README.md

---
task_categories:
- unconditional-image-generation
tags:
- vae
- diffusion-models
- imagenet
- image-generation
---

# REPA-E: Generated Samples and Artifacts

This repository hosts the generated image samples and artifacts associated with the paper [REPA-E: Unlocking VAE for End-to-End Tuning with Latent Diffusion Transformers](https://huggingface.co/papers/2504.10483). These samples are instrumental for quantitative evaluation of the REPA-E method, which enables end-to-end tuning of latent diffusion models and VAEs, achieving state-of-the-art image generation performance.

- 🌐 [Project Page](https://end2end-diffusion.github.io/)
- 📃 [Paper](https://huggingface.co/papers/2504.10483)
- 💻 [Code Repository](https://github.com/End2End-Diffusion/REPA-E)

## Overview

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.

**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. The generated `.npz` files for these evaluations can be found within this repository (e.g., under `labelsampling-equal-run1`).

## Sample Usage

This section shows how to load and use the REPA-E fine-tuned VAE (E2E-VAE) in latent diffusion training, demonstrating how E2E-VAE acts as a drop-in replacement for the original VAE, enabling significantly accelerated generation performance.

### ⚡️ Quickstart
```python
from diffusers import AutoencoderKL

# Load end-to-end tuned VAE (ImageNet VAE example)
vae = AutoencoderKL.from_pretrained("REPA-E/e2e-vavae-hf").to("cuda")

# Or load a text-to-image VAE
vae = AutoencoderKL.from_pretrained("REPA-E/e2e-flux-vae").to("cuda")

# Use in your pipeline with vae.encode(...) / vae.decode(...)
```

### 🧩 Complete Example
Full workflow for encoding and decoding images:
```python
from io import BytesIO
import requests
from diffusers import AutoencoderKLQwenImage
import numpy as np
import torch
from PIL import Image

response = requests.get("https://raw.githubusercontent.com/End2End-Diffusion/fuse-dit/main/assets/example.png")
device = "cuda"

image = torch.from_numpy(
    np.array(
        Image.open(BytesIO(response.content))
    )
).permute(2, 0, 1).unsqueeze(0).to(torch.float32) / 127.5 - 1
image = image.to(device)

vae = AutoencoderKLQwenImage.from_pretrained("REPA-E/e2e-qwenimage-vae").to(device)

# Add frame dimension (required for QwenImage VAE)
image_ = image.unsqueeze(2)

with torch.no_grad():
    latents = vae.encode(image_).latent_dist.sample()
    reconstructed = vae.decode(latents).sample

# Remove frame dimension
latents = latents.squeeze(2)
reconstructed = reconstructed.squeeze(2)
```

## Quantitative Results
Tables below report generation performance using gFID on 50k samples, with and without classifier-free guidance (CFG). We compare models trained end-to-end with **REPA-E** and models using a frozen REPA-E fine-tuned VAE (**E2E-VAE**). Lower is better. All linked checkpoints below are hosted on our [🤗 Hugging Face Hub](https://huggingface.co/REPA-E). To reproduce these results, download the respective checkpoints to the `pretrained` folder and run the evaluation script as detailed in the [GitHub repository](https://github.com/End2End-Diffusion/REPA-E/blob/main/README.md#5-generate-samples-and-run-evaluation).

#### A. End-to-End Training (REPA-E)
| Tokenizer | Generation Model | Epochs | gFID-50k ↓ | gFID-50k (CFG) ↓ |
|:---------|:----------------|:-----:|:----:|:---:|
| [**SD-VAE<sup>*</sup>**](https://huggingface.co/REPA-E/sdvae) | [**SiT-XL/2**](https://huggingface.co/REPA-E/sit-repae-sdvae) | 80 | 4.07 | 1.67<sup>a</sup> |
| [**IN-VAE<sup>*</sup>**](https://huggingface.co/REPA-E/invae) | [**SiT-XL/1**](https://huggingface.co/REPA-E/sit-repae-invae) | 80 | 4.09 | 1.61<sup>b</sup> |
| [**VA-VAE<sup>*</sup>**](https://huggingface.co/REPA-E/vavae) | [**SiT-XL/1**](https://huggingface.co/REPA-E/sit-repae-vavae) | 80 | 4.05 | 1.73<sup>c</sup> |

\* The "Tokenizer" column refers to the initial VAE used for joint REPA-E training. The final (jointly optimized) VAE is bundled within the generation model checkpoint. 

#### B. Traditional Latent Diffusion Model Training (Frozen VAE)
| Tokenizer | Generation Model | Method | Epochs | gFID-50k ↓ | gFID-50k (CFG) ↓ |
|:------|:---------|:----------------|:-----:|:----:|:---:|
| SD-VAE | SiT-XL/2 | SiT | 1400 | 8.30 | 2.06 |
| SD-VAE | SiT-XL/2 | REPA | 800 | 5.84 | 1.28 |
| VA-VAE | LightningDiT-XL/1 | LightningDiT | 800 | 2.05 | 1.25 |
| [**E2E-VAVAE (Ours)**](https://huggingface.co/REPA-E/e2e-vavae) | [**SiT-XL/1**](https://huggingface.co/REPA-E/sit-ldm-e2e-vavae) | REPA | 800 | **1.69** | **1.12**<sup>†</sup> |

In this setup, the VAE is kept frozen and only the generator is trained. Models using our E2E-VAE (fine-tuned via REPA-E) consistently outperform baselines such as SD-VAE and VA-VAE, achieving state-of-the-art performance when incorporating the REPA alignment objective.

**Note**: The results for the last three rows (REPA, LightningDiT, and E2E-VAE) are obtained using the class-balanced sampling protocol (50 images per class).

## Citation
If you find our work useful, please consider citing:

```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},
}
```