LCDiff: Line art colorization with coarse-to-fine diffusion and mask-guided voting

Abstract

Line art colorization is crucial in animation production. It aims to add colors to target line art based on reference color images. The process of colorization animation remains challenging due to inadequate handling of large movements between frames, error accumulation during sequential frame processing, and color fragmentation issues during pixel-level processing. To address this issue, we propose a novel LCDiff method for line art colorization. In our method, LCDiff first utilizes a coarse-to-fine framework combining preliminary color estimation and label map diffusion modules to address the inadequate handling of large movements. Then, we introduce a color correction pathway in diffusion model that prevents error accumulation in sequential processing. Additionally, we incorporate a mask-guided voting mechanism to resolve color fragmentation issues during pixel-level processing. Extensive experiments on synthetic and real-world datasets demonstrate that our method achieves impressive performance in line art colorization.