${\mathrm{T}}^3\mathrm{LIE}$: Retinex-inner Test-Time Training for Low-Light Image Enhancement

Low-light image enhancement remains challenging due to diverse illumination conditions and complex degradation patterns. Existing methods typically rely on offline-trained models with fixed parameters, limiting their ability to adapt to image-specific illumination during inference. To address this issue, we propose ${\mathrm{T}}^3\mathrm{LIE}$, a Retinex-inner Test-Time Training framework that performs illumination-aware adaptation by updating fast weights with self-supervised guidance. Specifically, ${\mathrm{T}}^3\mathrm{LIE}$ introduces a Retinex-based residual compensation formulation that decomposes enhancement into illumination compensation and illumination-dependent reflectance refinement. Based on this formulation, illumination representations are progressively adapted and further used to guide the inner-loop optimization of reflectance fast weights. By exploiting image-specific illumination priors, ${\mathrm{T}}^3\mathrm{LIE}$ enables adaptive enhancement while preserving scene details. Extensive experiments on multiple low-light benchmarks demonstrate state-of-the-art enhancement performance and improved generalization across diverse illumination conditions.