[3L4-OS-22b] 創作者と人工知能が創る創作の未来(2)

[3L4-OS-22b-04] Tone Pasting Using cGANs with Tone Feature Loss

〇Koki Tsubota¹, Kiyoharu Aizawa¹ (1. The University of Tokyo)

Keywords:Tone Pasting, conditional generative adversarial networks, Image Processing, Manga

Tone pasting is one of the processes of manga creation and there is a demand for automatic tone pasting. In this study, we tackle a task of automatic tone pasting of manga characters. Tone pasting is difficult because tones have characteristic patterns. It is hard to learn tone patterns for usual conditional generative adversarial networks (cGANs) which are combined with L₁ loss or perceptual loss. To train pasting tones in a tone pattern aware manner, we introduce tone feature loss to cGANs. Tone feature loss is the distance between tone features of target images and those of generated images. We performed experiments on two characters in Manga109 and showed our results are equal to or more visually appealing than those by a baseline method.