DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos

Project page · arXiv · Hugging Face (coming)

Release plan

• inference code
• online interaction code
• checkpoints
• training code
• custom data and preprocessing code
• full configurations

Installation

# 1. Python env (Python 3.10 + PyTorch 2.4.0 + CUDA 12.1 / 12.4 tested)
conda create -n deformmaster python=3.10 -y
conda activate deformmaster

# 2. Install Python deps + CUDA rasterizer submodules
bash install.sh

Inference

# Example: planar (cloth)
python inference.py --case_name my_mono_cloth --config configs/planar.yaml

Online interaction

python playground.py \
    --case_name my_mono_cloth \
    --output ./checkpoints \
    --base_path ./data \
    --gaussian_path ./checkpoints/gaussian \
    --server_port 7860

Citation

@article{li2026deformmaster,
      title={DeformMaster: An Interactive Physics-Neural World Model for Deformable Objects from Videos},
      author={Can Li and Zhoujian Li and Ren Li and Jie Gu and Lei Lei and Jingmin Chen and Lei Sun},
      year={2026},
      eprint={2605.09586},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.09586},
}

Acknowledgements

We thank the authors of PhysTwin, PGND, and 3D Gaussian Splatting.