Real-Time and Generalizable Physics-Based Animation
via Self-Supervised Neural Skinning

Yuanhang Lei, Baofei Liu, Xingxuan Li, Tao Cheng,
Zixuan Wang, Boming Zhao, Yujun Shen, Siyuan Huang,
Ruizhen Hu, Peter Yichen Chen, Hujun Bao, Zhaopeng Cui^†
_{PhysSkin++ authors · † Corresponding author}

PhysSkin++ · arXiv 2026   |   PhysSkin · CVPR 2026 Highlight

News   /   Papers   /   Overview   /   Method   /   Installation   /   Visualization   /   Citation

📰 News

• [2026-10-30] PhysSkin is now featured in PhysStudio, our open-source world simulator for physics-based animation. 🌏
• [2026-10-19] PhysSkin++ is now on arXiv, extending PhysSkin with interactive control, multi-object interaction, and rig-complementary animation.
• [2026-08-24] The PhysSkin code is now available!
• [2026-04-09] PhysSkin has been selected as a CVPR 2026 Highlight! 🎉

From static 3D geometry to interactive, physics-based motion.

▶ Watch the 90-second preview   ·   Explore all demos

📚 Papers

Work	Publication	Links
PhysSkin++	arXiv 2026	arXiv · Project · Hugging Face
PhysSkin	CVPR 2026 · Highlight	arXiv · Project

PhysSkin (CVPR 2026) authors

Yuanhang Lei, Tao Cheng, Xingxuan Li, Boming Zhao, Siyuan Huang, Ruizhen Hu, Peter Yichen Chen, Hujun Bao, and Zhaopeng Cui.

✨ Overview

PhysSkin++ learns continuous neural skinning fields from static geometry to build compact deformation subspaces. Physics-informed self-supervision enables real-time reduced simulation across diverse shapes without per-object network training or simulation trajectories for supervision.

The framework supports animation across meshes, point clouds, NeRFs, and 3D Gaussians.

🔬 Method

The framework combines a transformer-based shape encoder, a cross-attention field decoder, and physics-informed self-supervised objectives.

🛠️ Installation

The code was tested with Python 3.11, PyTorch 2.7.1, NumPy 1.26.4, PyTorch Lightning 1.6.4, and conflictfree 0.1.8.

python -m pip install "pip<24.1"
pip install torch==2.7.1 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

🗂️ Data

<dataset_root>/
└── 00000001/
    └── <model_id>/
        └── models/
            └── samples/
                ├── latents.npz
                └── internal_filled.npz

• latents.npz contains latents, a float32 array of shape (1, 256, 768).
• internal_filled.npz contains points, a float32 array of shape (N, 3) in approximately [-0.5, 0.5].

The loader rescales the points to approximately [-1, 1]. Data splits are provided in data/splits, and processed examples for visualization are included in data/examples.

🎬 Visualization

physskin_vis.ipynb

The notebook uses data/examples by default. A different sample or dataset can be selected with MODEL_ID and DATASET_ROOT.

🏋️ Training

DATASET_ROOT=/path/to/dataset \
GPU_IDS="0 1" \
bash scripts/train_physskin.sh

🧹 Data preprocessing

pip install -r data_process/requirements.txt

DATASET_ROOT=/path/to/dataset \
SAMPLING_PYTHON=/path/to/python \
RAYTRACING_PYTHON=/path/to/python \
MICHELANGELO_ROOT=/path/containing/Michelangelo \
ENCODER_CHECKPOINT=/path/to/shape_encoder.pth \
bash data_process/process_data.sh 00000001 0

📦 Checkpoint

The pretrained checkpoint is stored at checkpoints/physskin_rignet_epoch150.pt.

🙏 Acknowledgements

This project builds upon Simplicits, Michelangelo, and ConFIG. See THIRD_PARTY_NOTICES.md for details.

📝 Citation

📄 PhysSkin++

@misc{lei2026physskinplus,
    title         = {PhysSkin++: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning},
    author        = {Yuanhang Lei and Baofei Liu and Xingxuan Li and Tao Cheng and Zixuan Wang and Boming Zhao and Yujun Shen and Siyuan Huang and Ruizhen Hu and Peter Yichen Chen and Hujun Bao and Zhaopeng Cui},
    year          = {2026},
    eprint        = {2610.26067},
    archivePrefix = {arXiv},
    primaryClass  = {cs.CV},
    url           = {https://arxiv.org/abs/2610.26067}
}

📄 PhysSkin (CVPR 2026)

@InProceedings{Lei_2026_CVPR,
    author    = {Lei, Yuanhang and Cheng, Tao and Li, Xingxuan and Zhao, Boming and Huang, Siyuan and Hu, Ruizhen and Chen, Peter Yichen and Bao, Hujun and Cui, Zhaopeng},
    title     = {PhysSkin: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {32357-32366}
}