Research map / Image generation / Fast diffusion sampling
Realistic Noise Synthesis with Diffusion Models
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 RNSD, a diffusion-based method that synthesizes realistic RGB camera noise using camera-conditioned affine modulation (TCCAM), a multi-scale content-aware UNet module (MCAM), and an accelerated sampler (DIPS) to augment denoising training data.
Problem
Real-world paired noisy training data for deep denoisers is costly to collect, and existing noise synthesis methods (GAN-based or Gaussian-based) fail to model complex, spatially correlated RGB noise across camera settings.
Approach
Treats real noisy images as the diffusion target x0 so the DDPM samples realistic noise; conditions the UNet on camera settings with time-dependent affine modulation (TCCAM), injects multi-scale clean-image features (MCAM) for signal-correlated noise, and accelerates sampling with a distilled single-step model and a denser, decaying step schedule (DIPS).
Key results
- RNSD reduces AKLD to 0.117 and PGap to 0.54 on the SIDD validation set, versus 0.153 and 0.84 for the best prior method PNGAN.
- Augmenting SIDD small with RNSD noise raises DnCNN-B PSNR from 37.70 to 38.27 dB; DnCNN trained on RNSD synthetic noise reaches 38.11 dB vs 38.40 dB with real data.
- DIPS-Advanced reaches AKLD 0.122 at 5 steps versus DDIM's 0.356 at 5 steps, approaching the 1000-step DDPM result (0.117).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace baseline noise synthesis with RNSD | Noise synthesis methods · SIDD validation | AKLD: 0.153 -> 0.117 | Helped |
| replace baseline noise synthesis with RNSD | Noise synthesis methods · SIDD validation | PGap: 0.84 -> 0.54 | Helped |
| train DnCNN on RNSD-synthesized noise instead of the strongest listed baseline, NeCA | DnCNN · Synthetic-noise denoising | PSNR: 37.65 -> 38.11 | Helped |
| augment SIDD small noise samples with RNSD | DnCNN-B · SIDD validation | PSNR: 37.70 -> 38.27 | Helped |
| augment SIDD small noise samples with RNSD | RIDNet · SIDD validation | PSNR: 38.20 -> 38.74 | Helped |
| augment SIDD small noise samples with RNSD | NAFNet · SIDD validation | PSNR: 38.95 -> 39.56 | Helped |
| fine-tune with RNSD noise augmentation | RIDNet · SIDD | PSNR: 38.77 -> 39.10 | Helped |
| fine-tune with RNSD noise augmentation | NAFNet · DND | PSNR: 38.43 -> 39.05 | Helped |
| add camera-setting concatenation (ablation) | RNSD · Noise synthesis | AKLD: 0.169 -> 0.137 | Helped |
| add TCCAM (ablation) | RNSD · Noise synthesis | AKLD: 0.137 -> 0.126 | Helped |
| add MCAM to TCCAM (ablation) | RNSD · Noise synthesis | AKLD: 0.126 -> 0.117 | Helped |
| replace DDIM sampling with DIPS-Basic (ablation) | RNSD · Noise synthesis, 5 steps | AKLD: 0.356 -> 0.208 | Helped |
Surprising findings
- Plain DDIM with reduced steps degrades strongly (AKLD 0.507 at 3 steps), motivating a non-uniform, distillation-based sampler.
Limitations
- Results rely on SIDD small, a limited dataset with five smartphone cameras, so generalization to other sensors is uncertain.
- Reported numbers have internal inconsistencies (e.g., AKLD reduction described inconsistently between 0.169 and 0.130), and the DIPS-Advanced 4% loss claim is not clearly matched in the table.
Experimental setup
Datasets: SIDD small SIDD medium LSDIR DND
Benchmarks: SIDD validation set DND
Metrics: AKLD PGap PSNR SSIM
Baselines: C2N sRGB2Flow DANet GRDN PNGAN NeCA NoiseFlow GMDCN DDIM
Builds on: DDPM DDIM Deep Image Prior FiLM-style affine modulation
Closest papers by meaning
- sRGB Real Noise Modeling via Noise-Aware Sampling with Normalizing Flows (ICLR 2024 · similarity 0.88)
- Diffusion Priors for Variational Likelihood Estimation and Image Denoising (NeurIPS 2024 · similarity 0.87)
- Compressed Image Generation with Denoising Diffusion Codebook Models (ICML 2025 · similarity 0.87)
- Relay Diffusion: Unifying diffusion process across resolutions for image synthesis (ICLR 2024 · similarity 0.87)
- Noise Modeling in One Hour: Minimizing Preparation Efforts for Self-supervised Low-Light RAW Image Denoising (CVPR 2025 · similarity 0.87)
- NoiseDiffusion: Correcting Noise for Image Interpolation with Diffusion Models beyond Spherical Linear Interpolation (ICLR 2024 · 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.