MambaLIE: Scene Light Intensity-Boosted Low-Light Image Enhancement with State Space Model

This paper introduces MambaLIE, a low-light image enhancement method built on a state space model. It models scene light intensity and gates that signal with the low-light input to guide enhancement. Its Locally Enhanced State Space Model (LESSM) combines an SSM branch for long-range dependencies with linear-time complexity and a local-enhancement branch for fine-grained representations. According to the abstract, MambaLIE outperforms CNN- and Transformer-based methods across four synthetic benchmarks and five real-world benchmarks in accuracy, speed, and model size, targeting deployment on resource-constrained devices.