SAR‐Flow: Spatially Adaptive Rectified Flow for Efficient Low‐Light Image Enhancement

ABSTRACT

Existing flow‐based models for low‐light image enhancement are fundamentally constrained by the isotropic Gaussian assumption, severely overlooking the inherent spatially non‐uniform SNR degradation characteristics in low‐light images. Furthermore, training via MLE often yields highly non‐linear probability transport trajectories, which significantly increased the computational overhead of ordinary differential equation solvers. To circumvent these limitations, this paper departs from the conventional standard Gaussian initialization and introduces a heterogeneous manifold initialization strategy. This strategy leverages the local SNR features of the image to construct a spatially inhomogeneous noise field, explicitly encoding it into the second‐order moment of the initial state. Concurrently, we reformulate the enhancement process as a velocity field‐driven ODE system. By exploiting the rectification properties of flow matching, we encourage the transport paths to become approximately linear in the latent space, enabling their approximation by piecewise linear numerical integration. Extensive experiments demonstrate that SAR‐Flow achieves a superior trade‐off between perceptual quality and fidelity across multiple benchmark datasets. Benefiting from the piecewise linear transport trajectories, SAR‐Flow accelerates training convergence by approximately 21% compared to previous flow‐based models, thereby facilitating highly efficient low‐light image reconstruction.