Research map / Image generation / Image restoration
sRGB Real Noise Modeling via Noise-Aware Sampling with Normalizing Flows
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NAFlow, a single conditional normalizing flow with per-camera-configuration Gaussian latents and a Noise-Aware Sampling procedure that generates sRGB noise matching an input image without metadata, improving SIDD denoising with DnCNN.
Problem
Generating realistic sRGB real-world noise typically requires camera metadata (ISO, smartphone model) at sampling time, and existing models like NeCA need multiple separate models per camera; spatial noise correlation is also often not modeled.
Approach
A single conditional normalizing flow (conditioned on the clean image) maps noisy images to a mixture of learnable per-camera-configuration Gaussians in latent space; at inference, the input's latent likelihoods under each Gaussian define a mixture from which noise is sampled (Noise-Aware Sampling), removing the need for metadata. A multi-scale squeeze-based architecture captures spatially correla
Key results
- NAFlow average KLD 0.0305 and AKLD 0.1306 on SIDD-Validation, best among compared methods, with a single model of 1.1M parameters versus five NeCA-W sub-model sets.
- On SIDD Benchmark, DnCNN trained on NAFlow-generated noise reaches 37.22 dB PSNR / 0.935 SSIM, 0.39 dB above NeCA-W and close to real-noise training (37.63 dB).
- On SIDD+ (unseen smartphones), NAFlow achieves KLD 0.0494 vs 0.0629 for NeCA-W, without metadata at inference.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NAFlow rather than NeCA-W for synthetic noise generation | sRGB noise generation · SIDD-Validation average | KLD: 0.0342 -> 0.0305 | Helped |
| Use NAFlow rather than NeCA-W to generate training noise | DnCNN · SIDD Benchmark | PSNR: 36.82 -> 37.22 | Helped |
| Use NAFlow rather than NeCA-W to generate training noise | DnCNN · SIDD Benchmark | SSIM: 0.932 -> 0.935 | Helped |
| Use noise-aware sampling over randomly selected latent distributions (ablation) | NAFlow · SIDD-Validation | KLD: 0.4700 -> 0.0291 | Helped |
| Use all latent distributions rather than the highest-likelihood distribution alone (ablation) | NAFlow · SIDD-Validation | KLD: 0.0294 -> 0.0291 | Helped |
| Use multiple distinct latent distributions rather than a single standard normal (ablation) | NAFlow · SIDD-Validation | KLD: 0.0544 -> 0.0291 | Helped |
| Use multiple distinct latent distributions rather than a single standard normal (ablation) | NAFlow · SIDD-Validation | AKLD: 0.1657 -> 0.1293 | Helped |
| Learn both the Gaussian mean and covariance (ablation) | NAFlow · SIDD-Validation | KLD | Helped |
| Increase the number of scale-level blocks (ablation) | NAFlow · SIDD-Validation | KLD | Helped |
Surprising findings
- NeCA-W achieves lower KLD than NAFlow on several individual cameras (e.g., G4 and N6), with NAFlow best only on average.
- Rank-1 Gaussian sampling is more accurate than random Gaussian selection, and NAS with all components is best in KLD (0.0291) and AKLD (0.1293).
Limitations
- Trained only on SIDD, which has limited smartphone diversity and scene variety.
- Training became unstable for scale-level L>=4, limiting the multi-scale design.
- Per-camera results are mixed; NeCA-W is better on some cameras and PSNR still lags real-noise upper bound.
Experimental setup
Datasets: SIDD-Medium SIDD-Validation SIDD-Benchmark SIDD+ BSD500 DND
Benchmarks: SIDD-Validation SIDD+ SIDD Benchmark
Metrics: KLD AKLD PSNR SSIM
Baselines: C2N Flow-sRGB NeCA-S NeCA-W NoiseFlow
Builds on: Glow RealNVP conditional normalizing flow NeCA Flow-sRGB
Closest papers by meaning
- Diffusion-Based sRGB Real Noise Generation via Prompt-Driven Noise Representation Learning (CVPR 2026 · similarity 0.92)
- sRGB Real Noise Synthesizing With Neighboring Correlation-Aware Noise Model (CVPR 2023 · similarity 0.91)
- Learning to Generate Realistic Noisy Images via Pixel-level Noise-aware Adversarial Training (NeurIPS 2021 · similarity 0.91)
- Modeling sRGB Camera Noise With Normalizing Flows (CVPR 2022 · similarity 0.90)
- GAN2GAN: Generative Noise Learning for Blind Denoising with Single Noisy Images (ICLR 2021 · similarity 0.90)
- Realistic Noise Synthesis with Diffusion Models (AAAI 2025 · similarity 0.88)
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.