<p>In low-light conditions, image quality deteriorates, posing significant challenges for advanced computer vision tasks. The three main goals of Low-Light Image Enhancement (LLIE) are to enhance visual quality: restoring natural color and brightness, maintaining textures and edges, and minimizing noise and artifacts. Convolutional Neural Networks (CNNs) and Self-Attention (SA) processes, which have demonstrated excellent performance in low-level vision tasks, have been the focus of recent research on deep learning-based methodologies. CNNs excel at capturing local patterns with translational equivariance, while SA mechanisms are effective at modeling distant dependencies. However, both approaches have limitations—CNNs suffer from a restricted receptive field and limited feature diversity, while SA mechanisms struggle with local feature associations. To overcome these challenges, we propose a Multi-Level Integration and Disintegration Network (MIDNet) for LLIE. It leverages the strengths of both CNNs and SA by introducing an uneven dual-path architecture that facilitates mutual feature representation and progressive enhancement. This design enables effective decomposition and association of normal light and low-light features. Comprehensive assessments show that MIDNet outperforms state-of-the-art methods in LLIE across several benchmarks. The repository is available on GitHub at <a href="https://github.com/SATHISHMOTHE/MIDNet">https://github.com/SATHISHMOTHE/MIDNet</a></p>