The automated colorization of manga presents unique challenges due to its distinctive artistic style and complex visuals. While deep learning has shown promise in image colorization, existing approaches often struggle with consistency and artistic integrity of manga artwork. This paper presents a comparative analysis of two deep learning architectures for manga colorization: a modified U-Net with progressive dropout and a ResNet-based autoencoder with adaptive skip connections. We introduce a novel composite loss function that specifically addresses manga-specific challenges by incorporating structural and perceptual components. Experiments on a diverse manga dataset show that the ResNet-based model achieves higher color consistency and better stability, producing fewer artifacts in uniform areas. However, U-Net preserves fine details more effectively. These results provide insights into trade-offs between architectures, guiding practical implementations of manga colorization systems.
Cakiral et al. (Fri,) studied this question.