PhysRVG: Physics-Aware Unified Reinforcement Learning for Video Generative Models

This repository hosts the model weights for PhysRVG (ECCV 2026). PhysRVG leverages a unified reinforcement learning framework with verifiable rewards to improve rigid-body motion generation in video synthesis.

📌 Demos, training, and inference code are in the GitHub repository. This page only provides the checkpoints.

Contents

PhysRVG/
├── dit                       # PhysRVG DiT weights (used with --resume_from_checkpoint)
├── lora                      # LoRA weights for memory-efficient fine-tuning / inference
├── sam2.1-hiera-large        # SAM 2 model used to compute the verifiable reward
└── Wan2.2-TI2V-5B-Diffusers  # base text/image-to-video diffusion model

Usage

Download the weights into the ./models directory of the code repository:

huggingface-cli download HappyP4nda/PhysRVG --local-dir ./models

Then run inference (see the GitHub README for setup):

python inference.py --video_path data/example_videos/2/video.mp4

Citation

@article{PhysRVG2026,
  title={PhysRVG: Physics-Aware Unified Reinforcement Learning for Video Generative Models},
  author={Zhang, Qiyuan and Gong, Biao and Tan, Shuai and Zhang, Zheng and Shen, Yujun and Zhu, Xing and Li, Yuyuan and Yao, Kelu and Shen, Chunhua and Zou, Changqing},
  journal={ECCV 2026},
  year={2026}
}