From Images to Video: Temporal Stabilization for Low-Light Image Enhancement Methods

Low-light image enhancement (LLIE) methods are designed and trained on still images, yet video applications require temporal consistency that these methods cannot provide on their own. A common workaround is to apply an LLIE method independently to each video frame, but this introduces flickering and inconsistent brightness or color across the sequence, since frame-to-frame relationships are ignored. We propose a recurrent neural network that converts per-frame LLIE outputs into a temporally consistent video, effectively extending single-frame LLIE methods to the video domain without retraining or modifying them. Our network takes as input two contiguous low-light frames, the current frame enhanced by the chosen LLIE method, and the previously stabilized frame, and produces a stabilized output that preserves the underlying enhancement. We align frames through an explicit optical-flow-based module to prevent motion-induced ghosting, and we train the network with a forward–backward temporal consistency loss that avoids the progressive brightness degradation seen when only forward consistency is enforced, together with a content-preserving loss that maintains fidelity to the per-frame enhanced output. We evaluate our method on two low-light video datasets, each processed by eight different LLIE methods, and compare its performance against existing temporal consistency approaches. Results show that our method achieves the best trade-off between temporal consistency and fidelity to the per-frame enhanced video, generalizing effectively even to LLIE methods that were not used during training. Our model also has lower memory complexity than prior approaches, with comparable computational cost. These results confirm that existing LLIE methods can be effectively adapted for video use through our temporal stabilization framework.