Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,13 +1,12 @@
|
|
1 |
import os
|
2 |
import random
|
3 |
import tempfile
|
4 |
-
from typing import Any, List
|
5 |
|
6 |
import spaces
|
7 |
import gradio as gr
|
8 |
import numpy as np
|
9 |
import torch
|
10 |
-
# from gradio_image_prompter import ImagePrompter
|
11 |
from gradio_litmodel3d import LitModel3D
|
12 |
from huggingface_hub import snapshot_download
|
13 |
from PIL import Image
|
@@ -22,113 +21,52 @@ MAX_SEED = np.iinfo(np.int32).max
|
|
22 |
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "tmp")
|
23 |
DTYPE = torch.bfloat16
|
24 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
25 |
-
REPO_ID = "VAST-AI/DetailGen3D"
|
26 |
|
27 |
MARKDOWN = """
|
28 |
## Generating geometry details guided by reference image with [DetailGen3D](https://detailgen3d.github.io/DetailGen3D/)
|
29 |
-
1. Upload a detailed image of the frontal view and a coarse model. Then
|
30 |
-
2. If
|
31 |
-
3.
|
32 |
"""
|
|
|
33 |
EXAMPLES = [
|
34 |
-
[
|
|
|
|
|
|
|
|
|
|
|
35 |
]
|
36 |
-
# EXAMPLES = [
|
37 |
-
# [
|
38 |
-
# # {
|
39 |
-
# # "image": "assets/image/100.png",
|
40 |
-
# # },
|
41 |
-
# "assets/image/100.png",
|
42 |
-
# "assets/model/100.glb",
|
43 |
-
# 42,
|
44 |
-
# False,
|
45 |
-
# ],
|
46 |
-
# [
|
47 |
-
# {
|
48 |
-
# "image": "assets/image/503d193a-1b9b-4685-b05f-00ac82f93d7b.png",
|
49 |
-
# },
|
50 |
-
# "assets/image/503d193a-1b9b-4685-b05f-00ac82f93d7b.png",
|
51 |
-
# "assets/model/503d193a-1b9b-4685-b05f-00ac82f93d7b.glb",
|
52 |
-
# 42,
|
53 |
-
# False,
|
54 |
-
# ],
|
55 |
-
# [
|
56 |
-
# {
|
57 |
-
# "image": "assets/image/34933195-9c2c-4271-8d31-a28bc5348b7a.png",
|
58 |
-
# },
|
59 |
-
# "assets/model/34933195-9c2c-4271-8d31-a28bc5348b7a.glb",
|
60 |
-
# 42,
|
61 |
-
# False,
|
62 |
-
# ],
|
63 |
-
# [
|
64 |
-
# {
|
65 |
-
# "image": "assets/image/a5d09c66-1617-465c-aec9-431f48d9a7e1.png",
|
66 |
-
# },
|
67 |
-
# "assets/model/a5d09c66-1617-465c-aec9-431f48d9a7e1.glb",
|
68 |
-
# 42,
|
69 |
-
# False,
|
70 |
-
# ],
|
71 |
-
# [
|
72 |
-
# {
|
73 |
-
# "image": "assets/image/cb7e6c4a-b4dd-483c-9789-3d4887ee7434.png",
|
74 |
-
# },
|
75 |
-
# "assets/model/cb7e6c4a-b4dd-483c-9789-3d4887ee7434.glb",
|
76 |
-
# 42,
|
77 |
-
# False,
|
78 |
-
# ],
|
79 |
-
# [
|
80 |
-
# {
|
81 |
-
# "image": "assets/image/e799e6b4-3b47-40e0-befb-b156af8758ad.png",
|
82 |
-
# },
|
83 |
-
# "assets/model/instant3d/e799e6b4-3b47-40e0-befb-b156af8758ad.glb",
|
84 |
-
# 42,
|
85 |
-
# False,
|
86 |
-
# ],
|
87 |
-
# ]
|
88 |
|
89 |
os.makedirs(TMP_DIR, exist_ok=True)
|
90 |
-
|
91 |
local_dir = "pretrained_weights/DetailGen3D"
|
92 |
snapshot_download(repo_id=REPO_ID, local_dir=local_dir)
|
93 |
-
pipeline = DetailGen3DPipeline.from_pretrained(
|
94 |
-
local_dir
|
95 |
-
).to(DEVICE, dtype=DTYPE)
|
96 |
-
|
97 |
|
98 |
def load_mesh(mesh_path, num_pc=20480):
|
99 |
-
mesh = trimesh.load(mesh_path,force="mesh")
|
100 |
-
|
101 |
center = mesh.bounding_box.centroid
|
102 |
mesh.apply_translation(-center)
|
103 |
scale = max(mesh.bounding_box.extents)
|
104 |
mesh.apply_scale(1.9 / scale)
|
105 |
-
|
106 |
-
surface, face_indices = trimesh.sample.sample_surface(mesh, 1000000
|
107 |
normal = mesh.face_normals[face_indices]
|
108 |
-
|
109 |
rng = np.random.default_rng()
|
110 |
ind = rng.choice(surface.shape[0], num_pc, replace=False)
|
111 |
surface = torch.FloatTensor(surface[ind])
|
112 |
normal = torch.FloatTensor(normal[ind])
|
113 |
-
|
114 |
-
|
115 |
-
return surface
|
116 |
|
117 |
@torch.no_grad()
|
118 |
@torch.autocast(device_type=DEVICE)
|
119 |
-
def run_detailgen3d(
|
120 |
-
pipeline,
|
121 |
-
image,
|
122 |
-
mesh,
|
123 |
-
seed,
|
124 |
-
num_inference_steps,
|
125 |
-
guidance_scale,
|
126 |
-
):
|
127 |
surface = load_mesh(mesh)
|
128 |
-
|
129 |
batch_size = 1
|
130 |
|
131 |
-
#
|
132 |
box_min = np.array([-1.005, -1.005, -1.005])
|
133 |
box_max = np.array([1.005, 1.005, 1.005])
|
134 |
sampled_points, grid_size, bbox_size = generate_dense_grid_points(
|
@@ -137,25 +75,27 @@ def run_detailgen3d(
|
|
137 |
sampled_points = torch.FloatTensor(sampled_points).to(DEVICE, dtype=DTYPE)
|
138 |
sampled_points = sampled_points.unsqueeze(0).repeat(batch_size, 1, 1)
|
139 |
|
140 |
-
#
|
141 |
sample = pipeline.vae.encode(surface).latent_dist.sample()
|
142 |
-
occ = pipeline(
|
143 |
-
|
144 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
grid_logits = occ.view(grid_size).cpu().numpy()
|
146 |
-
vertices, faces, normals, _ = measure.marching_cubes(
|
147 |
-
grid_logits, 0, method="lewiner"
|
148 |
-
)
|
149 |
vertices = vertices / grid_size * bbox_size + box_min
|
150 |
-
|
151 |
-
return mesh
|
152 |
|
153 |
@spaces.GPU(duration=180)
|
154 |
-
@torch.no_grad()
|
155 |
-
@torch.autocast(device_type=DEVICE)
|
156 |
def run_refinement(
|
157 |
-
|
158 |
-
|
159 |
seed: int,
|
160 |
randomize_seed: bool = False,
|
161 |
num_inference_steps: int = 50,
|
@@ -163,92 +103,88 @@ def run_refinement(
|
|
163 |
):
|
164 |
if randomize_seed:
|
165 |
seed = random.randint(0, MAX_SEED)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
166 |
|
167 |
-
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
|
174 |
-
|
175 |
-
scene = run_detailgen3d(
|
176 |
-
pipeline,
|
177 |
-
rgb_image,
|
178 |
-
mesh,
|
179 |
-
seed,
|
180 |
-
num_inference_steps,
|
181 |
-
guidance_scale,
|
182 |
-
)
|
183 |
-
|
184 |
-
_, tmp_path = tempfile.mkstemp(suffix=".glb", prefix="detailgen3d_", dir=TMP_DIR)
|
185 |
-
scene.export(tmp_path)
|
186 |
-
|
187 |
-
torch.cuda.empty_cache()
|
188 |
-
|
189 |
-
return tmp_path, tmp_path, seed
|
190 |
|
191 |
-
# Demo
|
192 |
with gr.Blocks() as demo:
|
193 |
gr.Markdown(MARKDOWN)
|
194 |
|
195 |
with gr.Row():
|
196 |
with gr.Column():
|
197 |
with gr.Row():
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
with gr.Accordion("Generation Settings", open=False):
|
203 |
-
seed = gr.Slider(
|
204 |
-
label="Seed",
|
205 |
-
minimum=0,
|
206 |
-
maximum=MAX_SEED,
|
207 |
-
step=1,
|
208 |
-
value=0,
|
209 |
-
)
|
210 |
-
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
211 |
-
num_inference_steps = gr.Slider(
|
212 |
-
label="Number of inference steps",
|
213 |
-
minimum=1,
|
214 |
-
maximum=50,
|
215 |
-
step=1,
|
216 |
-
value=50,
|
217 |
)
|
218 |
-
|
219 |
-
label="
|
220 |
-
|
221 |
-
maximum=50.0,
|
222 |
-
step=0.1,
|
223 |
-
value=4.0,
|
224 |
)
|
225 |
-
gen_button = gr.Button("Generate details", variant="primary")
|
226 |
-
|
227 |
-
with gr.Column():
|
228 |
-
model_output = LitModel3D(label="Generated GLB", exposure=1.0, height=500,camera_position=(90,90,3))
|
229 |
-
download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
|
230 |
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
)
|
239 |
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
-
|
245 |
-
|
246 |
-
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
251 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
252 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
253 |
|
254 |
-
demo.launch()
|
|
|
1 |
import os
|
2 |
import random
|
3 |
import tempfile
|
4 |
+
from typing import Any, List
|
5 |
|
6 |
import spaces
|
7 |
import gradio as gr
|
8 |
import numpy as np
|
9 |
import torch
|
|
|
10 |
from gradio_litmodel3d import LitModel3D
|
11 |
from huggingface_hub import snapshot_download
|
12 |
from PIL import Image
|
|
|
21 |
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "tmp")
|
22 |
DTYPE = torch.bfloat16
|
23 |
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
|
24 |
+
REPO_ID = "VAST-AI/DetailGen3D"
|
25 |
|
26 |
MARKDOWN = """
|
27 |
## Generating geometry details guided by reference image with [DetailGen3D](https://detailgen3d.github.io/DetailGen3D/)
|
28 |
+
1. Upload a detailed image of the frontal view and a coarse model. Then click "Run" to generate the refined result.
|
29 |
+
2. If satisfied, download the result using the "Download GLB" button.
|
30 |
+
3. Increase CFG strength for better image consistency.
|
31 |
"""
|
32 |
+
|
33 |
EXAMPLES = [
|
34 |
+
[
|
35 |
+
"assets/image/100.png",
|
36 |
+
"assets/model/100.glb",
|
37 |
+
42,
|
38 |
+
False
|
39 |
+
]
|
40 |
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
|
42 |
os.makedirs(TMP_DIR, exist_ok=True)
|
|
|
43 |
local_dir = "pretrained_weights/DetailGen3D"
|
44 |
snapshot_download(repo_id=REPO_ID, local_dir=local_dir)
|
45 |
+
pipeline = DetailGen3DPipeline.from_pretrained(local_dir).to(DEVICE, dtype=DTYPE)
|
|
|
|
|
|
|
46 |
|
47 |
def load_mesh(mesh_path, num_pc=20480):
|
48 |
+
mesh = trimesh.load(mesh_path, force="mesh")
|
|
|
49 |
center = mesh.bounding_box.centroid
|
50 |
mesh.apply_translation(-center)
|
51 |
scale = max(mesh.bounding_box.extents)
|
52 |
mesh.apply_scale(1.9 / scale)
|
53 |
+
|
54 |
+
surface, face_indices = trimesh.sample.sample_surface(mesh, 1000000)
|
55 |
normal = mesh.face_normals[face_indices]
|
56 |
+
|
57 |
rng = np.random.default_rng()
|
58 |
ind = rng.choice(surface.shape[0], num_pc, replace=False)
|
59 |
surface = torch.FloatTensor(surface[ind])
|
60 |
normal = torch.FloatTensor(normal[ind])
|
61 |
+
return torch.cat([surface, normal], dim=-1).unsqueeze(0).cuda()
|
|
|
|
|
62 |
|
63 |
@torch.no_grad()
|
64 |
@torch.autocast(device_type=DEVICE)
|
65 |
+
def run_detailgen3d(pipeline, image, mesh, seed, num_inference_steps, guidance_scale):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
66 |
surface = load_mesh(mesh)
|
|
|
67 |
batch_size = 1
|
68 |
|
69 |
+
# Grid generation
|
70 |
box_min = np.array([-1.005, -1.005, -1.005])
|
71 |
box_max = np.array([1.005, 1.005, 1.005])
|
72 |
sampled_points, grid_size, bbox_size = generate_dense_grid_points(
|
|
|
75 |
sampled_points = torch.FloatTensor(sampled_points).to(DEVICE, dtype=DTYPE)
|
76 |
sampled_points = sampled_points.unsqueeze(0).repeat(batch_size, 1, 1)
|
77 |
|
78 |
+
# Pipeline execution
|
79 |
sample = pipeline.vae.encode(surface).latent_dist.sample()
|
80 |
+
occ = pipeline(
|
81 |
+
image,
|
82 |
+
latents=sample,
|
83 |
+
sampled_points=sampled_points,
|
84 |
+
guidance_scale=guidance_scale,
|
85 |
+
noise_aug_level=0,
|
86 |
+
num_inference_steps=num_inference_steps
|
87 |
+
).samples[0]
|
88 |
+
|
89 |
+
# Mesh processing
|
90 |
grid_logits = occ.view(grid_size).cpu().numpy()
|
91 |
+
vertices, faces, normals, _ = measure.marching_cubes(grid_logits, 0, method="lewiner")
|
|
|
|
|
92 |
vertices = vertices / grid_size * bbox_size + box_min
|
93 |
+
return trimesh.Trimesh(vertices.astype(np.float32), np.ascontiguousarray(faces))
|
|
|
94 |
|
95 |
@spaces.GPU(duration=180)
|
|
|
|
|
96 |
def run_refinement(
|
97 |
+
image_path: str,
|
98 |
+
mesh_path: str,
|
99 |
seed: int,
|
100 |
randomize_seed: bool = False,
|
101 |
num_inference_steps: int = 50,
|
|
|
103 |
):
|
104 |
if randomize_seed:
|
105 |
seed = random.randint(0, MAX_SEED)
|
106 |
+
|
107 |
+
try:
|
108 |
+
# Validate inputs
|
109 |
+
if not os.path.exists(image_path):
|
110 |
+
raise ValueError(f"Image path {image_path} not found")
|
111 |
+
if not os.path.exists(mesh_path):
|
112 |
+
raise ValueError(f"Mesh path {mesh_path} not found")
|
113 |
+
|
114 |
+
image = Image.open(image_path).convert("RGB")
|
115 |
+
scene = run_detailgen3d(
|
116 |
+
pipeline,
|
117 |
+
image,
|
118 |
+
mesh_path,
|
119 |
+
seed,
|
120 |
+
num_inference_steps,
|
121 |
+
guidance_scale,
|
122 |
+
)
|
123 |
|
124 |
+
# Save temporary result
|
125 |
+
_, tmp_path = tempfile.mkstemp(suffix=".glb", prefix="detailgen3d_", dir=TMP_DIR)
|
126 |
+
scene.export(tmp_path)
|
127 |
+
|
128 |
+
return tmp_path, tmp_path, seed
|
129 |
+
|
130 |
+
finally:
|
131 |
+
torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
132 |
|
133 |
+
# Demo interface
|
134 |
with gr.Blocks() as demo:
|
135 |
gr.Markdown(MARKDOWN)
|
136 |
|
137 |
with gr.Row():
|
138 |
with gr.Column():
|
139 |
with gr.Row():
|
140 |
+
image_input = gr.Image(
|
141 |
+
label="Reference Image",
|
142 |
+
type="filepath",
|
143 |
+
sources=["upload", "clipboard"],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
)
|
145 |
+
mesh_input = gr.Model3D(
|
146 |
+
label="Input Model",
|
147 |
+
camera_position=(90, 90, 3)
|
|
|
|
|
|
|
148 |
)
|
|
|
|
|
|
|
|
|
|
|
149 |
|
150 |
+
with gr.Accordion("Advanced Settings", open=False):
|
151 |
+
seed_input = gr.Slider(0, MAX_SEED, value=0, label="Seed")
|
152 |
+
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
153 |
+
steps_input = gr.Slider(1, 100, value=50, step=1, label="Inference Steps")
|
154 |
+
cfg_scale = gr.Slider(0.0, 20.0, value=4.0, step=0.1, label="CFG Scale")
|
155 |
+
|
156 |
+
run_btn = gr.Button("Generate", variant="primary")
|
|
|
157 |
|
158 |
+
with gr.Column():
|
159 |
+
model_output = LitModel3D(
|
160 |
+
label="Result Preview",
|
161 |
+
height=500,
|
162 |
+
camera_position=(90, 90, 3)
|
163 |
+
)
|
164 |
+
download_btn = gr.DownloadButton(
|
165 |
+
"Download GLB",
|
166 |
+
file_count="multiple",
|
167 |
+
interactive=False
|
168 |
+
)
|
169 |
+
|
170 |
+
# Examples section
|
171 |
+
gr.Examples(
|
172 |
+
examples=EXAMPLES,
|
173 |
+
inputs=[image_input, mesh_input, seed_input, randomize_seed],
|
174 |
+
outputs=[model_output, download_btn, seed_input],
|
175 |
+
fn=run_refinement,
|
176 |
+
cache_examples=False,
|
177 |
+
label="Example Inputs"
|
178 |
+
)
|
179 |
|
180 |
+
# Event handling
|
181 |
+
run_btn.click(
|
182 |
+
run_refinement,
|
183 |
+
inputs=[image_input, mesh_input, seed_input, randomize_seed, steps_input, cfg_scale],
|
184 |
+
outputs=[model_output, download_btn, seed_input]
|
185 |
+
).then(
|
186 |
+
lambda: gr.DownloadButton(interactive=True),
|
187 |
+
outputs=[download_btn]
|
188 |
+
)
|
189 |
|
190 |
+
demo.launch()
|