Research map / Image generation / Image restoration

DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors

AAAI 2025 · Image generation › Image restoration · approach: Low-light image enhancement · 15 citations (OpenAlex)

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 DPLUT, an unsupervised low-light image enhancement framework combining two 3D lookup tables (LLUT predicts pixel-wise curve parameters under zero-reference losses; NLUT is supervised by a pretrained diffusion model's outputs for noise suppression), achieving real-time 4K enhancement.

Problem

Low-light image enhancement methods are either efficient but weak, or high-quality but heavy and reliant on paired data, limiting practical real-time use on high-resolution images.

Approach

Replace RGB-to-RGB LUTs with an image-adaptive LUT that maps RGB to curve parameters (small 9-point table) trained with unsupervised losses, then use a second LUT (NLUT) trained with pseudo-references from a pretrained diffusion model to suppress amplified noise; the diffusion model is used only during NLUT training.

Key results

• DPLUT achieves 20.66 PSNR / 0.74 SSIM / 0.22 LPIPS on LOL, the best among unsupervised methods, and 21.27 PSNR on SICE and 18.91 PSNR on LSRW.
• With 0.078M parameters, DPLUT runs in 19.8 ms at 4K (3840x2160) on a Titan RTX, while Retinexformer and EnlightenGAN run out of memory at 4K.
• Ablation: replacing the diffusion prior with a conventional denoiser drops LOL PSNR from 20.66 to 20.11; removing NLUT drops it to 19.89.

Measured effects reported in the paper

Surprising findings

• Removing the exposure loss collapses LOL PSNR to 8.31 dB, showing the exposure term is essential.
• Increasing the number of DDIM denoising iterations beyond 4 lowers PSNR/SSIM even though LPIPS keeps improving slightly.

Limitations

• LLUT and NLUT are trained separately rather than jointly, as noted by authors as future work.
• Quality-oriented supervised baselines such as DiffLL and Retinexformer still achieve higher PSNR on some datasets (e.g., SICE and LSRW), so DPLUT only matches them partially.
• Diffusion model is used only at training time but still adds cost to NLUT training (300 iterations, 100 DDIM steps).

Experimental setup

Datasets: LOL SICE-Part2 LSRW ImageNet

Benchmarks: LOL SICE LSRW

Metrics: PSNR SSIM LPIPS runtime (ms) params (M)

Baselines: SCI ZeroDCE ZeroDCE++ RUAS PairLIE NeRCo CLIP-LIT EnlightenGAN Retinexformer DiffLL CUE 3DLUT

Builds on: 3D-LUT Zero-DCE DDIM denoising diffusion probabilistic models

Closest papers by meaning

• Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.90)
• AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.88)
• Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.87)
• Multi-Domain Multi-Scale Diffusion Model for Low-Light Image Enhancement (AAAI 2024 · similarity 0.87)
• Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.87)
• DarkIR: Robust Low-Light Image Restoration (CVPR 2025 · similarity 0.87)