---
pipeline_tag: unconditional-image-generation
---

# Distribution Matching Variational AutoEncoder (DMVAE)

This repository contains the official implementation of the paper [**"Distribution Matching Variational AutoEncoder"**](https://huggingface.co/papers/2512.07778).

![DMVAE Framework](https://github.com/sen-ye/dmvae/raw/main/figs/dmvae.png)

DMVAE introduces a novel approach to visual generative models by explicitly aligning the encoder's latent distribution with an arbitrary reference distribution via a distribution matching constraint. This method generalizes beyond the Gaussian prior of conventional VAEs, enabling alignment with distributions derived from self-supervised features, diffusion noise, or other prior distributions, leading to efficient and high-fidelity image synthesis.

For more details on the installation, training, and evaluation, please refer to the official [GitHub repository](https://github.com/sen-ye/dmvae).

## Citation

If you find this repository useful in your research or applications, please consider citing:

```text
@misc{dmvae,
      title={Distribution Matching Variational AutoEncoder}, 
      author={Sen Ye and Jianning Pei and Mengde Xu and Shuyang Gu and Chunyu Wang and Liwei Wang and Han Hu},
      year={2025},
      eprint={2512.07778},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.07778}, 
}
```