Abstract
Style transfer has been widely applied to give real-world images a newartistic look. However, given a stylized image, the attempts to use typicalstyle transfer methods for de-stylization or transferring it again into anotherstyle usually lead to artifacts or undesired results. We realize that theseissues are originated from the content inconsistency between the original imageand its stylized output. Therefore in this paper we advance to keep the contentinformation of the input image during process of style transfer by the power ofsteganography, with two approaches proposed: a two-stage model and anend-to-end model. We conduct extensive experiments to successfully verify thecapacity of our models, in which both of them are able to not only generatestylized images of quality comparable with the ones produced bystate-of-the-art style transfer methods, but also effectively eliminate theartifacts introduced in reconstructing original input from a stylized image aswell as performing multiple times of style transfer in series.