Spaces:
Running
on
Zero
Running
on
Zero
chat gpot
Browse files
app.py
CHANGED
|
@@ -1,22 +1,17 @@
|
|
| 1 |
-
# Version: 1.1.
|
| 2 |
-
#
|
| 3 |
-
#
|
| 4 |
-
#
|
| 5 |
-
#
|
| 6 |
-
#
|
| 7 |
-
#
|
| 8 |
-
# so the UI continues to function by passing the dictionary through output_buf.
|
| 9 |
-
# - Added minor safety checks and logging.
|
| 10 |
|
| 11 |
import gradio as gr
|
| 12 |
import spaces
|
| 13 |
-
|
| 14 |
import os
|
| 15 |
import shutil
|
| 16 |
os.environ['TOKENIZERS_PARALLELISM'] = 'true'
|
| 17 |
-
|
| 18 |
-
# os.environ.setdefault('SPCONV_ALGO', 'native') # Use setdefault to avoid overwriting if already set
|
| 19 |
-
os.environ['SPCONV_ALGO'] = 'native' # Direct set as per original
|
| 20 |
|
| 21 |
from typing import *
|
| 22 |
import torch
|
|
@@ -26,105 +21,80 @@ from easydict import EasyDict as edict
|
|
| 26 |
from trellis.pipelines import TrellisTextTo3DPipeline
|
| 27 |
from trellis.representations import Gaussian, MeshExtractResult
|
| 28 |
from trellis.utils import render_utils, postprocessing_utils
|
| 29 |
-
|
| 30 |
import traceback
|
| 31 |
import sys
|
| 32 |
|
| 33 |
-
|
| 34 |
MAX_SEED = np.iinfo(np.int32).max
|
| 35 |
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
|
| 36 |
os.makedirs(TMP_DIR, exist_ok=True)
|
| 37 |
|
| 38 |
|
| 39 |
def start_session(req: gr.Request):
|
| 40 |
-
"""Creates a temporary directory for the user session."""
|
| 41 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 42 |
os.makedirs(user_dir, exist_ok=True)
|
| 43 |
print(f"Started session, created directory: {user_dir}")
|
| 44 |
|
| 45 |
|
| 46 |
def end_session(req: gr.Request):
|
| 47 |
-
"""Removes the temporary directory for the user session."""
|
| 48 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 49 |
if os.path.exists(user_dir):
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
else:
|
| 56 |
print(f"Ended session, directory already removed: {user_dir}")
|
| 57 |
|
| 58 |
|
| 59 |
def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
|
| 60 |
-
"""Packs Gaussian and Mesh data into a serializable dictionary."""
|
| 61 |
-
# Ensure tensors are on CPU and converted to numpy before returning the dict
|
| 62 |
-
print("[pack_state] Packing state to dictionary...")
|
| 63 |
packed_data = {
|
| 64 |
'gaussian': {
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
'_xyz': gs._xyz.detach().cpu().numpy(),
|
| 68 |
-
'_features_dc': gs._features_dc.detach().cpu().numpy(),
|
| 69 |
-
'_scaling': gs._scaling.detach().cpu().numpy(),
|
| 70 |
-
'_rotation': gs._rotation.detach().cpu().numpy(),
|
| 71 |
-
'_opacity': gs._opacity.detach().cpu().numpy(),
|
| 72 |
},
|
| 73 |
'mesh': {
|
| 74 |
-
'vertices': mesh.vertices.detach().cpu().numpy(),
|
| 75 |
-
'faces': mesh.faces.detach().cpu().numpy(),
|
| 76 |
},
|
| 77 |
}
|
| 78 |
-
print(f"[pack_state] Dictionary created. Keys: {list(packed_data.keys())}, Gaussian points: {len(packed_data['gaussian']['_xyz'])}, Mesh vertices: {len(packed_data['mesh']['vertices'])}")
|
| 79 |
return packed_data
|
| 80 |
|
| 81 |
|
| 82 |
def unpack_state(state_dict: dict) -> Tuple[Gaussian, edict]:
|
| 83 |
-
"
|
| 84 |
-
print("[unpack_state] Unpacking state from dictionary...")
|
| 85 |
-
if not isinstance(state_dict, dict) or 'gaussian' not in state_dict or 'mesh' not in state_dict:
|
| 86 |
-
raise ValueError("Invalid state_dict structure passed to unpack_state.")
|
| 87 |
-
|
| 88 |
-
# Ensure the device is correctly set when unpacking
|
| 89 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
| 90 |
-
print(f"[unpack_state] Using device: {device}")
|
| 91 |
-
|
| 92 |
gauss_data = state_dict['gaussian']
|
| 93 |
mesh_data = state_dict['mesh']
|
| 94 |
-
|
| 95 |
-
# Recreate Gaussian object using parameters stored during packing
|
| 96 |
gs = Gaussian(
|
| 97 |
-
aabb=gauss_data.get('aabb'),
|
| 98 |
sh_degree=gauss_data.get('sh_degree'),
|
| 99 |
mininum_kernel_size=gauss_data.get('mininum_kernel_size'),
|
| 100 |
scaling_bias=gauss_data.get('scaling_bias'),
|
| 101 |
opacity_bias=gauss_data.get('opacity_bias'),
|
| 102 |
scaling_activation=gauss_data.get('scaling_activation'),
|
| 103 |
)
|
| 104 |
-
|
| 105 |
-
gs.
|
| 106 |
-
gs.
|
| 107 |
-
gs.
|
| 108 |
-
gs.
|
| 109 |
-
gs._opacity = torch.tensor(gauss_data['_opacity'], device=device, dtype=torch.float32)
|
| 110 |
-
print(f"[unpack_state] Gaussian unpacked. Points: {gs.get_xyz.shape[0]}")
|
| 111 |
-
|
| 112 |
-
# Recreate mesh object using edict for compatibility if needed elsewhere
|
| 113 |
mesh = edict(
|
| 114 |
-
vertices=torch.tensor(mesh_data['vertices'], device=device, dtype=torch.float32),
|
| 115 |
-
faces=torch.tensor(mesh_data['faces'], device=device, dtype=torch.int64),
|
| 116 |
)
|
| 117 |
-
print(f"[unpack_state] Mesh unpacked. Vertices: {mesh.vertices.shape[0]}, Faces: {mesh.faces.shape[0]}")
|
| 118 |
-
|
| 119 |
return gs, mesh
|
| 120 |
|
| 121 |
|
| 122 |
def get_seed(randomize_seed: bool, seed: int) -> int:
|
| 123 |
-
"""Gets a seed value, randomizing if requested."""
|
| 124 |
new_seed = np.random.randint(0, MAX_SEED) if randomize_seed else seed
|
| 125 |
-
|
| 126 |
-
return int(new_seed) # Ensure it's a standard int
|
| 127 |
-
|
| 128 |
|
| 129 |
@spaces.GPU
|
| 130 |
def text_to_3d(
|
|
@@ -135,331 +105,137 @@ def text_to_3d(
|
|
| 135 |
slat_guidance_strength: float,
|
| 136 |
slat_sampling_steps: int,
|
| 137 |
req: gr.Request,
|
| 138 |
-
) -> Tuple[dict, str]:
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 145 |
os.makedirs(user_dir, exist_ok=True)
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
try:
|
| 150 |
-
print("[text_to_3d] Running Trellis pipeline...")
|
| 151 |
-
outputs = pipeline.run(
|
| 152 |
-
prompt,
|
| 153 |
-
seed=seed,
|
| 154 |
-
formats=["gaussian", "mesh"], # Ensure both are generated
|
| 155 |
-
sparse_structure_sampler_params={
|
| 156 |
-
"steps": int(ss_sampling_steps), # Ensure steps are int
|
| 157 |
-
"cfg_strength": float(ss_guidance_strength),
|
| 158 |
-
},
|
| 159 |
-
slat_sampler_params={
|
| 160 |
-
"steps": int(slat_sampling_steps), # Ensure steps are int
|
| 161 |
-
"cfg_strength": float(slat_guidance_strength),
|
| 162 |
-
},
|
| 163 |
-
)
|
| 164 |
-
print("[text_to_3d] Pipeline run completed.")
|
| 165 |
-
except Exception as e:
|
| 166 |
-
print(f"❌ [text_to_3d] Pipeline error: {e}", file=sys.stderr)
|
| 167 |
-
traceback.print_exc()
|
| 168 |
-
# Return an empty dict and maybe an error indicator path or None?
|
| 169 |
-
# For now, re-raise to signal failure clearly upstream.
|
| 170 |
-
raise gr.Error(f"Trellis pipeline failed: {e}")
|
| 171 |
-
|
| 172 |
-
# --- Create Serializable State Dictionary --- VITAL CHANGE for API
|
| 173 |
-
# This dictionary holds the necessary data for later extraction.
|
| 174 |
-
try:
|
| 175 |
-
state_dict = pack_state(outputs['gaussian'][0], outputs['mesh'][0])
|
| 176 |
-
except Exception as e:
|
| 177 |
-
print(f"❌ [text_to_3d] pack_state error: {e}", file=sys.stderr)
|
| 178 |
-
traceback.print_exc()
|
| 179 |
-
raise gr.Error(f"Failed to pack state: {e}")
|
| 180 |
-
|
| 181 |
-
# --- Render Video Preview ---
|
| 182 |
-
try:
|
| 183 |
-
print("[text_to_3d] Rendering video preview...")
|
| 184 |
-
video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
|
| 185 |
-
video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
|
| 186 |
-
# Ensure video frames are uint8
|
| 187 |
-
video = [np.concatenate([v.astype(np.uint8), vg.astype(np.uint8)], axis=1) for v, vg in zip(video, video_geo)]
|
| 188 |
-
video_path = os.path.join(user_dir, 'sample.mp4')
|
| 189 |
-
imageio.mimsave(video_path, video, fps=15, quality=8) # Added quality setting
|
| 190 |
-
print(f"[text_to_3d] Video saved to: {video_path}")
|
| 191 |
-
except Exception as e:
|
| 192 |
-
print(f"❌ [text_to_3d] Video rendering/saving error: {e}", file=sys.stderr)
|
| 193 |
-
traceback.print_exc()
|
| 194 |
-
# Still return state_dict, but maybe signal video error? Return None for path.
|
| 195 |
-
video_path = None # Indicate video failure
|
| 196 |
-
|
| 197 |
-
# --- Cleanup and Return ---
|
| 198 |
-
# Clear CUDA cache if GPU was used
|
| 199 |
-
if torch.cuda.is_available():
|
| 200 |
-
torch.cuda.empty_cache()
|
| 201 |
-
print("[text_to_3d] Cleared CUDA cache.")
|
| 202 |
-
|
| 203 |
-
# --- Return Serializable Dictionary and Video Path --- VITAL CHANGE for API
|
| 204 |
-
print("[text_to_3d] Returning state dictionary and video path.")
|
| 205 |
return state_dict, video_path
|
| 206 |
|
| 207 |
-
|
| 208 |
-
@spaces.GPU(duration=120) # Increased duration slightly
|
| 209 |
def extract_glb(
|
| 210 |
-
state_dict: dict,
|
| 211 |
mesh_simplify: float,
|
| 212 |
texture_size: int,
|
| 213 |
req: gr.Request,
|
| 214 |
) -> Tuple[str, str]:
|
| 215 |
-
|
| 216 |
-
Extracts a GLB file from the provided 3D model state dictionary.
|
| 217 |
-
"""
|
| 218 |
-
print(f"[extract_glb] Received request. Simplify: {mesh_simplify}, Texture Size: {texture_size}")
|
| 219 |
-
if not isinstance(state_dict, dict):
|
| 220 |
-
print("❌ [extract_glb] Error: Invalid state_dict received (not a dictionary).")
|
| 221 |
-
raise gr.Error("Invalid state data received. Please generate the model first.")
|
| 222 |
-
|
| 223 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 224 |
os.makedirs(user_dir, exist_ok=True)
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
gs, mesh = unpack_state(state_dict)
|
| 230 |
-
except Exception as e:
|
| 231 |
-
print(f"❌ [extract_glb] unpack_state error: {e}", file=sys.stderr)
|
| 232 |
-
traceback.print_exc()
|
| 233 |
-
raise gr.Error(f"Failed to unpack state: {e}")
|
| 234 |
-
|
| 235 |
-
# --- Postprocessing and Export ---
|
| 236 |
-
try:
|
| 237 |
-
print("[extract_glb] Converting to GLB...")
|
| 238 |
-
glb = postprocessing_utils.to_glb(gs, mesh, simplify=float(mesh_simplify), texture_size=int(texture_size), verbose=True) # Verbose for debugging
|
| 239 |
-
glb_path = os.path.join(user_dir, 'sample.glb')
|
| 240 |
-
print(f"[extract_glb] Exporting GLB to: {glb_path}")
|
| 241 |
-
glb.export(glb_path)
|
| 242 |
-
print("[extract_glb] GLB exported successfully.")
|
| 243 |
-
except Exception as e:
|
| 244 |
-
print(f"❌ [extract_glb] GLB conversion/export error: {e}", file=sys.stderr)
|
| 245 |
-
traceback.print_exc()
|
| 246 |
-
raise gr.Error(f"Failed to extract GLB: {e}")
|
| 247 |
-
|
| 248 |
-
# --- Cleanup and Return ---
|
| 249 |
-
if torch.cuda.is_available():
|
| 250 |
-
torch.cuda.empty_cache()
|
| 251 |
-
print("[extract_glb] Cleared CUDA cache.")
|
| 252 |
-
|
| 253 |
-
# Return path twice for both Model3D and DownloadButton components
|
| 254 |
-
print("[extract_glb] Returning GLB path.")
|
| 255 |
return glb_path, glb_path
|
| 256 |
|
| 257 |
-
|
| 258 |
@spaces.GPU
|
| 259 |
def extract_gaussian(
|
| 260 |
-
state_dict: dict,
|
| 261 |
req: gr.Request
|
| 262 |
) -> Tuple[str, str]:
|
| 263 |
-
|
| 264 |
-
Extracts a PLY (Gaussian) file from the provided 3D model state dictionary.
|
| 265 |
-
"""
|
| 266 |
-
print("[extract_gaussian] Received request.")
|
| 267 |
-
if not isinstance(state_dict, dict):
|
| 268 |
-
print("❌ [extract_gaussian] Error: Invalid state_dict received (not a dictionary).")
|
| 269 |
-
raise gr.Error("Invalid state data received. Please generate the model first.")
|
| 270 |
-
|
| 271 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 272 |
os.makedirs(user_dir, exist_ok=True)
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
try:
|
| 277 |
-
gs, _ = unpack_state(state_dict) # Only need Gaussian part
|
| 278 |
-
except Exception as e:
|
| 279 |
-
print(f"❌ [extract_gaussian] unpack_state error: {e}", file=sys.stderr)
|
| 280 |
-
traceback.print_exc()
|
| 281 |
-
raise gr.Error(f"Failed to unpack state: {e}")
|
| 282 |
-
|
| 283 |
-
# --- Export PLY ---
|
| 284 |
-
try:
|
| 285 |
-
gaussian_path = os.path.join(user_dir, 'sample.ply')
|
| 286 |
-
print(f"[extract_gaussian] Saving PLY to: {gaussian_path}")
|
| 287 |
-
gs.save_ply(gaussian_path)
|
| 288 |
-
print("[extract_gaussian] PLY saved successfully.")
|
| 289 |
-
except Exception as e:
|
| 290 |
-
print(f"❌ [extract_gaussian] PLY saving error: {e}", file=sys.stderr)
|
| 291 |
-
traceback.print_exc()
|
| 292 |
-
raise gr.Error(f"Failed to extract Gaussian PLY: {e}")
|
| 293 |
-
|
| 294 |
-
# --- Cleanup and Return ---
|
| 295 |
-
if torch.cuda.is_available():
|
| 296 |
-
torch.cuda.empty_cache()
|
| 297 |
-
print("[extract_gaussian] Cleared CUDA cache.")
|
| 298 |
-
|
| 299 |
-
# Return path twice for both Model3D and DownloadButton components
|
| 300 |
-
print("[extract_gaussian] Returning PLY path.")
|
| 301 |
return gaussian_path, gaussian_path
|
| 302 |
|
| 303 |
-
|
| 304 |
# --- Gradio UI Definition ---
|
| 305 |
-
print("Setting up Gradio Blocks interface...")
|
| 306 |
with gr.Blocks(delete_cache=(600, 600), title="TRELLIS Text-to-3D") as demo:
|
| 307 |
gr.Markdown("""
|
| 308 |
# Text to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
|
| 309 |
-
* Type a text prompt and click "Generate" to create a 3D asset preview.
|
| 310 |
-
* Adjust extraction settings if desired.
|
| 311 |
-
* Click "Extract GLB" or "Extract Gaussian" to get the downloadable 3D file.
|
| 312 |
""")
|
| 313 |
|
| 314 |
-
#
|
| 315 |
-
# This hidden component will hold the dictionary returned by text_to_3d,
|
| 316 |
-
# acting as the state link between generation and extraction for the UI/API.
|
| 317 |
output_buf = gr.State()
|
| 318 |
|
| 319 |
with gr.Row():
|
| 320 |
-
with gr.Column(scale=1):
|
| 321 |
-
text_prompt = gr.Textbox(label="Text Prompt", lines=5
|
| 322 |
-
|
| 323 |
with gr.Accordion(label="Generation Settings", open=False):
|
| 324 |
seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
|
| 325 |
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
| 326 |
-
gr.Markdown("
|
| 327 |
-
|
| 328 |
-
|
| 329 |
-
|
| 330 |
-
gr.
|
| 331 |
-
|
| 332 |
-
slat_guidance_strength = gr.Slider(0.0, 15.0, label="Guidance Strength", value=7.5, step=0.1)
|
| 333 |
-
slat_sampling_steps = gr.Slider(10, 50, label="Sampling Steps", value=25, step=1)
|
| 334 |
-
|
| 335 |
generate_btn = gr.Button("Generate 3D Preview", variant="primary")
|
| 336 |
-
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
gr.
|
| 346 |
-
|
| 347 |
-
|
| 348 |
-
|
| 349 |
-
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
with gr.Row():
|
| 354 |
-
# Link download button visibility/interactivity to model_output potentially
|
| 355 |
-
download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
|
| 356 |
-
download_gs = gr.DownloadButton(label="Download Gaussian (PLY)", interactive=False)
|
| 357 |
-
|
| 358 |
-
# --- Event Handlers ---
|
| 359 |
-
print("Defining Gradio event handlers...")
|
| 360 |
-
|
| 361 |
-
# Handle session start/end
|
| 362 |
-
demo.load(start_session)
|
| 363 |
-
demo.unload(end_session)
|
| 364 |
-
|
| 365 |
-
# --- Generate Button Click Flow ---
|
| 366 |
-
# 1. Get Seed -> 2. Run text_to_3d -> 3. Enable extraction buttons
|
| 367 |
generate_event = generate_btn.click(
|
| 368 |
get_seed,
|
| 369 |
inputs=[randomize_seed, seed],
|
| 370 |
outputs=[seed],
|
| 371 |
-
api_name="get_seed" # Optional API name
|
| 372 |
).then(
|
| 373 |
text_to_3d,
|
| 374 |
inputs=[text_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
|
| 375 |
-
outputs=[output_buf, video_output],
|
| 376 |
-
api_name="text_to_3d"
|
| 377 |
).then(
|
| 378 |
-
lambda: (
|
| 379 |
-
|
| 380 |
-
gr.Button(interactive=True),
|
| 381 |
-
gr.DownloadButton(interactive=False), # Ensure download buttons are disabled initially
|
| 382 |
-
gr.DownloadButton(interactive=False)
|
| 383 |
-
),
|
| 384 |
-
outputs=[extract_glb_btn, extract_gs_btn, download_glb, download_gs], # Update interactivity
|
| 385 |
)
|
| 386 |
|
| 387 |
-
# --- Clear video/model outputs if prompt changes (optional, prevents confusion)
|
| 388 |
-
# text_prompt.change(lambda: (None, None, gr.Button(interactive=False), gr.Button(interactive=False)), outputs=[video_output, model_output, extract_glb_btn, extract_gs_btn])
|
| 389 |
-
|
| 390 |
-
# --- Extract GLB Button Click Flow ---
|
| 391 |
-
# 1. Run extract_glb -> 2. Update Model3D and Download Button
|
| 392 |
extract_glb_event = extract_glb_btn.click(
|
| 393 |
extract_glb,
|
| 394 |
-
inputs=[output_buf, mesh_simplify, texture_size],
|
| 395 |
-
outputs=[model_output, download_glb],
|
| 396 |
-
api_name="extract_glb"
|
| 397 |
).then(
|
| 398 |
-
lambda:
|
| 399 |
outputs=[download_glb],
|
| 400 |
)
|
| 401 |
|
| 402 |
-
# --- Extract Gaussian Button Click Flow ---
|
| 403 |
-
# 1. Run extract_gaussian -> 2. Update Model3D and Download Button
|
| 404 |
extract_gs_event = extract_gs_btn.click(
|
| 405 |
extract_gaussian,
|
| 406 |
-
inputs=[output_buf],
|
| 407 |
-
outputs=[model_output, download_gs],
|
| 408 |
-
api_name="extract_gaussian"
|
| 409 |
).then(
|
| 410 |
-
lambda:
|
| 411 |
outputs=[download_gs],
|
| 412 |
)
|
| 413 |
|
| 414 |
-
#
|
| 415 |
-
# This might be redundant if generate disables them, but adds safety
|
| 416 |
model_output.clear(
|
| 417 |
-
lambda: (
|
| 418 |
-
outputs=[download_glb, download_gs]
|
| 419 |
)
|
| 420 |
-
video_output.clear(
|
| 421 |
-
|
| 422 |
-
gr.Button(interactive=False),
|
| 423 |
-
gr.Button(interactive=False),
|
| 424 |
-
gr.DownloadButton(interactive=False),
|
| 425 |
-
gr.DownloadButton(interactive=False)
|
| 426 |
-
),
|
| 427 |
outputs=[extract_glb_btn, extract_gs_btn, download_glb, download_gs],
|
| 428 |
)
|
| 429 |
|
| 430 |
-
print("Gradio interface setup complete.")
|
| 431 |
-
|
| 432 |
-
|
| 433 |
-
# --- Launch the Gradio app ---
|
| 434 |
if __name__ == "__main__":
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
pipeline = TrellisTextTo3DPipeline.from_pretrained(
|
| 439 |
-
"JeffreyXiang/TRELLIS-text-xlarge",
|
| 440 |
-
# revision="main", # Specify if needed
|
| 441 |
-
torch_dtype=torch.float16 # Use float16 if GPU supports it for less memory
|
| 442 |
-
)
|
| 443 |
-
# Move to GPU if available
|
| 444 |
-
if torch.cuda.is_available():
|
| 445 |
-
pipeline = pipeline.to("cuda")
|
| 446 |
-
print("✅ Trellis pipeline loaded successfully to GPU.")
|
| 447 |
-
else:
|
| 448 |
-
print("⚠️ WARNING: CUDA not available, running on CPU (will be very slow).")
|
| 449 |
-
print("✅ Trellis pipeline loaded successfully to CPU.")
|
| 450 |
-
except Exception as e:
|
| 451 |
-
print(f"❌ Failed to load Trellis pipeline: {e}", file=sys.stderr)
|
| 452 |
-
traceback.print_exc()
|
| 453 |
-
# Exit if pipeline is critical for the app to run
|
| 454 |
-
print("❌ Exiting due to pipeline load failure.")
|
| 455 |
-
sys.exit(1)
|
| 456 |
-
|
| 457 |
-
print("Launching Gradio demo...")
|
| 458 |
-
# Set share=True if you need a public link (e.g., for testing from outside local network)
|
| 459 |
-
# Set server_name="0.0.0.0" to allow access from local network IP
|
| 460 |
-
demo.queue().launch( # Use queue for potentially long-running tasks
|
| 461 |
-
# server_name="0.0.0.0",
|
| 462 |
-
# share=False,
|
| 463 |
-
debug=True # Enable debug mode for more logs
|
| 464 |
)
|
| 465 |
-
|
|
|
|
|
|
| 1 |
+
# Version: 1.1.1 - Targeted signature and state fixes
|
| 2 |
+
# Applied:
|
| 3 |
+
# - Removed unsupported inputs/outputs kwargs on demo.load/unload
|
| 4 |
+
# - Converted NumPy arrays to lists in pack_state for JSON safety
|
| 5 |
+
# - Fixed indentation in Blocks event-handlers
|
| 6 |
+
# - Verified clear() callbacks use only callback + outputs
|
| 7 |
+
# - Bumped version, added comments at change sites
|
|
|
|
|
|
|
| 8 |
|
| 9 |
import gradio as gr
|
| 10 |
import spaces
|
|
|
|
| 11 |
import os
|
| 12 |
import shutil
|
| 13 |
os.environ['TOKENIZERS_PARALLELISM'] = 'true'
|
| 14 |
+
os.environ['SPCONV_ALGO'] = 'native'
|
|
|
|
|
|
|
| 15 |
|
| 16 |
from typing import *
|
| 17 |
import torch
|
|
|
|
| 21 |
from trellis.pipelines import TrellisTextTo3DPipeline
|
| 22 |
from trellis.representations import Gaussian, MeshExtractResult
|
| 23 |
from trellis.utils import render_utils, postprocessing_utils
|
|
|
|
| 24 |
import traceback
|
| 25 |
import sys
|
| 26 |
|
|
|
|
| 27 |
MAX_SEED = np.iinfo(np.int32).max
|
| 28 |
TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
|
| 29 |
os.makedirs(TMP_DIR, exist_ok=True)
|
| 30 |
|
| 31 |
|
| 32 |
def start_session(req: gr.Request):
|
|
|
|
| 33 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 34 |
os.makedirs(user_dir, exist_ok=True)
|
| 35 |
print(f"Started session, created directory: {user_dir}")
|
| 36 |
|
| 37 |
|
| 38 |
def end_session(req: gr.Request):
|
|
|
|
| 39 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 40 |
if os.path.exists(user_dir):
|
| 41 |
+
try:
|
| 42 |
+
shutil.rmtree(user_dir)
|
| 43 |
+
print(f"Ended session, removed directory: {user_dir}")
|
| 44 |
+
except OSError as e:
|
| 45 |
+
print(f"Error removing tmp directory {user_dir}: {e.strerror}", file=sys.stderr)
|
| 46 |
else:
|
| 47 |
print(f"Ended session, directory already removed: {user_dir}")
|
| 48 |
|
| 49 |
|
| 50 |
def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
|
| 51 |
+
"""Packs Gaussian and Mesh data into a JSON-serializable dictionary."""
|
|
|
|
|
|
|
| 52 |
packed_data = {
|
| 53 |
'gaussian': {
|
| 54 |
+
**{k: v for k, v in gs.init_params.items()},
|
| 55 |
+
# FIX: convert arrays to lists for JSON
|
| 56 |
+
'_xyz': gs._xyz.detach().cpu().numpy().tolist(),
|
| 57 |
+
'_features_dc': gs._features_dc.detach().cpu().numpy().tolist(),
|
| 58 |
+
'_scaling': gs._scaling.detach().cpu().numpy().tolist(),
|
| 59 |
+
'_rotation': gs._rotation.detach().cpu().numpy().tolist(),
|
| 60 |
+
'_opacity': gs._opacity.detach().cpu().numpy().tolist(),
|
| 61 |
},
|
| 62 |
'mesh': {
|
| 63 |
+
'vertices': mesh.vertices.detach().cpu().numpy().tolist(),
|
| 64 |
+
'faces': mesh.faces.detach().cpu().numpy().tolist(),
|
| 65 |
},
|
| 66 |
}
|
|
|
|
| 67 |
return packed_data
|
| 68 |
|
| 69 |
|
| 70 |
def unpack_state(state_dict: dict) -> Tuple[Gaussian, edict]:
|
| 71 |
+
print("[unpack_state] Unpacking state from dictionary... ")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 72 |
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
|
|
|
|
|
|
| 73 |
gauss_data = state_dict['gaussian']
|
| 74 |
mesh_data = state_dict['mesh']
|
|
|
|
|
|
|
| 75 |
gs = Gaussian(
|
| 76 |
+
aabb=gauss_data.get('aabb'),
|
| 77 |
sh_degree=gauss_data.get('sh_degree'),
|
| 78 |
mininum_kernel_size=gauss_data.get('mininum_kernel_size'),
|
| 79 |
scaling_bias=gauss_data.get('scaling_bias'),
|
| 80 |
opacity_bias=gauss_data.get('opacity_bias'),
|
| 81 |
scaling_activation=gauss_data.get('scaling_activation'),
|
| 82 |
)
|
| 83 |
+
gs._xyz = torch.tensor(np.array(gauss_data['_xyz']), device=device, dtype=torch.float32)
|
| 84 |
+
gs._features_dc = torch.tensor(np.array(gauss_data['_features_dc']), device=device, dtype=torch.float32)
|
| 85 |
+
gs._scaling = torch.tensor(np.array(gauss_data['_scaling']), device=device, dtype=torch.float32)
|
| 86 |
+
gs._rotation = torch.tensor(np.array(gauss_data['_rotation']), device=device, dtype=torch.float32)
|
| 87 |
+
gs._opacity = torch.tensor(np.array(gauss_data['_opacity']), device=device, dtype=torch.float32)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
mesh = edict(
|
| 89 |
+
vertices=torch.tensor(np.array(mesh_data['vertices']), device=device, dtype=torch.float32),
|
| 90 |
+
faces=torch.tensor(np.array(mesh_data['faces']), device=device, dtype=torch.int64),
|
| 91 |
)
|
|
|
|
|
|
|
| 92 |
return gs, mesh
|
| 93 |
|
| 94 |
|
| 95 |
def get_seed(randomize_seed: bool, seed: int) -> int:
|
|
|
|
| 96 |
new_seed = np.random.randint(0, MAX_SEED) if randomize_seed else seed
|
| 97 |
+
return int(new_seed)
|
|
|
|
|
|
|
| 98 |
|
| 99 |
@spaces.GPU
|
| 100 |
def text_to_3d(
|
|
|
|
| 105 |
slat_guidance_strength: float,
|
| 106 |
slat_sampling_steps: int,
|
| 107 |
req: gr.Request,
|
| 108 |
+
) -> Tuple[dict, str]:
|
| 109 |
+
outputs = pipeline.run(
|
| 110 |
+
prompt,
|
| 111 |
+
seed=seed,
|
| 112 |
+
formats=["gaussian", "mesh"],
|
| 113 |
+
sparse_structure_sampler_params={"steps": int(ss_sampling_steps), "cfg_strength": float(ss_guidance_strength)},
|
| 114 |
+
slat_sampler_params={"steps": int(slat_sampling_steps), "cfg_strength": float(slat_guidance_strength)},
|
| 115 |
+
)
|
| 116 |
+
state_dict = pack_state(outputs['gaussian'][0], outputs['mesh'][0])
|
| 117 |
+
video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
|
| 118 |
+
video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
|
| 119 |
+
video_combined = [np.concatenate([v.astype(np.uint8), vg.astype(np.uint8)], axis=1) for v, vg in zip(video, video_geo)]
|
| 120 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 121 |
os.makedirs(user_dir, exist_ok=True)
|
| 122 |
+
video_path = os.path.join(user_dir, 'sample.mp4')
|
| 123 |
+
imageio.mimsave(video_path, video_combined, fps=15, quality=8)
|
| 124 |
+
if torch.cuda.is_available(): torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 125 |
return state_dict, video_path
|
| 126 |
|
| 127 |
+
@spaces.GPU(duration=120)
|
|
|
|
| 128 |
def extract_glb(
|
| 129 |
+
state_dict: dict,
|
| 130 |
mesh_simplify: float,
|
| 131 |
texture_size: int,
|
| 132 |
req: gr.Request,
|
| 133 |
) -> Tuple[str, str]:
|
| 134 |
+
gs, mesh = unpack_state(state_dict)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 135 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 136 |
os.makedirs(user_dir, exist_ok=True)
|
| 137 |
+
glb = postprocessing_utils.to_glb(gs, mesh, simplify=float(mesh_simplify), texture_size=int(texture_size), verbose=True)
|
| 138 |
+
glb_path = os.path.join(user_dir, 'sample.glb')
|
| 139 |
+
glb.export(glb_path)
|
| 140 |
+
if torch.cuda.is_available(): torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
return glb_path, glb_path
|
| 142 |
|
|
|
|
| 143 |
@spaces.GPU
|
| 144 |
def extract_gaussian(
|
| 145 |
+
state_dict: dict,
|
| 146 |
req: gr.Request
|
| 147 |
) -> Tuple[str, str]:
|
| 148 |
+
gs, _ = unpack_state(state_dict)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 149 |
user_dir = os.path.join(TMP_DIR, str(req.session_hash))
|
| 150 |
os.makedirs(user_dir, exist_ok=True)
|
| 151 |
+
gaussian_path = os.path.join(user_dir, 'sample.ply')
|
| 152 |
+
gs.save_ply(gaussian_path)
|
| 153 |
+
if torch.cuda.is_available(): torch.cuda.empty_cache()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 154 |
return gaussian_path, gaussian_path
|
| 155 |
|
|
|
|
| 156 |
# --- Gradio UI Definition ---
|
|
|
|
| 157 |
with gr.Blocks(delete_cache=(600, 600), title="TRELLIS Text-to-3D") as demo:
|
| 158 |
gr.Markdown("""
|
| 159 |
# Text to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
|
|
|
|
|
|
|
|
|
|
| 160 |
""")
|
| 161 |
|
| 162 |
+
# State buffer
|
|
|
|
|
|
|
| 163 |
output_buf = gr.State()
|
| 164 |
|
| 165 |
with gr.Row():
|
| 166 |
+
with gr.Column(scale=1):
|
| 167 |
+
text_prompt = gr.Textbox(label="Text Prompt", lines=5)
|
|
|
|
| 168 |
with gr.Accordion(label="Generation Settings", open=False):
|
| 169 |
seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
|
| 170 |
randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
|
| 171 |
+
gr.Markdown("---\n**Stage 1**")
|
| 172 |
+
ss_guidance_strength = gr.Slider(0.0, 15.0, label="Guidance Strength", value=7.5, step=0.1)
|
| 173 |
+
ss_sampling_steps = gr.Slider(10, 50, label="Sampling Steps", value=25, step=1)
|
| 174 |
+
gr.Markdown("---\n**Stage 2**")
|
| 175 |
+
slat_guidance_strength = gr.Slider(0.0, 15.0, label="Guidance Strength", value=7.5, step=0.1)
|
| 176 |
+
slat_sampling_steps = gr.Slider(10, 50, label="Sampling Steps", value=25, step=1)
|
|
|
|
|
|
|
|
|
|
| 177 |
generate_btn = gr.Button("Generate 3D Preview", variant="primary")
|
| 178 |
+
with gr.Accordion(label="GLB Extraction Settings", open=True):
|
| 179 |
+
mesh_simplify = gr.Slider(0.9, 0.99, label="Simplify Factor", value=0.95, step=0.01)
|
| 180 |
+
texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
|
| 181 |
+
extract_glb_btn = gr.Button("Extract GLB", interactive=False)
|
| 182 |
+
extract_gs_btn = gr.Button("Extract Gaussian (PLY)", interactive=False)
|
| 183 |
+
download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
|
| 184 |
+
download_gs = gr.DownloadButton(label="Download Gaussian (PLY)", interactive=False)
|
| 185 |
+
with gr.Column(scale=1):
|
| 186 |
+
video_output = gr.Video(label="3D Preview", autoplay=True, loop=True)
|
| 187 |
+
model_output = gr.Model3D(label="Extracted Model Preview")
|
| 188 |
+
|
| 189 |
+
# --- Event handlers ---
|
| 190 |
+
demo.load(start_session) # FIX: remove inputs/outputs kwargs
|
| 191 |
+
demo.unload(end_session) # FIX: remove inputs/outputs kwargs
|
| 192 |
+
|
| 193 |
+
# Align indentation to one level under Blocks
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 194 |
generate_event = generate_btn.click(
|
| 195 |
get_seed,
|
| 196 |
inputs=[randomize_seed, seed],
|
| 197 |
outputs=[seed],
|
|
|
|
| 198 |
).then(
|
| 199 |
text_to_3d,
|
| 200 |
inputs=[text_prompt, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps],
|
| 201 |
+
outputs=[output_buf, video_output],
|
|
|
|
| 202 |
).then(
|
| 203 |
+
lambda: (extract_glb_btn.update(interactive=True), extract_gs_btn.update(interactive=True)),
|
| 204 |
+
outputs=[extract_glb_btn, extract_gs_btn],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 205 |
)
|
| 206 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 207 |
extract_glb_event = extract_glb_btn.click(
|
| 208 |
extract_glb,
|
| 209 |
+
inputs=[output_buf, mesh_simplify, texture_size],
|
| 210 |
+
outputs=[model_output, download_glb],
|
|
|
|
| 211 |
).then(
|
| 212 |
+
lambda: download_glb.update(interactive=True),
|
| 213 |
outputs=[download_glb],
|
| 214 |
)
|
| 215 |
|
|
|
|
|
|
|
| 216 |
extract_gs_event = extract_gs_btn.click(
|
| 217 |
extract_gaussian,
|
| 218 |
+
inputs=[output_buf],
|
| 219 |
+
outputs=[model_output, download_gs],
|
|
|
|
| 220 |
).then(
|
| 221 |
+
lambda: download_gs.update(interactive=True),
|
| 222 |
outputs=[download_gs],
|
| 223 |
)
|
| 224 |
|
| 225 |
+
# Clear callbacks
|
|
|
|
| 226 |
model_output.clear(
|
| 227 |
+
lambda: (download_glb.update(interactive=False), download_gs.update(interactive=False)),
|
| 228 |
+
outputs=[download_glb, download_gs],
|
| 229 |
)
|
| 230 |
+
video_output.clear(
|
| 231 |
+
lambda: (extract_glb_btn.update(interactive=False), extract_gs_btn.update(interactive=False), download_glb.update(interactive=False), download_gs.update(interactive=False)),
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 232 |
outputs=[extract_glb_btn, extract_gs_btn, download_glb, download_gs],
|
| 233 |
)
|
| 234 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 235 |
if __name__ == "__main__":
|
| 236 |
+
pipeline = TrellisTextTo3DPipeline.from_pretrained(
|
| 237 |
+
"JeffreyXiang/TRELLIS-text-xlarge",
|
| 238 |
+
torch_dtype=torch.float16
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 239 |
)
|
| 240 |
+
if torch.cuda.is_available(): pipeline = pipeline.to("cuda")
|
| 241 |
+
demo.queue().launch(debug=True)
|