Illumination-Aware Context Modeling for Low-Light Image Enhancement in Complex Real Scenes

Low-light image enhancement (LLIE) in real scenes is challenging due to complex natural illumination, where multiple light sources, shadows, and view-dependent reflections jointly induce spatially heterogeneous and cross-region coupled appearance changes. Many existing learning-based methods mainly rely on local cues or generic long-range dependencies, which may entangle illumination effects with intrinsic reflectance and consequently yield color shifts, detail loss, or artifacts on specular and high-contrast regions. In this paper, we propose Illumination-Aware Context Modeling (IACM), a framework that explicitly captures illumination priors and leverages them as global guidance throughout reconstruction. Specifically, an Illumination Estimation module predicts an illumination prior, which is embedded into a Transformer via an Illumination-Aware Position Encoding (IAPE). By coupling illumination priors with positional representations, the Transformer is encouraged to model illumination consistent global dependencies beyond pure spatial co-occurrence. To bridge global illumination context and fine-grained textures, we further introduce a multi-scale fusion mechanism based on modulation matrices, where Transformer features dynamically predict affine parameters to modulate CNN encoder features, enabling adaptive integration instead of static concatenation or addition. Extensive experiments on representative sRGB LLIE benchmarks demonstrate that the proposed method achieves state-of-the-art performance and improves robustness in real-world low-light scenarios. Code is available at https://anonymous.4open.science/r/NICM-0xAEE499A7.