Improving Universal Style Transfer using Sub-style Decomposition
Paraskevas Pegios, Nikolaos Passalis, Anastasios Tefas · 2020
Many universal Neural Style Transfer (NST) methods have been recently proposed. These methods are capable of transferring arbitrary styles in a style-agnostic manner by employing the appropriate feature transformations. Despite their apparent effectiveness when used to transfer a single style from relatively simple images, they are usually not capable of effectively handling more complex styles, producing artifacts, as well as reducing the quality of the synthesized textures in the stylized image. In this paper, we propose a novel universal style transfer method that separately models each sub-style that exists in a given style image, overcoming these limitations to a great extent. In this way, the proposed method is capable of detecting and semantically matching the sub-styles that exist both in the content and style images, improving the style transfer quality and reducing artifacts over state-of-the-art NST approaches, as demonstrated in this paper using both qualitative and quantitative experiments.