Generation With Nuanced Changes: Continuous Image-to-Image Translation With Adversarial Preferences

Most previous methods for continuous image-to-image translation resorted to binary attributes with restrictive description ability and thus cannot achieve satisfactory performance. Some works proposed to use fine-grained semantic information, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">relative attributes (RAs), preferences over pairs of images on the strength of a specified attribute</i>. However, they still failed to reconcile both goals for smooth translation and for high-quality generation simultaneously. In this work, we propose a new model continuous translation via adversarial preferences (CTAP) to coordinate these two goals for high-quality continuous translation based on RAs. In CTAP, we simultaneously train two modules: a generator that translates an input image to the desired image with smooth nuanced changes w.r.t. the interested attributes; and a ranker that executes adversarial preferences consisting of the input image and the desired image. Particularly, adversarial preferences involve an adversarial ranking process: 1) the ranker thinks no difference between the desired image and the input image in terms of the interested attributes; 2) the generator fools the ranker to believe the attributes of its output image changes as expect compared with the input image. RAs over pairs of real images are introduced to guide the ranker to rank image pairs regarding the interested attributes only. With an effective ranker, the generator would “win” the adversarial game by producing high-quality images that present smooth changes. The experiments on two face datasets and one shoe dataset demonstrate that our CTAP achieves state-of-art results in generating high-fidelity images which exhibit smooth changes over the interested attributes.