Research map / Image generation / Image restoration
Zero-Reference Low-Light Enhancement via Physical Quadruple Priors
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 a zero-reference low-light enhancement framework trained only on normal-light COCO images, mapping a physical quadruple illumination-invariant prior back to images with a ControlNet-style Stable Diffusion model plus a bypass decoder and a distilled lightweight variant.
Problem
Low-light image enhancement typically needs paired or unpaired low-light training data, and zero-reference methods are sensitive to training data distribution and illumination-specific hyperparameters, so they generalize poorly to unseen scenes.
Approach
An illumination-invariant "physical quadruple prior" (H, C, W, O) derived from Kubelka-Munk light transfer theory acts as an intermediary; a prior-to-image diffusion model is trained only on normal-light images to reconstruct images from the prior, and at test time low-light images are mapped to the prior and restored.
Key results
- On LOL, the method achieves 20.31 PSNR and 0.808 SSIM, outperforming most unsupervised and zero-reference baselines (e.g., SCI-LOL 16.97 PSNR).
- On MIT-Adobe FiveK, it reaches 18.51 PSNR, ahead of SCI trained on MIT (16.29), indicating robustness across datasets without low-light training data.
- The lightweight distilled version reduces parameters from 1.3G to 327.36k and runs 500x faster, processing 1024x1024 images in 0.03 s on a Tesla M40.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add H to the physical quadruple prior (ablation) | Ours · LOL | PSNR: 17.60 -> 20.25 | Helped |
| add C to the physical quadruple prior (ablation) | Ours · LOL | PSNR: 17.60 -> 20.25 | Helped |
| add W to the physical quadruple prior (ablation) | Ours · LOL | PSNR: 17.77 -> 20.25 | Helped |
| add O to the physical quadruple prior (ablation) | Ours · LOL | PSNR: 18.63 -> 20.25 | Helped |
| replace HS channels in HSV with the physical quadruple prior (ablation) | Ours · LOL | PSNR: 18.04 -> 20.25 | Helped |
| replace CIConv with the physical quadruple prior (ablation) | Ours · LOL | PSNR: 17.02 -> 20.25 | Helped |
| replace PairLIE reflectance with the physical quadruple prior (ablation) | Ours · LOL | PSNR: 20.16 -> 20.25 | Helped |
| use the bypass decoder instead of the SD decoder (ablation) | Ours · LOL | PSNR: 19.26 -> 20.25 | Helped |
| use the bypass decoder instead of the Consistency Decoder (ablation) | Ours · LOL | PSNR: 19.35 -> 20.25 | Helped |
| distill the full framework into a lightweight model (ablation) | Ours · LOL | PSNR: 20.31 -> 20.45 | Helped |
Surprising findings
- Supervised Retinexformer trained on LOL drops to 13.87 PSNR on MIT, showing supervised overfitting to training sets.
Limitations
- Full model is large (about 1.3G parameters) and requires multi-step diffusion inference.
- Some illumination-relevant information is discarded by the prior, requiring generative completion that can introduce randomness.
Experimental setup
Datasets: COCO-2017 LOL MIT-Adobe FiveK LIME NPE MEF DICM VV
Benchmarks: LOL MIT-Adobe FiveK LIME NPE MEF DICM VV
Metrics: PSNR SSIM LPIPS LOE BRISQUE NL
Baselines: Zero-DCE SCI RUAS PairLIE CLIP-LIT NeRCo EnlightenGAN Retinexformer
Builds on: Stable Diffusion ControlNet Kubelka-Munk theory CIConv Restormer
Closest papers by meaning
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.89)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.88)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.88)
- ZERO-IG: Zero-Shot Illumination-Guided Joint Denoising and Adaptive Enhancement for Low-Light Images (CVPR 2024 · similarity 0.88)
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.88)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 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.