Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
@@ -1,171 +1,136 @@
|
|
1 |
# MIT License
|
2 |
-
|
3 |
# Copyright (c) Microsoft
|
4 |
-
|
5 |
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
6 |
# of this software and associated documentation files (the "Software"), to deal
|
7 |
# in the Software without restriction, including without limitation the rights
|
8 |
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
9 |
-
# copies of the Software, and to permit persons to
|
10 |
-
#
|
11 |
-
|
12 |
-
#
|
13 |
-
#
|
14 |
-
|
15 |
-
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
16 |
-
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
17 |
-
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
18 |
-
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
19 |
-
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
20 |
-
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
21 |
-
# SOFTWARE.
|
22 |
|
23 |
# Copyright (c) [2025] [Microsoft]
|
24 |
-
# Copyright (c) [2025] [Chongjie Ye]
|
25 |
# SPDX-License-Identifier: MIT
|
26 |
# This file has been modified by Chongjie Ye on 2025/04/10
|
27 |
-
# Original file was released under MIT, with the full license text
|
|
|
28 |
# This modified file is released under the same license.
|
29 |
|
30 |
import gradio as gr
|
31 |
import os
|
32 |
|
33 |
# ---- Force CPU-only environment globally ----
|
34 |
-
os.environ["CUDA_VISIBLE_DEVICES"] = "-1" # hide
|
35 |
-
os.environ.setdefault("ATTN_BACKEND", "sdpa") # avoid xformers
|
36 |
os.environ.setdefault("SPCONV_ALGO", "native") # safe sparseconv algo
|
37 |
# ---------------------------------------------
|
|
|
38 |
from typing import *
|
39 |
import torch
|
40 |
import numpy as np
|
41 |
import tempfile
|
42 |
-
|
43 |
import zipfile
|
44 |
import types
|
|
|
45 |
|
46 |
# ---------------------------------------------------------------------------
|
47 |
-
#
|
48 |
-
# The original Hi3DGen implementation expects the `hi3dgen` Python package to
|
49 |
-
# reside alongside this app file. Hugging Face Spaces do not currently
|
50 |
-
# support uploading an entire folder via the web interface, so the `hi3dgen`
|
51 |
-
# source tree is bundled into a single `hi3dgen.zip` archive. On startup we
|
52 |
-
# extract this archive into the working directory if the `hi3dgen` package is
|
53 |
-
# not already present. This allows the rest of the code to `import hi3dgen` as
|
54 |
-
# normal.
|
55 |
# ---------------------------------------------------------------------------
|
56 |
-
|
57 |
def _ensure_hi3dgen_available():
|
58 |
-
"""Unpack hi3dgen.zip into the current directory if the hi3dgen package
|
59 |
-
is missing. This function is idempotent and safe to call multiple times.
|
60 |
-
"""
|
61 |
pkg_name = 'hi3dgen'
|
62 |
-
|
|
|
63 |
if os.path.isdir(pkg_dir):
|
64 |
return
|
65 |
-
archive_path = os.path.join(
|
66 |
-
if os.path.isfile(archive_path):
|
67 |
-
try:
|
68 |
-
with zipfile.ZipFile(archive_path, 'r') as zf:
|
69 |
-
zf.extractall(os.path.dirname(__file__))
|
70 |
-
except Exception as e:
|
71 |
-
raise RuntimeError(f"Failed to extract {archive_path}: {e}")
|
72 |
-
else:
|
73 |
raise FileNotFoundError(
|
74 |
-
f"Required archive {archive_path} is missing.
|
75 |
)
|
|
|
|
|
|
|
|
|
|
|
76 |
|
77 |
-
# Make sure the hi3dgen package is available before importing it
|
78 |
_ensure_hi3dgen_available()
|
79 |
|
80 |
# ---------------------------------------------------------------------------
|
81 |
-
# xformers stub
|
82 |
-
#
|
83 |
-
# Some modules in the Hi3DGen pipeline import `xformers.ops.memory_efficient_attention`
|
84 |
-
# to compute multi-head attention. The official `xformers` library is not
|
85 |
-
# installed in this Space (and requires GPU-only build), so we provide a
|
86 |
-
# minimal in-memory stub that exposes a compatible API backed by PyTorch's
|
87 |
-
# built-in scaled dot-product attention. This stub is lightweight and
|
88 |
-
# CPU-friendly. It registers both the `xformers` and `xformers.ops` modules
|
89 |
-
# in sys.modules so that subsequent imports succeed.
|
90 |
# ---------------------------------------------------------------------------
|
91 |
-
|
92 |
def _ensure_xformers_stub():
|
93 |
-
|
94 |
-
# ---------------------------------------------------------------------------
|
95 |
import sys
|
96 |
-
# If xformers is already available, do nothing.
|
97 |
if 'xformers.ops' in sys.modules:
|
98 |
return
|
99 |
import torch.nn.functional as F
|
100 |
-
|
101 |
-
xformers_mod = types.ModuleType('xformers')
|
102 |
ops_mod = types.ModuleType('xformers.ops')
|
103 |
|
104 |
def memory_efficient_attention(query, key, value, attn_bias=None):
|
105 |
-
"""
|
106 |
-
Fallback implementation of memory_efficient_attention for CPU environments.
|
107 |
-
This wraps torch.nn.functional.scaled_dot_product_attention.
|
108 |
-
"""
|
109 |
-
# PyTorch expects the attention mask (bias) to be additive with shape
|
110 |
-
# broadcastable to (batch, num_heads, seq_len_query, seq_len_key). If
|
111 |
-
# attn_bias is provided and is non-zero, pass it through; otherwise
|
112 |
-
# supply None to avoid unnecessary allocations.
|
113 |
return F.scaled_dot_product_attention(query, key, value, attn_bias)
|
114 |
|
115 |
-
# Populate the ops module with our fallback function
|
116 |
ops_mod.memory_efficient_attention = memory_efficient_attention
|
117 |
-
|
118 |
-
|
119 |
-
# Register modules
|
120 |
-
sys.modules['xformers'] = xformers_mod
|
121 |
sys.modules['xformers.ops'] = ops_mod
|
122 |
|
123 |
-
# Ensure the xformers stub is registered before importing Hi3DGen
|
124 |
_ensure_xformers_stub()
|
125 |
|
126 |
# ---------------------------------------------------------------------------
|
127 |
-
#
|
128 |
-
#
|
129 |
-
# Some utility functions and classes in the Hi3DGen codebase assume the
|
130 |
-
# presence of a CUDA device by default. Specifically, the function
|
131 |
-
# `construct_dense_grid` in `hi3dgen.representations.mesh.utils_cube` uses
|
132 |
-
# `device='cuda'` as its default argument, and the class
|
133 |
-
# `EnhancedMarchingCubes` in `hi3dgen.representations.mesh.cube2mesh` has
|
134 |
-
# a constructor that defaults to `device="cuda"`. On CPU-only Spaces, these
|
135 |
-
# defaults cause runtime errors when PyTorch attempts to allocate tensors on
|
136 |
-
# a non-existent GPU. To avoid this, we override the default arguments for
|
137 |
-
# these functions to use the CPU instead. If the patch cannot be applied
|
138 |
-
# (for example, if the module structure changes in a future version), we
|
139 |
-
# catch any exceptions and log a warning without stopping execution.
|
140 |
-
try:
|
141 |
-
from hi3dgen.representations.mesh import utils_cube
|
142 |
-
if hasattr(utils_cube.construct_dense_grid, '__defaults__'):
|
143 |
-
_defaults = list(utils_cube.construct_dense_grid.__defaults__ or ())
|
144 |
-
if _defaults and _defaults[-1] == 'cuda':
|
145 |
-
_defaults[-1] = 'cpu'
|
146 |
-
utils_cube.construct_dense_grid.__defaults__ = tuple(_defaults)
|
147 |
-
|
148 |
-
from hi3dgen.representations.mesh.cube2mesh import EnhancedMarchingCubes
|
149 |
-
if hasattr(EnhancedMarchingCubes.__init__, '__defaults__'):
|
150 |
-
_mc_defaults = list(EnhancedMarchingCubes.__init__.__defaults__ or ())
|
151 |
-
if _mc_defaults and _mc_defaults[-1] == 'cuda':
|
152 |
-
_mc_defaults[-1] = 'cpu'
|
153 |
-
EnhancedMarchingCubes.__init__.__defaults__ = tuple(_mc_defaults)
|
154 |
-
except Exception as _e:
|
155 |
-
print(f"Warning: failed to apply CPU device overrides: {_e}")
|
156 |
-
|
157 |
from hi3dgen.pipelines import Hi3DGenPipeline
|
158 |
import trimesh
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
159 |
MAX_SEED = np.iinfo(np.int32).max
|
160 |
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
|
161 |
WEIGHTS_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'weights')
|
162 |
os.makedirs(TMP_DIR, exist_ok=True)
|
163 |
os.makedirs(WEIGHTS_DIR, exist_ok=True)
|
164 |
|
|
|
|
|
|
|
165 |
def cache_weights(weights_dir: str) -> dict:
|
166 |
-
import os
|
167 |
from huggingface_hub import snapshot_download
|
168 |
-
|
169 |
os.makedirs(weights_dir, exist_ok=True)
|
170 |
model_ids = [
|
171 |
"Stable-X/trellis-normal-v0-1",
|
@@ -175,49 +140,62 @@ def cache_weights(weights_dir: str) -> dict:
|
|
175 |
cached_paths = {}
|
176 |
for model_id in model_ids:
|
177 |
print(f"Caching weights for: {model_id}")
|
178 |
-
# Check if the model is already cached
|
179 |
local_path = os.path.join(weights_dir, model_id.split("/")[-1])
|
180 |
if os.path.exists(local_path):
|
181 |
print(f"Already cached at: {local_path}")
|
182 |
cached_paths[model_id] = local_path
|
183 |
continue
|
184 |
-
# Download the model and cache it
|
185 |
print(f"Downloading and caching model: {model_id}")
|
186 |
-
|
187 |
-
|
|
|
|
|
|
|
188 |
cached_paths[model_id] = local_path
|
189 |
print(f"Cached at: {local_path}")
|
190 |
-
|
191 |
return cached_paths
|
192 |
|
|
|
|
|
|
|
193 |
def preprocess_mesh(mesh_prompt):
|
194 |
print("Processing mesh")
|
195 |
trimesh_mesh = trimesh.load_mesh(mesh_prompt)
|
196 |
-
|
197 |
-
|
|
|
198 |
|
199 |
def preprocess_image(image):
|
200 |
if image is None:
|
201 |
return None
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
|
|
|
|
|
|
|
|
208 |
if image is None:
|
209 |
return None, None, None
|
210 |
|
211 |
if seed == -1:
|
212 |
seed = np.random.randint(0, MAX_SEED)
|
213 |
-
|
214 |
image = hi3dgen_pipeline.preprocess_image(image, resolution=1024)
|
215 |
-
normal_image = normal_predictor(
|
|
|
|
|
|
|
|
|
|
|
216 |
|
217 |
outputs = hi3dgen_pipeline.run(
|
218 |
normal_image,
|
219 |
seed=seed,
|
220 |
-
formats=["mesh"
|
221 |
preprocess_image=False,
|
222 |
sparse_structure_sampler_params={
|
223 |
"steps": ss_sampling_steps,
|
@@ -229,80 +207,71 @@ def generate_3d(image, seed=-1,
|
|
229 |
},
|
230 |
)
|
231 |
generated_mesh = outputs['mesh'][0]
|
232 |
-
|
233 |
-
# Save outputs
|
234 |
-
import datetime
|
235 |
-
|
236 |
|
|
|
237 |
output_id = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
|
238 |
os.makedirs(os.path.join(TMP_DIR, output_id), exist_ok=True)
|
239 |
mesh_path = f"{TMP_DIR}/{output_id}/mesh.glb"
|
240 |
-
|
241 |
-
# Export mesh
|
242 |
-
trimesh_mesh = generated_mesh.to_trimesh(transform_pose=True)
|
243 |
|
|
|
244 |
trimesh_mesh.export(mesh_path)
|
245 |
|
246 |
return normal_image, mesh_path, mesh_path
|
247 |
|
248 |
def convert_mesh(mesh_path, export_format):
|
249 |
-
"""Download the mesh in the selected format."""
|
250 |
if not mesh_path:
|
251 |
return None
|
252 |
-
|
253 |
-
# Create a temporary file to store the mesh data
|
254 |
temp_file = tempfile.NamedTemporaryFile(suffix=f".{export_format}", delete=False)
|
255 |
temp_file_path = temp_file.name
|
256 |
-
|
257 |
-
new_mesh_path = mesh_path.replace(".glb", f".{export_format}")
|
258 |
mesh = trimesh.load_mesh(mesh_path)
|
259 |
-
mesh.export(temp_file_path)
|
260 |
-
|
261 |
-
return temp_file_path # Return the path to the temporary file
|
262 |
|
263 |
-
#
|
|
|
|
|
264 |
with gr.Blocks(css="footer {visibility: hidden}") as demo:
|
265 |
gr.Markdown(
|
266 |
"""
|
267 |
<h1 style='text-align: center;'>Hi3DGen: High-fidelity 3D Geometry Generation from Images via Normal Bridging</h1>
|
268 |
<p style='text-align: center;'>
|
269 |
<strong>V0.1, Introduced By
|
270 |
-
<a href="https://gaplab.cuhk.edu.cn/" target="_blank">GAP Lab</a>
|
271 |
-
<a href="https://www.nvsgames.cn/" target="_blank">Game-AIGC Team</a>
|
272 |
</p>
|
273 |
"""
|
274 |
)
|
275 |
-
|
276 |
with gr.Row():
|
277 |
gr.Markdown("""
|
278 |
-
|
279 |
-
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
291 |
-
|
292 |
-
|
293 |
|
294 |
with gr.Row():
|
295 |
with gr.Column(scale=1):
|
296 |
with gr.Tabs():
|
297 |
-
|
298 |
with gr.Tab("Single Image"):
|
299 |
with gr.Row():
|
300 |
image_prompt = gr.Image(label="Image Prompt", image_mode="RGBA", type="pil")
|
301 |
normal_output = gr.Image(label="Normal Bridge", image_mode="RGBA", type="pil")
|
302 |
-
|
303 |
with gr.Tab("Multiple Images"):
|
304 |
-
gr.Markdown(
|
305 |
-
|
|
|
|
|
306 |
with gr.Accordion("Advanced Settings", open=False):
|
307 |
seed = gr.Slider(-1, MAX_SEED, label="Seed", value=0, step=1)
|
308 |
gr.Markdown("#### Stage 1: Sparse Structure Generation")
|
@@ -313,15 +282,14 @@ with gr.Blocks(css="footer {visibility: hidden}") as demo:
|
|
313 |
with gr.Row():
|
314 |
slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
|
315 |
slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=6, step=1)
|
316 |
-
|
317 |
with gr.Group():
|
318 |
with gr.Row():
|
319 |
gen_shape_btn = gr.Button("Generate Shape", size="lg", variant="primary")
|
320 |
-
|
321 |
-
# Right column - Output
|
322 |
with gr.Column(scale=1):
|
323 |
with gr.Column():
|
324 |
-
model_output = gr.Model3D(label="3D Model Preview (Each model is
|
325 |
with gr.Column():
|
326 |
export_format = gr.Dropdown(
|
327 |
choices=["obj", "glb", "ply", "stl"],
|
@@ -335,11 +303,11 @@ with gr.Blocks(css="footer {visibility: hidden}") as demo:
|
|
335 |
inputs=[image_prompt],
|
336 |
outputs=[image_prompt]
|
337 |
)
|
338 |
-
|
339 |
gen_shape_btn.click(
|
340 |
generate_3d,
|
341 |
inputs=[
|
342 |
-
image_prompt, seed,
|
343 |
ss_guidance_strength, ss_sampling_steps,
|
344 |
slat_guidance_strength, slat_sampling_steps
|
345 |
],
|
@@ -348,15 +316,13 @@ with gr.Blocks(css="footer {visibility: hidden}") as demo:
|
|
348 |
lambda: gr.Button(interactive=True),
|
349 |
outputs=[download_btn],
|
350 |
)
|
351 |
-
|
352 |
-
|
353 |
def update_download_button(mesh_path, export_format):
|
354 |
if not mesh_path:
|
355 |
return gr.File.update(value=None, interactive=False)
|
356 |
-
|
357 |
download_path = convert_mesh(mesh_path, export_format)
|
358 |
return download_path
|
359 |
-
|
360 |
export_format.change(
|
361 |
update_download_button,
|
362 |
inputs=[model_output, export_format],
|
@@ -365,31 +331,19 @@ with gr.Blocks(css="footer {visibility: hidden}") as demo:
|
|
365 |
lambda: gr.Button(interactive=True),
|
366 |
outputs=[download_btn],
|
367 |
)
|
368 |
-
|
369 |
examples = None
|
370 |
|
371 |
gr.Markdown(
|
372 |
"""
|
373 |
**Acknowledgments**: Hi3DGen is built on the shoulders of giants. We would like to express our gratitude to the open-source research community and the developers of these pioneering projects:
|
374 |
-
- **3D Modeling:** Our 3D Model is finetuned from the SOTA open-source 3D foundation model [Trellis](https://github.com/microsoft/TRELLIS)
|
375 |
-
- **Normal Estimation:**
|
376 |
-
|
377 |
-
**Your contributions and collaboration push the boundaries of 3D modeling!**
|
378 |
"""
|
379 |
)
|
380 |
|
|
|
|
|
|
|
381 |
if __name__ == "__main__":
|
382 |
-
#
|
383 |
-
cache_weights(WEIGHTS_DIR)
|
384 |
-
|
385 |
-
hi3dgen_pipeline = Hi3DGenPipeline.from_pretrained("weights/trellis-normal-v0-1")
|
386 |
-
hi3dgen_pipeline.cuda()
|
387 |
-
|
388 |
-
# Initialize normal predictor
|
389 |
-
try:
|
390 |
-
normal_predictor = torch.hub.load(os.path.join(torch.hub.get_dir(), 'hugoycj_StableNormal_main'), "StableNormal_turbo", yoso_version='yoso-normal-v1-8-1', source='local', local_cache_dir='./weights', pretrained=True)
|
391 |
-
except:
|
392 |
-
normal_predictor = torch.hub.load("hugoycj/StableNormal", "StableNormal_turbo", trust_repo=True, yoso_version='yoso-normal-v1-8-1', local_cache_dir='./weights')
|
393 |
-
|
394 |
-
# Launch the app
|
395 |
-
demo.launch(share=False, server_name="0.0.0.0")
|
|
|
1 |
# MIT License
|
|
|
2 |
# Copyright (c) Microsoft
|
|
|
3 |
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
4 |
# of this software and associated documentation files (the "Software"), to deal
|
5 |
# in the Software without restriction, including without limitation the rights
|
6 |
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
7 |
+
# copies of the Software, and to permit persons to do so, subject to the following conditions:
|
8 |
+
# The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
|
9 |
+
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
10 |
+
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
11 |
+
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
# Copyright (c) [2025] [Microsoft]
|
14 |
+
# Copyright (c) [2025] [Chongjie Ye]
|
15 |
# SPDX-License-Identifier: MIT
|
16 |
# This file has been modified by Chongjie Ye on 2025/04/10
|
17 |
+
# Original file was released under MIT, with the full license text available at:
|
18 |
+
# https://github.com/atong01/conditional-flow-matching/blob/1.0.7/LICENSE
|
19 |
# This modified file is released under the same license.
|
20 |
|
21 |
import gradio as gr
|
22 |
import os
|
23 |
|
24 |
# ---- Force CPU-only environment globally ----
|
25 |
+
os.environ["CUDA_VISIBLE_DEVICES"] = "-1" # hide GPUs from torch
|
26 |
+
os.environ.setdefault("ATTN_BACKEND", "sdpa") # avoid xformers path
|
27 |
os.environ.setdefault("SPCONV_ALGO", "native") # safe sparseconv algo
|
28 |
# ---------------------------------------------
|
29 |
+
|
30 |
from typing import *
|
31 |
import torch
|
32 |
import numpy as np
|
33 |
import tempfile
|
|
|
34 |
import zipfile
|
35 |
import types
|
36 |
+
import importlib
|
37 |
|
38 |
# ---------------------------------------------------------------------------
|
39 |
+
# Ensure bundled hi3dgen sources are available (extracted from hi3dgen.zip)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
40 |
# ---------------------------------------------------------------------------
|
|
|
41 |
def _ensure_hi3dgen_available():
|
|
|
|
|
|
|
42 |
pkg_name = 'hi3dgen'
|
43 |
+
here = os.path.dirname(__file__)
|
44 |
+
pkg_dir = os.path.join(here, pkg_name)
|
45 |
if os.path.isdir(pkg_dir):
|
46 |
return
|
47 |
+
archive_path = os.path.join(here, f"{pkg_name}.zip")
|
48 |
+
if not os.path.isfile(archive_path):
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
raise FileNotFoundError(
|
50 |
+
f"Required archive {archive_path} is missing. Upload hi3dgen.zip next to app.py."
|
51 |
)
|
52 |
+
try:
|
53 |
+
with zipfile.ZipFile(archive_path, 'r') as zf:
|
54 |
+
zf.extractall(here)
|
55 |
+
except Exception as e:
|
56 |
+
raise RuntimeError(f"Failed to extract {archive_path}: {e}")
|
57 |
|
|
|
58 |
_ensure_hi3dgen_available()
|
59 |
|
60 |
# ---------------------------------------------------------------------------
|
61 |
+
# xformers stub (CPU-friendly fallback for xformers.ops.memory_efficient_attention)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
62 |
# ---------------------------------------------------------------------------
|
|
|
63 |
def _ensure_xformers_stub():
|
|
|
|
|
64 |
import sys
|
|
|
65 |
if 'xformers.ops' in sys.modules:
|
66 |
return
|
67 |
import torch.nn.functional as F
|
68 |
+
xf_mod = types.ModuleType('xformers')
|
|
|
69 |
ops_mod = types.ModuleType('xformers.ops')
|
70 |
|
71 |
def memory_efficient_attention(query, key, value, attn_bias=None):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
return F.scaled_dot_product_attention(query, key, value, attn_bias)
|
73 |
|
|
|
74 |
ops_mod.memory_efficient_attention = memory_efficient_attention
|
75 |
+
xf_mod.ops = ops_mod
|
76 |
+
sys.modules['xformers'] = xf_mod
|
|
|
|
|
77 |
sys.modules['xformers.ops'] = ops_mod
|
78 |
|
|
|
79 |
_ensure_xformers_stub()
|
80 |
|
81 |
# ---------------------------------------------------------------------------
|
82 |
+
# Import pipeline AFTER stubbing xformers, then patch CUDA-hotspots to CPU
|
83 |
+
# ---------------------------------------------------------------------------
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
from hi3dgen.pipelines import Hi3DGenPipeline
|
85 |
import trimesh
|
86 |
+
|
87 |
+
# ---- Force CPU inside hi3dgen (avoid any CUDA paths) ----
|
88 |
+
print("[PATCH] Applying CPU monkey-patches to hi3dgen")
|
89 |
+
# 1) utils_cube.construct_dense_grid(..., device=...) -> force CPU
|
90 |
+
uc = importlib.import_module("hi3dgen.representations.mesh.utils_cube")
|
91 |
+
if not hasattr(uc, "_CPU_PATCHED"):
|
92 |
+
_orig_construct_dense_grid = uc.construct_dense_grid
|
93 |
+
def _construct_dense_grid_cpu(res, device=None):
|
94 |
+
return _orig_construct_dense_grid(res, device="cpu")
|
95 |
+
uc.construct_dense_grid = _construct_dense_grid_cpu
|
96 |
+
uc._CPU_PATCHED = True
|
97 |
+
print("[PATCH] utils_cube.construct_dense_grid -> CPU")
|
98 |
+
|
99 |
+
# 2) cube2mesh.EnhancedMarchingCubes default device -> force CPU
|
100 |
+
cm = importlib.import_module("hi3dgen.representations.mesh.cube2mesh")
|
101 |
+
M = cm.EnhancedMarchingCubes
|
102 |
+
if not hasattr(M, "_CPU_PATCHED"):
|
103 |
+
_orig_init = M.__init__
|
104 |
+
def _init_cpu(self, res, device=None):
|
105 |
+
return _orig_init(self, res, device=torch.device("cpu"))
|
106 |
+
M.__init__ = _init_cpu
|
107 |
+
M._CPU_PATCHED = True
|
108 |
+
print("[PATCH] cube2mesh.EnhancedMarchingCubes.__init__ -> CPU")
|
109 |
+
|
110 |
+
# 3) Belt & suspenders: coerce torch.arange(device='cuda') to CPU if any call slips through
|
111 |
+
if not hasattr(torch, "_ARANGE_CPU_PATCHED"):
|
112 |
+
_orig_arange = torch.arange
|
113 |
+
def _arange_cpu(*args, **kwargs):
|
114 |
+
dev = kwargs.get("device", None)
|
115 |
+
if dev is not None and str(dev).startswith("cuda"):
|
116 |
+
kwargs["device"] = "cpu"
|
117 |
+
return _orig_arange(*args, **kwargs)
|
118 |
+
torch.arange = _arange_cpu
|
119 |
+
torch._ARANGE_CPU_PATCHED = True
|
120 |
+
print("[PATCH] torch.arange(device='cuda') -> CPU")
|
121 |
+
# ----------------------------------------------------------
|
122 |
+
|
123 |
MAX_SEED = np.iinfo(np.int32).max
|
124 |
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
|
125 |
WEIGHTS_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'weights')
|
126 |
os.makedirs(TMP_DIR, exist_ok=True)
|
127 |
os.makedirs(WEIGHTS_DIR, exist_ok=True)
|
128 |
|
129 |
+
# ---------------------------------------------------------------------------
|
130 |
+
# Weights caching
|
131 |
+
# ---------------------------------------------------------------------------
|
132 |
def cache_weights(weights_dir: str) -> dict:
|
|
|
133 |
from huggingface_hub import snapshot_download
|
|
|
134 |
os.makedirs(weights_dir, exist_ok=True)
|
135 |
model_ids = [
|
136 |
"Stable-X/trellis-normal-v0-1",
|
|
|
140 |
cached_paths = {}
|
141 |
for model_id in model_ids:
|
142 |
print(f"Caching weights for: {model_id}")
|
|
|
143 |
local_path = os.path.join(weights_dir, model_id.split("/")[-1])
|
144 |
if os.path.exists(local_path):
|
145 |
print(f"Already cached at: {local_path}")
|
146 |
cached_paths[model_id] = local_path
|
147 |
continue
|
|
|
148 |
print(f"Downloading and caching model: {model_id}")
|
149 |
+
local_path = snapshot_download(
|
150 |
+
repo_id=model_id,
|
151 |
+
local_dir=os.path.join(weights_dir, model_id.split("/")[-1]),
|
152 |
+
force_download=False
|
153 |
+
)
|
154 |
cached_paths[model_id] = local_path
|
155 |
print(f"Cached at: {local_path}")
|
|
|
156 |
return cached_paths
|
157 |
|
158 |
+
# ---------------------------------------------------------------------------
|
159 |
+
# Pre/Post processing and generation
|
160 |
+
# ---------------------------------------------------------------------------
|
161 |
def preprocess_mesh(mesh_prompt):
|
162 |
print("Processing mesh")
|
163 |
trimesh_mesh = trimesh.load_mesh(mesh_prompt)
|
164 |
+
out_path = mesh_prompt + '.glb'
|
165 |
+
trimesh_mesh.export(out_path)
|
166 |
+
return out_path
|
167 |
|
168 |
def preprocess_image(image):
|
169 |
if image is None:
|
170 |
return None
|
171 |
+
return hi3dgen_pipeline.preprocess_image(image, resolution=1024)
|
172 |
+
|
173 |
+
def generate_3d(
|
174 |
+
image,
|
175 |
+
seed: int = -1,
|
176 |
+
ss_guidance_strength: float = 3,
|
177 |
+
ss_sampling_steps: int = 50,
|
178 |
+
slat_guidance_strength: float = 3,
|
179 |
+
slat_sampling_steps: int = 6,
|
180 |
+
):
|
181 |
if image is None:
|
182 |
return None, None, None
|
183 |
|
184 |
if seed == -1:
|
185 |
seed = np.random.randint(0, MAX_SEED)
|
186 |
+
|
187 |
image = hi3dgen_pipeline.preprocess_image(image, resolution=1024)
|
188 |
+
normal_image = normal_predictor(
|
189 |
+
image,
|
190 |
+
resolution=768,
|
191 |
+
match_input_resolution=True,
|
192 |
+
data_type='object'
|
193 |
+
)
|
194 |
|
195 |
outputs = hi3dgen_pipeline.run(
|
196 |
normal_image,
|
197 |
seed=seed,
|
198 |
+
formats=["mesh"],
|
199 |
preprocess_image=False,
|
200 |
sparse_structure_sampler_params={
|
201 |
"steps": ss_sampling_steps,
|
|
|
207 |
},
|
208 |
)
|
209 |
generated_mesh = outputs['mesh'][0]
|
|
|
|
|
|
|
|
|
210 |
|
211 |
+
import datetime
|
212 |
output_id = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
|
213 |
os.makedirs(os.path.join(TMP_DIR, output_id), exist_ok=True)
|
214 |
mesh_path = f"{TMP_DIR}/{output_id}/mesh.glb"
|
|
|
|
|
|
|
215 |
|
216 |
+
trimesh_mesh = generated_mesh.to_trimesh(transform_pose=True)
|
217 |
trimesh_mesh.export(mesh_path)
|
218 |
|
219 |
return normal_image, mesh_path, mesh_path
|
220 |
|
221 |
def convert_mesh(mesh_path, export_format):
|
|
|
222 |
if not mesh_path:
|
223 |
return None
|
|
|
|
|
224 |
temp_file = tempfile.NamedTemporaryFile(suffix=f".{export_format}", delete=False)
|
225 |
temp_file_path = temp_file.name
|
|
|
|
|
226 |
mesh = trimesh.load_mesh(mesh_path)
|
227 |
+
mesh.export(temp_file_path)
|
228 |
+
return temp_file_path
|
|
|
229 |
|
230 |
+
# ---------------------------------------------------------------------------
|
231 |
+
# UI
|
232 |
+
# ---------------------------------------------------------------------------
|
233 |
with gr.Blocks(css="footer {visibility: hidden}") as demo:
|
234 |
gr.Markdown(
|
235 |
"""
|
236 |
<h1 style='text-align: center;'>Hi3DGen: High-fidelity 3D Geometry Generation from Images via Normal Bridging</h1>
|
237 |
<p style='text-align: center;'>
|
238 |
<strong>V0.1, Introduced By
|
239 |
+
<a href="https://gaplab.cuhk.edu.cn/" target="_blank">GAP Lab</a> (CUHKSZ) and
|
240 |
+
<a href="https://www.nvsgames.cn/" target="_blank">Game-AIGC Team</a> (ByteDance)</strong>
|
241 |
</p>
|
242 |
"""
|
243 |
)
|
244 |
+
|
245 |
with gr.Row():
|
246 |
gr.Markdown("""
|
247 |
+
<p align="center">
|
248 |
+
<a title="Website" href="https://stable-x.github.io/Hi3DGen/" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
|
249 |
+
<img src="https://www.obukhov.ai/img/badges/badge-website.svg">
|
250 |
+
</a>
|
251 |
+
<a title="arXiv" href="https://stable-x.github.io/Hi3DGen/hi3dgen_paper.pdf" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
|
252 |
+
<img src="https://www.obukhov.ai/img/badges/badge-pdf.svg">
|
253 |
+
</a>
|
254 |
+
<a title="Github" href="https://github.com/Stable-X/Hi3DGen" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
|
255 |
+
<img src="https://img.shields.io/github/stars/Stable-X/Hi3DGen?label=GitHub%20%E2%98%85&logo=github&color=C8C" alt="badge-github-stars">
|
256 |
+
</a>
|
257 |
+
<a title="Social" href="https://x.com/ychngji6" target="_blank" rel="noopener noreferrer" style="display: inline-block;">
|
258 |
+
<img src="https://www.obukhov.ai/img/badges/badge-social.svg" alt="social">
|
259 |
+
</a>
|
260 |
+
</p>
|
261 |
+
""")
|
262 |
|
263 |
with gr.Row():
|
264 |
with gr.Column(scale=1):
|
265 |
with gr.Tabs():
|
|
|
266 |
with gr.Tab("Single Image"):
|
267 |
with gr.Row():
|
268 |
image_prompt = gr.Image(label="Image Prompt", image_mode="RGBA", type="pil")
|
269 |
normal_output = gr.Image(label="Normal Bridge", image_mode="RGBA", type="pil")
|
|
|
270 |
with gr.Tab("Multiple Images"):
|
271 |
+
gr.Markdown(
|
272 |
+
"<div style='text-align: center; padding: 40px; font-size: 24px;'>Multiple Images functionality is coming soon!</div>"
|
273 |
+
)
|
274 |
+
|
275 |
with gr.Accordion("Advanced Settings", open=False):
|
276 |
seed = gr.Slider(-1, MAX_SEED, label="Seed", value=0, step=1)
|
277 |
gr.Markdown("#### Stage 1: Sparse Structure Generation")
|
|
|
282 |
with gr.Row():
|
283 |
slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
|
284 |
slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=6, step=1)
|
285 |
+
|
286 |
with gr.Group():
|
287 |
with gr.Row():
|
288 |
gen_shape_btn = gr.Button("Generate Shape", size="lg", variant="primary")
|
289 |
+
|
|
|
290 |
with gr.Column(scale=1):
|
291 |
with gr.Column():
|
292 |
+
model_output = gr.Model3D(label="3D Model Preview (Each model is ~40MB; may take ~1 min to load)")
|
293 |
with gr.Column():
|
294 |
export_format = gr.Dropdown(
|
295 |
choices=["obj", "glb", "ply", "stl"],
|
|
|
303 |
inputs=[image_prompt],
|
304 |
outputs=[image_prompt]
|
305 |
)
|
306 |
+
|
307 |
gen_shape_btn.click(
|
308 |
generate_3d,
|
309 |
inputs=[
|
310 |
+
image_prompt, seed,
|
311 |
ss_guidance_strength, ss_sampling_steps,
|
312 |
slat_guidance_strength, slat_sampling_steps
|
313 |
],
|
|
|
316 |
lambda: gr.Button(interactive=True),
|
317 |
outputs=[download_btn],
|
318 |
)
|
319 |
+
|
|
|
320 |
def update_download_button(mesh_path, export_format):
|
321 |
if not mesh_path:
|
322 |
return gr.File.update(value=None, interactive=False)
|
|
|
323 |
download_path = convert_mesh(mesh_path, export_format)
|
324 |
return download_path
|
325 |
+
|
326 |
export_format.change(
|
327 |
update_download_button,
|
328 |
inputs=[model_output, export_format],
|
|
|
331 |
lambda: gr.Button(interactive=True),
|
332 |
outputs=[download_btn],
|
333 |
)
|
334 |
+
|
335 |
examples = None
|
336 |
|
337 |
gr.Markdown(
|
338 |
"""
|
339 |
**Acknowledgments**: Hi3DGen is built on the shoulders of giants. We would like to express our gratitude to the open-source research community and the developers of these pioneering projects:
|
340 |
+
- **3D Modeling:** Our 3D Model is finetuned from the SOTA open-source 3D foundation model [Trellis](https://github.com/microsoft/TRELLIS); inspired by [Rodin], [Tripo], and [Dora].
|
341 |
+
- **Normal Estimation:** Builds on [StableNormal] and [GenPercept].
|
|
|
|
|
342 |
"""
|
343 |
)
|
344 |
|
345 |
+
# ---------------------------------------------------------------------------
|
346 |
+
# Entry
|
347 |
+
# ---------------------------------------------------------------------------
|
348 |
if __name__ == "__main__":
|
349 |
+
# Cache model w
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|