README.md

---
title: README
emoji: ⚡
colorFrom: indigo
colorTo: red
sdk: static
pinned: false
---

<h1 align="center"> REPA-E: Unlocking VAE for End-to-End Tuning of Latent Diffusion Transformers </h1>

<p align="center">
  <a href="https://scholar.google.com.au/citations?user=GQzvqS4AAAAJ" target="_blank">Xingjian&nbsp;Leng</a><sup>1*</sup> &ensp; <b>&middot;</b> &ensp;
  <a href="https://1jsingh.github.io/" target="_blank">Jaskirat&nbsp;Singh</a><sup>1*</sup> &ensp; <b>&middot;</b> &ensp;
  <a href="https://hou-yz.github.io/" target="_blank">Yunzhong&nbsp;Hou</a><sup>1</sup> &ensp; <b>&middot;</b> &ensp;
  <a href="https://people.csiro.au/X/Z/Zhenchang-Xing/" target="_blank">Zhenchang&nbsp;Xing</a><sup>2</sup>&ensp; <b>&middot;</b> &ensp;
  <a href="https://www.sainingxie.com/" target="_blank">Saining&nbsp;Xie</a><sup>3</sup>&ensp; <b>&middot;</b> &ensp;
  <a href="https://zheng-lab-anu.github.io/" target="_blank">Liang&nbsp;Zheng</a><sup>1</sup>&ensp;
</p>

<p align="center">
  <sup>1</sup> Australian National University &emsp; <sup>2</sup>Data61-CSIRO &emsp; <sup>3</sup>New York University &emsp; <br>
  <sub><sup>*</sup>Equal Contribution&emsp;</sub>
</p>

<p align="center">
  <a href="https://End2End-Diffusion.github.io">🌐 Project Page</a> &ensp;
  <a href="https://huggingface.co/REPA-E">🤗 Models</a> &ensp;
  <a href="https://arxiv.org/abs/2504.10483">📃 Paper</a> &ensp;
  <br>
  <!-- <a href="https://paperswithcode.com/sota/image-generation-on-imagenet-256x256?p=repa-e-unlocking-vae-for-end-to-end-tuning-of"><img src="https://img.shields.io/endpoint.svg?url=https://paperswithcode.com/badge/repa-e-unlocking-vae-for-end-to-end-tuning-of/image-generation-on-imagenet-256x256" alt="PWC"></a> -->
</p>


<p align="center">
  <img src="https://github.com/End2End-Diffusion/REPA-E/raw/main/assets/vis-examples.jpg" width="100%" alt="teaser">
</p>

---

We address a fundamental question: ***Can latent diffusion models and their VAE tokenizer be trained end-to-end?*** While training both components jointly with standard diffusion loss is observed to be ineffective — often degrading final performance — we show that this limitation can be overcome using a simple representation-alignment (REPA) loss. Our proposed method, **REPA-E**, enables stable and effective joint training of both the VAE and the diffusion model.

<p align="center">
  <img src="https://github.com/End2End-Diffusion/REPA-E/raw/main/assets/overview.jpg" width="100%" alt="teaser">
</p>

**REPA-E** significantly accelerates training — achieving over **17×** speedup compared to REPA and **45×** over the vanilla training recipe. Interestingly, end-to-end tuning also improves the VAE itself: the resulting **E2E-VAE** provides better latent structure and serves as a **drop-in replacement** for existing VAEs (e.g., SD-VAE), improving convergence and generation quality across diverse LDM architectures. Our method achieves state-of-the-art FID scores on ImageNet 256×256: **1.12** with CFG and **1.69** without CFG.

 
## Usage and Training

Please refer our [Github Repo](https://github.com/End2End-Diffusion/REPA-E) for detailed notes on end-to-end training and inference using REPA-E.

## 📚 Citation

```bibtex
@article{leng2025repae,
  title={REPA-E: Unlocking VAE for End-to-End Tuning with Latent Diffusion Transformers},
  author={Xingjian Leng and Jaskirat Singh and Yunzhong Hou and Zhenchang Xing and Saining Xie and Liang Zheng},
  year={2025},
  journal={arXiv preprint arXiv:2504.10483},
}
```