• Sep 14, 2026: 🔥🔥 ComfyUI integration is now available 🎨. Many thanks to Terry Jia (@jtydhr88) for the wonderful implementation!
• Sep 2, 2026: 🔥🔥 Released inference codes 📟, checkpoints 🤗, arXiv 📚, and project page 🏠.

• Single-object prior for complex compositional scenes. We demonstrate that a single-object generative prior can be leveraged for compositional meshification of very complex scenes with hundreds of objects.
• Adapting Single-Object Prior. We finetune Pixal3D to consume occluded multiple-view input.
• Benchmark. We release the UE-MeshyScene dataset with per-frame, per-instance annotations.
• Application to Marble. The same pipeline turns a Marble 3DGS world into object-level compositional meshes.

Please first follow the installation guide of TRELLIS.2 to set up the base environment.

NATTEN_CUDA_ARCH is the compute capability of your GPU. Print yours with

python -c "import torch; print('%d.%d' % torch.cuda.get_device_capability())"

# H100; replace 9.0 with your own compute capability
NATTEN_CUDA_ARCH="9.0" NATTEN_N_WORKERS=8 pip install natten==0.21.0 --no-build-isolation
pip install https://github.com/LDYang694/Storages/releases/download/20260430/utils3d-0.0.2-py3-none-any.whl
pip install peft pillow imageio imageio-ffmpeg tqdm easydict opencv-python-headless trimesh transformers==4.57.1 zstandard kornia timm diffusers accelerate gradio plyfile matplotlib scikit-image scikit-learn fpsample iopath pycocotools ftfy
pip install "setuptools<81"

hf download AlayaLab/WorldSculpt --local-dir ./pretrained
hf download TencentARC/Pixal3D --local-dir ./pretrained/Pixal3D

# UE-MeshyScene
hf download AlayaLab/WorldSculpt_data --include "UE-MeshyScene/*" --local-dir ./input --repo-type=dataset
cd ./input/UE-MeshyScene
tar xvf scene*.tar
cd ../..

# Marble DEMO data
hf download AlayaLab/WorldSculpt_data --include "Marble/*" --local-dir ./input --repo-type=dataset
cd ./input/Marble
tar xvf Marble.tar.gz
cd ../..

Every option is passed explicitly — the script has no defaults, so a run is fully described by its own command line. Step 4 extracts each dataset into its own subdirectory, so INPUT_ROOT points at that subdirectory, not at ./input (the script reads $INPUT_ROOT/<scene>/transforms.json).

# UE-MeshyScene, scene_00001
INPUT_ROOT=./input/UE-MeshyScene OUTPUT_ROOT=./output CKPT_ROOT=./pretrained \
SAMPLER=official SS_STEP=15000 SHAPE_STEP=15000 \
RENDER=1 FACE_BUDGET=1000000 GPU=0 \
./inference.sh scene_00001

# Marble demo data, marble_serene_living_room_countryside_view
INPUT_ROOT=./input/Marble OUTPUT_ROOT=./output CKPT_ROOT=./pretrained \
SAMPLER=official SS_STEP=15000 SHAPE_STEP=15000 \
RENDER=1 FACE_BUDGET=1000000 GPU=0 \
./inference.sh marble_serene_living_room_countryside_view

A community ComfyUI implementation of WorldSculpt is available at ComfyUI-WorldSculpt, developed by Terry Jia (@jtydhr88).

It provides node-based workflows for running WorldSculpt within ComfyUI. Please refer to the repository for installation instructions, example workflows, and the latest usage details.

We thank Pixal3D, TRELLIS.2, and DINOv3 for their wonderful work and open-source repositories.

We also thank Terry Jia for developing and open-sourcing the community ComfyUI integration for WorldSculpt.

If you find this work useful, please cite:

@misc{niu2026worldsculptgeneratingcompositionalworlds,
      title={WorldSculpt: Generating Compositional Worlds from Grounded Videos}, 
      author={Muyao Niu and Jixuan He and Ruihan Yu and Lian Fu and Yonghao Yu and Zheng-Hui Huang and Yifan Zhan and Fengbo Lan and Yongtao Ge and Yinqiang Zheng and Kaipeng Zhang and Zhixiang Wang},
      year={2026},
      eprint={2609.05416},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.05416}, 
}