FaithSR

Abstract

Despite active researches on image super-resolution, reconstructing high-resolution images that preserve the features of an image from low-resolution images remains as a challenging task. While recent super-resolution networks focus on creation of high-quality images, producing images that are faithful to features of original input image is essential. In this paper, we present FaithSR, a super-resolution network that focuses on formulating faithful latent vectors which leads to both high-quality and high-fidelity images. We enlarge the effect of precise attributes by taking advantage of our novel encoder loss, latent vector pool, Bayesian approach, and latent operation. We show that FaithSR outperforms previous works by evaluation on both quantitative and qualitative aspects. Our network can be applied to variety of other computer-vision tasks, e.g., inversion, face frontalization and conditional image synthesis.