Remove temp dir and gr.State
Browse files
app.py
CHANGED
@@ -1,8 +1,8 @@
|
|
1 |
import os
|
2 |
import pathlib
|
3 |
import shlex
|
4 |
-
import shutil
|
5 |
import subprocess
|
|
|
6 |
|
7 |
os.environ["SPCONV_ALGO"] = "native"
|
8 |
|
@@ -29,25 +29,12 @@ from trellis.representations import Gaussian, MeshExtractResult
|
|
29 |
from trellis.utils import postprocessing_utils, render_utils
|
30 |
|
31 |
MAX_SEED = np.iinfo(np.int32).max
|
32 |
-
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "tmp")
|
33 |
-
os.makedirs(TMP_DIR, exist_ok=True)
|
34 |
-
|
35 |
|
36 |
pipeline = TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS-image-large")
|
37 |
pipeline.cuda()
|
38 |
pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))) # Preload rembg
|
39 |
|
40 |
|
41 |
-
def start_session(req: gr.Request):
|
42 |
-
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
43 |
-
os.makedirs(user_dir, exist_ok=True)
|
44 |
-
|
45 |
-
|
46 |
-
def end_session(req: gr.Request):
|
47 |
-
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
48 |
-
shutil.rmtree(user_dir)
|
49 |
-
|
50 |
-
|
51 |
def preprocess_image(image: Image.Image) -> Image.Image:
|
52 |
"""Preprocess the input image.
|
53 |
|
@@ -73,24 +60,26 @@ def preprocess_images(images: list[tuple[Image.Image, str]]) -> list[Image.Image
|
|
73 |
return [pipeline.preprocess_image(image) for image in images]
|
74 |
|
75 |
|
76 |
-
def
|
77 |
-
|
78 |
"gaussian": {
|
79 |
**gs.init_params,
|
80 |
-
"_xyz": gs._xyz
|
81 |
-
"_features_dc": gs._features_dc
|
82 |
-
"_scaling": gs._scaling
|
83 |
-
"_rotation": gs._rotation
|
84 |
-
"_opacity": gs._opacity
|
85 |
},
|
86 |
"mesh": {
|
87 |
-
"vertices": mesh.vertices
|
88 |
-
"faces": mesh.faces
|
89 |
},
|
90 |
}
|
|
|
91 |
|
92 |
|
93 |
-
def
|
|
|
94 |
gs = Gaussian(
|
95 |
aabb=state["gaussian"]["aabb"],
|
96 |
sh_degree=state["gaussian"]["sh_degree"],
|
@@ -99,15 +88,15 @@ def unpack_state(state: dict) -> tuple[Gaussian, EasyDict, str]:
|
|
99 |
opacity_bias=state["gaussian"]["opacity_bias"],
|
100 |
scaling_activation=state["gaussian"]["scaling_activation"],
|
101 |
)
|
102 |
-
gs._xyz =
|
103 |
-
gs._features_dc =
|
104 |
-
gs._scaling =
|
105 |
-
gs._rotation =
|
106 |
-
gs._opacity =
|
107 |
|
108 |
mesh = EasyDict(
|
109 |
-
vertices=
|
110 |
-
faces=
|
111 |
)
|
112 |
|
113 |
return gs, mesh
|
@@ -126,8 +115,7 @@ def image_to_3d(
|
|
126 |
ss_sampling_steps: int,
|
127 |
slat_guidance_strength: float,
|
128 |
slat_sampling_steps: int,
|
129 |
-
|
130 |
-
) -> tuple[dict, str]:
|
131 |
"""Convert an image to a 3D model.
|
132 |
|
133 |
Args:
|
@@ -139,11 +127,9 @@ def image_to_3d(
|
|
139 |
slat_sampling_steps (int): The number of sampling steps for structured latent generation.
|
140 |
|
141 |
Returns:
|
142 |
-
|
143 |
str: The path to the video of the 3D model.
|
144 |
"""
|
145 |
-
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
146 |
-
|
147 |
outputs = pipeline.run(
|
148 |
image,
|
149 |
seed=seed,
|
@@ -162,69 +148,55 @@ def image_to_3d(
|
|
162 |
video = render_utils.render_video(outputs["gaussian"][0], num_frames=120)["color"]
|
163 |
video_geo = render_utils.render_video(outputs["mesh"][0], num_frames=120)["normal"]
|
164 |
video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
|
165 |
-
|
166 |
-
|
167 |
-
|
168 |
-
|
169 |
-
|
|
|
|
|
|
|
|
|
170 |
|
171 |
|
172 |
@spaces.GPU(duration=90)
|
173 |
def extract_glb(
|
174 |
-
|
175 |
mesh_simplify: float,
|
176 |
texture_size: int,
|
177 |
-
|
178 |
-
) -> tuple[str, str]:
|
179 |
"""Extract a GLB file from the 3D model.
|
180 |
|
181 |
Args:
|
182 |
-
|
183 |
mesh_simplify (float): The mesh simplification factor.
|
184 |
texture_size (int): The texture resolution.
|
185 |
|
186 |
Returns:
|
187 |
str: The path to the extracted GLB file.
|
188 |
"""
|
189 |
-
|
190 |
-
gs, mesh = unpack_state(state)
|
191 |
glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
|
192 |
-
glb_path = os.path.join(user_dir, "sample.glb")
|
193 |
-
glb.export(glb_path)
|
194 |
torch.cuda.empty_cache()
|
195 |
-
|
|
|
|
|
196 |
|
197 |
|
198 |
@spaces.GPU
|
199 |
-
def extract_gaussian(
|
200 |
"""Extract a Gaussian file from the 3D model.
|
201 |
|
202 |
Args:
|
203 |
-
|
204 |
|
205 |
Returns:
|
206 |
str: The path to the extracted Gaussian file.
|
207 |
"""
|
208 |
-
|
209 |
-
|
210 |
-
|
211 |
-
|
212 |
-
torch.cuda.empty_cache()
|
213 |
-
return gaussian_path, gaussian_path
|
214 |
-
|
215 |
-
|
216 |
-
def prepare_multi_example() -> list[Image.Image]:
|
217 |
-
multi_case = list(set([i.split("_")[0] for i in os.listdir("assets/example_multi_image")]))
|
218 |
-
images = []
|
219 |
-
for case in multi_case:
|
220 |
-
_images = []
|
221 |
-
for i in range(1, 4):
|
222 |
-
img = Image.open(f"assets/example_multi_image/{case}_{i}.png")
|
223 |
-
W, H = img.size
|
224 |
-
img = img.resize((int(W / H * 512), 512))
|
225 |
-
_images.append(np.array(img))
|
226 |
-
images.append(Image.fromarray(np.concatenate(_images, axis=1)))
|
227 |
-
return images
|
228 |
|
229 |
|
230 |
with gr.Blocks(delete_cache=(600, 600)) as demo:
|
@@ -279,11 +251,7 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
279 |
video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
|
280 |
model_output = gr.Model3D(label="Extracted GLB/Gaussian", height=300)
|
281 |
|
282 |
-
|
283 |
-
download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
|
284 |
-
download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False)
|
285 |
-
|
286 |
-
output_buf = gr.State()
|
287 |
|
288 |
examples = gr.Examples(
|
289 |
examples=sorted(pathlib.Path("assets/example_image").glob("*.png")),
|
@@ -294,10 +262,6 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
294 |
examples_per_page=64,
|
295 |
)
|
296 |
|
297 |
-
# Handlers
|
298 |
-
demo.load(start_session)
|
299 |
-
demo.unload(end_session)
|
300 |
-
|
301 |
image_prompt.upload(
|
302 |
fn=preprocess_image,
|
303 |
inputs=image_prompt,
|
@@ -318,40 +282,21 @@ with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
318 |
slat_guidance_strength,
|
319 |
slat_sampling_steps,
|
320 |
],
|
321 |
-
outputs=[
|
322 |
).then(
|
323 |
fn=lambda: (gr.Button(interactive=True), gr.Button(interactive=True)),
|
324 |
outputs=[extract_glb_btn, extract_gs_btn],
|
|
|
325 |
)
|
326 |
|
327 |
video_output.clear(
|
328 |
fn=lambda: (gr.Button(interactive=False), gr.Button(interactive=False)),
|
329 |
outputs=[extract_glb_btn, extract_gs_btn],
|
|
|
330 |
)
|
331 |
|
332 |
-
extract_glb_btn.click(
|
333 |
-
|
334 |
-
inputs=[output_buf, mesh_simplify, texture_size],
|
335 |
-
outputs=[model_output, download_glb],
|
336 |
-
).then(
|
337 |
-
fn=lambda: gr.Button(interactive=True),
|
338 |
-
outputs=[download_glb],
|
339 |
-
)
|
340 |
-
|
341 |
-
extract_gs_btn.click(
|
342 |
-
fn=extract_gaussian,
|
343 |
-
inputs=[output_buf],
|
344 |
-
outputs=[model_output, download_gs],
|
345 |
-
).then(
|
346 |
-
fn=lambda: gr.Button(interactive=True),
|
347 |
-
outputs=[download_gs],
|
348 |
-
)
|
349 |
-
|
350 |
-
model_output.clear(
|
351 |
-
fn=lambda: gr.Button(interactive=False),
|
352 |
-
outputs=[download_glb],
|
353 |
-
)
|
354 |
-
|
355 |
|
356 |
if __name__ == "__main__":
|
357 |
demo.launch(mcp_server=True)
|
|
|
1 |
import os
|
2 |
import pathlib
|
3 |
import shlex
|
|
|
4 |
import subprocess
|
5 |
+
import tempfile
|
6 |
|
7 |
os.environ["SPCONV_ALGO"] = "native"
|
8 |
|
|
|
29 |
from trellis.utils import postprocessing_utils, render_utils
|
30 |
|
31 |
MAX_SEED = np.iinfo(np.int32).max
|
|
|
|
|
|
|
32 |
|
33 |
pipeline = TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS-image-large")
|
34 |
pipeline.cuda()
|
35 |
pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8))) # Preload rembg
|
36 |
|
37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
def preprocess_image(image: Image.Image) -> Image.Image:
|
39 |
"""Preprocess the input image.
|
40 |
|
|
|
60 |
return [pipeline.preprocess_image(image) for image in images]
|
61 |
|
62 |
|
63 |
+
def save_state_to_file(gs: Gaussian, mesh: MeshExtractResult, output_path: str) -> None:
|
64 |
+
state = {
|
65 |
"gaussian": {
|
66 |
**gs.init_params,
|
67 |
+
"_xyz": gs._xyz,
|
68 |
+
"_features_dc": gs._features_dc,
|
69 |
+
"_scaling": gs._scaling,
|
70 |
+
"_rotation": gs._rotation,
|
71 |
+
"_opacity": gs._opacity,
|
72 |
},
|
73 |
"mesh": {
|
74 |
+
"vertices": mesh.vertices,
|
75 |
+
"faces": mesh.faces,
|
76 |
},
|
77 |
}
|
78 |
+
torch.save(state, output_path)
|
79 |
|
80 |
|
81 |
+
def load_state_from_file(state_path: str) -> tuple[Gaussian, EasyDict]:
|
82 |
+
state = torch.load(state_path)
|
83 |
gs = Gaussian(
|
84 |
aabb=state["gaussian"]["aabb"],
|
85 |
sh_degree=state["gaussian"]["sh_degree"],
|
|
|
88 |
opacity_bias=state["gaussian"]["opacity_bias"],
|
89 |
scaling_activation=state["gaussian"]["scaling_activation"],
|
90 |
)
|
91 |
+
gs._xyz = state["gaussian"]["_xyz"]
|
92 |
+
gs._features_dc = state["gaussian"]["_features_dc"]
|
93 |
+
gs._scaling = state["gaussian"]["_scaling"]
|
94 |
+
gs._rotation = state["gaussian"]["_rotation"]
|
95 |
+
gs._opacity = state["gaussian"]["_opacity"]
|
96 |
|
97 |
mesh = EasyDict(
|
98 |
+
vertices=state["mesh"]["vertices"],
|
99 |
+
faces=state["mesh"]["faces"],
|
100 |
)
|
101 |
|
102 |
return gs, mesh
|
|
|
115 |
ss_sampling_steps: int,
|
116 |
slat_guidance_strength: float,
|
117 |
slat_sampling_steps: int,
|
118 |
+
) -> tuple[str, str]:
|
|
|
119 |
"""Convert an image to a 3D model.
|
120 |
|
121 |
Args:
|
|
|
127 |
slat_sampling_steps (int): The number of sampling steps for structured latent generation.
|
128 |
|
129 |
Returns:
|
130 |
+
str: The path to the pickle file that contains the state of the generated 3D model.
|
131 |
str: The path to the video of the 3D model.
|
132 |
"""
|
|
|
|
|
133 |
outputs = pipeline.run(
|
134 |
image,
|
135 |
seed=seed,
|
|
|
148 |
video = render_utils.render_video(outputs["gaussian"][0], num_frames=120)["color"]
|
149 |
video_geo = render_utils.render_video(outputs["mesh"][0], num_frames=120)["normal"]
|
150 |
video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
|
151 |
+
|
152 |
+
with (
|
153 |
+
tempfile.NamedTemporaryFile(suffix=".pth", delete=False) as state_file,
|
154 |
+
tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as video_file,
|
155 |
+
):
|
156 |
+
save_state_to_file(outputs["gaussian"][0], outputs["mesh"][0], state_file.name)
|
157 |
+
torch.cuda.empty_cache()
|
158 |
+
imageio.mimsave(video_file.name, video, fps=15)
|
159 |
+
return state_file.name, video_file.name
|
160 |
|
161 |
|
162 |
@spaces.GPU(duration=90)
|
163 |
def extract_glb(
|
164 |
+
state_path: str,
|
165 |
mesh_simplify: float,
|
166 |
texture_size: int,
|
167 |
+
) -> str:
|
|
|
168 |
"""Extract a GLB file from the 3D model.
|
169 |
|
170 |
Args:
|
171 |
+
state_path (str): The path to the pickle file that contains the state of the generated 3D model.
|
172 |
mesh_simplify (float): The mesh simplification factor.
|
173 |
texture_size (int): The texture resolution.
|
174 |
|
175 |
Returns:
|
176 |
str: The path to the extracted GLB file.
|
177 |
"""
|
178 |
+
gs, mesh = load_state_from_file(state_path)
|
|
|
179 |
glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
|
|
|
|
|
180 |
torch.cuda.empty_cache()
|
181 |
+
with tempfile.NamedTemporaryFile(suffix=".glb", delete=False) as glb_file:
|
182 |
+
glb.export(glb_file.name)
|
183 |
+
return glb_file.name
|
184 |
|
185 |
|
186 |
@spaces.GPU
|
187 |
+
def extract_gaussian(state_path: str) -> str:
|
188 |
"""Extract a Gaussian file from the 3D model.
|
189 |
|
190 |
Args:
|
191 |
+
state_path (str): The path to the pickle file that contains the state of the generated 3D model.
|
192 |
|
193 |
Returns:
|
194 |
str: The path to the extracted Gaussian file.
|
195 |
"""
|
196 |
+
gs, _ = load_state_from_file(state_path)
|
197 |
+
with tempfile.NamedTemporaryFile(suffix=".ply", delete=False) as gaussian_file:
|
198 |
+
gs.save_ply(gaussian_file.name)
|
199 |
+
return gaussian_file.name
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
200 |
|
201 |
|
202 |
with gr.Blocks(delete_cache=(600, 600)) as demo:
|
|
|
251 |
video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
|
252 |
model_output = gr.Model3D(label="Extracted GLB/Gaussian", height=300)
|
253 |
|
254 |
+
state_file_path = gr.Textbox(visible=False)
|
|
|
|
|
|
|
|
|
255 |
|
256 |
examples = gr.Examples(
|
257 |
examples=sorted(pathlib.Path("assets/example_image").glob("*.png")),
|
|
|
262 |
examples_per_page=64,
|
263 |
)
|
264 |
|
|
|
|
|
|
|
|
|
265 |
image_prompt.upload(
|
266 |
fn=preprocess_image,
|
267 |
inputs=image_prompt,
|
|
|
282 |
slat_guidance_strength,
|
283 |
slat_sampling_steps,
|
284 |
],
|
285 |
+
outputs=[state_file_path, video_output],
|
286 |
).then(
|
287 |
fn=lambda: (gr.Button(interactive=True), gr.Button(interactive=True)),
|
288 |
outputs=[extract_glb_btn, extract_gs_btn],
|
289 |
+
api_name=False,
|
290 |
)
|
291 |
|
292 |
video_output.clear(
|
293 |
fn=lambda: (gr.Button(interactive=False), gr.Button(interactive=False)),
|
294 |
outputs=[extract_glb_btn, extract_gs_btn],
|
295 |
+
api_name=False,
|
296 |
)
|
297 |
|
298 |
+
extract_glb_btn.click(fn=extract_glb, inputs=[state_file_path, mesh_simplify, texture_size], outputs=model_output)
|
299 |
+
extract_gs_btn.click(fn=extract_gaussian, inputs=state_file_path, outputs=model_output)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
300 |
|
301 |
if __name__ == "__main__":
|
302 |
demo.launch(mcp_server=True)
|