+
raise ValueError(
|
164 |
+
f"Do not support weight dtype: {config.weight_dtype} during training"
|
165 |
+
)
|
166 |
+
|
167 |
+
image_encoder.to(weight_dtype)
|
168 |
+
vae.to(weight_dtype)
|
169 |
+
unet.to(weight_dtype)
|
170 |
+
id_linear.to(weight_dtype)
|
171 |
+
net_arcface.requires_grad_(False).to(weight_dtype)
|
172 |
+
|
173 |
+
pipe = LQ2VideoLongSVDPipeline(
|
174 |
+
unet=unet,
|
175 |
+
image_encoder=image_encoder,
|
176 |
+
vae=vae,
|
177 |
+
scheduler=val_noise_scheduler,
|
178 |
+
feature_extractor=None
|
179 |
+
|
180 |
+
)
|
181 |
+
pipe = pipe.to("cuda", dtype=unet.dtype)
|
182 |
+
|
183 |
+
def gen(args,pipe):
|
184 |
+
save_dir = f"{BASE_DIR}/{args.output_dir}"
|
185 |
+
os.makedirs(save_dir,exist_ok=True)
|
186 |
+
|
187 |
+
seed_input = args.seed
|
188 |
+
seed_everything(seed_input)
|
189 |
+
|
190 |
+
video_path = args.input_path
|
191 |
+
task_ids = args.task_ids
|
192 |
+
|
193 |
+
if 2 in task_ids and args.mask_path is not None:
|
194 |
+
mask_path = args.mask_path
|
195 |
+
mask = Image.open(mask_path).convert("L")
|
196 |
+
mask_array = np.array(mask)
|
197 |
+
|
198 |
+
white_positions = mask_array == 255
|
199 |
+
|
200 |
+
print('task_ids:',task_ids)
|
201 |
+
task_prompt = [0,0,0]
|
202 |
+
for i in range(3):
|
203 |
+
if i in task_ids:
|
204 |
+
task_prompt[i] = 1
|
205 |
+
print("task_prompt:",task_prompt)
|
206 |
+
|
207 |
+
video_name = video_path.split('/')[-1]
|
208 |
+
# print(video_name)
|
209 |
+
|
210 |
+
if os.path.exists(os.path.join(save_dir, "result_frames", video_name[:-4])):
|
211 |
+
print(os.path.join(save_dir, "result_frames", video_name[:-4]))
|
212 |
+
# continue
|
213 |
+
|
214 |
+
cap = decord.VideoReader(video_path, fault_tol=1)
|
215 |
+
total_frames = len(cap)
|
216 |
+
T = total_frames #
|
217 |
+
print("total_frames:",total_frames)
|
218 |
+
step=1
|
219 |
+
drive_idx_start = 0
|
220 |
+
drive_idx_list = list(range(drive_idx_start, drive_idx_start + T * step, step))
|
221 |
+
assert len(drive_idx_list) == T
|
222 |
+
|
223 |
+
# Crop faces from the video for further processing
|
224 |
+
bbox_list = []
|
225 |
+
frame_interval = 5
|
226 |
+
for frame_count, drive_idx in enumerate(drive_idx_list):
|
227 |
+
if frame_count % frame_interval != 0:
|
228 |
+
continue
|
229 |
+
frame = cap[drive_idx].asnumpy()
|
230 |
+
_, _, bboxes_list = align_instance(frame[:,:,[2,1,0]], maxface=True)
|
231 |
+
if bboxes_list==[]:
|
232 |
+
continue
|
233 |
+
x1, y1, ww, hh = bboxes_list[0]
|
234 |
+
x2, y2 = x1 + ww, y1 + hh
|
235 |
+
bbox = [x1, y1, x2, y2]
|
236 |
+
bbox_list.append(bbox)
|
237 |
+
bbox = get_union_bbox(bbox_list)
|
238 |
+
bbox_s = process_bbox(bbox, expand_radio=0.4, height=frame.shape[0], width=frame.shape[1])
|
239 |
+
|
240 |
+
imSameIDs = []
|
241 |
+
vid_gt = []
|
242 |
+
for i, drive_idx in enumerate(drive_idx_list):
|
243 |
+
frame = cap[drive_idx].asnumpy()
|
244 |
+
imSameID = Image.fromarray(frame)
|
245 |
+
imSameID = crop_resize_img(imSameID, bbox_s)
|
246 |
+
imSameID = imSameID.resize((512,512))
|
247 |
+
if 1 in task_ids:
|
248 |
+
imSameID = imSameID.convert("L") # Convert to grayscale
|
249 |
+
imSameID = imSameID.convert("RGB")
|
250 |
+
image_array = np.array(imSameID)
|
251 |
+
if 2 in task_ids and args.mask_path is not None:
|
252 |
+
image_array[white_positions] = [255, 255, 255] # mask for inpainting task
|
253 |
+
vid_gt.append(np.float32(image_array/255.))
|
254 |
+
imSameIDs.append(imSameID)
|
255 |
+
|
256 |
+
vid_lq = [(torch.from_numpy(frame).permute(2,0,1) - 0.5) / 0.5 for frame in vid_gt]
|
257 |
+
|
258 |
+
val_data = dict(
|
259 |
+
pixel_values_vid_lq = torch.stack(vid_lq,dim=0),
|
260 |
+
# pixel_values_ref_img=self.to_tensor(target_image),
|
261 |
+
# pixel_values_ref_concat_img=self.to_tensor(imSrc2),
|
262 |
+
task_ids=task_ids,
|
263 |
+
task_id_input=torch.tensor(task_prompt),
|
264 |
+
total_frames=total_frames,
|
265 |
+
)
|
266 |
+
|
267 |
+
window_overlap=0
|
268 |
+
inter_frame_list = get_overlap_slide_window_indices(val_data["total_frames"],config.data.n_sample_frames,window_overlap)
|
269 |
+
|
270 |
+
lq_frames = val_data["pixel_values_vid_lq"]
|
271 |
+
task_ids = val_data["task_ids"]
|
272 |
+
task_id_input = val_data["task_id_input"]
|
273 |
+
height, width = val_data["pixel_values_vid_lq"].shape[-2:]
|
274 |
+
|
275 |
+
print("Generating the first clip...")
|
276 |
+
output = pipe(
|
277 |
+
lq_frames[inter_frame_list[0]].to("cuda").to(weight_dtype), # lq
|
278 |
+
None, # ref concat
|
279 |
+
torch.zeros((1, len(inter_frame_list[0]), 49, 1024)).to("cuda").to(weight_dtype),# encoder_hidden_states
|
280 |
+
task_id_input.to("cuda").to(weight_dtype),
|
281 |
+
height=height,
|
282 |
+
width=width,
|
283 |
+
num_frames=len(inter_frame_list[0]),
|
284 |
+
decode_chunk_size=config.decode_chunk_size,
|
285 |
+
noise_aug_strength=config.noise_aug_strength,
|
286 |
+
min_guidance_scale=config.min_appearance_guidance_scale,
|
287 |
+
max_guidance_scale=config.max_appearance_guidance_scale,
|
288 |
+
overlap=config.overlap,
|
289 |
+
frames_per_batch=len(inter_frame_list[0]),
|
290 |
+
num_inference_steps=50,
|
291 |
+
i2i_noise_strength=config.i2i_noise_strength,
|
292 |
+
)
|
293 |
+
video = output.frames
|
294 |
+
|
295 |
+
ref_img_tensor = video[0][:,-1]
|
296 |
+
ref_img = (video[0][:,-1] *0.5+0.5).clamp(0,1) * 255.
|
297 |
+
ref_img = ref_img.permute(1,2,0).cpu().numpy().astype(np.uint8)
|
298 |
+
|
299 |
+
pts5 = align_instance(ref_img[:,:,[2,1,0]], maxface=True)[0][0]
|
300 |
+
|
301 |
+
warp_mat = get_affine_transform(pts5, mean_face_lm5p_256 * height/256)
|
302 |
+
ref_img = cv2.warpAffine(np.array(Image.fromarray(ref_img)), warp_mat, (height, width), flags=cv2.INTER_CUBIC)
|
303 |
+
ref_img = to_tensor(ref_img).to("cuda").to(weight_dtype)
|
304 |
+
|
305 |
+
save_image(ref_img*0.5 + 0.5,f"{save_dir}/ref_img_align.png")
|
306 |
+
|
307 |
+
ref_img = F.interpolate(ref_img.unsqueeze(0)[:, :, 0:224, 16:240], size=[112, 112], mode='bilinear')
|
308 |
+
_, id_feature_conv = net_arcface(ref_img)
|
309 |
+
id_embedding = id_linear(id_feature_conv)
|
310 |
+
|
311 |
+
print('Generating all video clips...')
|
312 |
+
video = pipe(
|
313 |
+
lq_frames.to("cuda").to(weight_dtype), # lq
|
314 |
+
ref_img_tensor.to("cuda").to(weight_dtype),
|
315 |
+
id_embedding.unsqueeze(1).repeat(1, len(lq_frames), 1, 1).to("cuda").to(weight_dtype), # encoder_hidden_states
|
316 |
+
task_id_input.to("cuda").to(weight_dtype),
|
317 |
+
height=height,
|
318 |
+
width=width,
|
319 |
+
num_frames=val_data["total_frames"],#frame_num,
|
320 |
+
decode_chunk_size=config.decode_chunk_size,
|
321 |
+
noise_aug_strength=config.noise_aug_strength,
|
322 |
+
min_guidance_scale=config.min_appearance_guidance_scale,
|
323 |
+
max_guidance_scale=config.max_appearance_guidance_scale,
|
324 |
+
overlap=config.overlap,
|
325 |
+
frames_per_batch=config.data.n_sample_frames,
|
326 |
+
num_inference_steps=config.num_inference_steps,
|
327 |
+
i2i_noise_strength=config.i2i_noise_strength,
|
328 |
+
).frames
|
329 |
+
|
330 |
+
|
331 |
+
video = (video*0.5 + 0.5).clamp(0, 1)
|
332 |
+
video = torch.cat([video.to(device="cuda")], dim=0).cpu()
|
333 |
+
save_videos_grid(video, f"{save_dir}/{video_name[:-4]}_{seed_input}_gen.mp4", n_rows=1, fps=25)
|
334 |
+
|
335 |
+
lq_frames = lq_frames.permute(1,0,2,3).unsqueeze(0)
|
336 |
+
lq_frames = (lq_frames * 0.5 + 0.5).clamp(0, 1).to(device="cuda").cpu()
|
337 |
+
save_videos_grid(lq_frames, f"{save_dir}/{video_name[:-4]}_{seed_input}_ori.mp4", n_rows=1, fps=25)
|
338 |
+
|
339 |
+
if args.restore_frames:
|
340 |
+
video = video.squeeze(0)
|
341 |
+
os.makedirs(os.path.join(save_dir, "result_frames", f"{video_name[:-4]}_{seed_input}"),exist_ok=True)
|
342 |
+
print(os.path.join(save_dir, "result_frames", video_name[:-4]))
|
343 |
+
for i in range(video.shape[1]):
|
344 |
+
save_frames_path = os.path.join(f"{save_dir}/result_frames", f"{video_name[:-4]}_{seed_input}", f'{i:08d}.png')
|
345 |
+
save_image(video[:,i], save_frames_path)
|
346 |
+
|
347 |
+
|
348 |
+
def get_overlap_slide_window_indices(video_length, window_size, window_overlap):
|
349 |
+
inter_frame_list = []
|
350 |
+
for j in range(0, video_length, window_size-window_overlap):
|
351 |
+
inter_frame_list.append( [e % video_length for e in range(j, min(j + window_size, video_length))] )
|
352 |
+
|
353 |
+
return inter_frame_list
|
354 |
+
|
355 |
+
|
356 |
+
|
357 |
+
def random_seed():
|
358 |
+
return random.randint(0, 10000)
|
359 |
+
|
360 |
+
def infer(lq_sequence, task_name, mask, seed):
|
361 |
|
362 |
unique_id = str(uuid.uuid4())
|
363 |
output_dir = f"results_{unique_id}"
|
364 |
|
365 |
+
task_mapping = {
|
366 |
+
"BFR": 0,
|
367 |
+
"Colorization": 1,
|
368 |
+
"Inpainting": 2
|
369 |
+
}
|
370 |
+
|
371 |
+
task_ids = [task_mapping[task] for task in task_name if task in task_mapping]
|
372 |
+
# task_id = ",".join(task_ids)
|
373 |
|
374 |
try:
|
375 |
+
parser = argparse.ArgumentParser()
|
376 |
+
args = parser.parse_args()
|
377 |
+
args.task_ids = task_ids
|
378 |
+
args.input_path = f"{lq_sequence}"
|
379 |
+
args.output_dir = f"{output_dir}"
|
380 |
+
args.mask_path = f"{mask}"
|
381 |
+
args.seed = int(seed)
|
382 |
+
args.restore_frames = False
|
383 |
+
|
384 |
+
gen(args,pipe)
|
|
|
385 |
|
386 |
# Search for the mp4 file in a subfolder of output_dir
|
387 |
+
output_video = glob(os.path.join(output_dir,"*gen.mp4"))
|
388 |
+
face_region_video = glob(os.path.join(output_dir,"*ori.mp4"))
|
389 |
+
# print(face_region_video,output_video)
|
390 |
|
391 |
if output_video:
|
392 |
output_video_path = output_video[0] # Get the first match
|
393 |
+
face_region_video_path = face_region_video[0] # Get the first match
|
394 |
else:
|
395 |
output_video_path = None
|
396 |
+
face_region_video = None
|
397 |
|
398 |
+
print(output_video_path,face_region_video_path)
|
399 |
+
torch.cuda.empty_cache()
|
400 |
+
return face_region_video_path,output_video_path
|
401 |
|
402 |
except subprocess.CalledProcessError as e:
|
403 |
+
torch.cuda.empty_cache()
|
404 |
raise gr.Error(f"Error during inference: {str(e)}")
|
405 |
|
406 |
css="""
|
|
|
424 |
<a href="https://arxiv.org/pdf/2501.01235">
|
425 |
<img src='https://img.shields.io/badge/ArXiv-Paper-red'>
|
426 |
</a>
|
|
|
|
|
|
|
|
|
|
|
|
|
427 |
</div>
|
428 |
""")
|
429 |
with gr.Row():
|
430 |
with gr.Column():
|
431 |
input_seq = gr.Video(label="Video LQ")
|
432 |
+
task_name = gr.CheckboxGroup(
|
433 |
label="Task",
|
434 |
+
choices=["BFR", "Colorization", "Inpainting"],
|
435 |
+
value=["BFR"] # default
|
436 |
)
|
437 |
+
mask_input = gr.Image(type="filepath",label="Inpainting Mask")
|
438 |
+
with gr.Row():
|
439 |
+
seed_input = gr.Number(label="Seed", value=77, precision=0)
|
440 |
+
random_seed_btn = gr.Button("🎲",scale=1,elem_id="dice-btn")
|
441 |
+
submit_btn = gr.Button("Submit", variant="primary")
|
442 |
+
clear_btn = gr.Button("Clear")
|
443 |
with gr.Column():
|
444 |
+
output_face = gr.Video(label="Face Region Input")
|
445 |
output_res = gr.Video(label="Restored")
|
446 |
gr.Examples(
|
447 |
examples = [
|
448 |
+
["./assert/lq/lq1.mp4", ["BFR"],None],
|
449 |
+
["./assert/lq/lq2.mp4", ["BFR", "Colorization"],None],
|
450 |
+
["./assert/lq/lq3.mp4", ["BFR", "Colorization", "Inpainting"],"./assert/mask/lq3.png"]
|
451 |
],
|
452 |
+
inputs = [input_seq, task_name, mask_input]
|
453 |
)
|
454 |
+
|
455 |
+
random_seed_btn.click(
|
456 |
+
fn=random_seed,
|
457 |
+
inputs=[],
|
458 |
+
outputs=seed_input
|
459 |
+
)
|
460 |
+
|
461 |
+
|
462 |
submit_btn.click(
|
463 |
fn = infer,
|
464 |
+
inputs = [input_seq, task_name, mask_input,seed_input],
|
465 |
+
outputs = [output_face,output_res]
|
466 |
+
)
|
467 |
+
clear_btn.click(
|
468 |
+
fn=lambda: [None,["BFR"],None,77,None,None],
|
469 |
+
inputs=None,
|
470 |
+
outputs=[input_seq, task_name, mask_input, seed_input, output_face, output_res]
|
471 |
)
|
472 |
|
473 |
+
demo.queue().launch(show_api=False, show_error=True, server_port=1203)
|
infer.py
CHANGED
@@ -33,10 +33,11 @@ warnings.filterwarnings("ignore")
|
|
33 |
|
34 |
import decord
|
35 |
import cv2
|
36 |
-
from src.dataset.dataset import get_affine_transform, mean_face_lm5p_256
|
37 |
|
38 |
BASE_DIR = '.'
|
39 |
|
|
|
40 |
def main(config,args):
|
41 |
if 'CUDA_VISIBLE_DEVICES' in os.environ:
|
42 |
cuda_visible_devices = os.environ['CUDA_VISIBLE_DEVICES']
|
@@ -179,13 +180,33 @@ def main(config,args):
|
|
179 |
drive_idx_list = list(range(drive_idx_start, drive_idx_start + T * step, step))
|
180 |
assert len(drive_idx_list) == T
|
181 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
182 |
imSameIDs = []
|
183 |
vid_gt = []
|
184 |
for i, drive_idx in enumerate(drive_idx_list):
|
185 |
frame = cap[drive_idx].asnumpy()
|
186 |
imSameID = Image.fromarray(frame)
|
187 |
-
|
188 |
imSameID = imSameID.resize((512,512))
|
|
|
|
|
|
|
189 |
image_array = np.array(imSameID)
|
190 |
if 2 in task_ids and args.mask_path is not None:
|
191 |
image_array[white_positions] = [255, 255, 255] # mask for inpainting task
|
@@ -241,7 +262,7 @@ def main(config,args):
|
|
241 |
ref_img = cv2.warpAffine(np.array(Image.fromarray(ref_img)), warp_mat, (height, width), flags=cv2.INTER_CUBIC)
|
242 |
ref_img = to_tensor(ref_img).to("cuda").to(weight_dtype)
|
243 |
|
244 |
-
save_image(ref_img*0.5 + 0.5,f"{save_dir}/ref_img_align.png")
|
245 |
|
246 |
ref_img = F.interpolate(ref_img.unsqueeze(0)[:, :, 0:224, 16:240], size=[112, 112], mode='bilinear')
|
247 |
_, id_feature_conv = net_arcface(ref_img)
|
@@ -269,8 +290,11 @@ def main(config,args):
|
|
269 |
|
270 |
video = (video*0.5 + 0.5).clamp(0, 1)
|
271 |
video = torch.cat([video.to(device="cuda")], dim=0).cpu()
|
272 |
-
|
273 |
-
|
|
|
|
|
|
|
274 |
|
275 |
if args.restore_frames:
|
276 |
video = video.squeeze(0)
|
|
|
33 |
|
34 |
import decord
|
35 |
import cv2
|
36 |
+
from src.dataset.dataset import get_affine_transform, mean_face_lm5p_256, get_union_bbox, process_bbox, crop_resize_img
|
37 |
|
38 |
BASE_DIR = '.'
|
39 |
|
40 |
+
|
41 |
def main(config,args):
|
42 |
if 'CUDA_VISIBLE_DEVICES' in os.environ:
|
43 |
cuda_visible_devices = os.environ['CUDA_VISIBLE_DEVICES']
|
|
|
180 |
drive_idx_list = list(range(drive_idx_start, drive_idx_start + T * step, step))
|
181 |
assert len(drive_idx_list) == T
|
182 |
|
183 |
+
# Crop faces from the video for further processing
|
184 |
+
bbox_list = []
|
185 |
+
frame_interval = 5
|
186 |
+
for frame_count, drive_idx in enumerate(drive_idx_list):
|
187 |
+
if frame_count % frame_interval != 0:
|
188 |
+
continue
|
189 |
+
frame = cap[drive_idx].asnumpy()
|
190 |
+
_, _, bboxes_list = align_instance(frame[:,:,[2,1,0]], maxface=True)
|
191 |
+
if bboxes_list==[]:
|
192 |
+
continue
|
193 |
+
x1, y1, ww, hh = bboxes_list[0]
|
194 |
+
x2, y2 = x1 + ww, y1 + hh
|
195 |
+
bbox = [x1, y1, x2, y2]
|
196 |
+
bbox_list.append(bbox)
|
197 |
+
bbox = get_union_bbox(bbox_list)
|
198 |
+
bbox_s = process_bbox(bbox, expand_radio=0.4, height=frame.shape[0], width=frame.shape[1])
|
199 |
+
|
200 |
imSameIDs = []
|
201 |
vid_gt = []
|
202 |
for i, drive_idx in enumerate(drive_idx_list):
|
203 |
frame = cap[drive_idx].asnumpy()
|
204 |
imSameID = Image.fromarray(frame)
|
205 |
+
imSameID = crop_resize_img(imSameID, bbox_s)
|
206 |
imSameID = imSameID.resize((512,512))
|
207 |
+
if 1 in task_ids:
|
208 |
+
imSameID = imSameID.convert("L") # Convert to grayscale
|
209 |
+
imSameID = imSameID.convert("RGB")
|
210 |
image_array = np.array(imSameID)
|
211 |
if 2 in task_ids and args.mask_path is not None:
|
212 |
image_array[white_positions] = [255, 255, 255] # mask for inpainting task
|
|
|
262 |
ref_img = cv2.warpAffine(np.array(Image.fromarray(ref_img)), warp_mat, (height, width), flags=cv2.INTER_CUBIC)
|
263 |
ref_img = to_tensor(ref_img).to("cuda").to(weight_dtype)
|
264 |
|
265 |
+
# save_image(ref_img*0.5 + 0.5,f"{save_dir}/ref_img_align.png")
|
266 |
|
267 |
ref_img = F.interpolate(ref_img.unsqueeze(0)[:, :, 0:224, 16:240], size=[112, 112], mode='bilinear')
|
268 |
_, id_feature_conv = net_arcface(ref_img)
|
|
|
290 |
|
291 |
video = (video*0.5 + 0.5).clamp(0, 1)
|
292 |
video = torch.cat([video.to(device="cuda")], dim=0).cpu()
|
293 |
+
save_videos_grid(video, f"{save_dir}/{video_name[:-4]}_{seed_input}_gen.mp4", n_rows=1, fps=25)
|
294 |
+
|
295 |
+
lq_frames = lq_frames.permute(1,0,2,3).unsqueeze(0)
|
296 |
+
lq_frames = (lq_frames * 0.5 + 0.5).clamp(0, 1).to(device="cuda").cpu()
|
297 |
+
save_videos_grid(lq_frames, f"{save_dir}/{video_name[:-4]}_{seed_input}_ori.mp4", n_rows=1, fps=25)
|
298 |
|
299 |
if args.restore_frames:
|
300 |
video = video.squeeze(0)
|
src/dataset/dataset.py
CHANGED
@@ -48,3 +48,71 @@ def get_affine_transform(target_face_lm5p, mean_lm5p):
|
|
48 |
mat_warp[1][2] = mat23[3]
|
49 |
|
50 |
return mat_warp
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
mat_warp[1][2] = mat23[3]
|
49 |
|
50 |
return mat_warp
|
51 |
+
|
52 |
+
def get_union_bbox(bboxes):
|
53 |
+
bboxes = np.array(bboxes)
|
54 |
+
min_x = np.min(bboxes[:, 0])
|
55 |
+
min_y = np.min(bboxes[:, 1])
|
56 |
+
max_x = np.max(bboxes[:, 2])
|
57 |
+
max_y = np.max(bboxes[:, 3])
|
58 |
+
return np.array([min_x, min_y, max_x, max_y])
|
59 |
+
|
60 |
+
|
61 |
+
def process_bbox(bbox, expand_radio, height, width):
|
62 |
+
|
63 |
+
def expand(bbox, ratio, height, width):
|
64 |
+
|
65 |
+
bbox_h = bbox[3] - bbox[1]
|
66 |
+
bbox_w = bbox[2] - bbox[0]
|
67 |
+
|
68 |
+
expand_x1 = max(bbox[0] - ratio * bbox_w, 0)
|
69 |
+
expand_y1 = max(bbox[1] - ratio * bbox_h, 0)
|
70 |
+
expand_x2 = min(bbox[2] + ratio * bbox_w, width)
|
71 |
+
expand_y2 = min(bbox[3] + ratio * bbox_h, height)
|
72 |
+
|
73 |
+
return [expand_x1,expand_y1,expand_x2,expand_y2]
|
74 |
+
|
75 |
+
def to_square(bbox_src, bbox_expend, height, width):
|
76 |
+
|
77 |
+
h = bbox_expend[3] - bbox_expend[1]
|
78 |
+
w = bbox_expend[2] - bbox_expend[0]
|
79 |
+
c_h = (bbox_expend[1] + bbox_expend[3]) / 2
|
80 |
+
c_w = (bbox_expend[0] + bbox_expend[2]) / 2
|
81 |
+
|
82 |
+
c = min(h, w) / 2
|
83 |
+
|
84 |
+
c_src_h = (bbox_src[1] + bbox_src[3]) / 2
|
85 |
+
c_src_w = (bbox_src[0] + bbox_src[2]) / 2
|
86 |
+
|
87 |
+
s_h, s_w = 0, 0
|
88 |
+
if w < h:
|
89 |
+
d = abs((h - w) / 2)
|
90 |
+
s_h = min(d, abs(c_src_h-c_h))
|
91 |
+
s_h = s_h if c_src_h > c_h else s_h * (-1)
|
92 |
+
else:
|
93 |
+
d = abs((h - w) / 2)
|
94 |
+
s_w = min(d, abs(c_src_w-c_w))
|
95 |
+
s_w = s_w if c_src_w > c_w else s_w * (-1)
|
96 |
+
|
97 |
+
|
98 |
+
c_h = (bbox_expend[1] + bbox_expend[3]) / 2 + s_h
|
99 |
+
c_w = (bbox_expend[0] + bbox_expend[2]) / 2 + s_w
|
100 |
+
|
101 |
+
square_x1 = c_w - c
|
102 |
+
square_y1 = c_h - c
|
103 |
+
square_x2 = c_w + c
|
104 |
+
square_y2 = c_h + c
|
105 |
+
|
106 |
+
return [round(square_x1), round(square_y1), round(square_x2), round(square_y2)]
|
107 |
+
|
108 |
+
|
109 |
+
bbox_expend = expand(bbox, expand_radio, height=height, width=width)
|
110 |
+
processed_bbox = to_square(bbox, bbox_expend, height=height, width=width)
|
111 |
+
|
112 |
+
return processed_bbox
|
113 |
+
|
114 |
+
|
115 |
+
def crop_resize_img(img, bbox):
|
116 |
+
x1, y1, x2, y2 = bbox
|
117 |
+
img = img.crop((x1, y1, x2, y2))
|
118 |
+
return img
|