Closing the Domain Gap in Manga Colorization via Aligned Paired Dataset

Abstract

This paper addresses the challenge of artwork colorizationby proposing a benchmark for manga colorization usingreal black-and-white and colorized image pairs. Colorimages are widely recognized for their ability to captureattention and improve memory retention, yet the manualprocess of colorization is labor-intensive. Deep learningmethods for supervised image-to-image translation offer apromising solution, relying on aligned pairs of black-andwhiteand color images for training. However, these pairsare often generated synthetically, introducing a domain gapthat limits model performance. To address this, we explorethe use of real data, proposing a method for creating suchdatasets. Our benchmarks reveal that models trained onreal data significantly outperform those trained on syntheticpairs. Furthermore, we present a pipeline for text removaland panel segmentation, streamlining the comic colorizationprocess. These contributions aim to enhance the generalizationand applicability of deep learning models forartwork colorization.