PhysiGen: Integrating Collision-Aware Physical Constraints for High-Fidelity Human-Human Interaction Generation

Accepted to ICASSP 2026

Nan Lei, Yuan-Ming Li, Ling-An Zeng, Liang Xu, Zhi-Wei Xia, Hui-Wen Huang, Fa-Ting Hong†, Wei-Shi Zheng

^{Sun Yat-sen University  |  The Hong Kong University of Science and Technology  |  Shanghai Jiao Tong University  |  Guilin University of Electronic Technology}

📢 News

• [2025.12.11] 🎉 Check out our new work IRG-MotionLLM with     ——the first model supporting natively text-motion interleaved reasoning for text-to-motion generation, with advanced performance and emerging cross-task, cross-model synergies!
  • 📄 Paper  |  💻 Code
• [2026.01.18] 🎉 PhysiGen is accepted to ICASSP 2026!
• 🚧 Code and models coming soon...

---

📝 TODO

• ✅ Release core computation code
• ✅ Release core evaluation code
• ⬜ Release full training and inference pipeline
• ⬜ Release bounding box parameter files
• ⬜ Release pretrained model checkpoints

---

🔍 Overview

PhysiGen is a computationally efficient optimization strategy that explicitly integrates collision-aware physical constraints into human-human interaction generation.

Generating realistic two-person interaction sequences remains challenging due to pervasive body interpenetration — a problem that spans from data acquisition to generated results. Existing approaches either ignore this issue or rely on computationally expensive mesh-level SDF losses (e.g., inflating training time from 3 days → 14 days).

PhysiGen addresses this by:

• 🔷 Simplifying high-resolution human body meshes into geometric primitives (cylinders/cuboids) for efficient collision detection
• 🔷 Computing physics-inspired guidance directions via antipodal point construction to resolve penetration
• 🔷 Integrating seamlessly into existing models as a plug-and-play module — no architectural changes required

Key results on InterHuman:

Model	Collision Distance ↓	Collision Rate ↓	Top-1 R-Precision ↑
InterGen	3.905	0.2270	0.371
InterGen + PhysiGen(from scratch)	1.836	0.1878	0.485
in2IN	3.142	0.1863	0.455
in2IN + PhysiGen (adaption)	2.005	0.1503	0.481

---

🛠️ Installation

Clone the repository and set up the environment:

git clone https://github.com/iSEE-Laboratory/PhysiGen.git
cd PhysiGen

We provide a Conda environment configuration file.

conda env create -f environment.yml
conda activate physigen

Note: the package sdf-pytorch==0.0.1 is installed from a local path: /code/collision/sdf0/sdf:

cd /code/collision/sdf0/sdf
python3 setup.py install

---

Datasets

We evaluate on two datasets:

• InterHuman — 7,779 two-person interaction sequences with text annotations. Download from InterGen.
• Inter-X — 11,388 interaction sequences using SMPL-X. Download from Inter-X.

---

▶️ Usage

🏋️ Training & Evaluation

We provide the core implementation for collision computation and collision testing in the collision directory.
The corresponding function calls can be found in main.py and eval_interhuman_coll.py.

---

📊 Results

Qualitative Comparison

Other Model	Other Model	Other Model	Other Model	Other Model	Other Model	Other Model
Ours	Ours	Ours	Ours	Ours	Ours	Ours

PhysiGen significantly reduces interpenetration while maintaining semantic consistency with the text prompt. Red dashed boxes highlight severe collision artifacts in baseline methods.

Computational Cost

Method	Memory (MB)	Time per batch (s)
Baseline	15,107	—
PhysiGen (50×19 pts)	20,219	0.053
SDF Loss (128 pts)	24,152	0.352
SDF Loss (6890 pts)	24,156	3.734
pts means the num of the points.

---

📄 Citation

If you find this work useful, please consider citing:

@article{Lei2026PhysigenIC,
  title={Physigen: Integrating Collision-Aware Physical Constraints for High-Fidelity Human-Human Interaction Generation},
  author={Nan Lei and Yuan-Ming Li and Ling-an Zeng and Liangliang Xu and Zhi-Wei Xia and Huihui Huang and Fa-Ting Hong and Wei-Shi Zheng},
  journal={ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2026},
  url={https://api.semanticscholar.org/CorpusID:287685790}
}

---

📜 License

This project is released under the Apache License 2.0.

---

🙏 Acknowledgement

We thank the following open-source projects for their contributions:

• InterGen — two-person interaction generation framework and InterHuman dataset
• Inter-X — large-scale human interaction dataset
• in2IN — individual-aware interaction generation
• TIMotion — temporal and interactive motion generation framework
• multiperson — SDF-based loss implementation

---

^{If you have any questions, please open an issue or contact us at lein7@mail2.sysu.edu.cn}