CaiYuanhao/PhyGDPO_demo

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

Dataset Description

Here are the demo videos for PhyGDPO, our physics-aware groupwise direct preference optimization framework for physically consistent text-to-video generation.

This repository contains 124 MP4 videos, including comparisons with Seedance 2.0, Wan2.1-T2V-14B, OpenAI Sora2, Google DeepMind Veo3.1, VideoDPO, and PhyT2V, as well as DPO framework comparisons and ablation studies. The following GIFs show selected qualitative results. Please visit our project page for all comparisons and prompts.

(1) A gymnast drops from the parallel bars and lands safely on the mat below.
Ours	OpenAI Sora 2	Google Veo 3.1
(2) A soccer player runs, plants their foot, and drop-kicks a soccer ball high into the air, with the ball visibly arcing through the air.
Ours	OpenAI Sora 2	Google Veo 3.1
(3) A person wearing a helmet performs a handspring over a platform.
Ours	OpenAI Sora 2	Google Veo 3.1
(4) A person plays squash on an indoor court.
Ours	OpenAI Sora 2	Google Veo 3.1

Dataset Structure

Directory	Videos	Description
Main_Comparison/	112	16 scenes, each containing videos from 7 methods, including Ours.
DPO_comparison/	4	Wan2.1-T2V-1.3B baseline, VideoDPO, Flow-DPO, and PhyGDPO.
Breakdown_ablation/	4	Baseline and progressive additions of LoRA-SR, groupwise modeling, and physics-guided rewarding.
ablation_lorasr/	4	Baseline, LoRA-SFT, and PhyGDPO without and with LoRA-SR.

This repository provides selected demonstration videos. For training data, please use the separate dataset linked below. These qualitative examples complement the quantitative evaluation in our paper.

Github Code Link

For implementation and usage instructions, please refer to our GitHub repository:

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

Huggingface Model Link

We release PhyGDPO models based on Wan2.1-1.3B and Wan2.1-14B at:

https://huggingface.co/CaiYuanhao/PhyGDPO

Huggingface Training Dataset Link

The training data is available at:

https://huggingface.co/datasets/CaiYuanhao/PhyGDPO

Project Page Link

For more video results, corresponding prompts, and side-by-side comparisons, please refer to our project page:

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

Meta AI Research Report

https://ai.meta.com/research/publications/phygdpo-physics-aware-groupwise-direct-preference-optimization-for-physically-consistent-text-to-video-generation/

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}
}