Enhancing temporal consistency in video super-resolution using optical flow-guided deep networks

This paper presents an optical flow-guided deep learning framework for video super-resolution (VSR) that enhances spatial detail and temporal consistency. While recent VSR methods incorporate temporal propagation, achieving stable motion alignment under occlusion and rapid motion remains challenging. The proposed approach employs RAFT-based optical flow to align adjacent low-resolution frames with a reference frame. These motion-compensated features are fused using a confidence-weighted attention mechanism and refined through a residual-in-residual dense block (RRDB) network to produce high-resolution outputs. A composite loss combining pixel-wise, perceptual, and temporal terms jointly optimizes spatial fidelity and motion coherence. Experiments on Vid4 and REDS benchmarks demonstrate that the proposed model attains higher PSNR and SSIM with lower LPIPS compared to state-of-the-art methods, while maintaining reduced computational complexity. Qualitative results confirm its ability to preserve textures, suppress artefacts, and ensure smooth frame transitions, making it suitable for real-world video enhancement applications.