Update app.py
Browse files
app.py
CHANGED
@@ -1,29 +1,219 @@
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
-
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
)
|
26 |
-
|
27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
if __name__ == "__main__":
|
29 |
-
|
|
|
|
1 |
+
import torch
|
2 |
+
import spaces
|
3 |
import gradio as gr
|
4 |
+
import os
|
5 |
+
import numpy as np
|
6 |
+
import trimesh
|
7 |
+
import mcubes
|
8 |
+
import imageio
|
9 |
+
from PIL import Image
|
10 |
+
from transformers import AutoModel, AutoConfig
|
11 |
+
from rembg import remove, new_session
|
12 |
+
from functools import partial
|
13 |
+
import kiui
|
14 |
+
from gradio_litmodel3d import LitModel3D
|
15 |
+
|
16 |
+
class VFusion3DGenerator:
|
17 |
+
def __init__(self, model_name="facebook/vfusion3d"):
|
18 |
+
"""
|
19 |
+
Initialize the VFusion3D model
|
20 |
+
|
21 |
+
Args:
|
22 |
+
model_name (str): Hugging Face model identifier
|
23 |
+
"""
|
24 |
+
self.config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
|
25 |
+
self.model = AutoModel.from_pretrained(model_name, trust_remote_code=True)
|
26 |
+
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
27 |
+
self.model.to(self.device)
|
28 |
+
self.model.eval()
|
29 |
+
|
30 |
+
# Background removal session
|
31 |
+
self.rembg_session = new_session("isnet-general-use")
|
32 |
+
|
33 |
+
def preprocess_image(self, image, source_size=512):
|
34 |
+
"""
|
35 |
+
Preprocess input image for VFusion3D model
|
36 |
+
|
37 |
+
Args:
|
38 |
+
image (PIL.Image): Input image
|
39 |
+
source_size (int): Target image size
|
40 |
+
|
41 |
+
Returns:
|
42 |
+
torch.Tensor: Preprocessed image tensor
|
43 |
+
"""
|
44 |
+
rembg_remove = partial(remove, session=self.rembg_session)
|
45 |
+
image = np.array(image)
|
46 |
+
image = rembg_remove(image)
|
47 |
+
mask = rembg_remove(image, only_mask=True)
|
48 |
+
image = kiui.op.recenter(image, mask, border_ratio=0.20)
|
49 |
+
|
50 |
+
image = torch.tensor(image).permute(2, 0, 1).unsqueeze(0) / 255.0
|
51 |
+
if image.shape[1] == 4:
|
52 |
+
image = image[:, :3, ...] * image[:, 3:, ...] + (1 - image[:, 3:, ...])
|
53 |
+
|
54 |
+
image = torch.nn.functional.interpolate(
|
55 |
+
image,
|
56 |
+
size=(source_size, source_size),
|
57 |
+
mode='bicubic',
|
58 |
+
align_corners=True
|
59 |
+
)
|
60 |
+
return torch.clamp(image, 0, 1)
|
61 |
+
|
62 |
+
def generate_3d_output(self, image, output_type='mesh', render_size=384, mesh_size=512):
|
63 |
+
"""
|
64 |
+
Generate 3D output (mesh or video) from input image
|
65 |
+
|
66 |
+
Args:
|
67 |
+
image (PIL.Image): Input image
|
68 |
+
output_type (str): Type of output ('mesh' or 'video')
|
69 |
+
render_size (int): Rendering size
|
70 |
+
mesh_size (int): Mesh generation size
|
71 |
+
|
72 |
+
Returns:
|
73 |
+
str: Path to generated file
|
74 |
+
"""
|
75 |
+
# Preprocess image
|
76 |
+
image = self.preprocess_image(image).to(self.device)
|
77 |
+
|
78 |
+
# Default camera settings (you might want to adjust these)
|
79 |
+
source_camera = self._get_default_source_camera(batch_size=1).to(self.device)
|
80 |
+
|
81 |
+
with torch.no_grad():
|
82 |
+
# Forward pass
|
83 |
+
planes = self.model(image, source_camera)
|
84 |
+
|
85 |
+
if output_type == 'mesh':
|
86 |
+
return self._generate_mesh(planes, mesh_size)
|
87 |
+
elif output_type == 'video':
|
88 |
+
return self._generate_video(planes, render_size)
|
89 |
+
|
90 |
+
def _generate_mesh(self, planes, mesh_size=512):
|
91 |
+
"""
|
92 |
+
Generate 3D mesh from neural planes
|
93 |
+
|
94 |
+
Args:
|
95 |
+
planes: Neural representation planes
|
96 |
+
mesh_size (int): Size of the mesh grid
|
97 |
+
|
98 |
+
Returns:
|
99 |
+
str: Path to saved mesh file
|
100 |
+
"""
|
101 |
+
from skimage import measure
|
102 |
+
import numpy as np
|
103 |
+
import trimesh
|
104 |
+
|
105 |
+
# Use scikit-image's marching cubes instead of mcubes
|
106 |
+
grid_out = self.model.synthesizer.forward_grid(planes=planes, grid_size=mesh_size)
|
107 |
+
|
108 |
+
# Extract the sigma grid and threshold
|
109 |
+
sigma_grid = grid_out['sigma'].float().squeeze(0).squeeze(-1).cpu().numpy()
|
110 |
+
|
111 |
+
# Use marching cubes from scikit-image
|
112 |
+
vtx, faces, _, _ = measure.marching_cubes(sigma_grid, level=1.0)
|
113 |
+
|
114 |
+
# Normalize vertices
|
115 |
+
vtx = vtx / (mesh_size - 1) * 2 - 1
|
116 |
+
|
117 |
+
# Color vertices
|
118 |
+
vtx_tensor = torch.tensor(vtx, dtype=torch.float32, device=self.device).unsqueeze(0)
|
119 |
+
vtx_colors = self.model.synthesizer.forward_points(planes, vtx_tensor)['rgb'].float().squeeze(0).cpu().numpy()
|
120 |
+
vtx_colors = (vtx_colors * 255).astype(np.uint8)
|
121 |
+
|
122 |
+
# Create and save mesh
|
123 |
+
mesh = trimesh.Trimesh(vertices=vtx, faces=faces, vertex_colors=vtx_colors)
|
124 |
+
mesh_path = "generated_mesh.obj"
|
125 |
+
mesh.export(mesh_path, 'obj')
|
126 |
+
return mesh_path
|
127 |
+
def _generate_video(self, planes, render_size=384, fps=30):
|
128 |
+
"""
|
129 |
+
Generate rotating video from neural planes
|
130 |
+
|
131 |
+
Args:
|
132 |
+
planes: Neural representation planes
|
133 |
+
render_size (int): Size of rendered frames
|
134 |
+
fps (int): Frames per second
|
135 |
+
|
136 |
+
Returns:
|
137 |
+
str: Path to saved video file
|
138 |
+
"""
|
139 |
+
render_cameras = self._get_default_render_cameras(batch_size=1).to(self.device)
|
140 |
+
frames = []
|
141 |
+
|
142 |
+
for i in range(0, render_cameras.shape[1], 1):
|
143 |
+
frame_chunk = self.model.synthesizer(
|
144 |
+
planes,
|
145 |
+
render_cameras[:, i:i + 1],
|
146 |
+
render_size,
|
147 |
+
render_size,
|
148 |
+
0,
|
149 |
+
0
|
150 |
+
)
|
151 |
+
frames.append(frame_chunk['images_rgb'])
|
152 |
+
|
153 |
+
frames = torch.cat(frames, dim=1)
|
154 |
+
frames = frames.squeeze(0)
|
155 |
+
frames = (frames.permute(0, 2, 3, 1).cpu().numpy() * 255).astype(np.uint8)
|
156 |
+
|
157 |
+
video_path = "generated_video.mp4"
|
158 |
+
imageio.mimwrite(video_path, frames, fps=fps)
|
159 |
+
|
160 |
+
return video_path
|
161 |
+
|
162 |
+
def _get_default_source_camera(self, batch_size=1):
|
163 |
+
"""Generate default source camera parameters"""
|
164 |
+
# Implement camera generation logic here
|
165 |
+
# This is a placeholder and should match the original implementation
|
166 |
+
pass
|
167 |
+
|
168 |
+
def _get_default_render_cameras(self, batch_size=1):
|
169 |
+
"""Generate default render camera parameters"""
|
170 |
+
# Implement render camera generation logic here
|
171 |
+
# This is a placeholder and should match the original implementation
|
172 |
+
pass
|
173 |
+
|
174 |
+
# Create Gradio Interface
|
175 |
+
def create_vfusion3d_interface():
|
176 |
+
generator = VFusion3DGenerator()
|
177 |
+
|
178 |
+
with gr.Blocks() as demo:
|
179 |
+
with gr.Row():
|
180 |
+
with gr.Column():
|
181 |
+
gr.Markdown("# VFusion3D Model Converter")
|
182 |
+
input_image = gr.Image(type="pil", label="Upload Image")
|
183 |
+
|
184 |
+
with gr.Row():
|
185 |
+
mesh_btn = gr.Button("Generate 3D Mesh")
|
186 |
+
video_btn = gr.Button("Generate Rotation Video")
|
187 |
+
|
188 |
+
mesh_output = gr.File(label="3D Mesh (.obj)")
|
189 |
+
video_output = gr.File(label="Rotation Video")
|
190 |
+
|
191 |
+
with gr.Column():
|
192 |
+
model_viewer = LitModel3D(
|
193 |
+
label="3D Model Preview",
|
194 |
+
scale=1.0,
|
195 |
+
interactive=True
|
196 |
+
)
|
197 |
+
|
198 |
+
# Button click events
|
199 |
+
mesh_btn.click(
|
200 |
+
fn=lambda img: (
|
201 |
+
generator.generate_3d_output(img, output_type='mesh'),
|
202 |
+
generator.generate_3d_output(img, output_type='mesh')
|
203 |
+
),
|
204 |
+
inputs=input_image,
|
205 |
+
outputs=[mesh_output, model_viewer]
|
206 |
+
)
|
207 |
+
|
208 |
+
video_btn.click(
|
209 |
+
fn=lambda img: generator.generate_3d_output(img, output_type='video'),
|
210 |
+
inputs=input_image,
|
211 |
+
outputs=video_output
|
212 |
+
)
|
213 |
+
|
214 |
+
return demo
|
215 |
+
|
216 |
+
# Launch the interface
|
217 |
if __name__ == "__main__":
|
218 |
+
demo = create_vfusion3d_interface()
|
219 |
+
demo.launch()
|