Comicolorization : Semi-automatic Manga Colorization

Given a monochrome manga and reference images as inputs, a colorization tool, "Comicolorization", generates a plausible color version. Although colorization research has a long history, no approaches have not focused on the colorization of an entire manga. We present three contributions. First, our system receives a whole manga page (not manga panels) as an input. Our method is the first system to colorize an entire manga page at once. Second, our tool generates a colorized manga using vivid colors and by maintaining the color coherency between colorized results. We extract a color feature from an inputted reference image. In addition, the color feature helps to inference by Convolutional Neural Networks. Our approach employs adversarial loss while training the colorization networks to encourage the effect of given color information. Finally, our tool allows users to edit specific regions using target colors by adjusting color features and adding color dots interactively.