DuCol: Text-Tag Adaptive Colorization of Dual-Character Line Art

Abstract

Automatic colorization techniques often struggle with dual-character line art, particularly in areas such as color coordination between characters, handling complex scenes and interactions, and meeting personalized colorization requirements. To address these challenges, we introduce DuCol, a novel framework specifically designed for the colorization of dual-character line art. DuCol leverages text-based inputs to accommodate personalized color preferences and integrates a Text-Labeled Adaptive Colorization (TAC) Module to ensure global color assignment, effectively harmonizing colors between characters. Furthermore, the model utilizes detailed segmentation information from a skeleton graph to enable precise boundary detection, resolving interactions between characters and preventing color bleeding or ambiguity. Extensive experiments on a large-scale illustration dataset demonstrate that DuCol’s superiority in dual-character line art colorization, establishing it as a leading solution in this domain.