<p>Low-light image capture often suffers from low brightness, amplified noise, and detail loss, degrading visual quality and downstream tasks. To address the high computational cost and complexity of existing methods, we propose SARA-Net, a lightweight and adaptive low-light enhancement framework based on the SCI model that improves image quality without relying on scene priors. We introduce a Multi-Scale Channel Attention Enhancement module for better feature integration and color-structure consistency, enabling the model to handle complex illumination. Additionally, a novel Channel-Coupled Adaptive Normalization strategy models cross-channel statistics via learnable affine parameters and coupling matrices, enhancing color fidelity and structural representation with minimal overhead. To tackle brightness bias, edge blurring, and color distortion, we refine the smoothness loss and add luminance and color consistency constraints, improving brightness recovery, structure, and color accuracy. Experiments on the LSRW dataset show SARA-Net achieves over 22% PSNR improvement over SCI and delivers competitive results on multiple benchmarks. It also demonstrates strong generalization in downstream tasks like low-light face detection, proving its practical value.</p>