---
license: apache-2.0
pipeline_tag: image-text-to-text
---

# HermesFlow

Official Repository of the paper: *[HermesFlow](https://github.com/Gen-Verse/HermesFlow)*.
<p align="left">
  <a href='https://arxiv.org/abs/2502.12148'>
  <img src='https://img.shields.io/badge/Arxiv-2502.12148-A42C25?style=flat&logo=arXiv&logoColor=A42C25'></a> 
  <a href='https://github.com/Gen-Verse/HermesFlow'>
    <img src='https://img.shields.io/badge/GitHub-Code-black?style=flat&logo=github&logoColor=white'></a> 
</p>
<img src="./pipeline.png" style="zoom:100%;" />

<img src="./image.png" style="zoom:100%;" />

<img src="./image-1.png" style="zoom:100%;" />

## News🔥🔥🔥

* Feb.18, 2025. Our checkpoints are publicly available on [HuggingFace Repo](https://huggingface.co/Gen-Verse/HermesFlow).

## Introduction

HermesFlow is a general alignment framework for multimodal LLMs, which cruate homologous preference data itself and utilize self-play iterative optimization with Pair-DPO to seamlessly close the gap between multimodal understanding and generation.

## Citation

```
@article{yang2025hermesflow,
  title={HermesFlow: Seamlessly Closing the Gap in Multimodal Understanding and Generation},
  author={Yang, Ling and Zhang, Xinchen and Tian, Ye and Shang, Chenming and Xu, Minghao and Zhang, Wentao and Cui, Bin},
  journal={arXiv preprint arXiv:2502.12148},
  year={2025}
}
```