---
library_name: diffusers
pipeline_tag: image-to-image
tags:
  - diffusers
  - safetensors
  - image-editing
  - instruct-pix2pix
  - magicbrush
  - unsafe2safe
  - safe-attention
---

# Unsafe2Safe Checkpoint - Unofficial Reproduction

This repository contains an unofficial Safe Attention based UNet checkpoint trained by following the methodology described in the Unsafe2Safe paper.

This is an independent reproduction and is not affiliated with or endorsed by the original paper authors. Implementation details and model behavior may differ from the authors' unreleased implementation and weights.

[Unofficial implementation: https://github.com/JovanHengGhimHong/Unsafe2Safe_Unofficial_Reproduction](https://github.com/JovanHengGhimHong/Unsafe2Safe_Unofficial_Reproduction)

## Original Paper

**Unsafe2Safe: Controllable Image Anonymization for Downstream Utility**

Mih Dinh, SouYoung Jin

Paper: `https://arxiv.org/abs/2603.28605`

Please cite the original paper when using this checkpoint or implementation.

## Citations

```bibtex
@article{dinh2026unsafe2safe,
  title = {Unsafe2Safe: Controllable Image Anonymization for Downstream Utility},
  author = {Dinh, Mih and Jin, SouYoung},
  journal = {CVPR 2026},
  year = {2026}
}

@inproceedings{zhang2023magicbrush,
  title = {MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing},
  author = {Zhang, Kai and Mo, Lingbo and Chen, Wenhu and Sun, Huan and Su, Yu},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2023}
}

@inproceedings{brooks2023instructpix2pix,
  title = {InstructPix2Pix: Learning to Follow Image Editing Instructions},
  author = {Brooks, Tim and Holynski, Aleksander and Efros, Alexei A.},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2023}
}
```

## Disclaimer

This model does not guarantee complete or irreversible anonymization. Privacy-sensitive information may remain visible or inferable from image content, text, background context, metadata, or model-generated artifacts. Outputs should be independently evaluated before use in privacy-critical applications.