Research map / Image generation / Image restoration
Noise2Noise: Learning Image Restoration without Clean Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace clean targets with independent zero-mean noisy targets | RED30 Gaussian denoising · Gaussian denoising, σ=25 | PSNR: 31.63±0.02 dB; performance remains equally good | No clear effect |
| Replace clean targets with noisy targets | U-Net Poisson denoising · Poisson denoising | PSNR: 30.59±0.02 dB -> 30.57±0.02 dB | No clear effect |
| Use separate corrupted masks for input and target | U-Net Bernoulli denoising · Bernoulli denoising | PSNR: 31.85±0.03 dB -> 32.02±0.03 dB | Helped |
| Replace clean targets with noisy targets at 2000 epochs | Monte Carlo denoiser · Monte Carlo denoising, 64 spp | PSNR: 31.83 dB; noisy targets gave 0.5 dB less | Hurt |
| Train noisy-target model for 4000 instead of 2000 epochs (ablation) | Monte Carlo denoiser · Monte Carlo denoising, 64 spp | PSNR: 31.83 dB at 4000 epochs | Helped |
| Replace clean targets with independently undersampled targets | MRI reconstruction network · MRI reconstruction, IXI validation | PSNR: 31.77 dB -> 31.74 dB | No clear effect |
| Reuse noisy realizations to create more noisy/noisy training pairs | Noise2Noise Gaussian denoising · Finite capture budget, 100 latents | PSNR: notably better results (several .1s of dB) | Helped |
| Increase the number of clean latents while using two noisy realizations per latent | Noise2Noise Gaussian denoising · Finite capture budget, 1000 latents | PSNR: even better results (again, by several .1s of dB) | Helped |
| Use spatially correlated brown noise rather than independent pixel noise (ablation) | Noise2Noise Gaussian denoising · Correlated Gaussian noise | convergence speed: eventual quality similar (within 0.1 dB) | Hurt |
| Use L1 loss instead of L2 for text corruption, targeting the median (ablation) | Text-removal network · Text removal | restoration quality | Helped |
| Use L0-style annealed loss rather than L1 for random-valued impulse noise (ablation) | Impulse-noise restoration network · Random-valued impulse noise, up to 90% corrupted pixels | restoration bias: 90% | Helped |
| Use L1 rather than L2 for random-valued impulse noise (ablation) | Impulse-noise restoration network · Random-valued impulse noise | restoration quality: fewer than 50% | Mixed |
Closest papers by meaning
- UNSURE: self-supervised learning with Unknown Noise level and Stein's Unbiased Risk Estimate (ICLR 2025 · similarity 0.87)
- Denoising as Adaptation: Noise-Space Domain Adaptation for Image Restoration (ICLR 2025 · similarity 0.87)
- Stochastic Deep Restoration Priors for Imaging Inverse Problems (ICML 2025 · similarity 0.86)
- Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise (NeurIPS 2024 · similarity 0.86)
- RestoreGrad: Signal Restoration Using Conditional Denoising Diffusion Models with Jointly Learned Prior (ICML 2025 · similarity 0.86)
- SeNM-VAE: Semi-Supervised Noise Modeling with Hierarchical Variational Autoencoder (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.