[ECCV 2026] PhyGDPO: Physics-Aware Groupwise Direct Preference Optimization for Physically Consistent Text-to-Video Generation

Dataset Description

This dataset supports the DPO training algorithm, PhyGDPO, described in our ECCV paper. It is re-implemented based on a public dataset, VidGen-1M. This dataset contains 135k text-video pairs that are rich in physics interactions and phenomena in total. Please also download VidGen-1M if you are going to use the whole dataset PhyVidGen-135K with the explicit physics extended prompts written by Gemini-2.5-pro in the json_file.zip. Besides, we also construct 17k video groups that support our reinforcement learning algorithm, PhyGDPO, based on our data construction pipeline, PhyAugPipe, as

Each group contains three losing video cases generated by Wan-2.1-1.3B or Wan-2.1-14B. Here is a data example:

Github Code Link

This dataset is intended to be used together with our code. Please refer to the GitHub repository below for more detailed instructions.

https://github.com/caiyuanhao1998/Open-PhyGDPO

Huggingface Model Link

We also release three models based on Wan2.1-1.3B and Wan2.1-14B in the following link:

https://huggingface.co/CaiYuanhao/PhyGDPO

Project Page Link

For more video customization results, please refer to our project page:

https://caiyuanhao1998.github.io/project/PhyGDPO/

Arxiv Paper Link

For more technical details, please refer to our ECCV 2026 paper:

https://arxiv.org/abs/2512.24551

知乎中文解读

https://zhuanlan.zhihu.com/p/2060200098131350370

Citation

If you find our code, data, and models useful, please consider citing our paper:

@inproceedings{phygdpo,
  title={PhyGDPO: Physics-Aware Groupwise Direct Preference Optimization for Physically Consistent Text-to-Video Generation},
  author={Cai, Yuanhao and Li, Kunpeng and Jia, Menglin and Wang, Jialiang and Sun, Junzhe and Liang, Feng and Chen, Weifeng and Juefei-Xu, Felix and Wang, Chu and Thabet, Ali and Dai, Xiaoliang and Ju, Xuan and Yuille, Alan and Hou, Ji},
  booktitle={ECCV},
  year={2026}
}