README.md

---
license: mit
---

<h2 align="center">Scale Space Diffusion [CVPR 2026]</h2>

<p align="center">
  <a href="https://soumik-kanad.github.io/">Soumik Mukhopadhyay</a><sup>*</sup>,&nbsp;
  <a href="https://prateksha.github.io/">Prateksha Udhayanan</a><sup>*</sup>,&nbsp;
  <a href="https://abhinavsh.info/">Abhinav Shrivastava</a>&nbsp;
</p>
<p align="center">
  <sup></sup>University of Maryland, College Park &nbsp; 
  <br>
  <sup>*</sup>Equal contribution.
</p>


<p align="center">
  <a href="https://prateksha.github.io/projects/scale-space-diffusion/"><img src="https://img.shields.io/badge/Website-Project_Page-2ea44f" /></a>
  &nbsp;
  <a href="https://arxiv.org/abs/2603.08709"><img src="https://img.shields.io/badge/arXiv-2603.08709-b31b1b.svg" /></a>
  &nbsp;
  <a href="https://github.com/prateksha/ScaleSpaceDiffusion"><img src="https://img.shields.io/badge/GitHub-Code-blue" /></a>
</p>
Scale Space Diffusion is a pixel-space diffusion model that integrates scale-space theory by combining Gaussian noise with linear degradations (downsampling). It introduces Flexi-UNet for resolution-aware denoising.

<table>
  <tr>
    <td align="center">
      <img src="https://raw.githubusercontent.com/prateksha/ScaleSpaceDiffusion/main/assets/teaser.png"
           alt="Scale Space Diffusion teaser" width="85%">
    </td>
    <td align="center">
      <img src="https://raw.githubusercontent.com/prateksha/ScaleSpaceDiffusion/main/assets/ssd_gif.gif"
           alt="Scale Space Diffusion sampling animation" width="100%">
    </td>
  </tr>
</table>


This model release accompanies the official implementation of **Scale Space Diffusion**.

## Model Details

The released checkpoints include SSD models trained on CelebA and ImageNet:

- CelebA-64: 2L and 4L Flexi-UNet checkpoints
- CelebA-128: 3L and 5L Flexi-UNet checkpoints
- CelebA-256: 3L and 6L Flexi-UNet checkpoints
- ImageNet-64: 2L Flexi-UNet checkpoint

## Citation
```
@article{mukhopadhyay2026scale,
  title={Scale Space Diffusion},
  author={Mukhopadhyay, Soumik and Udhayanan, Prateksha and Shrivastava, Abhinav},
  journal={arXiv preprint arXiv:2603.08709},
  year={2026}
}
```