Research map / Image generation / Image restoration
Low-Light Image Enhancement via Generative Perceptual 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 GPP-LLIE, a transformer-based diffusion model for low-light image enhancement guided by global and local perceptual priors extracted from LLaVA via a sigmoid-quantized good/poor token-probability score.
Problem
Low-light image enhancement struggles to generalize to diverse real-world illumination conditions and to produce realistic, visually appealing results; existing diffusion LLIE methods use U-Net backbones and lack perceptual guidance.
Approach
Extract global and local perceptual priors from a VLM (LLaVA fine-tuned on Q-Instruct) by prompting it to rate contrast, visibility and sharpness, converting good/poor token logits into sigmoid scores; use these priors to modulate layer norm (GPP-LN) and guide channel attention (LPP-Attn) in a DiT-style diffusion backbone.
Key results
- On LOL, FID 36.73 vs. best prior 48.11 (DiffLL), about 23.6% lower; LPIPS 0.081 vs. PyDiff 0.099.
- On LOL-v2-synthetic, FID 9.74 vs. 13.26 (Reti-Diff), PSNR 30.17 vs. 29.27 (PyDiff).
- On real-world datasets (MEF, LIME, DICM, NPE), best mean NIQE 3.67 vs. 3.92 for the second-best method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the local perceptual prior using LPP-Attn (ablation) | GPP-LLIE · LOL | FID: 49.83 -> 36.73 | Helped |
| add the local perceptual prior using LPP-Attn (ablation) | GPP-LLIE · LOL | LPIPS: 0.103 -> 0.081 | Helped |
| replace spatial feature transform with LPP-Attn (ablation) | GPP-LLIE · LOL | FID: 47.18 -> 36.73 | Helped |
| replace spatial feature transform with LPP-Attn (ablation) | GPP-LLIE · LOL | LPIPS: 0.100 -> 0.081 | Helped |
| add the global perceptual prior and GPP-LN (ablation) | Variant 1 · LOL | FID: 61.36 -> 49.83 | Helped |
| add the global perceptual prior and GPP-LN (ablation) | Variant 1 · LOL | LPIPS: 0.113 -> 0.103 | Helped |
| use GPP-LN instead of directly adding the global score to the noised latent feature (ablation) | Variant 1 · LOL | FID: 58.36 -> 49.83 | Helped |
| use GPP-LN instead of directly adding the global score to the noised latent feature (ablation) | Variant 1 · LOL | LPIPS: 0.111 -> 0.103 | Helped |
Limitations
- Real-world evaluation relies only on no-reference NIQE without ground truth.
- Reliance on VLM token probabilities and patchwise assessment adds extraction cost not quantified.
Experimental setup
Datasets: LOL LOL-v2-real LOL-v2-synthetic Q-Instruct
Benchmarks: LOL LOL-v2-real LOL-v2-synthetic MEF LIME DICM NPE
Metrics: FID LPIPS DISTS PSNR NIQE
Baselines: KinD MIRNet DRBN SNR-Net URetinex-Net LLFlow RQLLIE Retinexformer CUE CLEDiff LL-SKF Reti-Diff
Builds on: DiT LLaVA Q-Instruct StableSR Diff-Retinex
Closest papers by meaning
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.87)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.84)
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.83)
- UniRestore: Unified Perceptual and Task-Oriented Image Restoration Model Using Diffusion Prior (CVPR 2025 · similarity 0.83)
- Fourier Priors-Guided Diffusion for Zero-Shot Joint Low-Light Enhancement and Deblurring (CVPR 2024 · similarity 0.83)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.82)
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.