Research map / Image generation / Image restoration
Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement
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 LASQ, an unsupervised low-light image enhancement framework that models luminance transitions as hierarchical power-law distributions sampled via MCMC and embeds this sampling into a diffusion forward process, with an optional GAN variant (LASQ++) when normal-light references are available.
Problem
Low-light image enhancement struggles to balance reconstruction fidelity with cross-scenario generalization; deterministic pixel-level mappings overfit paired data and degrade when normal-light references are unavailable.
Approach
Reformulates LLIE as statistical sampling over hierarchical luminance adaptation operators (power-law gamma-like curves) drawn from a truncated Gaussian via hierarchical MCMC, and uses these sampled luminance-corrected images to guide the forward and reverse steps of a diffusion model without requiring normal-light references.
Key results
- On LOLv1, LASQ reports PSNR 20.375, SSIM 0.814, LPIPS 0.191, outperforming all unsupervised baselines in the table including LightenDiffusion (20.453 PSNR is higher for LightenDiffusion, so LASQ is not best on PSNR).
- On the no-reference VV dataset, LASQ reports NIQE 2.777 and PI 2.623, versus LightenDiffusion at 2.941 NIQE.
- Ablation: replacing adaptive MCMC luminance adjustment with fixed gamma-like curves drops LOLv1 PSNR from 20.375 to 16.741.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| incorporate unpaired normal-light references | LASQ · LOLv1 and LSRW | PSNR: 20.375 -> 20.481; 18.137 -> 18.584 | Mixed |
| incorporate unpaired normal-light references | LASQ · DICM, NPE, and VV | NIQE: 3.715 -> 3.723; 3.571 -> 3.601; 2.777 -> 2.850 | Hurt |
| replace adaptive MCMC luminance adaptation with fixed hand-crafted functions (ablation) | LASQ · LOLv1 test | PSNR: 20.375 -> 16.741 | Hurt |
| limit adaptive MCMC sampling to two layers, omitting intermediate strata (ablation) | LASQ · LOLv1 test | PSNR: 20.375 -> 19.139 | Hurt |
| vary alpha across 0.05 / 0.15 / 0.3 / 0.6 (ablation) | LASQ · LOLv1 | PSNR: 17.81 / 18.10 / 17.92 / 17.84 | No clear effect |
| vary eta across 0.1 / 1.0 / 3.0 / 6.0 (ablation) | LASQ · LOLv1 | PSNR: 17.85 / 18.35 / 18.17 / 17.95 | No clear effect |
| vary lambda_d across 0.1 / 1.0 / 10 / 20 (ablation) | LASQ · LOLv1 | PSNR: 17.82 / 18.04 / 17.85 / 17.87 | No clear effect |
| vary lambda_g across 0.001 / 0.005 / 0.01 / 0.1 (ablation) | LASQ · LOLv1 | PSNR: 17.76 / 18.22 / 18.16 / 17.88 | No clear effect |
Surprising findings
- LASQ++ (with unpaired normal-light references) slightly degrades no-reference metrics on several datasets relative to LASQ, which the authors attribute to overfitting.
Limitations
- The power-law/MCMC formulation is heuristic and the paper gives few formal derivations; key formulas (e.g., luminance scalar) are deferred to an unclear appendix.
- Inference FLOPs (219.75 G) and memory (6.5 GB) are high compared with non-diffusion baselines.
- Several details are inconsistent across the text (e.g., A800 vs A100 hardware, LSGAN vs BCE discriminator, LSAQ vs LASQ naming).
Experimental setup
Datasets: LOLv1 LSRW MEF
Benchmarks: LOLv1 LSRW DICM NPE VV LIME MEF
Metrics: PSNR SSIM LPIPS NIQE PI
Baselines: RetinexNet KinD++ LCDPNet URetinexNet SMG PyDiff Zero-DCE EnlightenGAN SCI PairLIE SCL-LLE LightenDiffusion
Builds on: denoising diffusion implicit models Zero-DCE LightenDiffusion KinD
Closest papers by meaning
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.88)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.88)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.87)
- Efficient Diffusion as Low Light Enhancer (CVPR 2025 · similarity 0.87)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.87)
- Luminance-GS: Adapting 3D Gaussian Splatting to Challenging Lighting Conditions with View-Adaptive Curve Adjustment (CVPR 2025 · similarity 0.86)
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.