Self-diffusion for Solving Inverse Problems

NeurIPS 2025 | OpenReview | PDF

Authors: Guanxiong Luo, Shoujin Huang, Yanlong Yang

TL;DR: Self-diffusion solves inverse problems without the need of pretrained generative models via a self-contained iterative process that alternates between noising and denoising steps to progressively refine its estimate of the solution.

Keywords: inverse problems, computational imaging, image reconstruction, diffusion models

This repository contains the implementation of Self-diffusion for solving inverse problems. The code is organized as follows:

• sdi.py: Main implementation of the Self-diffusion algorithm.
• denoise.py: Denoising network and training utilities.
• utils.py: Helper functions for data loading and processing.
• simulation.py: Simulation scripts for generating synthetic data.
• run_*.sh: Bash scripts to run experiments on different tasks (MRI, general inverse problems, hyperparameter tuning, etc.).
• net/: Neural network architectures (UNet, extra deep UNet, etc.).
• misc/: Miscellaneous files including sample data and images.

Quick Start

To run Self-diffusion on a sample inverse problem (e.g., MRI reconstruction), use:

./run_mri.sh

Or for a general inverse problem:

./run_general.sh

Citation

If you use this code or find the paper useful, please cite:

@inproceedings{
  luo2025selfdiffusion,
  title={Self-diffusion for Solving Inverse Problems},
  author={Guanxiong Luo and Shoujin Huang},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
  url={https://openreview.net/forum?id=5g9qls1V7Q}
}

or

@article{
  luo2025selfdiffusionsolvinginverseproblems,
  title={Self-diffusion for Solving Inverse Problems}, 
  author={Guanxiong Luo and Shoujin Huang and Yanlong Yang},
  year={2025},
  eprint={2510.21417},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2510.21417}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Contact

For questions, please contact the authors via OpenReview or open an issue in this repository.