MFLab: decoupled luminance-chrominance modeling with uncertainty-aware flow refinement for low-light image enhancement

Abstract

Low-light image enhancement aims to improve image quality under insufficient illumination while recovering underlying structural and texture information. Existing approaches, ranging from conventional image processing techniques to deep generative models, have shown promising performance in brightness enhancement and detail restoration. However, two challenges remain. Diffusion-based methods typically incur considerable computational cost due to their iterative sampling process, while models operating directly in the sRGB domain often suffer from color distortion and structural hallucinations in severely degraded real-world scenes. To address these issues, we propose MFLab, an uncertainty-aware low-light image enhancement framework. The proposed method performs enhancement in the CIELab color space, where luminance restoration and chrominance reconstruction are explicitly decoupled, reducing the coupling between brightness and color information. A Mamba-based backbone provides efficient long-range feature modeling, while Flow Matching is introduced for generative detail refinement. Furthermore, an uncertainty-aware strategy is used to identify spatially unreliable regions and guide constrained refinement. Extensive experiments on six paired benchmarks and additional real-world unpaired datasets demonstrate competitive restoration performance under diverse low-light conditions. On the LOL-v1 dataset, MFLab achieves a PSNR of 24.89 dB, an SSIM of 0.873, and an LPIPS score of 0.081, indicating a favorable balance between perceptual quality and structural fidelity.