ARM-Net: Adaptive Radiance Modulation for Unsupervised Low-Light Image Enhancement

We introduce ARM-Net (Adaptive Radiance Modulation Network), a lightweight, self-supervised framework for low-light image enhancement that requires no paired training data. At the heart of ARM-Net is a recursive enhancement strategy called Recursive Intensity Modulation (RIM), where the network predicts a spatially-adaptive radiance adjustment tensor that is iteratively applied to the input image to progressively recover visibility and contrast. The network architecture is fully convolutional and preserves spatial resolution through a series of Spatially-Invariant Enhancement Blocks (SIEBs), enabling fine-grained illumination control without introducing artifacts. To guide training in the absence of reference images, we design a set of Self-Consistent Photometric Regularizers, which enforce exposure balance, edge-aware smoothness, color stability, and gradient compactness. Experimental results demonstrate that ARM-Net produces visually compelling enhancements and achieves strong quantitative performance on standard low-light datasets, all while maintaining real-time efficiency.