$\color{orange}{\textbf{{[CVPR 2026]}}}$ Part²GS: Part-aware Modeling of Articulated Objects using 3D Gaussian Splatting

Tianjiao Yu · Vedant Shah · Muntasir Wahed · Ying Shen · Kiet A. Nguyen · Ismini Lourentzou

University of Illinois Urbana-Champaign

🔥 Highlights

Part²GS reconstructs articulated 3D digital twins from multi-view images, recovering high-fidelity geometry and physically consistent motion — with no part labels, kinematic templates, or predefined part libraries.

• 🧩 Emergent part discovery. Every Gaussian carries a learnable part-identity embedding, so part structure emerges directly from geometry, motion, and physics.
• 🧲 Collision-free articulation. A field of repel points applies localized repulsive forces at part boundaries, stabilizing joint trajectories and keeping parts cleanly separated.
• ⚙️ Physics-grounded motion. Contact, velocity-consistency, and vector-field-alignment losses constrain each part to rigid-body-consistent, physically plausible motion.
• 🏆 State-of-the-art accuracy. Up to 10× lower Chamfer Distance on movable parts across PARIS, ArtGS-Multi, and DTA-Multi.

🌟 Installation

We follow the ArtGS setup. All Python dependencies are listed in requirements.txt.

conda create -n Part2GS python=3.10
conda activate Part2GS
conda install pytorch==2.4.1 torchvision==0.19.1 torchaudio==2.4.1 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

Note: PyTorch3D and the PointNet++ CUDA ops are also required — see ArtGS for installation details.

🌟 Dataset

Download data from PARIS and ARTGS-MULTI

🌟 Training Pipeline

The pipeline runs in three stages — canonical Gaussian initialization, joint-type prediction, and full-model training. Each stage can be launched on its own, or run end-to-end with full_pipe.sh.

# Step 1: canonical Gaussian initialization
bash scripts/coarse.sh
# Step 2: joint-type prediction
bash scripts/predict.sh
# Step 3: full-model training
bash scripts/train.sh

Run the complete pipeline:

bash scripts/full_pipe.sh

🌟 Citation

If you find Part²GS useful for your research, please consider citing:

@inproceedings{yu2026part2gs,
  title     = {{Part\textsuperscript{2}GS: Part-aware Modeling of Articulated Objects using 3D Gaussian Splatting}},
  author    = {Yu, Tianjiao and Shah, Vedant and Wahed, Muntasir and Shen, Ying and Nguyen, Kiet A. and Lourentzou, Ismini},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {18913--18923},
  year      = {2026}
}

🌟 Acknowledgement

This codebase builds on ArtGS, PARIS, Reart, and Lab4D. We thank the authors for open-sourcing their awesome projects.

Check out our related work on part-level 3D generation, DreamPartGen (ECCV'26), also led by Tianjiao (Joey) Yu, and ArtGS's follow-up VideoArtGS if you are interested in articulated object generation.