SwiftLight: One-Step Low-Light Image Enhancement via Trajectory-Aware Distillation and Global-Local Consistency

Diffusion models have shown promising performance in low-light image enhancement. However, iterative sampling incurs substantial computation, especially for high-resolution images. Aggressive acceleration can degrade restoration quality, while independent patch processing may produce inconsistent illumination. To address these challenges, we propose SwiftLight, a trajectory-aware distillation framework for one-step low-light image enhancement. Specifically, we introduce a three-stage training strategy. In the first stage, we distill a multi-step teacher into a few-step model through segment-wise consistency learning. In the second stage, the few-step model supervises a one-step student that starts from the low-light input. Teacher trajectory cues and input darkness guide spatial supervision in both stages to retain important restoration corrections. In the third stage, adversarial refinement is incorporated into one-step training to improve texture realism. Furthermore, we design a lightweight consistency head to calibrate independently enhanced patches using a full-image prediction. Adaptive low-frequency correction and center-weighted fusion align illumination and color while retaining local details. Experiments demonstrate strong perceptual quality with one-step patch inference. Code will be made publicly available.