Mitigating data replication in text-to-audio generative diffusion models through anti-memorization guidance

Francisco Messina, Francesca Ronchini, Luca Comanducci, Paolo Bestagini, Fabio Antonacci

This repository accompains the the paper Mitigating data replication in text-to-audio generative diffusion models through anti-memorization guidance. It builds on Stability AI’s Stable Audio Open 1.0 and adds Anti-Memorization Guidance (AMG) during sampling.

• In the folder code you can find the code used to perform the experiments and also the plots included in the paper.
• In the folder docs you can find the additional material, which is also nicely presented in the accompanying webpage References to the paper and base model are at the end of this document.

Citation

If you use this code in academic work, please cite:

@inproceedings{messina2026mitigating,
  title={Mitigating data replication in text-to-audio generative diffusion models through anti-memorization guidance},
  author={Messina, Francisco and Ronchini, Francesca and Comanducci, Luca and Bestagini, Paolo and Antonacci, Fabio},
  booktitle={ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={15742--15746},
  year={2026},
  organization={IEEE}
}

Stable Audio Open:

@inproceedings{evans2025stable,
  title={Stable audio open},
  author={Evans, Zach and Parker, Julian D and Carr, CJ and Zukowski, Zack and Taylor, Josiah and Pons, Jordi},
  booktitle={ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={1--5},
  year={2025},
  organization={IEEE}
}

---

For questions or reproducibility details (e.g., exact c1/c2/c3 and scheduling configurations used for the paper experiments), you can inspect amg_infer.py in this repository, the AMG logic within stable_audio_tools/inference/amg_generation.py, and the reference paper.