Say Yes to the Dress: Shape and Style Transfer Using Conditional GANs

Abstract

Objects are defined by their shape and visual style. Previous work into image manipulation has generally altered the stylistic appearance of a whole image, while maintaining the image content and object shapes. In this paper we transfer both the shape and style of chosen objects between images, leaving the remaining areas unaltered. To tackle this problem, we propose a two stage method, where each stage contains a generative adversarial network, that will alter the shape and style of objects in a subject image to reflect a donor image. We demonstrate the effectiveness of our method by transferring clothing between images.