A Unified Framework for Diffusion Model Unlearning with f-Divergence

Nicola Novello, Federico Fontana, Luigi Cinque, Deniz Gündüz, Andrea M. Tonello

Official repository of the paper "A Unified Framework for Diffusion Model Unlearning with f-Divergence" published at ICML 2026.

$f$-DMU is a unified framework for Diffusion Model Unlearning based on $f$-divergence. It comprises two classes of objective functions: i) ''closed-form losses'' (best choice for most scenarios) characterized by a good erasure-preservation trade-off; ii) ''variational losses'' that lead to a more aggressive erasure.

---

💻 How to run the code

The section will be completed soon...

Closed-Form Losses

Variational Losses

---

🤓 Guidelines

The section will be completed soon...

Closed-Form Losses

Variational Losses

---

📝 Reference

If you use the code for your research, please cite our paper:

@article{novello2026unified,
  title={A Unified Framework for Diffusion Model Unlearning with f-Divergence},
  author={Novello, Nicola and Fontana, Federico and Cinque, Luigi and Gunduz, Deniz and Tonello, Andrea M},
  journal={International Conference on Machine Learning},
  year={2026}
}

📋 Acknowledgments

The implementation is based on / inspired by:

• https://github.com/nupurkmr9/concept-ablation
• https://github.com/yongliang-wu/DoCo

---

📧 Contact

nicola.novello@aau.at