GAN Prior Based Null-Space Learning for Consistent Super-resolution

Consistency and realness have always been the two critical issues of image super-resolution. While has dramatically improved with use GAN prior, state-of-the-art methods still suffer inconsistencies in local structures colors (e.g., tooth eyes). In this paper, we show that these can be analytically eliminated by learning only null-space component while fixing range-space part. Further, design a pooling-based decomposition (PD), universal range-null space for super-resolution tasks, which is concise, fast, parameter-free. PD easily applied to Prior based SR eliminate their inconsistencies, neither compromise nor bring extra parameters or computational costs. Besides, our ablation studies reveal replace pixel-wise losses training achieve better generalization performance when facing unseen downsamplings even real-world degradation. Experiments refreshes speeds up convergence 2~10 times.