import spaces
import gradio as gr
import os
import sys
from glob import glob
import time
from typing import Any, Union
import numpy as np
import torch
import uuid
import shutil
print(f'torch version:{torch.__version__}')
import trimesh
import glob
from huggingface_hub import snapshot_download
from PIL import Image
from accelerate.utils import set_seed
import subprocess
import importlib, site, sys
# Re-discover all .pth/.egg-link files
for sitedir in site.getsitepackages():
site.addsitedir(sitedir)
# Clear caches so importlib will pick up new modules
importlib.invalidate_caches()
def sh(cmd): subprocess.check_call(cmd, shell=True)
def install_cuda_toolkit():
CUDA_TOOLKIT_URL = "https://developer.download.nvidia.com/compute/cuda/12.6.0/local_installers/cuda_12.6.0_560.28.03_linux.run"
CUDA_TOOLKIT_FILE = "/tmp/%s" % os.path.basename(CUDA_TOOLKIT_URL)
subprocess.check_call(["wget", "-q", CUDA_TOOLKIT_URL, "-O", CUDA_TOOLKIT_FILE])
subprocess.check_call(["chmod", "+x", CUDA_TOOLKIT_FILE])
subprocess.check_call([CUDA_TOOLKIT_FILE, "--silent", "--toolkit"])
os.environ["CUDA_HOME"] = "/usr/local/cuda"
os.environ["PATH"] = "%s/bin:%s" % (os.environ["CUDA_HOME"], os.environ["PATH"])
os.environ["LD_LIBRARY_PATH"] = "%s/lib:%s" % (
os.environ["CUDA_HOME"],
"" if "LD_LIBRARY_PATH" not in os.environ else os.environ["LD_LIBRARY_PATH"],
)
# add for compiler header lookup
os.environ["CPATH"] = f"{os.environ['CUDA_HOME']}/include" + (
f":{os.environ['CPATH']}" if "CPATH" in os.environ else ""
)
# Fix: arch_list[-1] += '+PTX'; IndexError: list index out of range
os.environ["TORCH_CUDA_ARCH_LIST"] = "8.9;9.0"
print("==> finished installation")
print("installing cuda toolkit")
install_cuda_toolkit()
print("finished")
os.environ["PARTCRAFTER_PROCESSED"] = f"{os.getcwd()}/proprocess_results"
def sh(cmd_list, extra_env=None):
env = os.environ.copy()
if extra_env:
env.update(extra_env)
subprocess.check_call(cmd_list, env=env)
# install with FORCE_CUDA=1
sh(["pip", "install", "diso"], {"FORCE_CUDA": "1"})
# sh(["pip", "install", "torch-cluster", "-f", "https://data.pyg.org/whl/torch-2.7.0+126.html"])
# tell Python to re-scan site-packages now that the egg-link exists
import importlib, site; site.addsitedir(site.getsitepackages()[0]); importlib.invalidate_caches()
from src.utils.data_utils import get_colored_mesh_composition, scene_to_parts, load_surfaces
from src.utils.render_utils import render_views_around_mesh, render_normal_views_around_mesh, make_grid_for_images_or_videos, export_renderings, explode_mesh
from src.pipelines.pipeline_partcrafter import PartCrafterPipeline
from src.utils.image_utils import prepare_image
from src.models.briarmbg import BriaRMBG
# Constants
MAX_NUM_PARTS = 16
DEVICE = "cuda"
DTYPE = torch.float16
# Download and initialize models
partcrafter_weights_dir = "pretrained_weights/PartCrafter"
rmbg_weights_dir = "pretrained_weights/RMBG-1.4"
snapshot_download(repo_id="wgsxm/PartCrafter", local_dir=partcrafter_weights_dir)
snapshot_download(repo_id="briaai/RMBG-1.4", local_dir=rmbg_weights_dir)
rmbg_net = BriaRMBG.from_pretrained(rmbg_weights_dir).to(DEVICE)
rmbg_net.eval()
pipe: PartCrafterPipeline = PartCrafterPipeline.from_pretrained(partcrafter_weights_dir).to(DEVICE, DTYPE)
def first_file_from_dir(directory, ext):
files = glob.glob(os.path.join(directory, f"*.{ext}"))
return sorted(files)[0] if files else None
def get_duration(
image_path,
num_parts,
seed,
num_tokens,
num_inference_steps,
guidance_scale,
use_flash_decoder,
rmbg,
session_id,
progress,
):
duration_seconds = 75
if num_parts > 10:
duration_seconds = 120
elif num_parts > 5:
duration_seconds = 90
return int(duration_seconds)
@spaces.GPU(duration=140)
def gen_model_n_video(image_path: str,
num_parts: int,
progress=gr.Progress(track_tqdm=True),):
model_path = run_partcrafter(image_path, num_parts=num_parts, progress=progress)
video_path = gen_video(model_path)
return model_path, video_path
@spaces.GPU()
def gen_video(model_path):
if model_path is None:
gr.Info("You must craft the 3d parts first")
return None
export_dir = os.path.dirname(model_path)
merged = trimesh.load(model_path)
preview_path = os.path.join(export_dir, "rendering.gif")
num_views = 36
radius = 4
fps = 7
rendered_images = render_views_around_mesh(
merged,
num_views=num_views,
radius=radius,
)
export_renderings(
rendered_images,
preview_path,
fps=fps,
)
return preview_path
@spaces.GPU(duration=get_duration)
@torch.no_grad()
def run_partcrafter(image_path: str,
num_parts: int = 1,
seed: int = 0,
num_tokens: int = 1024,
num_inference_steps: int = 50,
guidance_scale: float = 7.0,
use_flash_decoder: bool = False,
rmbg: bool = True,
session_id = None,
progress=gr.Progress(track_tqdm=True),):
"""
Generate structured 3D meshes from a 2D image using the PartCrafter pipeline.
This function takes a single 2D image as input and produces a set of part-based 3D meshes,
using compositional latent diffusion with attention to structure and part separation.
Optionally removes the background using a pretrained background removal model (RMBG),
and outputs a merged object mesh.
Args:
image_path (str): Path to the input image file on disk.
num_parts (int, optional): Number of distinct parts to decompose the object into. Defaults to 1.
seed (int, optional): Random seed for reproducibility. Defaults to 0.
num_tokens (int, optional): Number of tokens used during latent encoding. Higher values yield finer detail. Defaults to 1024.
num_inference_steps (int, optional): Number of diffusion inference steps. More steps improve quality but increase runtime. Defaults to 50.
guidance_scale (float, optional): Classifier-free guidance scale. Higher values emphasize adherence to conditioning. Defaults to 7.0.
use_flash_decoder (bool, optional): Whether to use FlashAttention in the decoder for performance. Defaults to False.
rmbg (bool, optional): Whether to apply background removal before processing. Defaults to True.
session_id (str, optional): Optional session ID to manage export paths. If not provided, a random UUID is generated.
progress (gr.Progress, optional): Gradio progress object for visual feedback. Automatically handled by Gradio.
Returns:
Tuple[str, str, str, str]:
- `merged_path` (str): File path to the merged full object mesh (`object.glb`).
Notes:
- This function utilizes HuggingFace pretrained weights for both part generation and background removal.
- The final output includes merged model parts to visualize object structure.
- Generation time depends on the number of parts and inference parameters.
"""
max_num_expanded_coords = 1e9
if session_id is None:
session_id = uuid.uuid4().hex
if rmbg:
img_pil = prepare_image(image_path, bg_color=np.array([1.0, 1.0, 1.0]), rmbg_net=rmbg_net)
else:
img_pil = Image.open(image_path)
set_seed(seed)
start_time = time.time()
outputs = pipe(
image=[img_pil] * num_parts,
attention_kwargs={"num_parts": num_parts},
num_tokens=num_tokens,
generator=torch.Generator(device=pipe.device).manual_seed(seed),
num_inference_steps=num_inference_steps,
guidance_scale=guidance_scale,
max_num_expanded_coords=max_num_expanded_coords,
use_flash_decoder=use_flash_decoder,
).meshes
duration = time.time() - start_time
print(f"Generation time: {duration:.2f}s")
# Ensure no None outputs
for i, mesh in enumerate(outputs):
if mesh is None:
outputs[i] = trimesh.Trimesh(vertices=[[0,0,0]], faces=[[0,0,0]])
export_dir = os.path.join(os.environ["PARTCRAFTER_PROCESSED"], session_id)
# If it already exists, delete it (and all its contents)
if os.path.exists(export_dir):
shutil.rmtree(export_dir)
os.makedirs(export_dir, exist_ok=True)
parts = []
for idx, mesh in enumerate(outputs):
part = os.path.join(export_dir, f"part_{idx:02}.glb")
mesh.export(part)
parts.append(part)
# Merge and color
merged = get_colored_mesh_composition(outputs)
split_mesh = explode_mesh(merged)
merged_path = os.path.join(export_dir, "object.glb")
merged.export(merged_path)
return merged_path
def cleanup(request: gr.Request):
sid = request.session_hash
if sid:
d1 = os.path.join(os.environ["PARTCRAFTER_PROCESSED"], sid)
shutil.rmtree(d1, ignore_errors=True)
def start_session(request: gr.Request):
return request.session_hash
def build_demo():
css = """
#col-container {
margin: 0 auto;
max-width: 1560px;
}
"""
theme = gr.themes.Ocean()
with gr.Blocks(css=css, theme=theme) as demo:
session_state = gr.State()
demo.load(start_session, outputs=[session_state])
with gr.Column(elem_id="col-container"):
gr.HTML(
"""
PartCrafter – Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers
HF Space by :
The 3D Preview might take a few seconds to load the 3D model
"""
)
with gr.Row():
output_model = gr.Model3D(label="Merged 3D Object", height=512, interactive=False)
video_output = gr.Image(label="Split Preview", height=512)
with gr.Row():
with gr.Column():
examples = gr.Examples(
examples=[
[
"assets/images/np5_b81f29e567ea4db48014f89c9079e403.png",
5,
],
[
"assets/images/np7_1c004909dedb4ebe8db69b4d7b077434.png",
7,
],
[
"assets/images/np16_dino.png",
16,
],
[
"assets/images/np13_39c0fa16ed324b54a605dcdbcd80797c.png",
13,
],
],
inputs=[input_image, num_parts],
outputs=[output_model, video_output],
fn=gen_model_n_video,
cache_examples=True
)
run_button.click(fn=run_partcrafter,
inputs=[input_image, num_parts, seed, num_tokens, num_steps,
guidance, flash_decoder, remove_bg, session_state],
outputs=[output_model])
video_button.click(fn=gen_video,
inputs=[output_model],
outputs=[video_output])
return demo
if __name__ == "__main__":
demo = build_demo()
demo.unload(cleanup)
demo.queue()
demo.launch(mcp_server=True, ssr_mode=False)