# Noise2Noise Revisited: Training Pair Distributions Dominate Loss Choice in Self-Supervised Denoising

arXiv: https://arxiv.org/abs/2609.16788 · HTML: https://mlsanity.com/paper/2609.16788/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 19 of its 36 references that are in the map.
- First posted on arXiv 15 Sep 2026.

## Builds on (most cited first, 19 of 19)

- [U-Net: Convolutional Networks for Biomedical Image Segmentation](https://mlsanity.com/paper/1505.04597/) (2015)
- [The Unreasonable Effectiveness of Deep Features as a Perceptual Metric](https://mlsanity.com/paper/1801.03924/) (2018)
- [Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising](https://mlsanity.com/paper/1608.03981/) (2016)
- [Noise2Noise: Learning Image Restoration without Clean Data](https://mlsanity.com/paper/1803.04189/) (2018)
- [Noise2Void - Learning Denoising from Single Noisy Images](https://mlsanity.com/paper/1811.10980/) (2018)
- [Noise2Self: Blind Denoising by Self-Supervision](https://mlsanity.com/paper/1901.11365/) (2019)
- [Toward Convolutional Blind Denoising of Real Photographs](https://mlsanity.com/paper/1807.04686/) (2018)
- [Unprocessing Images for Learned Raw Denoising](https://mlsanity.com/paper/1811.11127/) (2018)
- [Benchmarking Denoising Algorithms with Real Photographs](https://mlsanity.com/paper/1707.01313/) (2017)
- [Neighbor2Neighbor: Self-Supervised Denoising from Single Noisy Images](https://mlsanity.com/paper/2101.02824/) (2021)
- [High-Quality Self-Supervised Deep Image Denoising](https://mlsanity.com/paper/1901.10277/) (2019)
- [Noisier2Noise: Learning to Denoise from Unpaired Noisy Data](https://mlsanity.com/paper/1910.11908/) (2019)
- [AP-BSN: Self-Supervised Denoising for Real-World Images via Asymmetric PD and Blind-Spot Network](https://mlsanity.com/paper/2203.11799/) (2022)
- [Blind2Unblind: Self-Supervised Image Denoising with Visible Blind Spots](https://mlsanity.com/paper/2203.06967/) (2022)
- [Noise2Same: Optimizing A Self-Supervised Bound for Image Denoising](https://mlsanity.com/paper/2010.11971/) (2020)
- [CVF-SID: Cyclic multi-Variate Function for Self-Supervised Image Denoising by Disentangling Noise from Image](https://mlsanity.com/paper/2203.13009/) (2022)
- [Positive2Negative: Breaking the Information-Lossy Barrier in Self-Supervised Single Image Denoising](https://mlsanity.com/paper/2412.16460/) (2024)
- [Next-Scale Prediction: A Self-Supervised Approach for Real-World Image Denoising](https://mlsanity.com/paper/2512.21038/) (2025)
- [Learning to Recorrupt: Noise Distribution Agnostic Self-Supervised Image Denoising](https://mlsanity.com/paper/2603.25869/) (2026)

## Abstract (arXiv)

Noise2Noise (N2N) trains denoisers on pairs of independently corrupted observations, eliminating clean references. We stress-test two natural conjectures about why the L1 loss outperforms L2 here. First, the hypothesis that the L1 loss confers robustness via parameter sparsity confuses the loss with Lasso regularization: an explicit Lasso penalty produces the predicted sparsity yet fails to reproduce L1's cross-noise behavior, while L1- and L2-trained weight distributions are indistinguishable. Second, the population optima of the two losses coincide exactly for symmetric signal posteriors and nearly so for concentrated ones. Measured differences are therefore dominated by optimization dynamics (bounded-influence gradients), which we probe with gradient statistics and contaminated-target training. On Kodak24 with five synthetic noise families, the L1 loss holds a statistically significant edge over L2, below 1 dB PSNR, holding across three seeds on 13 of the 14 noise columns. On real camera noise the loss is not the decisive variable in distribution: on official SIDD validation blocks, synthetic-Gaussian-trained N2N models gain only 0.8 to 3.7 dB over the noisy input regardless of loss, while retraining on SIDD's own noisy pairs, never reading ground truth, gains 9.4 to 11.0 dB, far ahead of BM3D. All metrics are on raw network outputs, and the study makes no leaderboard claim. The training pair distribution, not the loss, carries the inductive bias. That design rule applies wherever clean references are unobtainable, from microscopy to industrial inspection sensors.