[CVPR 2026] ArtLLM: Generating Articulated Assets via 3D LLM

Official implementation for ArtLLM: Generating Articulated Assets via 3D LLM.

Penghao Wang, Siyuan Xie, Hongyu Yan, Xianghui Yang, Jingwei Huang†, Chunchao Guo†, Jiayuan Gu†

CVPR 2026

| Project Page | Paper | Code | Model |

Installation

Create a new environment:

conda create -n artllm python=3.11
conda activate artllm

First install CUDA, our code is evaluated on CUDA 12.4. Then install the following dependencies:

# PyTorch for CUDA 12.4
pip install torch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 --index-url https://download.pytorch.org/whl/cu124

# Sparse Conv
pip install spconv-cu124
# torch scatter
pip install torch-scatter -f https://data.pyg.org/whl/torch-2.4.1+cu124.html
# torch cluster
pip install torch-cluster -f https://data.pyg.org/whl/torch-2.4.1+cu124.html

# FlashAttention
pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu12torch2.4cxx11abiFALSE-cp311-cp311-linux_x86_64.whl

Install ArtLLM:

pip install -e .

If you plan to run the XPart geometry generation module, also install its dependencies as needed:

pip install -r XPart/requirements.txt

Inference

We have provided 2 example input meshes under test/input.

Layout and Articulation Prediction

python inference.py

Before inference, update the configuration variables near the bottom of the script:

• model_path: path to the trained ArtLLM checkpoint;
• glb_folder: folder containing input .glb meshes;
• output_path: folder for generated results;
• code_template_path: usually preprocess/code_template_pcd2artbbox_limit.txt;
• art_axis_dir_codebook_path: path to the .npz direction codebook;
• test_names: optional list of object names to evaluate, or None to process all meshes.

Part Geometry Generation

cd XPart
python generate_artllm.py --test_folder ../test

URDF convert

python scripts/convert_to_urdf.py

Training

Current repo contains the corresponding training code, detailed instructions will be released soon.

Data Preprocess

We have provided preprocess scripts under preprocess for training, detailed instructions will be released soon.

Acknowledgement

Our code is based on SpatialLM, and uses XPart for high-fidelity part geometry generation.

We gratefully acknowledge the invaluable discussion and feedback provided by Chunshi Wang, Junliang Ye, Yunhan Yang from the Tencent Hunyuan3D Team, Xinyu Lian from Shanghai AI Lab, and Kaixin Yao, Zhehao Shen from ShanghaiTech University.

Citation

If you find our work useful in your research, please consider citing our paper:

@inproceedings{wang2026artllm,
  title={ArtLLM: Generating Articulated Assets via 3D LLM},
  author={Wang, Penghao and Xie, Siyuan and Yan, Hongyu and Yang, Xianghui and Huang, Jingwei and Guo, Chunchao and Gu, Jiayuan},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={34281--34291},
  year={2026}
}