Video Super-Resolution

Video super-resolution (VSR) uses information across multiple frames to recover details that are missing or degraded, supporting clearer video on limited-bandwidth devices and in settings such as surveillance and medical imaging. Methods commonly align and fuse frames with convolutional, recurrent, or transformer networks; newer diffusion-based approaches generate realistic detail, while causal and streaming designs aim to reduce latency. Research spans standard benchmarks such as REDS4 and Vid4 as well as real-world and task-specific tests, with methods addressing motion blur, compression, arbitrary scales, and downstream recognition. A central tension is balancing perceptual realism and temporal stability against fidelity to the input, computational cost, and reliable evaluation beyond pixel-based scores such as PSNR.