Anime Style Space Exploration Using Metric Learning and Generative Adversarial Networks

  • 2018-05-21 11:43:12
  • Sitao Xiang, Hao Li
  • 22

Abstract

Deep learning-based style transfer between images has recently become apopular area of research. A common way of encoding "style" is through a featurerepresentation based on the Gram matrix of features extracted by somepre-trained neural network or some other form of feature statistics. Such adefinition is based on an arbitrary human decision and may not best capturewhat a style really is. In trying to gain a better understanding of "style", wepropose a metric learning-based method to explicitly encode the style of anartwork. In particular, our definition of style captures the differencesbetween artists, as shown by classification performances, and such that thestyle representation can be interpreted, manipulated and visualized throughstyle-conditioned image generation through a Generative Adversarial Network. Weemploy this method to explore the style space of anime portrait illustrations.

 

Quick Read (beta)

loading the full paper ...