license: mit
library_name: diffusers
pipeline_tag: image-to-image
EQ-VAE: Equivariance Regularized Latent Space for Improved Generative Image Modeling
Arxiv: https://arxiv.org/abs/2502.09509 Project Page: https://eq-vae.github.io/ Code: https://github.com/zelaki/eqvae
EQ-VAE regularizes the latent space of pretrained autoencoders by enforcing equivariance under scaling and rotation transformations.
Model Description
This model is a regularized version of SD-VAE. We finetune it with EQ-VAE regularization for 5 epochs on OpenImages.
Model Usage
These weights are intended to be used with the EQ-VAE codebase or the CompVis Stable Diffusion codebase. If you are looking for the model to use with the 🧨 diffusers library, come here.
Quick Start with 🧨 Diffusers
If you just want to use EQ-VAE to speed up 🚀 the training on your diffusion model, you can use our HuggingFace checkpoints 🤗. We provide two models: eq-vae and eq-vae-ema.
from diffusers import AutoencoderKL
eqvae = AutoencoderKL.from_pretrained("zelaki/eq-vae")
If you are looking for the weights in the original LDM format you can find them here: eq-vae-ldm, eq-vae-ema-ldm
Metrics
Reconstruction performance of eq-vae-ema on Imagenet Validation Set.
| Metric | Score |
|---|---|
| FID | 0.82 |
| PSNR | 25.95 |
| LPIPS | 0.141 |
| SSIM | 0.72 |
Acknowledgement
This code is mainly built upon LDM and fastDiT.
Citation
@inproceedings{
kouzelis2025eqvae,
title={{EQ}-{VAE}: Equivariance Regularized Latent Space for Improved Generative Image Modeling},
author={Theodoros Kouzelis and Ioannis Kakogeorgiou and Spyros Gidaris and Nikos Komodakis},
booktitle={Forty-second International Conference on Machine Learning},
year={2025},
url={https://openreview.net/forum?id=UWhW5YYLo6}
}