Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement
Authors: Derong Kong, Zhixiong Yang, Shengxi Li, Shuaifeng Zhi, Li Liu, Zhen Liu, Jingyuan Xia
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate that LASQ, when integrated with a vanilla diffusion model, achieves state-of-the-art performance on non-reference datasets while attaining comparable performance to reference-dependent methods on normal-light benchmark datasets. Furthermore, LASQ exhibits versatile compatibility: it seamlessly adapts to scenarios where normal-light references are available, delivering superior domain-specific enhancement alongside unparalleled cross-dataset generalization capabilities. |
| Researcher Affiliation | Academia | 1College of Electronic Science and Technology, National University of Defense Technology 2College of Electronic and Information Engineering, Beihang University |
| Pseudocode | Yes | Algorithm 1 The LASQ pipeline. |
| Open Source Code | Yes | The code is available at: https://github.com/XYLGroup/LASQ. |
| Open Datasets | Yes | We validated our approach using both paired and unpaired low-light benchmarks. For the paired evaluation, we used the LOLv1 (22) and LSRW (32) test sets each comprising matched low- and normal-illumination image pairs and reported restoration fidelity via PSNR and SSIM (33), alongside the full-reference perceptual score LPIPS (34). To assess performance in the absence of ground-truth references, we then tested on the unpaired LIME (35), DICM (36), NPE (37), and VV (38) collections, measuring perceptual quality with the no-reference NIQE (39) and PI (40) metrics. |
| Dataset Splits | No | The paper mentions "LOLv1 and LSRW test sets", "unpaired LIME, DICM, NPE, and VV collections", and "manually curated training splits augmented via random cropping" for LOLv1, LSWR, and MEF. However, specific percentages or sample counts for these splits are not explicitly provided within the paper's text or appendices, making it difficult to reproduce the exact data partitioning without further external information. |
| Hardware Specification | Yes | All experiments were conducted on a high-performance computing node equipped with four NVIDIA A100 80GB GPUs interconnected via NVLink. The system specifications are as follows: OS: Ubuntu 22.04 LTS with Linux 5.15 kernel CPU: Dual AMD EPYC 7763 64-Core @ 2.45GHz (128 cores/256 threads) GPU Interconnect: NVLink 3.0 (600GB/s bisectional bandwidth) Memory: 1TB DDR4 ECC @ 3200MHz Storage: 16TB NVMe SSD RAID (3.5GB/s sustained read) Accelerators: 4 NVIDIA A100 80GB (FP32: 19.5 TFLOPS, FP16: 312 TFLOPS) |
| Software Dependencies | Yes | All experiments are carried out on a cluster of four NVIDIA A800 GPUs under Python 3.9 and Py Torch 2.0, with a fixed batch size of 16. |
| Experiment Setup | Yes | All experiments are carried out on a cluster of four NVIDIA A800 GPUs under Python 3.9 and Py Torch 2.0, with a fixed batch size of 16. We employ the Adam optimizer (30), setting the denoising diffusion process learning rate to 2 10 5, while using a sampling ratio k = 3. The hyperparameters λd, λg and λGAN (if activated) are set to 0.9, 0.005 and 0.7 respectively. Noise estimation during diffusion training is performed using the U-Net (31) architecture with T = 1000 time steps. In the luminance adaptation framework, the power-law adjustment parameters α, η, and δ governing local contrast enhancement are initialized to 2, 0.1, and 0.01 respectively, while the MCMC sampling process employs an adaptive step size λ = 0.2 to balance exploration-exploitation dynamics across hierarchy levels. The temporal mapping function ψ(t) synchronizes N = 100 hierarchical guidance levels with the T-step diffusion through linear interpolation. The truncated Gaussian distribution for LAO sampling is bounded by γmin and γmax derived dynamically from image statistics, ensuring physically plausible luminance adjustments. |