Open-PhyGDPO

KO | EN

gitlite — search

#javascript #python #typescript #deep-learning #python3 #nextjs #react #html #css #flutter #ai #computer-vision

★ 69 Open GitHub ↗

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

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

&nbsp;

[![arXiv](https://img.shields.io/badge/paper-arxiv-179bd3)](https://arxiv.org/abs/2512.24551)
[![project](https://img.shields.io/badge/project-page-green)](https://caiyuanhao1998.github.io/project/PhyGDPO/)
[![hf](https://img.shields.io/badge/hugging-face-green)](https://huggingface.co/datasets/CaiYuanhao/PhyGDPO_demo)
<h3>PhyGDPO: Physics-Aware Groupwise Direct Preference <br> Optimization for Physically Consistent Text-to-Video Generation</h3>

<tr>
    <td colspan="3" align="center" style="border:0;padding:6px 10px;font-style:italic;">
      (1) A gymnast drops from the parallel bars and lands safely on the mat below.
    </td>
  </tr>
  <tr>
    <td style="border:0;padding:10px;">
      <img src="img/72_ours.gif" width="250" height="140">
    </td>
    <td style="border:0;padding:10px;">
      <img src="img/72_sora.gif" width="250" height="140">
    </td>
    <td style="border:0;padding:10px;">
      <img src="img/72_veo.gif"  width="250" height="140">
    </td>
  </tr>

<tr>
    <td align="center" style="border:0;padding-top:6px;font-weight:700;">Ours</td>
    <td align="center" style="border:0;padding-top:6px;font-weight:700;">OpenAI Sora2</td>
    <td align="center" style="border:0;padding-top:6px;font-weight:700;">Google Veo3.1</td>
  </tr>

<tr>
    <td colspan="3" align="center" style="border:0;padding:6px 10px;font-style:italic;">
      (2) A soccer player runs, plants their foot, and drop kicks a soccer ball high into the air, the ball arcing visibly.
    </td>
  </tr>
  <tr>
    <td style="border:0;padding:10px;">
      <img src="img/195_ours.gif" width="250" height="140">
    </td>
    <td style="border:0;padding:10px;">
      <img src="img/195_sora.gif" width="250" height="140">
    </td>
    <td style="border:0;padding:10px;">
      <img src="img/195_veo.gif" width="250" height="140">
    </td>
  </tr>

<tr>
    <td colspan="3" align="center" style="border:0;padding:6px 10px;font-style:italic;">
      (3) A person wearing a helmet performs a handspring over a platform.
    </td>
  </tr>
  <tr>
    <td style="border:0;padding:10px;">
      <img src="img/wild_1_ours.gif" width="250" height="140">
    </td>
    <td style="border:0;padding:10px;">
      <img src="img/wild_1_sora.gif" width="250" height="140">
    </td>
    <td style="border:0;padding:10px;">
      <img src="img/wild_1_veo.gif" width="250" height="140">
    </td>
  </tr>

<tr>
    <td colspan="3" align="center" style="border:0;padding:6px 10px;font-style:italic;">
      (4) A person plays squash on an indoor court.
    </td>
  </tr>
  <tr>
    <td style="border:0;padding:10px;">
      <img src="img/wild_2_ours.gif" width="250" height="140">
    </td>
    <td style="border:0;padding:10px;">
      <img src="img/wild_2_sora.gif" width="250" height="140">
    </td>
    <td style="border:0;padding:10px;">
      <img src="img/wild_2_veo.gif" width="250" height="140">
    </td>
  </tr>

<tr>
    <td align="center" style="border:0;padding-top:6px;font-weight:700;">Ours</td>
    <td align="center" style="border:0;padding-top:6px;font-weight:700;">OpenAI Sora2</td>
    <td align="center" style="border:0;padding-top:6px;font-weight:700;">Google Veo3.1</td>
  </tr>

</table>
</p>

&nbsp;

</div>

### 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 :)

&nbsp;

<p align="center"><strong>Figure 1:</strong> Our data construction pipeline</p>

&nbsp;

<p align="center"><strong>Figure 2:</strong> The overall framework of our OmniVCus</p>

### News
- **2026.06.18 :** Our paper has been accepted by ECCV 2026. Code, data, and models will be released soon. Stay tuned. 🚀 
- **2025.12.31 :** Our paper is on [arxiv](https://arxiv.org/abs/2512.24551) now. Code, data, and models will be made publicly available, stay tuned. 🚀 
- **2025.12.31 :** Our [project page](https://caiyuanhao1998.github.io/project/PhyGDPO/) has been built up. Feel free to check the video generation results on the project page.

## Citation
If you find our repo useful, please consider citing our paper:
```sh
@article{cai2025phygdpo,
  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 others},
  journal={arXiv preprint arXiv:2512.24551},
  year={2025}
}
```

Open-PhyGDPO

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README