Research map / Image generation / Image restoration
Zero-Shot Low-Light Image Enhancement via Latent 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 a zero-shot low-light image enhancement framework that uses a pre-trained Stable Diffusion latent diffusion model with an exposure-aware bright-channel degradation model and latent regularization with adaptive guidance, requiring no training data.
Problem
Low-light image enhancement typically requires large paired or unpaired training datasets, limiting practical deployment and generalization.
Approach
Reformulates LLIE as zero-shot posterior sampling with a pre-trained latent diffusion model, using a bright-channel-based exposure-adaptive degradation model, a latent regularization term (encoder-decoder fixed-point consistency with low-pass filtering), a channel consistency loss, and an adaptive guidance scale computed from latent step differences.
Key results
- On LOL-v1, the zero-shot method reaches 19.82 PSNR / 0.841 SSIM / 0.242 LPIPS, significantly above other zero-shot methods (best prior ZeroIG 17.63 PSNR).
- On LOL-v2 (real), the method reaches 19.95 PSNR, outperforming zero-shot baselines (GDP 14.36, ZeroIG 15.66) and approaching QuadPrior (20.24).
- Ablation shows the full regularization set (Lmeas, Linv, Llatent, low-pass G) gives the best results, e.g., PSNR 19.15 vs 12.47 for Lmeas alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the exposure-aware bright-channel degradation model instead of constant illumination (ablation) | Ours · LOL-v1 qualitative comparison | visual quality | Helped |
| Use adaptive rather than constant guidance scale (st = 0.7) (ablation) | Ours · LOL-v1 qualitative comparison | visual quality | Helped |
| Add Lmeas to the variant with Linv and Llatent (ablation) | Ours · LOL-v1, 256×256 | PSNR: 19.18 -> 19.15 | No clear effect |
| Add Linv to the variant with Lmeas and Llatent (ablation) | Ours · LOL-v1, 256×256 | PSNR: 18.88 -> 19.15 | Helped |
| Add Llatent to the variant with Lmeas and Linv (ablation) | Ours · LOL-v1, 256×256 | PSNR: 10.57 -> 19.15 | Helped |
| Add G to Llatent (ablation) | Ours · LOL-v1, 256×256 | PSNR: 18.91 -> 19.15 | No clear effect |
| Add G to Llatent (ablation) | Ours · LOL-v1, 256×256 | LPIPS: 0.411 -> 0.452 | Hurt |
| Use the proposed method instead of the strongest zero-shot baseline | Ours · LOL-v1 | PSNR: 17.63 -> 19.82 | Helped |
| Use the proposed method instead of the strongest zero-shot baseline | Ours · LOL-v2 | PSNR: 15.66 -> 19.95 | Helped |
Surprising findings
- Pure measurement consistency (Lmeas alone) yields poor results, with low-pass filtered latent regularization being essential for preserving naturalness.
Limitations
- Does not surpass the best supervised methods on SSIM and LPIPS.
- Sampling with 1000 DDIM steps and per-image optimization is computationally heavy; ablations are run at 256x256 for speed.
Experimental setup
Datasets: LOL-v1 LOL-v2
Benchmarks: LOL-v1 LOL-v2
Metrics: PSNR SSIM LPIPS
Baselines: KinD DRBN KinD++ RetinexNet URetinexNet SNR-Aware DiffLL GSAD EnlightenGAN PairLIE NeRCo CLIP-LIT
Builds on: Stable Diffusion Chung et al. 2023 (diffusion posterior sampling) bright channel prior DDIM
Closest papers by meaning
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.91)
- DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors (AAAI 2025 · similarity 0.90)
- Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement (AAAI 2024 · similarity 0.89)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.89)
- Efficient Diffusion as Low Light Enhancer (CVPR 2025 · similarity 0.89)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (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.