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

Michael Hobley

Abstract:

Abstract. Objects are defined by their shape and visual style. Previous
work into image manipulation has generally altered the stylistic appear-
ance of a whole image, while maintaining the image content and object
shapes. In this paper we transfer both the shape and style of chosen ob-
jects 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.