Reproduce “FiRe: Fixed-points of Restoration Priors for Solving Inverse Problems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Matthieu Terris, Ulugbek S. Kamilov, Thomas Moreau
The FiRe framework introduces a method to leverage diverse pre-trained image restoration models as implicit priors for solving inverse problems, moving beyond traditional denoising. It resolves distribution mismatch and fixed-point instability by utilizing the stable fixed-point property of a restoration model composed with its training degradation (R 7 D), resulting in improved image reconstruction quality across various tasks.
Key insights
- Natural images exhibit stable fixed-point behavior under the composition of a restoration model with its training degradation (R 7 D), which preserves naturalness unlike direct iteration of R.
- The action of R 7 D can be formally linked to the gradient of the squared distance to a prox-regular set, providing a robust theoretical foundation for using diverse restoration networks as implicit priors.
- Explicitly applying the degradation D before restoration R during iterations addresses the distribution mismatch problem, critical for effectively integrating general restoration models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.