Research map / Image generation / Image restoration
Physics-based Noise Modeling for Extreme Low-light Photography
Kaixuan Wei, Ying Fu, Yinqiang Zheng, Jiaolong Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 Gaussian shot-noise approximation with true Poisson (ablation) | G+P · SID Sony (x250) | PSNR: 35.67 -> 37.27 | Improved |
| replace Gaussian read noise with Tukey lambda model (ablation) | G+P* · SID Sony (x250) | PSNR: 37.27 -> 38.20 | Improved |
| add color-wise bias noise B (ablation) | G*+P* · SID Sony (x250) | PSNR: 38.20 -> 40.44 | Improved |
| add row noise R (ablation) | G*+P*+B · SID Sony (x250) | PSNR: 40.44 -> 40.55 | Improved |
| add uniform quantization noise U (ablation) | G*+P*+B+R · SID Sony (x250) | PSNR: 40.55 -> 40.60 | Improved |
| add Gaussian-approx photon noise P to Gaussian read noise (ablation) | G · SID Sony (x250) | PSNR: 34.83 -> 35.67 | Improved |
| add color bias B (ablation) | G*+P* · SID Sony (x300) | PSNR: 34.56 -> 36.29 | Improved |
| train on synthetic noise instead of paired real data | denoiser U-Net · SID Sony (x250) | PSNR: 40.59 -> 40.60 | No clear effect |
| train on synthetic noise instead of paired real data | denoiser U-Net · ELD CanonEOS70D (x200) | PSNR: 37.32 -> 38.78 | Improved |
| train on synthetic noise instead of paired real data | denoiser U-Net · SID Fuji (x250) | PSNR: 38.56 -> 38.33 | Hurt |
| use physics-based model instead of G+P | denoiser U-Net · SID Fuji (x250) | PSNR: 34.86 -> 38.33 | Improved |
| train multi-frame fastDVDnet on synthetic dynamic video (D) instead of static video (S) (ablation) | fastDVDnet · synthetic dynamic video | PSNR: 28.47 -> 37.48 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Noise2Noise — Noise2Noise: Learning Image Restoration without Clean Data (ICML 2018) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- A Physics-Based Noise Formation Model for Extreme Low-Light Raw Denoising (CVPR 2020 · similarity 0.95)
- Physics-Guided ISO-Dependent Sensor Noise Modeling for Extreme Low-Light Photography (CVPR 2023 · similarity 0.93)
- Dark Noise Diffusion: Noise Synthesis for Low-Light Image Denoising (TPAMI 2025 · similarity 0.92)
- Towards General Low-Light Raw Noise Synthesis and Modeling (ICCV 2023 · similarity 0.91)
- Learning Physics-Informed Noise Models from Dark Frames for Low-Light Raw Image Denoising (TPAMI 2026 · similarity 0.90)
- Noise Modeling in One Hour: Minimizing Preparation Efforts for Self-supervised Low-Light RAW Image Denoising (CVPR 2025 · similarity 0.87)
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.
