Rethinking the Adaptiveness of Conditional Flow A Novel Distribution Adaptation Strategy for Image Super-Resolution

The conditional normalized flow model can alleviate the ill-posed nature of super-resolution problems by learning the conditional distribution of the output for a given low-resolution input, which allows multiple predictions for a given low-resolution image. However, the model may generate confusing artifacts in the high-frequency region of the image. While giving a smaller reduction in temperature would solve this problem, it would also smooth out the model's output, which would lead to a decrease in perceptual quality. In this paper, we propose a novel conditional normalizing flow-based distribution adaptation strategy for image Super-Resolution. More specifically, we demonstrate that the part of the latent variable that differs significantly between the latent variables of HR and the Bicubic of LR, contains mainly high-frequency information of the image. We adopt a Bicubic image of LR and a continuous threshold function to evaluate different temperatures in different latent variables. In this way, We can further alleviate generation of confusing artifacts without reducing perceptual quality. Extensive experiments show that our model can outperform state-of-the-art methods and generate more visually favorable results.