Towards Diverse Anime Face Generation: Active Label Completion and Style Feature Network

Abstract

It is interesting to use an anime face as personal virtual image to replace the traditional sequence code. To generate diverse anime faces, this paper proposes a style-gender based anime GAN (SGA-GAN), where the gender is directly conditioned to ensure the gender differentiation, and style features serve as a condition to guarantee the style diversity. To extract style features, we train a style feature network (SFN) as a multi-task classifier to simultaneously fulfill gender classification, style classification, and image quality estimation. To make full use of available data, partly labeled or unlabeled, during the SFN training, we propose a label completion method to actively complete the missing gender or style labels. The active label completion is essentially a weakly-supervised learning process through ensembling three distinct classifiers to improve the generalization capability. Experiments verify that the active label completion can improve the model accuracy and the style feature as a condition can make better the diversity of generated anime faces.