Spaces:
Running
on
Zero
Running
on
Zero
update
Browse files- app.py +4 -0
- assets/instruction.md +4 -4
- assets/title.md +1 -0
- gradio_tabs/animation.py +154 -53
- gradio_tabs/img_edit.py +87 -27
- gradio_tabs/vid_edit.py +157 -64
- networks/generator.py +40 -1
app.py
CHANGED
@@ -7,6 +7,10 @@ from gradio_tabs.vid_edit import vid_edit
|
|
7 |
from gradio_tabs.img_edit import img_edit
|
8 |
from networks.generator import Generator
|
9 |
|
|
|
|
|
|
|
|
|
10 |
device = torch.device("cuda")
|
11 |
gen = Generator(size=512, motion_dim=40, scale=2).to(device)
|
12 |
ckpt_path = hf_hub_download(repo_id="YaohuiW/LIA-X", filename="lia-x.pt")
|
|
|
7 |
from gradio_tabs.img_edit import img_edit
|
8 |
from networks.generator import Generator
|
9 |
|
10 |
+
# Optimize torch.compile performance
|
11 |
+
torch.set_float32_matmul_precision('high') # Enable TensorFloat32 for better performance
|
12 |
+
torch._dynamo.config.cache_size_limit = 64 # Increase cache size to reduce recompilations
|
13 |
+
|
14 |
device = torch.device("cuda")
|
15 |
gen = Generator(size=512, motion_dim=40, scale=2).to(device)
|
16 |
ckpt_path = hf_hub_download(repo_id="YaohuiW/LIA-X", filename="lia-x.pt")
|
assets/instruction.md
CHANGED
@@ -3,18 +3,18 @@
|
|
3 |
* **Image Animation**
|
4 |
|
5 |
- Upload `Source Image` and `Driving Video`
|
6 |
-
-
|
7 |
-
|
8 |
|
9 |
* **Image Editing**
|
10 |
|
11 |
- Upload `Source Image`
|
12 |
-
|
13 |
|
14 |
* **Video Editing**
|
15 |
|
16 |
- Upload `Video`
|
17 |
-
|
18 |
- Use `Generate` button to obtain `Edited Video`
|
19 |
|
20 |
**NOTE: we recommend to crop both input images and videos using provided [tools](https://github.com/wyhsirius/LIA-X/tree/main) for better results**
|
|
|
3 |
* **Image Animation**
|
4 |
|
5 |
- Upload `Source Image` and `Driving Video`
|
6 |
+
- Using sliders in the `Control Panel` to edit image
|
7 |
+
- Use `Animate` button to obtain `Animated Video`
|
8 |
|
9 |
* **Image Editing**
|
10 |
|
11 |
- Upload `Source Image`
|
12 |
+
- Using sliders in the `Control Panel` to edit image
|
13 |
|
14 |
* **Video Editing**
|
15 |
|
16 |
- Upload `Video`
|
17 |
+
- Using sliders in the `Control Panel` to edit image
|
18 |
- Use `Generate` button to obtain `Edited Video`
|
19 |
|
20 |
**NOTE: we recommend to crop both input images and videos using provided [tools](https://github.com/wyhsirius/LIA-X/tree/main) for better results**
|
assets/title.md
CHANGED
@@ -1,4 +1,5 @@
|
|
1 |
<font size=7><center>LIA-X: Interpretable Latent Portrait Animator</center></font>
|
|
|
2 |
<div style="display: flex;align-items: center;justify-content: center">
|
3 |
[<a href="https://arxiv.org/abs/2508.09959">Technical Report</a>] | [<a href="https://wyhsirius.github.io/LIA-X-project/">Project Page</a>] | [<a href="https://github.com/wyhsirius/LIA-X">Code</a>]
|
4 |
</div>
|
|
|
1 |
<font size=7><center>LIA-X: Interpretable Latent Portrait Animator</center></font>
|
2 |
+
<font size=5><center>Toward Interactive Portrait Animation and Editing</center></font>
|
3 |
<div style="display: flex;align-items: center;justify-content: center">
|
4 |
[<a href="https://arxiv.org/abs/2508.09959">Technical Report</a>] | [<a href="https://wyhsirius.github.io/LIA-X-project/">Project Page</a>] | [<a href="https://github.com/wyhsirius/LIA-X">Code</a>]
|
5 |
</div>
|
gradio_tabs/animation.py
CHANGED
@@ -36,64 +36,78 @@ labels_v = [
|
|
36 |
13, 24, 17, 26
|
37 |
]
|
38 |
|
39 |
-
|
40 |
def load_image(img, size):
|
41 |
|
42 |
img = Image.open(img).convert('RGB')
|
43 |
w, h = img.size
|
44 |
img = img.resize((size, size))
|
45 |
img = np.asarray(img)
|
|
|
46 |
img = np.transpose(img, (2, 0, 1)) # 3 x 256 x 256
|
47 |
|
48 |
return img / 255.0, w, h
|
49 |
|
50 |
|
|
|
51 |
def img_preprocessing(img_path, size):
|
52 |
-
img, w, h = load_image(img_path, size)
|
53 |
img = torch.from_numpy(img).unsqueeze(0).float() # [0, 1]
|
54 |
imgs_norm = (img - 0.5) * 2.0 # [-1, 1]
|
55 |
|
56 |
return imgs_norm, w, h
|
57 |
|
58 |
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
|
|
|
|
|
|
|
|
|
64 |
return transform(img)
|
65 |
|
66 |
|
67 |
def resize_back(img, w, h):
|
68 |
-
transform
|
69 |
-
|
70 |
-
])
|
71 |
-
|
72 |
return transform(img)
|
73 |
-
|
74 |
|
75 |
def vid_preprocessing(vid_path, size):
|
76 |
vid_dict = torchvision.io.read_video(vid_path, pts_unit='sec')
|
77 |
-
vid = vid_dict[0].permute(0, 3, 1, 2)
|
78 |
fps = vid_dict[2]['video_fps']
|
79 |
vid_norm = (vid / 255.0 - 0.5) * 2.0 # [-1, 1]
|
80 |
|
81 |
-
vid_norm = torch.cat([
|
82 |
-
|
83 |
-
], dim=1)
|
|
|
84 |
|
85 |
return vid_norm, fps
|
86 |
|
87 |
|
88 |
def img_denorm(img):
|
89 |
-
img = img.clamp(-1, 1)
|
90 |
img = (img - img.min()) / (img.max() - img.min())
|
91 |
|
92 |
return img
|
93 |
|
94 |
|
95 |
def vid_denorm(vid):
|
96 |
-
vid = vid.clamp(-1, 1)
|
97 |
vid = (vid - vid.min()) / (vid.max() - vid.min())
|
98 |
|
99 |
return vid
|
@@ -101,24 +115,32 @@ def vid_denorm(vid):
|
|
101 |
|
102 |
def img_postprocessing(image, w, h):
|
103 |
|
104 |
-
|
105 |
-
image = image.permute(0, 2, 3, 1)
|
106 |
-
edited_image = img_denorm(image)
|
107 |
-
img_output = (edited_image[0].numpy() * 255).astype(np.uint8)
|
108 |
|
109 |
-
|
110 |
-
|
111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
|
114 |
|
115 |
def vid_postprocessing(video, w, h, fps):
|
116 |
-
# video:
|
|
|
|
|
|
|
117 |
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
|
|
122 |
|
123 |
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_file:
|
124 |
imageio.mimwrite(temp_file.name, vid_np, fps=fps, codec='libx264', quality=8)
|
@@ -126,15 +148,59 @@ def vid_postprocessing(video, w, h, fps):
|
|
126 |
|
127 |
|
128 |
def animation(gen, chunk_size, device):
|
129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
130 |
@spaces.GPU
|
131 |
-
@torch.
|
132 |
def edit_media(image, *selected_s):
|
133 |
|
134 |
image_tensor, w, h = img_preprocessing(image, 512)
|
135 |
image_tensor = image_tensor.to(device)
|
136 |
|
137 |
-
|
|
|
138 |
|
139 |
# de-norm
|
140 |
edited_image = img_postprocessing(edited_image_tensor, w, h)
|
@@ -142,16 +208,38 @@ def animation(gen, chunk_size, device):
|
|
142 |
return edited_image
|
143 |
|
144 |
@spaces.GPU
|
145 |
-
@torch.
|
146 |
def animate_media(image, video, *selected_s):
|
147 |
|
148 |
image_tensor, w, h = img_preprocessing(image, 512)
|
149 |
vid_target_tensor, fps = vid_preprocessing(video, 512)
|
150 |
image_tensor = image_tensor.to(device)
|
151 |
-
video_target_tensor = vid_target_tensor.to(device)
|
152 |
-
|
153 |
-
animated_video = gen.animate_batch(image_tensor, video_target_tensor, labels_v, selected_s, chunk_size)
|
154 |
-
edited_image = animated_video[:,:,0,:,:]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
155 |
|
156 |
# postprocessing
|
157 |
animated_video = vid_postprocessing(animated_video, w, h, fps)
|
@@ -162,7 +250,7 @@ def animation(gen, chunk_size, device):
|
|
162 |
def clear_media():
|
163 |
return None, None, *([0] * len(labels_k))
|
164 |
|
165 |
-
|
166 |
with gr.Tab("Image Animation"):
|
167 |
|
168 |
inputs_s = []
|
@@ -202,11 +290,10 @@ def animation(gen, chunk_size, device):
|
|
202 |
with gr.Row():
|
203 |
with gr.Column(scale=1):
|
204 |
with gr.Row(): # Buttons now within a single Row
|
205 |
-
edit_btn = gr.Button("Edit", elem_id="button_edit",)
|
206 |
-
clear_btn = gr.Button("Clear", elem_id="button_clear")
|
207 |
-
with gr.Row():
|
208 |
animate_btn = gr.Button("Animate", elem_id="button_animate")
|
209 |
-
|
|
|
210 |
|
211 |
|
212 |
with gr.Column(scale=1):
|
@@ -221,7 +308,7 @@ def animation(gen, chunk_size, device):
|
|
221 |
#video_output.render()
|
222 |
video_output = gr.Video(label="Output Video", elem_id="output_vid", width=512)#.render()
|
223 |
|
224 |
-
with gr.Accordion("Control Panel", open=True):
|
225 |
with gr.Tab("Head"):
|
226 |
with gr.Row():
|
227 |
for k in labels_k[:3]:
|
@@ -251,20 +338,34 @@ def animation(gen, chunk_size, device):
|
|
251 |
for k in labels_k[12:14]:
|
252 |
slider = gr.Slider(minimum=-0.2, maximum=0.2, value=0, label=k, elem_id="slider_"+str(k))
|
253 |
inputs_s.append(slider)
|
|
|
|
|
|
|
|
|
|
|
|
|
254 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
255 |
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
-
|
260 |
-
|
261 |
-
|
|
|
262 |
|
263 |
animate_btn.click(
|
264 |
fn=animate_media,
|
265 |
inputs=[image_input, video_input] + inputs_s,
|
266 |
outputs=[image_output, video_output],
|
267 |
-
|
268 |
)
|
269 |
|
270 |
clear_btn.click(
|
@@ -280,14 +381,14 @@ def animation(gen, chunk_size, device):
|
|
280 |
['./data/source/macron.png', './data/driving/driving1.mp4', 0.14,0,-0.26,-0.29,-0.11,0,-0.13,-0.18,0,0,0,0,-0.02,0.07],
|
281 |
['./data/source/portrait3.png', './data/driving/driving1.mp4', -0.03,0.21,-0.31,-0.12,-0.11,0,-0.05,-0.16,0,0,0,0,-0.02,0.07],
|
282 |
['./data/source/einstein.png','./data/driving/driving2.mp4',-0.31,0,0,0.16,0.08,0,-0.07,0,0.13,0,0,0,0,0],
|
283 |
-
|
284 |
0.087, 0, 0, 0, 0, 0],
|
285 |
['./data/source/portrait2.png','./data/driving/driving8.mp4',0,0,-0.25,0,0,0,0,0,0,0.126,0,0,0,0],
|
286 |
|
287 |
],
|
288 |
-
|
289 |
inputs=[image_input, video_input] + inputs_s,
|
290 |
-
|
291 |
)
|
292 |
|
293 |
|
|
|
36 |
13, 24, 17, 26
|
37 |
]
|
38 |
|
39 |
+
@torch.compiler.allow_in_graph
|
40 |
def load_image(img, size):
|
41 |
|
42 |
img = Image.open(img).convert('RGB')
|
43 |
w, h = img.size
|
44 |
img = img.resize((size, size))
|
45 |
img = np.asarray(img)
|
46 |
+
img = np.copy(img)
|
47 |
img = np.transpose(img, (2, 0, 1)) # 3 x 256 x 256
|
48 |
|
49 |
return img / 255.0, w, h
|
50 |
|
51 |
|
52 |
+
@torch.compiler.allow_in_graph
|
53 |
def img_preprocessing(img_path, size):
|
54 |
+
img, w, h = load_image(img_path, size) # [0, 1]
|
55 |
img = torch.from_numpy(img).unsqueeze(0).float() # [0, 1]
|
56 |
imgs_norm = (img - 0.5) * 2.0 # [-1, 1]
|
57 |
|
58 |
return imgs_norm, w, h
|
59 |
|
60 |
|
61 |
+
# Pre-compile resize transforms for better performance
|
62 |
+
resize_transform_cache = {}
|
63 |
+
|
64 |
+
def get_resize_transform(size):
|
65 |
+
"""Get cached resize transform - creates once, reuses many times"""
|
66 |
+
if size not in resize_transform_cache:
|
67 |
+
# Only create the transform if it doesn't exist in cache
|
68 |
+
resize_transform_cache[size] = torchvision.transforms.Resize(
|
69 |
+
size,
|
70 |
+
interpolation=torchvision.transforms.InterpolationMode.BILINEAR,
|
71 |
+
antialias=True
|
72 |
+
)
|
73 |
+
return resize_transform_cache[size]
|
74 |
|
75 |
+
|
76 |
+
def resize(img, size):
|
77 |
+
"""Use cached resize transform"""
|
78 |
+
transform = get_resize_transform((size, size))
|
79 |
return transform(img)
|
80 |
|
81 |
|
82 |
def resize_back(img, w, h):
|
83 |
+
"""Use cached resize transform for back operation"""
|
84 |
+
transform = get_resize_transform((h, w))
|
|
|
|
|
85 |
return transform(img)
|
86 |
+
|
87 |
|
88 |
def vid_preprocessing(vid_path, size):
|
89 |
vid_dict = torchvision.io.read_video(vid_path, pts_unit='sec')
|
90 |
+
vid = vid_dict[0].permute(0, 3, 1, 2) # tchw
|
91 |
fps = vid_dict[2]['video_fps']
|
92 |
vid_norm = (vid / 255.0 - 0.5) * 2.0 # [-1, 1]
|
93 |
|
94 |
+
#vid_norm = torch.cat([
|
95 |
+
# resize(vid_norm[i:i+1, :, :, :], size).unsqueeze(1) for i in range(vid.size(0))
|
96 |
+
#], dim=1)
|
97 |
+
vid_norm = resize(vid_norm, size) # tchw
|
98 |
|
99 |
return vid_norm, fps
|
100 |
|
101 |
|
102 |
def img_denorm(img):
|
103 |
+
img = img.clamp(-1, 1)
|
104 |
img = (img - img.min()) / (img.max() - img.min())
|
105 |
|
106 |
return img
|
107 |
|
108 |
|
109 |
def vid_denorm(vid):
|
110 |
+
vid = vid.clamp(-1, 1)
|
111 |
vid = (vid - vid.min()) / (vid.max() - vid.min())
|
112 |
|
113 |
return vid
|
|
|
115 |
|
116 |
def img_postprocessing(image, w, h):
|
117 |
|
118 |
+
img = resize_back(image, w, h)
|
|
|
|
|
|
|
119 |
|
120 |
+
# Denormalize ON GPU (avoid early CPU transfer)
|
121 |
+
img = img.clamp(-1, 1) # Still on GPU
|
122 |
+
img = (img - img.min()) / (img.max() - img.min()) # Still on GPU
|
123 |
+
|
124 |
+
# Single optimized CPU transfer
|
125 |
+
img = img.squeeze(0).permute(1, 2, 0).contiguous() # contiguous() for fast transfer
|
126 |
+
img_output = (img.cpu().numpy() * 255).astype(np.uint8) # Single CPU transfer
|
127 |
+
|
128 |
+
# return the Numpy array directly, since Gradio supports it
|
129 |
+
return img_output
|
130 |
|
131 |
|
132 |
|
133 |
def vid_postprocessing(video, w, h, fps):
|
134 |
+
# video: TCHW
|
135 |
+
|
136 |
+
t,c,_,_ = video.size()
|
137 |
+
vid = resize_back(video, w, h)
|
138 |
|
139 |
+
vid = vid.clamp(-1, 1)
|
140 |
+
vid = (vid - vid.min()) / (vid.max() - vid.min())
|
141 |
+
|
142 |
+
vid = rearrange(vid, "t c h w -> t h w c") # T H W C
|
143 |
+
vid_np = (vid.cpu().numpy() * 255).astype('uint8')
|
144 |
|
145 |
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as temp_file:
|
146 |
imageio.mimwrite(temp_file.name, vid_np, fps=fps, codec='libx264', quality=8)
|
|
|
148 |
|
149 |
|
150 |
def animation(gen, chunk_size, device):
|
151 |
+
|
152 |
+
@torch.compile
|
153 |
+
def compiled_enc_img(image_tensor, selected_s):
|
154 |
+
"""Compiled version of just the model inference"""
|
155 |
+
return gen.enc_img(image_tensor, labels_v, selected_s)
|
156 |
+
|
157 |
+
@torch.compile
|
158 |
+
def compiled_dec_img(z_s2r, alpha_r2s, feat_rgb):
|
159 |
+
"""Compiled version of just the model inference"""
|
160 |
+
return gen.dec_img(z_s2r, alpha_r2s, feat_rgb)
|
161 |
+
|
162 |
+
@torch.compile
|
163 |
+
def compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch):
|
164 |
+
"""Compiled version of animate_batch for animation tab"""
|
165 |
+
return gen.dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch)
|
166 |
+
|
167 |
+
# Pre-warm the compiled model with dummy data to reduce first-run compilation time
|
168 |
+
def _warmup_model():
|
169 |
+
"""Pre-warm the model compilation with representative shapes"""
|
170 |
+
print("[img_edit] Pre-warming model compilation...")
|
171 |
+
dummy_image = torch.randn(1, 3, 512, 512, device=device)
|
172 |
+
dummy_video = torch.randn(chunk_size, 3, 512, 512, device=device)
|
173 |
+
dummy_selected_s = [0.0] * len(labels_v)
|
174 |
+
|
175 |
+
try:
|
176 |
+
with torch.inference_mode():
|
177 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(dummy_image, dummy_selected_s)
|
178 |
+
_ = compiled_dec_img(z_s2r, alpha_r2s, feat_rgb)
|
179 |
+
print("[img_edit] Model pre-warming completed successfully")
|
180 |
+
except Exception as e:
|
181 |
+
print(f"[img_edit] Model pre-warming failed (will compile on first use): {e}")
|
182 |
+
|
183 |
+
try:
|
184 |
+
with torch.inference_mode():
|
185 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(dummy_image, dummy_selected_s)
|
186 |
+
_ = compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, dummy_video[0], dummy_video)
|
187 |
+
print("[img_animation] Model pre-warming completed successfully")
|
188 |
+
except Exception as e:
|
189 |
+
print(f"[img_animation] Model pre-warming failed (will compile on first use): {e}")
|
190 |
+
|
191 |
+
# Pre-warm the model
|
192 |
+
_warmup_model()
|
193 |
+
|
194 |
+
|
195 |
@spaces.GPU
|
196 |
+
@torch.inference_mode()
|
197 |
def edit_media(image, *selected_s):
|
198 |
|
199 |
image_tensor, w, h = img_preprocessing(image, 512)
|
200 |
image_tensor = image_tensor.to(device)
|
201 |
|
202 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(image_tensor, selected_s)
|
203 |
+
edited_image_tensor = compiled_dec_img(z_s2r, alpha_r2s, feat_rgb)
|
204 |
|
205 |
# de-norm
|
206 |
edited_image = img_postprocessing(edited_image_tensor, w, h)
|
|
|
208 |
return edited_image
|
209 |
|
210 |
@spaces.GPU
|
211 |
+
@torch.inference_mode()
|
212 |
def animate_media(image, video, *selected_s):
|
213 |
|
214 |
image_tensor, w, h = img_preprocessing(image, 512)
|
215 |
vid_target_tensor, fps = vid_preprocessing(video, 512)
|
216 |
image_tensor = image_tensor.to(device)
|
217 |
+
video_target_tensor = vid_target_tensor.to(device) #tchw
|
218 |
+
|
219 |
+
#animated_video = gen.animate_batch(image_tensor, video_target_tensor, labels_v, selected_s, chunk_size)
|
220 |
+
#edited_image = animated_video[:,:,0,:,:]
|
221 |
+
|
222 |
+
img_start = video_target_tensor[0:1,:,:,:]
|
223 |
+
#vid_target_tensor_batch = rearrange(video_target_tensor, 'b t c h w -> (b t) c h w')
|
224 |
+
|
225 |
+
res = []
|
226 |
+
t = video_target_tensor.size(1)
|
227 |
+
chunks = t // chunk_size
|
228 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(image_tensor, selected_s)
|
229 |
+
#z_s2r, alpha_r2s, feat_rgb = gen.enc_img(image_tensor, labels_v, selected_s)
|
230 |
+
for i in range(chunks+1):
|
231 |
+
if i == chunks:
|
232 |
+
img_target = vid_target_tensor[i*chunk_size:, :, :, :]
|
233 |
+
img_animated = compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target)
|
234 |
+
#img_animated_batch = gen.dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch)
|
235 |
+
else:
|
236 |
+
img_target = vid_target_tensor[i*chunk_size:(i+1)*chunk_size, :, :, :]
|
237 |
+
img_animated = compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target)
|
238 |
+
#img_animated_batch = gen.dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch)
|
239 |
+
|
240 |
+
res.append(img_animated)
|
241 |
+
animated_video = torch.cat(res, dim=0) # TCHW
|
242 |
+
edited_image = animated_video[0:1,:,:,:]
|
243 |
|
244 |
# postprocessing
|
245 |
animated_video = vid_postprocessing(animated_video, w, h, fps)
|
|
|
250 |
def clear_media():
|
251 |
return None, None, *([0] * len(labels_k))
|
252 |
|
253 |
+
|
254 |
with gr.Tab("Image Animation"):
|
255 |
|
256 |
inputs_s = []
|
|
|
290 |
with gr.Row():
|
291 |
with gr.Column(scale=1):
|
292 |
with gr.Row(): # Buttons now within a single Row
|
293 |
+
#edit_btn = gr.Button("Edit", elem_id="button_edit",)
|
|
|
|
|
294 |
animate_btn = gr.Button("Animate", elem_id="button_animate")
|
295 |
+
with gr.Row():
|
296 |
+
clear_btn = gr.Button("Clear", elem_id="button_clear")
|
297 |
|
298 |
|
299 |
with gr.Column(scale=1):
|
|
|
308 |
#video_output.render()
|
309 |
video_output = gr.Video(label="Output Video", elem_id="output_vid", width=512)#.render()
|
310 |
|
311 |
+
with gr.Accordion("Control Panel (Using Sliders to Edit Image)", open=True):
|
312 |
with gr.Tab("Head"):
|
313 |
with gr.Row():
|
314 |
for k in labels_k[:3]:
|
|
|
338 |
for k in labels_k[12:14]:
|
339 |
slider = gr.Slider(minimum=-0.2, maximum=0.2, value=0, label=k, elem_id="slider_"+str(k))
|
340 |
inputs_s.append(slider)
|
341 |
+
|
342 |
+
for slider in inputs_s:
|
343 |
+
slider.change(
|
344 |
+
fn=edit_media,
|
345 |
+
inputs=[image_input] + inputs_s,
|
346 |
+
outputs=[image_output],
|
347 |
|
348 |
+
show_progress='hidden',
|
349 |
+
|
350 |
+
trigger_mode='always_last',
|
351 |
+
|
352 |
+
# currently we have a latency around 450ms
|
353 |
+
stream_every=0.5
|
354 |
+
)
|
355 |
|
356 |
+
|
357 |
+
#edit_btn.click(
|
358 |
+
# fn=edit_media,
|
359 |
+
# inputs=[image_input] + inputs_s,
|
360 |
+
# outputs=[image_output],
|
361 |
+
# show_progress=True
|
362 |
+
#)
|
363 |
|
364 |
animate_btn.click(
|
365 |
fn=animate_media,
|
366 |
inputs=[image_input, video_input] + inputs_s,
|
367 |
outputs=[image_output, video_output],
|
368 |
+
show_progress=True
|
369 |
)
|
370 |
|
371 |
clear_btn.click(
|
|
|
381 |
['./data/source/macron.png', './data/driving/driving1.mp4', 0.14,0,-0.26,-0.29,-0.11,0,-0.13,-0.18,0,0,0,0,-0.02,0.07],
|
382 |
['./data/source/portrait3.png', './data/driving/driving1.mp4', -0.03,0.21,-0.31,-0.12,-0.11,0,-0.05,-0.16,0,0,0,0,-0.02,0.07],
|
383 |
['./data/source/einstein.png','./data/driving/driving2.mp4',-0.31,0,0,0.16,0.08,0,-0.07,0,0.13,0,0,0,0,0],
|
384 |
+
['./data/source/portrait1.png', './data/driving/driving4.mp4', 0, 0, -0.17, -0.19, 0.25, 0, 0, -0.086,
|
385 |
0.087, 0, 0, 0, 0, 0],
|
386 |
['./data/source/portrait2.png','./data/driving/driving8.mp4',0,0,-0.25,0,0,0,0,0,0,0.126,0,0,0,0],
|
387 |
|
388 |
],
|
389 |
+
fn=animate_media,
|
390 |
inputs=[image_input, video_input] + inputs_s,
|
391 |
+
outputs=[image_output, video_output],
|
392 |
)
|
393 |
|
394 |
|
gradio_tabs/img_edit.py
CHANGED
@@ -37,69 +37,115 @@ labels_v = [
|
|
37 |
]
|
38 |
|
39 |
|
|
|
40 |
def load_image(img, size):
|
41 |
img = Image.open(img).convert('RGB')
|
42 |
w, h = img.size
|
43 |
img = img.resize((size, size))
|
44 |
img = np.asarray(img)
|
|
|
45 |
img = np.transpose(img, (2, 0, 1)) # 3 x 256 x 256
|
46 |
|
47 |
return img / 255.0, w, h
|
48 |
|
49 |
|
|
|
50 |
def img_preprocessing(img_path, size):
|
51 |
-
img, w, h = load_image(img_path, size)
|
52 |
img = torch.from_numpy(img).unsqueeze(0).float() # [0, 1]
|
53 |
imgs_norm = (img - 0.5) * 2.0 # [-1, 1]
|
54 |
|
55 |
return imgs_norm, w, h
|
56 |
|
57 |
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
|
|
|
|
|
|
|
|
|
63 |
return transform(img)
|
64 |
|
65 |
|
66 |
def resize_back(img, w, h):
|
67 |
-
transform
|
68 |
-
|
69 |
-
])
|
70 |
-
|
71 |
return transform(img)
|
72 |
|
73 |
|
74 |
def img_denorm(img):
|
75 |
-
img = img.clamp(-1, 1)
|
76 |
img = (img - img.min()) / (img.max() - img.min())
|
77 |
|
78 |
return img
|
79 |
|
80 |
|
81 |
-
def img_postprocessing(
|
82 |
|
83 |
-
|
84 |
-
image = image.permute(0, 2, 3, 1)
|
85 |
-
|
86 |
-
|
|
|
87 |
|
88 |
-
with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as temp_file:
|
89 |
-
|
90 |
-
|
|
|
91 |
|
92 |
|
93 |
def img_edit(gen, device):
|
94 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
@spaces.GPU
|
96 |
-
@torch.
|
97 |
def edit_img(image, *selected_s):
|
98 |
|
99 |
image_tensor, w, h = img_preprocessing(image, 512)
|
100 |
image_tensor = image_tensor.to(device)
|
101 |
|
102 |
-
|
|
|
103 |
|
104 |
# de-norm
|
105 |
edited_image = img_postprocessing(edited_image_tensor, w, h)
|
@@ -136,7 +182,7 @@ def img_edit(gen, device):
|
|
136 |
with gr.Row():
|
137 |
with gr.Column(scale=1):
|
138 |
with gr.Row(): # Buttons now within a single Row
|
139 |
-
edit_btn = gr.Button("Edit")
|
140 |
clear_btn = gr.Button("Clear")
|
141 |
#with gr.Row():
|
142 |
# animate_btn = gr.Button("Generate")
|
@@ -150,7 +196,7 @@ def img_edit(gen, device):
|
|
150 |
image_output = gr.Image(label="Output Image", type='numpy', interactive=False, width=512)
|
151 |
|
152 |
|
153 |
-
with gr.Accordion("Control Panel", open=True):
|
154 |
with gr.Tab("Head"):
|
155 |
with gr.Row():
|
156 |
for k in labels_k[:3]:
|
@@ -181,15 +227,29 @@ def img_edit(gen, device):
|
|
181 |
slider = gr.Slider(minimum=-0.2, maximum=0.2, value=0, label=k)
|
182 |
inputs_s.append(slider)
|
183 |
|
184 |
-
|
185 |
-
|
186 |
fn=edit_img,
|
187 |
inputs=[image_input] + inputs_s,
|
188 |
outputs=[image_output],
|
189 |
-
|
190 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
|
192 |
clear_btn.click(
|
193 |
fn=clear_media,
|
194 |
outputs=[image_output] + inputs_s
|
195 |
-
)
|
|
|
37 |
]
|
38 |
|
39 |
|
40 |
+
@torch.compiler.allow_in_graph
|
41 |
def load_image(img, size):
|
42 |
img = Image.open(img).convert('RGB')
|
43 |
w, h = img.size
|
44 |
img = img.resize((size, size))
|
45 |
img = np.asarray(img)
|
46 |
+
img = np.copy(img)
|
47 |
img = np.transpose(img, (2, 0, 1)) # 3 x 256 x 256
|
48 |
|
49 |
return img / 255.0, w, h
|
50 |
|
51 |
|
52 |
+
@torch.compiler.allow_in_graph
|
53 |
def img_preprocessing(img_path, size):
|
54 |
+
img, w, h = load_image(img_path, size) # [0, 1]
|
55 |
img = torch.from_numpy(img).unsqueeze(0).float() # [0, 1]
|
56 |
imgs_norm = (img - 0.5) * 2.0 # [-1, 1]
|
57 |
|
58 |
return imgs_norm, w, h
|
59 |
|
60 |
|
61 |
+
# Pre-compile resize transforms for better performance
|
62 |
+
resize_transform_cache = {}
|
63 |
+
|
64 |
+
def get_resize_transform(size):
|
65 |
+
"""Get cached resize transform - creates once, reuses many times"""
|
66 |
+
if size not in resize_transform_cache:
|
67 |
+
# Only create the transform if it doesn't exist in cache
|
68 |
+
resize_transform_cache[size] = torchvision.transforms.Resize(
|
69 |
+
size,
|
70 |
+
interpolation=torchvision.transforms.InterpolationMode.BILINEAR,
|
71 |
+
antialias=True
|
72 |
+
)
|
73 |
+
return resize_transform_cache[size]
|
74 |
|
75 |
+
|
76 |
+
def resize(img, size):
|
77 |
+
"""Use cached resize transform"""
|
78 |
+
transform = get_resize_transform((size, size))
|
79 |
return transform(img)
|
80 |
|
81 |
|
82 |
def resize_back(img, w, h):
|
83 |
+
"""Use cached resize transform for back operation"""
|
84 |
+
transform = get_resize_transform((h, w))
|
|
|
|
|
85 |
return transform(img)
|
86 |
|
87 |
|
88 |
def img_denorm(img):
|
89 |
+
img = img.clamp(-1, 1)
|
90 |
img = (img - img.min()) / (img.max() - img.min())
|
91 |
|
92 |
return img
|
93 |
|
94 |
|
95 |
+
def img_postprocessing(img, w, h):
|
96 |
|
97 |
+
img = resize_back(img, w, h)
|
98 |
+
#image = image.permute(0, 2, 3, 1)
|
99 |
+
img = img_denorm(img)
|
100 |
+
img = img.squeeze(0).permute(1, 2, 0).contiguous() # contiguous() for fast transfer
|
101 |
+
img_output = (img.cpu().numpy() * 255).astype(np.uint8)
|
102 |
|
103 |
+
#with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as temp_file:
|
104 |
+
# imageio.imwrite(temp_file.name, img_output, quality=8)
|
105 |
+
# return temp_file.name
|
106 |
+
return img_output
|
107 |
|
108 |
|
109 |
def img_edit(gen, device):
|
110 |
|
111 |
+
@torch.compile
|
112 |
+
def compiled_enc_img(image_tensor, selected_s):
|
113 |
+
"""Compiled version of just the model inference"""
|
114 |
+
return gen.enc_img(image_tensor, labels_v, selected_s)
|
115 |
+
|
116 |
+
@torch.compile
|
117 |
+
def compiled_dec_img(z_s2r, alpha_r2s, feat_rgb):
|
118 |
+
"""Compiled version of just the model inference"""
|
119 |
+
return gen.dec_img(z_s2r, alpha_r2s, feat_rgb)
|
120 |
+
|
121 |
+
|
122 |
+
# Pre-warm the compiled model with dummy data to reduce first-run compilation time
|
123 |
+
def _warmup_model():
|
124 |
+
"""Pre-warm the model compilation with representative shapes"""
|
125 |
+
print("[img_edit] Pre-warming model compilation...")
|
126 |
+
dummy_image = torch.randn(1, 3, 512, 512, device=device)
|
127 |
+
dummy_selected_s = [0.0] * len(labels_v)
|
128 |
+
|
129 |
+
try:
|
130 |
+
with torch.inference_mode():
|
131 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(dummy_image, dummy_selected_s)
|
132 |
+
_ = compiled_dec_img(z_s2r, alpha_r2s, feat_rgb)
|
133 |
+
print("[img_edit] Model pre-warming completed successfully")
|
134 |
+
except Exception as e:
|
135 |
+
print(f"[img_edit] Model pre-warming failed (will compile on first use): {e}")
|
136 |
+
|
137 |
+
# Pre-warm the model
|
138 |
+
_warmup_model()
|
139 |
+
|
140 |
@spaces.GPU
|
141 |
+
@torch.inference_mode()
|
142 |
def edit_img(image, *selected_s):
|
143 |
|
144 |
image_tensor, w, h = img_preprocessing(image, 512)
|
145 |
image_tensor = image_tensor.to(device)
|
146 |
|
147 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(image_tensor, selected_s)
|
148 |
+
edited_image_tensor = compiled_dec_img(z_s2r, alpha_r2s, feat_rgb)
|
149 |
|
150 |
# de-norm
|
151 |
edited_image = img_postprocessing(edited_image_tensor, w, h)
|
|
|
182 |
with gr.Row():
|
183 |
with gr.Column(scale=1):
|
184 |
with gr.Row(): # Buttons now within a single Row
|
185 |
+
#edit_btn = gr.Button("Edit")
|
186 |
clear_btn = gr.Button("Clear")
|
187 |
#with gr.Row():
|
188 |
# animate_btn = gr.Button("Generate")
|
|
|
196 |
image_output = gr.Image(label="Output Image", type='numpy', interactive=False, width=512)
|
197 |
|
198 |
|
199 |
+
with gr.Accordion("Control Panel (Using Sliders to Edit Image)", open=True):
|
200 |
with gr.Tab("Head"):
|
201 |
with gr.Row():
|
202 |
for k in labels_k[:3]:
|
|
|
227 |
slider = gr.Slider(minimum=-0.2, maximum=0.2, value=0, label=k)
|
228 |
inputs_s.append(slider)
|
229 |
|
230 |
+
for slider in inputs_s:
|
231 |
+
slider.change(
|
232 |
fn=edit_img,
|
233 |
inputs=[image_input] + inputs_s,
|
234 |
outputs=[image_output],
|
235 |
+
|
236 |
+
show_progress='hidden',
|
237 |
+
|
238 |
+
trigger_mode='always_last',
|
239 |
+
|
240 |
+
# currently we have a latency around 450ms
|
241 |
+
stream_every=0.5
|
242 |
+
)
|
243 |
+
|
244 |
+
|
245 |
+
#edit_btn.click(
|
246 |
+
# fn=edit_img,
|
247 |
+
# inputs=[image_input] + inputs_s,
|
248 |
+
# outputs=[image_output],
|
249 |
+
# show_progress=True
|
250 |
+
#)
|
251 |
|
252 |
clear_btn.click(
|
253 |
fn=clear_media,
|
254 |
outputs=[image_output] + inputs_s
|
255 |
+
)
|
gradio_tabs/vid_edit.py
CHANGED
@@ -37,92 +37,118 @@ labels_v = [
|
|
37 |
]
|
38 |
|
39 |
|
|
|
40 |
def load_image(img, size):
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
|
|
46 |
img = np.transpose(img, (2, 0, 1)) # 3 x 256 x 256
|
47 |
|
48 |
-
return img / 255.0
|
49 |
|
50 |
|
|
|
51 |
def img_preprocessing(img_path, size):
|
52 |
-
img = load_image(img_path, size)
|
53 |
img = torch.from_numpy(img).unsqueeze(0).float() # [0, 1]
|
54 |
imgs_norm = (img - 0.5) * 2.0 # [-1, 1]
|
55 |
|
56 |
-
return imgs_norm
|
57 |
|
|
|
|
|
58 |
|
59 |
-
def
|
60 |
-
transform
|
61 |
-
|
62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
63 |
|
|
|
|
|
|
|
64 |
return transform(img)
|
65 |
|
66 |
|
67 |
def resize_back(img, w, h):
|
68 |
-
transform
|
69 |
-
|
70 |
-
])
|
71 |
-
|
72 |
return transform(img)
|
73 |
-
|
74 |
|
75 |
def vid_preprocessing(vid_path, size):
|
76 |
vid_dict = torchvision.io.read_video(vid_path, pts_unit='sec')
|
77 |
-
vid = vid_dict[0].permute(0, 3, 1, 2)
|
78 |
_,_,_,h,w = vid.size()
|
79 |
fps = vid_dict[2]['video_fps']
|
80 |
vid_norm = (vid / 255.0 - 0.5) * 2.0 # [-1, 1]
|
81 |
|
82 |
-
vid_norm =
|
83 |
-
resize(vid_norm[:, i, :, :, :], size).unsqueeze(1) for i in range(vid.size(1))
|
84 |
-
], dim=1)
|
85 |
|
86 |
return vid_norm, fps, w, h
|
87 |
|
88 |
|
89 |
def img_denorm(img):
|
90 |
-
img = img.clamp(-1, 1)
|
91 |
img = (img - img.min()) / (img.max() - img.min())
|
92 |
|
93 |
return img
|
94 |
|
95 |
|
96 |
def vid_denorm(vid):
|
97 |
-
vid = vid.clamp(-1, 1)
|
98 |
vid = (vid - vid.min()) / (vid.max() - vid.min())
|
99 |
|
100 |
return vid
|
101 |
|
102 |
|
103 |
def img_postprocessing(image, w, h):
|
104 |
-
image = resize_back(image, w, h)
|
105 |
-
image = image.permute(0, 2, 3, 1)
|
106 |
-
edited_image = img_denorm(image)
|
107 |
-
img_output = (edited_image[0].numpy() * 255).astype(np.uint8)
|
108 |
|
109 |
-
|
110 |
-
|
111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
|
114 |
def vid_all_save(vid_d, vid_a, w, h, fps):
|
|
|
|
|
115 |
|
116 |
-
|
117 |
-
vid_d_batch = resize_back(
|
118 |
-
vid_a_batch = resize_back(
|
119 |
-
|
120 |
-
vid_d = rearrange(vid_d_batch, "
|
121 |
-
vid_a = rearrange(vid_a_batch, "
|
122 |
-
vid_all = torch.cat([vid_d, vid_a], dim=
|
123 |
|
124 |
-
vid_a_np = (vid_denorm(vid_a
|
125 |
-
vid_all_np = (vid_denorm(vid_all
|
126 |
|
127 |
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as output_path:
|
128 |
imageio.mimwrite(output_path.name, vid_a_np, fps=fps, codec='libx264', quality=8)
|
@@ -134,16 +160,59 @@ def vid_all_save(vid_d, vid_a, w, h, fps):
|
|
134 |
|
135 |
|
136 |
def vid_edit(gen, chunk_size, device):
|
137 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
138 |
@spaces.GPU
|
139 |
-
@torch.
|
140 |
def edit_img(video, *selected_s):
|
141 |
|
142 |
-
|
143 |
-
|
144 |
-
image_tensor = video_target_tensor[:,0,:,:,:]
|
145 |
|
146 |
-
|
|
|
147 |
|
148 |
# de-norm
|
149 |
edited_image = img_postprocessing(edited_image_tensor, w, h)
|
@@ -151,21 +220,35 @@ def vid_edit(gen, chunk_size, device):
|
|
151 |
return edited_image
|
152 |
|
153 |
@spaces.GPU
|
154 |
-
@torch.
|
155 |
def edit_vid(video, *selected_s):
|
156 |
|
157 |
video_target_tensor, fps, w, h = vid_preprocessing(video, 512)
|
158 |
video_target_tensor = video_target_tensor.to(device)
|
159 |
|
160 |
-
|
161 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
162 |
|
163 |
# de-norm
|
164 |
-
animated_video, animated_all_video = vid_all_save(
|
165 |
edited_image = img_postprocessing(edited_image_tensor, w, h)
|
166 |
|
167 |
-
return edited_image, animated_video, animated_all_video
|
168 |
-
|
169 |
|
170 |
def clear_media():
|
171 |
return None, None, None, *([0] * len(labels_k))
|
@@ -210,7 +293,7 @@ def vid_edit(gen, chunk_size, device):
|
|
210 |
video_all_output = gr.Video(label="Videos", elem_id="output_vid_all")
|
211 |
|
212 |
with gr.Column(scale=1):
|
213 |
-
with gr.Accordion("Control Panel", open=True):
|
214 |
with gr.Tab("Head"):
|
215 |
with gr.Row():
|
216 |
for k in labels_k[:3]:
|
@@ -244,17 +327,27 @@ def vid_edit(gen, chunk_size, device):
|
|
244 |
with gr.Row():
|
245 |
with gr.Column(scale=1):
|
246 |
with gr.Row(): # Buttons now within a single Row
|
247 |
-
edit_btn = gr.Button("Edit",elem_id="button_edit")
|
248 |
-
clear_btn = gr.Button("Clear",elem_id="button_clear")
|
249 |
-
with gr.Row():
|
250 |
animate_btn = gr.Button("Generate",elem_id="button_generate")
|
251 |
-
|
252 |
-
|
253 |
-
|
254 |
-
|
255 |
-
|
256 |
-
|
257 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
|
259 |
animate_btn.click(
|
260 |
fn=edit_vid,
|
@@ -280,9 +373,9 @@ def vid_edit(gen, chunk_size, device):
|
|
280 |
['./data/driving/driving9.mp4', 0, 0, 0, 0, 0, 0, 0,
|
281 |
0, 0, 0, 0, 0, -0.1, 0.07],
|
282 |
],
|
283 |
-
|
284 |
inputs=[video_input] + inputs_s,
|
285 |
-
|
286 |
)
|
287 |
|
288 |
|
|
|
37 |
]
|
38 |
|
39 |
|
40 |
+
@torch.compiler.allow_in_graph
|
41 |
def load_image(img, size):
|
42 |
+
|
43 |
+
img = Image.open(img).convert('RGB')
|
44 |
+
w, h = img.size
|
45 |
+
img = img.resize((size, size))
|
46 |
+
img = np.asarray(img)
|
47 |
+
img = np.copy(img)
|
48 |
img = np.transpose(img, (2, 0, 1)) # 3 x 256 x 256
|
49 |
|
50 |
+
return img / 255.0, w, h
|
51 |
|
52 |
|
53 |
+
@torch.compiler.allow_in_graph
|
54 |
def img_preprocessing(img_path, size):
|
55 |
+
img, w, h = load_image(img_path, size) # [0, 1]
|
56 |
img = torch.from_numpy(img).unsqueeze(0).float() # [0, 1]
|
57 |
imgs_norm = (img - 0.5) * 2.0 # [-1, 1]
|
58 |
|
59 |
+
return imgs_norm, w, h
|
60 |
|
61 |
+
# Pre-compile resize transforms for better performance
|
62 |
+
resize_transform_cache = {}
|
63 |
|
64 |
+
def get_resize_transform(size):
|
65 |
+
"""Get cached resize transform - creates once, reuses many times"""
|
66 |
+
if size not in resize_transform_cache:
|
67 |
+
# Only create the transform if it doesn't exist in cache
|
68 |
+
resize_transform_cache[size] = torchvision.transforms.Resize(
|
69 |
+
size,
|
70 |
+
interpolation=torchvision.transforms.InterpolationMode.BILINEAR,
|
71 |
+
antialias=True
|
72 |
+
)
|
73 |
+
return resize_transform_cache[size]
|
74 |
|
75 |
+
def resize(img, size):
|
76 |
+
"""Use cached resize transform"""
|
77 |
+
transform = get_resize_transform((size, size))
|
78 |
return transform(img)
|
79 |
|
80 |
|
81 |
def resize_back(img, w, h):
|
82 |
+
"""Use cached resize transform for back operation"""
|
83 |
+
transform = get_resize_transform((h, w))
|
|
|
|
|
84 |
return transform(img)
|
85 |
+
|
86 |
|
87 |
def vid_preprocessing(vid_path, size):
|
88 |
vid_dict = torchvision.io.read_video(vid_path, pts_unit='sec')
|
89 |
+
vid = vid_dict[0].permute(0, 3, 1, 2) # tchw
|
90 |
_,_,_,h,w = vid.size()
|
91 |
fps = vid_dict[2]['video_fps']
|
92 |
vid_norm = (vid / 255.0 - 0.5) * 2.0 # [-1, 1]
|
93 |
|
94 |
+
vid_norm = resize(vid_norm, size)
|
|
|
|
|
95 |
|
96 |
return vid_norm, fps, w, h
|
97 |
|
98 |
|
99 |
def img_denorm(img):
|
100 |
+
img = img.clamp(-1, 1)
|
101 |
img = (img - img.min()) / (img.max() - img.min())
|
102 |
|
103 |
return img
|
104 |
|
105 |
|
106 |
def vid_denorm(vid):
|
107 |
+
vid = vid.clamp(-1, 1)
|
108 |
vid = (vid - vid.min()) / (vid.max() - vid.min())
|
109 |
|
110 |
return vid
|
111 |
|
112 |
|
113 |
def img_postprocessing(image, w, h):
|
|
|
|
|
|
|
|
|
114 |
|
115 |
+
img = resize_back(image, w, h)
|
116 |
+
|
117 |
+
# Denormalize ON GPU (avoid early CPU transfer)
|
118 |
+
img = img_denorm(img)
|
119 |
+
|
120 |
+
# Single optimized CPU transfer
|
121 |
+
img = img.squeeze(0).permute(1, 2, 0).contiguous() # contiguous() for fast transfer
|
122 |
+
img_output = (img.cpu().numpy() * 255).astype(np.uint8) # Single CPU transfer
|
123 |
+
|
124 |
+
# return the Numpy array directly, since Gradio supports it
|
125 |
+
return img_output
|
126 |
+
|
127 |
+
|
128 |
+
def process_first_frame(vid_path, size):
|
129 |
+
vid_dict = torchvision.io.read_video(vid_path, start_pts=0, end_pts=0, pts_unit='sec')
|
130 |
+
img = vid_dict[0].permute(0, 3, 1, 2) # bchw
|
131 |
+
_, _, h, w = img.size()
|
132 |
+
img_norm = (img / 255.0 - 0.5) * 2.0 # [-1, 1]
|
133 |
+
img_norm = resize(img_norm, size)
|
134 |
+
|
135 |
+
return img_norm, w, h
|
136 |
|
137 |
|
138 |
def vid_all_save(vid_d, vid_a, w, h, fps):
|
139 |
+
# vid_d: tchw
|
140 |
+
# vid_a: tchw
|
141 |
|
142 |
+
t, c, _, _ = vid_d.size()
|
143 |
+
vid_d_batch = resize_back(vid_d, w, h)
|
144 |
+
vid_a_batch = resize_back(vid_a, w, h)
|
145 |
+
|
146 |
+
vid_d = rearrange(vid_d_batch, "t c h w -> t h w c") # T H W C
|
147 |
+
vid_a = rearrange(vid_a_batch, "t c h w -> t h w c") # T H W C
|
148 |
+
vid_all = torch.cat([vid_d, vid_a], dim=2)
|
149 |
|
150 |
+
vid_a_np = (vid_denorm(vid_a).cpu().numpy() * 255).astype('uint8')
|
151 |
+
vid_all_np = (vid_denorm(vid_all).cpu().numpy() * 255).astype('uint8')
|
152 |
|
153 |
with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as output_path:
|
154 |
imageio.mimwrite(output_path.name, vid_a_np, fps=fps, codec='libx264', quality=8)
|
|
|
160 |
|
161 |
|
162 |
def vid_edit(gen, chunk_size, device):
|
163 |
+
|
164 |
+
@torch.compile
|
165 |
+
def compiled_enc_img(image_tensor, selected_s):
|
166 |
+
"""Compiled version of just the model inference"""
|
167 |
+
return gen.enc_img(image_tensor, labels_v, selected_s)
|
168 |
+
|
169 |
+
@torch.compile
|
170 |
+
def compiled_dec_img(z_s2r, alpha_r2s, feat_rgb):
|
171 |
+
"""Compiled version of just the model inference"""
|
172 |
+
return gen.dec_img(z_s2r, alpha_r2s, feat_rgb)
|
173 |
+
|
174 |
+
@torch.compile
|
175 |
+
def compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch):
|
176 |
+
"""Compiled version of animate_batch for animation tab"""
|
177 |
+
return gen.dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch)
|
178 |
+
|
179 |
+
# Pre-warm the compiled model with dummy data to reduce first-run compilation time
|
180 |
+
def _warmup_model():
|
181 |
+
"""Pre-warm the model compilation with representative shapes"""
|
182 |
+
print("[img_edit] Pre-warming model compilation...")
|
183 |
+
dummy_image = torch.randn(1, 3, 512, 512, device=device)
|
184 |
+
dummy_video = torch.randn(chunk_size, 3, 512, 512, device=device)
|
185 |
+
dummy_selected_s = [0.0] * len(labels_v)
|
186 |
+
|
187 |
+
try:
|
188 |
+
with torch.inference_mode():
|
189 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(dummy_image, dummy_selected_s)
|
190 |
+
_ = compiled_dec_img(z_s2r, alpha_r2s, feat_rgb)
|
191 |
+
print("[img_edit] Model pre-warming completed successfully")
|
192 |
+
except Exception as e:
|
193 |
+
print(f"[img_edit] Model pre-warming failed (will compile on first use): {e}")
|
194 |
+
|
195 |
+
try:
|
196 |
+
with torch.inference_mode():
|
197 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(dummy_image, dummy_selected_s)
|
198 |
+
_ = compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, dummy_video[0], dummy_video)
|
199 |
+
print("[img_animation] Model pre-warming completed successfully")
|
200 |
+
except Exception as e:
|
201 |
+
print(f"[img_animation] Model pre-warming failed (will compile on first use): {e}")
|
202 |
+
|
203 |
+
# Pre-warm the model
|
204 |
+
_warmup_model()
|
205 |
+
|
206 |
+
|
207 |
@spaces.GPU
|
208 |
+
@torch.inference_mode()
|
209 |
def edit_img(video, *selected_s):
|
210 |
|
211 |
+
image_tensor, w, h = process_first_frame(video, 512)
|
212 |
+
image_tensor = image_tensor.to(device)
|
|
|
213 |
|
214 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(image_tensor, selected_s)
|
215 |
+
edited_image_tensor = compiled_dec_img(z_s2r, alpha_r2s, feat_rgb)
|
216 |
|
217 |
# de-norm
|
218 |
edited_image = img_postprocessing(edited_image_tensor, w, h)
|
|
|
220 |
return edited_image
|
221 |
|
222 |
@spaces.GPU
|
223 |
+
@torch.inference_mode()
|
224 |
def edit_vid(video, *selected_s):
|
225 |
|
226 |
video_target_tensor, fps, w, h = vid_preprocessing(video, 512)
|
227 |
video_target_tensor = video_target_tensor.to(device)
|
228 |
|
229 |
+
img_start = video_target_tensor[0:1, :, :, :]
|
230 |
+
|
231 |
+
res = []
|
232 |
+
t = video_target_tensor.size(1)
|
233 |
+
chunks = t // chunk_size
|
234 |
+
z_s2r, alpha_r2s, feat_rgb = compiled_enc_img(img_start, selected_s)
|
235 |
+
for i in range(chunks + 1):
|
236 |
+
if i == chunks:
|
237 |
+
img_target_batch = vid_target_tensor_batch[i * chunk_size:, :, :, :]
|
238 |
+
img_animated_batch = compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target)
|
239 |
+
else:
|
240 |
+
img_target_batch = vid_target_tensor_batch[i * chunk_size:(i + 1) * chunk_size, :, :, :]
|
241 |
+
img_animated_batch = compiled_dec_vid(z_s2r, alpha_r2s, feat_rgb, img_start, img_target)
|
242 |
+
|
243 |
+
res.append(img_animated_batch)
|
244 |
+
edited_video_tensor = torch.cat(res, dim=0) # TCHW
|
245 |
+
edited_image_tensor = edited_video_tensor[0:1,:,:,:]
|
246 |
|
247 |
# de-norm
|
248 |
+
animated_video, animated_all_video = vid_all_save(vid_target_tensor_batch, edited_video_tensor, w, h, fps)
|
249 |
edited_image = img_postprocessing(edited_image_tensor, w, h)
|
250 |
|
251 |
+
return edited_image, animated_video, animated_all_video
|
|
|
252 |
|
253 |
def clear_media():
|
254 |
return None, None, None, *([0] * len(labels_k))
|
|
|
293 |
video_all_output = gr.Video(label="Videos", elem_id="output_vid_all")
|
294 |
|
295 |
with gr.Column(scale=1):
|
296 |
+
with gr.Accordion("Control Panel (Using Sliders to Edit Image)", open=True):
|
297 |
with gr.Tab("Head"):
|
298 |
with gr.Row():
|
299 |
for k in labels_k[:3]:
|
|
|
327 |
with gr.Row():
|
328 |
with gr.Column(scale=1):
|
329 |
with gr.Row(): # Buttons now within a single Row
|
330 |
+
#edit_btn = gr.Button("Edit",elem_id="button_edit")
|
|
|
|
|
331 |
animate_btn = gr.Button("Generate",elem_id="button_generate")
|
332 |
+
clear_btn = gr.Button("Clear",elem_id="button_clear")
|
333 |
+
|
334 |
+
for slider in inputs_s:
|
335 |
+
slider.change(
|
336 |
+
fn=edit_img,
|
337 |
+
inputs=[video_input] + inputs_s,
|
338 |
+
outputs=[image_output],
|
339 |
+
show_progress='hidden',
|
340 |
+
trigger_mode='always_last',
|
341 |
+
# currently we have a latency around 450ms
|
342 |
+
stream_every=0.5
|
343 |
+
)
|
344 |
+
|
345 |
+
#edit_btn.click(
|
346 |
+
# fn=edit_img,
|
347 |
+
# inputs=[video_input] + inputs_s,
|
348 |
+
# outputs=[image_output],
|
349 |
+
# show_progress=True
|
350 |
+
#)
|
351 |
|
352 |
animate_btn.click(
|
353 |
fn=edit_vid,
|
|
|
373 |
['./data/driving/driving9.mp4', 0, 0, 0, 0, 0, 0, 0,
|
374 |
0, 0, 0, 0, 0, -0.1, 0.07],
|
375 |
],
|
376 |
+
fn=edit_vid,
|
377 |
inputs=[video_input] + inputs_s,
|
378 |
+
outputs=[image_output, video_output, video_all_output],
|
379 |
)
|
380 |
|
381 |
|
networks/generator.py
CHANGED
@@ -17,6 +17,12 @@ class Generator(nn.Module):
|
|
17 |
self.enc = Encoder(style_dim, motion_dim, scale)
|
18 |
self.dec = Decoder(style_dim, motion_dim, scale)
|
19 |
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
def get_alpha(self, x):
|
21 |
return self.enc.enc_motion(x)
|
22 |
|
@@ -83,7 +89,7 @@ class Generator(nn.Module):
|
|
83 |
vid_target_recon = rearrange(vid_target_recon, 'b t c h w -> b c t h w')
|
84 |
|
85 |
return vid_target_recon # BCTHW
|
86 |
-
|
87 |
def edit_vid(self, vid_target, d_l, v_l):
|
88 |
|
89 |
img_source = vid_target[:, 0, :, :, :]
|
@@ -195,3 +201,36 @@ class Generator(nn.Module):
|
|
195 |
|
196 |
return vid_target_recon
|
197 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
self.enc = Encoder(style_dim, motion_dim, scale)
|
18 |
self.dec = Decoder(style_dim, motion_dim, scale)
|
19 |
|
20 |
+
@property
|
21 |
+
def device(self):
|
22 |
+
if self._device is None:
|
23 |
+
self._device = next(self.parameters()).device
|
24 |
+
return self._device
|
25 |
+
|
26 |
def get_alpha(self, x):
|
27 |
return self.enc.enc_motion(x)
|
28 |
|
|
|
89 |
vid_target_recon = rearrange(vid_target_recon, 'b t c h w -> b c t h w')
|
90 |
|
91 |
return vid_target_recon # BCTHW
|
92 |
+
|
93 |
def edit_vid(self, vid_target, d_l, v_l):
|
94 |
|
95 |
img_source = vid_target[:, 0, :, :, :]
|
|
|
201 |
|
202 |
return vid_target_recon
|
203 |
|
204 |
+
def enc_img(self, img_source, d_l, v_l):
|
205 |
+
"""Core edit_img logic without timing - can be compiled"""
|
206 |
+
z_s2r, feat_rgb = self.enc.enc_2r(img_source)
|
207 |
+
alpha_r2s = self.enc.enc_r2t(z_s2r)
|
208 |
+
|
209 |
+
# Create tensor directly on the same device as alpha_r2s
|
210 |
+
v_l_tensor = torch.tensor(v_l, device=alpha_r2s.device, dtype=alpha_r2s.dtype).unsqueeze(0)
|
211 |
+
alpha_r2s[:, d_l] = alpha_r2s[:, d_l] + v_l_tensor
|
212 |
+
|
213 |
+
return z_s2r, alpha_r2s, feat_rgb
|
214 |
+
|
215 |
+
def dec_img(self, z_s2r, alpha_r2s, feat_rgb):
|
216 |
+
return self.dec(z_s2r, [alpha_r2s], feat_rgb)
|
217 |
+
|
218 |
+
|
219 |
+
def dec_vid(self, z_s2r, alpha_r2s, feat_rgb, img_start, img_target_batch):
|
220 |
+
# z_s2r: BC
|
221 |
+
# alpha_r2s: BC
|
222 |
+
# feat: BCHW
|
223 |
+
# alpha_start: BC
|
224 |
+
|
225 |
+
bs = img_target_batch.size(0)
|
226 |
+
alpha_start = self.get_alpha(img_start)
|
227 |
+
|
228 |
+
alpha_start_r = repeat(alpha_start, 'b c -> (repeat b) c', repeat=bs)
|
229 |
+
alpha_r2s_r = repeat(alpha_r2s, 'b c -> (repeat b) c', repeat=bs)
|
230 |
+
feat_rgb_r = [repeat(feat, 'b c h w -> (repeat b) c h w', repeat=bs) for feat in feat_rgb]
|
231 |
+
z_s2r_r = repeat(z_s2r, 'b c -> (repeat b) c', repeat=bs)
|
232 |
+
|
233 |
+
alpha = self.enc.enc_transfer_vid(alpha_r2s_r, img_target_batch, alpha_start_r)
|
234 |
+
img_batch_recon = self.dec(z_s2r_r, alpha, feat_rgb_r) # bs x 3 x h x w
|
235 |
+
|
236 |
+
return img_batch_recon
|