Generate an articulated, simulation-ready 3D object from a single-view input.

SIGGRAPH Asia 2026 · Conditionally Accepted

Authors: Qingming Liu^1,2, Xinyue Yao¹, Shuyuan Zhang¹, Yueci Deng^1,2, Guiliang Liu¹, Zhen Liu^1,†, Kui Jia^1,2

¹The Chinese University of Hong Kong, Shenzhen    ²DexForce Technology

^†Corresponding author

Given a single image, PAct generates an articulated 3D object by predicting a part-decomposed structure, synthesizing high-fidelity part geometry and appearance, and estimating articulation parameters for physics-based simulation. This repository currently focuses on inference with the PActPipeline, while the full training stack is being cleaned up for release. This README walks through environment setup, running inference, tuning parameters, and our roadmap for opening the remaining pieces.

Dates reflect our best-effort plan; we will update this table and tag releases in the repo as milestones land.

1. Clone

   git clone https://github.com/PAct-project/PAct.git
   cd PAct

2. Conda environment (recommended, we follow the TRELLIS and OmniPart repo)

   conda env create -f PAct_env.yml
   conda activate PAct
   pip install git+https://github.com/facebookresearch/detectron2.git

python app.py

For convenience, we provide a Hugging Face demo that also allows downloading exported URDFs. The exported URDF files can be interactively viewed in VS-Code with URDF Visualizer.

Call the batch inference script with your config and overrides:

python infer_imgs.py \
  --data_dir assets/real_world_examples \
  --outdir outputs/real_world \
  --batch_size 2 \
  --save_glb --export_arti_objects 

Results (videos, GLBs, Gaussian splats, logs) are written under the --outdir folder in subdirectories named with your sampling configuration and random seed. Generation process of an object typically takes ~15s, comparable to TRELLIS; exporting a mesh is optional, but the subsequent textured-mesh step can be significantly more time-consuming.

infer_imgs.py exposes every previously hard-coded hyperparameter as a CLI flag. Important options are summarized below (see infer_imgs.py for the full list):

Every argument can also be specified inside the JSON config; CLI values take precedence.

Each inference batch produces:

• grid_vids_samples_videos_*: Articulation and exploded-part video mosaics.
• grids_cond_vis_*: Conditioning image grids.
• *_arti_animation*.mp4, *exploded_part*.mp4/png: Per-object renders when mosaics are disabled.
• run_command.txt: Command provenance for reproducibility.
• exported_arti_objects: Optional GLB/Gaussian assets and articulation info if the corresponding flags are enabled.

Use the helper script to convert every object.json in an exported inference run into URDF files ( must set --export_arti_objects in Sec. 4.1):

python scripts/batch_json_to_urdf.py \
	--exported_art_objs_dir outputs/<run_name>/exported_arti_objects

Each generated URDF is placed next to its source metadata as <object_name>_fromJson2urdf.urdf.

TBD

TBD

If you build upon this work, please cite the PAct paper:

@article{liu2026pact,
    title   = {PAct: Part-Decomposed Single-View Articulated Object Generation},
    author  = {Liu, Qingming and Yao, Xinyue and Zhang, Shuyuan and Deng, Yueci and Liu, Guiliang and Liu, Zhen and Jia, Kui},
    journal = {arXiv preprint arXiv:2602.14965},
    year    = {2026}
}

And we sincerely thank the authors of TRELLIS and OmniPart, whose codes were used in our work.

@article{xiang2024structured,
    title   = {Structured 3D Latents for Scalable and Versatile 3D Generation},
    author  = {Xiang, Jianfeng and Lv, Zelong and Xu, Sicheng and Deng, Yu and Wang, Ruicheng and Zhang, Bowen and Chen, Dong and Tong, Xin and Yang, Jiaolong},
    journal = {arXiv preprint arXiv:2412.01506},
    year    = {2024}
}.

@article{yang2025omnipart,
        title={Omnipart: Part-aware 3d generation with semantic decoupling and structural cohesion},
        author={Yang, Yunhan and Zhou, Yufan and Guo, Yuan-Chen and Zou, Zi-Xin and Huang, Yukun and Liu, Ying-Tian and Xu, Hao and Liang, Ding and Cao, Yan-Pei and Liu, Xihui},
        journal={arXiv preprint arXiv:2507.06165},
        year={2025}
}