<p>The combination of Retinex theory and Mamba has provided innovative solutions for low-light image processing. However, these methods often demand substantial computational resources and extended training and inference times. In response, we propose EffRetMamba, which employs jump sampling to significantly reduce the sampling time of the Mamba module, thereby decreasing the model’s inference time. To prevent performance degradation, we introduce a frozen ResNet-18 based on Retinex decomposition to extract image illumination features at various scales, enhancing the model’s capacity to learn detailed illumination information. In addition, since jump connections in U-shaped architectures often encounter the issue of field mismatch, this paper presents MGAFusion, a novel multi-scale feature fusion module. MGAFusion integrates global spatial and channel attention to adaptively combine low-level features, which contain detailed information, with high-level features, which encompass global information, based on learned weights. Extensive experiments conducted on the LOL and LSRW datasets demonstrate that the proposed method achieves superior image enhancement while utilizing fewer parameters and requiring less inference time.</p>