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

(1) A gymnast drops from the parallel bars and lands safely on the mat below.
Ours	OpenAI Sora2	Google Veo3.1
(2) A soccer player runs, plants their foot, and drop kicks a soccer ball high into the air, the ball arcing visibly.
Ours	OpenAI Sora2	Google Veo3.1
(3) A person wearing a helmet performs a handspring over a platform.
Ours	OpenAI Sora2	Google Veo3.1
(4) A person plays squash on an indoor court.
Ours	OpenAI Sora2	Google Veo3.1

 

Introduction

This is a re-implementation of our work "PhyGDPO: Physics-Aware Groupwise Direct Preference Optimization for Physically Consistent Text-to-Video Generation" using public datasets and re-trained model based on public codes. In this work, we present a data construction pipeline, PhyAugPipe, that can create data pairs and a new direct preference optimization framework, PhyGDPO, for physically plausile text-to-video generation. If you find our repo useful, please give it a star ⭐ and consider citing our paper. Thank you :)

News

• 2026.08.31 : The code, data, and models of PhyGDPO are now fully open-sourced. Welcome to use them and share your feedback! 🚀
• 2026.01.04 : Our paper is on arXiv now. 🚀
• 2026.01.03 : Our project page has been built up. Feel free to check the video generation results on the project page.

1. Data Construction

We explain how to construct the training data from the original T2V data in the folder PhyAugPipe. Please enter this folder and refer to the README.md file for detailed instructions.

2. Model Training and Inference

We use the Wan2.1-T2V-14B as the base model and post-train it with our PhyGDPO framework. Our code is developed based on the DiffSynth GitHub repository. Please enter the folder DiffSynth and refer to README.md for detailed instruction of model training and inference.

3. Citation

@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}
}