---
license: mit
task_categories:
- image-to-image
language:
- en
tags:
- image-compression
- diffusion
- gaussian-blur
- zero-cost-encoder
---

# Diffusion-Based Image Compression with Zero-Cost Encoders

This repository contains model checkpoints for the paper:
**"Classical Transformations as Zero-Cost Encoders for Diffusion-Based Image Compression: A Study with Gaussian Blur"**
(Submitted to BMVC 2026)

## Checkpoints

We release the σ=1 checkpoint for each domain — the practical operating point achieving 20-28% compression with strong perceptual quality across all domains.

| Domain | File | PSNR | LPIPS | Compression |
|--------|------|------|-------|-------------|
| MURA X-ray | `xray_sigma1/cond_step_600000.pt` | 48.81 dB | 0.006 | 26.5% |
| BraTS MRI | `brats_sigma1/cond_step_600000.pt` | 49.20 dB | 0.002 | 28.3% |
| CelebA | `celeba_sigma1/cond_step_582000.pt` | 42.01 dB | 0.004 | 20.3% |
| Buildings | `buildings_sigma1/cond_step_600000.pt` | 38.50 dB | 0.007 | 23.7% |

## Usage

Evaluation scripts are provided in the supplementary material of the paper. To run inference on a single image:

```bash
KMP_DUPLICATE_LIB_OK=TRUE python eval_single.py /path/to/image.png
```

The script auto-detects the dataset from the path and uses the appropriate checkpoint.

## Datasets

- **MURA X-ray**: https://stanfordmlgroup.github.io/competitions/mura/
- **BraTS**: https://www.synapse.org/Synapse:syn51156910
- **CelebA**: https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
- **LSUN Buildings**: https://www.yf.io/p/lsun

## Requirements

```
torch
torchvision
Pillow
numpy
scikit-image
lpips
```