<p>Low-light image enhancement aims to improve the quality of images captured under low-light conditions. In previous research, traditional methods and deep learning methods based on Retinex theory have achieved commendable outcomes. However, they often neglect the degradation in low-light images and lack extraction of illumination features; standard attention mechanisms and feed-forward networks also suffer from bottlenecks in computational efficiency and redundant feature removal. To address these issues, we propose <b>LU</b>minance-guided Retine<b>X</b>-based Transformer (<b>LUXFormer</b>), which efficiently removes image degradation while achieving natural and realistic enhancement results. Firstly, we design a spatial-frequency feature enhancer to extract and fuse illumination features, achieving high-quality illumination modeling. Secondly, we propose an illumination-guided sparse attention to address the computational complexity bottleneck of standard self-attention mechanisms through a Top-k selection strategy. Finally, we introduce a refined feed-forward network to further enhance the model’s ability to remove redundant information and capture key features. Extensive experiments on multiple datasets demonstrate that our model outperforms existing mainstream methods. Specifically, our method surpasses the latest state-of-the-art method on the LOLv1 dataset by 2.20dB in PSNR, 0.008 in SSIM, and 0.007 in LPIPS. Furthermore, results in the low-light object detection show that our method significantly improves the performance of downstream tasks, outperforming existing methods on the ExDark dataset with a mean Average Precision (mAP) of 77.26%, demonstrating its superiority and competitiveness.</p>