PAct: Part-Decomposed Single-View Articulated Object Generation

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

SIGGRAPH Asia 2026 · Conditionally Accepted

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

PAct turns a single input image into a part-decomposed, articulated 3D object ready for physics-based simulation.

Abstract

Given a single-view input, PAct generates complete part-decomposed geometry and infers the kinematic structure, producing physically plausible articulated 3D objects. It predicts a part-decomposed structure, synthesizes high-fidelity part geometry and appearance, and estimates articulation parameters for physics-based simulation.

By reasoning about objects at the level of movable parts rather than as a single fused shape, PAct produces assets whose joints, hinges, and sliding components can be actuated directly — enabling downstream use in simulation, robotics, and interactive 3D content. The pipeline outputs GLB meshes, Gaussian splats, and articulation metadata, and can export URDF descriptions for physics engines.

What PAct Does

Part decomposition. Predicts a part-decomposed structure from a single view instead of a single fused mesh.
High-fidelity parts. Synthesizes detailed per-part geometry and appearance.
Articulation estimation. Estimates joint types and articulation parameters for each movable part.
Simulation-ready export. Exports GLB meshes, Gaussian splats, articulation metadata, and URDF for physics engines.

BibTeX

@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}
}