Haiyu Wu
commited on
Commit
·
28a6a8a
1
Parent(s):
7aa5b58
update
Browse files
app.py
CHANGED
@@ -10,6 +10,23 @@ from sixdrepnet.model import SixDRepNet
|
|
10 |
import pixel_generator.vec2face.model_vec2face as model_vec2face
|
11 |
MAX_SEED = np.iinfo(np.int32).max
|
12 |
import torch
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
|
15 |
def sample_nearby_vectors(base_vector, epsilons=[0.3, 0.5, 0.7], percentages=[0.4, 0.4, 0.2]):
|
@@ -62,7 +79,7 @@ def initialize_models():
|
|
62 |
return generator, id_model, pose_model, quality_model
|
63 |
|
64 |
|
65 |
-
def image_generation(input_image, quality, use_target_pose, pose, dimension):
|
66 |
generator, id_model, pose_model, quality_model = initialize_models()
|
67 |
|
68 |
generated_images = []
|
@@ -77,12 +94,10 @@ def image_generation(input_image, quality, use_target_pose, pose, dimension):
|
|
77 |
if not use_target_pose:
|
78 |
features = []
|
79 |
norm = np.linalg.norm(feature, 2, 1, True)
|
80 |
-
for i in np.arange(0, 4.8,
|
81 |
updated_feature = feature
|
82 |
updated_feature[0][dimension] = feature[0][dimension] + i
|
83 |
-
|
84 |
updated_feature = updated_feature / np.linalg.norm(updated_feature, 2, 1, True) * norm
|
85 |
-
|
86 |
features.append(updated_feature)
|
87 |
features = torch.tensor(np.vstack(features)).float()
|
88 |
if quality > 25:
|
@@ -90,7 +105,7 @@ def image_generation(input_image, quality, use_target_pose, pose, dimension):
|
|
90 |
else:
|
91 |
_, _, images, *_ = generator(features)
|
92 |
else:
|
93 |
-
features = torch.repeat_interleave(torch.tensor(feature),
|
94 |
features = sample_nearby_vectors(features, [0.7], [1]).float()
|
95 |
if quality > 25 or pose > 20:
|
96 |
images, _ = generator.gen_image(features, quality_model, id_model, pose_model=pose_model,
|
@@ -99,12 +114,13 @@ def image_generation(input_image, quality, use_target_pose, pose, dimension):
|
|
99 |
_, _, images, *_ = generator(features)
|
100 |
|
101 |
images = ((images.permute(0, 2, 3, 1).detach().cpu().numpy() + 1) / 2 * 255).astype(np.uint8)
|
102 |
-
for image in images:
|
103 |
generated_images.append(Image.fromarray(image))
|
|
|
104 |
return generated_images
|
105 |
|
106 |
|
107 |
-
def process_input(image_input, num1, num2, num3, num4, random_seed, target_quality, use_target_pose, target_pose):
|
108 |
# Ensure all dimension numbers are within [0, 512)
|
109 |
num1, num2, num3, num4 = [max(0, min(int(n), 511)) for n in [num1, num2, num3, num4]]
|
110 |
|
@@ -118,13 +134,14 @@ def process_input(image_input, num1, num2, num3, num4, random_seed, target_quali
|
|
118 |
input_data = Image.open(image_input)
|
119 |
input_data = np.array(input_data.resize((112, 112)))
|
120 |
|
121 |
-
generated_images = image_generation(input_data, target_quality, use_target_pose, target_pose, [num1, num2, num3, num4])
|
122 |
|
123 |
return generated_images
|
124 |
|
|
|
125 |
def select_image(value, images):
|
126 |
# Convert the float value (0 to 4) to an integer index (0 to 9)
|
127 |
-
index = int(value /
|
128 |
return images[index]
|
129 |
|
130 |
def toggle_inputs(use_pose):
|
@@ -147,9 +164,9 @@ def main():
|
|
147 |
<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/HaiyuWu/vec2face' target='_blank'><b>Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors</b></a>.<br>
|
148 |
|
149 |
How to use:<br>
|
150 |
-
1. Upload an image with a cropped face image or directly click <b>Submit</b> button,
|
151 |
2. You can control the image quality, image pose, and modify the values in the target dimensions to change the output images.
|
152 |
-
3. The output results will shown
|
153 |
4. Since the demo is CPU-based, higher quality and larger pose need longer time to run.
|
154 |
5. Enjoy! 😊
|
155 |
"""
|
@@ -167,9 +184,9 @@ def main():
|
|
167 |
|
168 |
with gr.Row():
|
169 |
num1 = gr.Number(label="Dimension 1", value=0, minimum=0, maximum=511, step=1)
|
170 |
-
num2 = gr.Number(label="Dimension 2", value=
|
171 |
-
num3 = gr.Number(label="Dimension 3", value=
|
172 |
-
num4 = gr.Number(label="Dimension 4", value=
|
173 |
|
174 |
random_seed = gr.Number(label="Random Seed", value=42, minimum=0, maximum=MAX_SEED, step=1)
|
175 |
target_quality = gr.Slider(label="Minimum Quality", minimum=22, maximum=35, step=1, value=24)
|
@@ -191,9 +208,10 @@ def main():
|
|
191 |
|
192 |
with gr.Column():
|
193 |
gallery = gr.Image(label="Generated Image")
|
|
|
194 |
incremental_value_slider = gr.Slider(
|
195 |
label="Result of dimension modification or results of pose images",
|
196 |
-
minimum=0, maximum=4, step=
|
197 |
)
|
198 |
gr.Markdown("""
|
199 |
- These values are added to the dimensions (before normalization), **please ignore it if pose editing is on**.
|
@@ -208,14 +226,35 @@ def main():
|
|
208 |
generated_images = gr.State([])
|
209 |
|
210 |
submit.click(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
211 |
fn=process_input,
|
212 |
inputs=[image_file, num1, num2, num3, num4, random_seed, target_quality, use_target_pose, target_pose],
|
213 |
outputs=[generated_images]
|
|
|
|
|
|
|
|
|
214 |
).then(
|
215 |
fn=select_image,
|
216 |
inputs=[incremental_value_slider, generated_images],
|
217 |
outputs=[gallery]
|
218 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
219 |
|
220 |
incremental_value_slider.change(
|
221 |
fn=select_image,
|
@@ -244,4 +283,4 @@ def main():
|
|
244 |
|
245 |
|
246 |
if __name__ == "__main__":
|
247 |
-
main()
|
|
|
10 |
import pixel_generator.vec2face.model_vec2face as model_vec2face
|
11 |
MAX_SEED = np.iinfo(np.int32).max
|
12 |
import torch
|
13 |
+
from time import time
|
14 |
+
|
15 |
+
|
16 |
+
def clear_image():
|
17 |
+
return None
|
18 |
+
|
19 |
+
|
20 |
+
def clear_generation_time():
|
21 |
+
return ""
|
22 |
+
|
23 |
+
|
24 |
+
def generating():
|
25 |
+
return "Generating images..."
|
26 |
+
|
27 |
+
|
28 |
+
def done():
|
29 |
+
return "Done!"
|
30 |
|
31 |
|
32 |
def sample_nearby_vectors(base_vector, epsilons=[0.3, 0.5, 0.7], percentages=[0.4, 0.4, 0.2]):
|
|
|
79 |
return generator, id_model, pose_model, quality_model
|
80 |
|
81 |
|
82 |
+
def image_generation(input_image, quality, use_target_pose, pose, dimension, progress=gr.Progress()):
|
83 |
generator, id_model, pose_model, quality_model = initialize_models()
|
84 |
|
85 |
generated_images = []
|
|
|
94 |
if not use_target_pose:
|
95 |
features = []
|
96 |
norm = np.linalg.norm(feature, 2, 1, True)
|
97 |
+
for i in progress.tqdm(np.arange(0, 4.8, 2), desc="Generating images"):
|
98 |
updated_feature = feature
|
99 |
updated_feature[0][dimension] = feature[0][dimension] + i
|
|
|
100 |
updated_feature = updated_feature / np.linalg.norm(updated_feature, 2, 1, True) * norm
|
|
|
101 |
features.append(updated_feature)
|
102 |
features = torch.tensor(np.vstack(features)).float()
|
103 |
if quality > 25:
|
|
|
105 |
else:
|
106 |
_, _, images, *_ = generator(features)
|
107 |
else:
|
108 |
+
features = torch.repeat_interleave(torch.tensor(feature), 3, dim=0)
|
109 |
features = sample_nearby_vectors(features, [0.7], [1]).float()
|
110 |
if quality > 25 or pose > 20:
|
111 |
images, _ = generator.gen_image(features, quality_model, id_model, pose_model=pose_model,
|
|
|
114 |
_, _, images, *_ = generator(features)
|
115 |
|
116 |
images = ((images.permute(0, 2, 3, 1).detach().cpu().numpy() + 1) / 2 * 255).astype(np.uint8)
|
117 |
+
for image in progress.tqdm(images, desc="Processing images"):
|
118 |
generated_images.append(Image.fromarray(image))
|
119 |
+
|
120 |
return generated_images
|
121 |
|
122 |
|
123 |
+
def process_input(image_input, num1, num2, num3, num4, random_seed, target_quality, use_target_pose, target_pose, progress=gr.Progress()):
|
124 |
# Ensure all dimension numbers are within [0, 512)
|
125 |
num1, num2, num3, num4 = [max(0, min(int(n), 511)) for n in [num1, num2, num3, num4]]
|
126 |
|
|
|
134 |
input_data = Image.open(image_input)
|
135 |
input_data = np.array(input_data.resize((112, 112)))
|
136 |
|
137 |
+
generated_images = image_generation(input_data, target_quality, use_target_pose, target_pose, [num1, num2, num3, num4], progress)
|
138 |
|
139 |
return generated_images
|
140 |
|
141 |
+
|
142 |
def select_image(value, images):
|
143 |
# Convert the float value (0 to 4) to an integer index (0 to 9)
|
144 |
+
index = int(value / 2)
|
145 |
return images[index]
|
146 |
|
147 |
def toggle_inputs(use_pose):
|
|
|
164 |
<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/HaiyuWu/vec2face' target='_blank'><b>Vec2Face: Scaling Face Dataset Generation with Loosely Constrained Vectors</b></a>.<br>
|
165 |
|
166 |
How to use:<br>
|
167 |
+
1. Upload an image with a cropped face image or directly click <b>Submit</b> button, three images will be shown on the right.
|
168 |
2. You can control the image quality, image pose, and modify the values in the target dimensions to change the output images.
|
169 |
+
3. The output results will shown three results of dimension modification or pose images.
|
170 |
4. Since the demo is CPU-based, higher quality and larger pose need longer time to run.
|
171 |
5. Enjoy! 😊
|
172 |
"""
|
|
|
184 |
|
185 |
with gr.Row():
|
186 |
num1 = gr.Number(label="Dimension 1", value=0, minimum=0, maximum=511, step=1)
|
187 |
+
num2 = gr.Number(label="Dimension 2", value=50, minimum=0, maximum=511, step=1)
|
188 |
+
num3 = gr.Number(label="Dimension 3", value=100, minimum=0, maximum=511, step=1)
|
189 |
+
num4 = gr.Number(label="Dimension 4", value=200, minimum=0, maximum=511, step=1)
|
190 |
|
191 |
random_seed = gr.Number(label="Random Seed", value=42, minimum=0, maximum=MAX_SEED, step=1)
|
192 |
target_quality = gr.Slider(label="Minimum Quality", minimum=22, maximum=35, step=1, value=24)
|
|
|
208 |
|
209 |
with gr.Column():
|
210 |
gallery = gr.Image(label="Generated Image")
|
211 |
+
generation_time = gr.Textbox(label="Generation Status")
|
212 |
incremental_value_slider = gr.Slider(
|
213 |
label="Result of dimension modification or results of pose images",
|
214 |
+
minimum=0, maximum=4, step=2, value=0
|
215 |
)
|
216 |
gr.Markdown("""
|
217 |
- These values are added to the dimensions (before normalization), **please ignore it if pose editing is on**.
|
|
|
226 |
generated_images = gr.State([])
|
227 |
|
228 |
submit.click(
|
229 |
+
fn=clear_image,
|
230 |
+
inputs=[],
|
231 |
+
outputs=[gallery]
|
232 |
+
).then(
|
233 |
+
fn=generating,
|
234 |
+
inputs=[],
|
235 |
+
outputs=[generation_time]
|
236 |
+
).then(
|
237 |
fn=process_input,
|
238 |
inputs=[image_file, num1, num2, num3, num4, random_seed, target_quality, use_target_pose, target_pose],
|
239 |
outputs=[generated_images]
|
240 |
+
).then(
|
241 |
+
fn=done,
|
242 |
+
inputs=[],
|
243 |
+
outputs=[generation_time]
|
244 |
).then(
|
245 |
fn=select_image,
|
246 |
inputs=[incremental_value_slider, generated_images],
|
247 |
outputs=[gallery]
|
248 |
)
|
249 |
+
# submit.click(
|
250 |
+
# fn=process_input,
|
251 |
+
# inputs=[image_file, num1, num2, num3, num4, random_seed, target_quality, use_target_pose, target_pose],
|
252 |
+
# outputs=[generated_images]
|
253 |
+
# ).then(
|
254 |
+
# fn=select_image,
|
255 |
+
# inputs=[incremental_value_slider, generated_images],
|
256 |
+
# outputs=[gallery]
|
257 |
+
# )
|
258 |
|
259 |
incremental_value_slider.change(
|
260 |
fn=select_image,
|
|
|
283 |
|
284 |
|
285 |
if __name__ == "__main__":
|
286 |
+
main()
|