Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising
Paper • 2603.25869 • Published
Noise distribution agnostic self-supervised image denoising.
Brayan Monroy · Jorge Bacca · Julian Tachella
L2R jointly learns a denoiser and a monotonic recorruptor using noisy images alone.
Eight RGB DRUNet checkpoints cover log-gamma, Laplace, correlated Gaussian, Poisson–Gaussian, and Poisson noise. manifest.json contains their settings and SHA-256 hashes. mfcn_init.pth initializes the monotonic network as a Gaussian identity map.
From the code repository, with dependencies installed and test caches available:
hf download bemc22/learning2recorrupt --local-dir pretrained
python test.py --checkpoint-dir pretrained
@article{monroy2026learning,
title={Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising},
author={Monroy, Brayan and Bacca, Jorge and Tachella, Juli{\'a}n},
journal={arXiv preprint arXiv:2603.25869},
year={2026}
}