Research map / Image generation / Image restoration
AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image 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 AGLLDiff, a training-free, unsupervised low-light enhancement method that guides a pre-trained ImageNet diffusion model with attribute losses (exposure, Fourier phase structure, Retinex reflectance color) and a dynamic guidance scheme.
Problem
Real-world low-light image enhancement lacks paired training data and accurate degradation models, so supervised and degradation-model-based diffusion methods generalize poorly.
Approach
Instead of modeling the degradation, guide diffusion sampling with differentiable losses on desired normal-light attributes (spatially variant exposure, Fourier phase structure, and Retinex reflectance color), with dynamic guidance scale and number of gradient steps derived from the intermediate image change magnitude.
Key results
- On LOLv1 AGLLDiff reaches 21.81 PSNR / 0.84 SSIM / 0.15 LPIPS, best among unsupervised methods (PairLIE 19.46 PSNR).
- On DICM, AGLLDiff gets NIQE 3.57 and BRISQUE 19.13, best among unsupervised methods across the five unpaired benchmarks.
- On SICE, AGLLDiff reaches 22.12 PSNR, outperforming several supervised methods such as DiffLL and PyDiff and exceeding the unsupervised baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AGLLDiff instead of the strongest listed unsupervised baseline on LOLv1 | Real-world low-light enhancement · LOLv1 | PSNR: 19.46 -> 21.81 | Helped |
| Add exposure attribute guidance (ablation) | AGLLDiff · LOLv1 | exposure control | Helped |
| Add structure attribute guidance (ablation) | AGLLDiff · LOLv1 | structure clarity | Helped |
| Add color attribute guidance (ablation) | AGLLDiff · LOLv1 | objective measures | Helped |
| Add attribute guidance rather than relying only on the pretrained diffusion model (ablation) | AGLLDiff · LOLv1 | low-light enhancement fidelity | Helped |
| Use dynamic guidance scale instead of a constant guidance scale (ablation) | AGLLDiff · LOLv1 | fidelity and clarity | Helped |
| Use dynamic gradient steps instead of constant gradient steps (ablation) | AGLLDiff · LOLv1 | artifact and noise removal | Helped |
| Increase the noise addition/removal iteration steps beyond 10 (ablation) | AGLLDiff · LOLv1 | sampling time | Hurt |
Surprising findings
- AGLLDiff's PSNR on LOLv1 (21.81) is below supervised DiffLL and PyDiff, and its LPIPS is worse than PyDiff on LOLv1 and Retinexformer on LOLv2-synthetic.
- Without attribute guidance, the pre-trained diffusion model degenerates into unconditional image generation as the noise level increases.
Limitations
- Sampling is slow; the authors state acceleration via advanced techniques is future work.
- Only three attributes (exposure, structure, color) are modeled; more underlying high-quality attributes remain unexplored.
Experimental setup
Datasets: LOLv1 LOLv2-synthetic SICE LIME NPE MEF DICM VV ImageNet
Benchmarks: LOLv1 LOLv2-synthetic SICE LIME NPE MEF DICM VV
Metrics: PSNR SSIM LPIPS NIQE BRISQUE PI
Baselines: SDD LECARM MBLLEN RetinexNet DSLR DRBN DiffLL PyDiff CUE Retinexformer EnlightenGAN RUAS
Builds on: classifier guidance DDPM Retinex theory SDEdit-style noise-then-denoise
Closest papers by meaning
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.90)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.89)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.88)
- DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors (AAAI 2025 · similarity 0.88)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.87)
- Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement (AAAI 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.