User-Guided Anime Line Art Colorization with Spatially-adaptive Normalization

Line art plays an essential role in the process of anime creation. Colorization of the line art images is a tough task for the reason that neither grayscale values nor coloring semantic information exists in grayscale line art images. In this paper, we propose a novel line art colorization architecture with spatially-adaptive normalization for user-guide colorization. Our method can obtain high-quality colorized anime images by reserving more semantic information of inputs. Specifically, we integrate spatially-adaptive normalization block as well as spectral normalization, and the hybrid normalization enables us to train the network robustly while preserving the semantic information as much as possible. Our model is based on U-net architecture, which is regarded as a successful feature extraction architecture and is widely applied in the field of image segmentation and image generation. Also, We propose an anime data processing paradigm, combining the colored picture and the synthetic line art to generate simulated strokes in line with human intuition, which is conducive to making the generated images more realistic and perform on in-the-wild data better. With the proposed model, we evaluate a collected anime dataset and compare it with the existing method. The experimental results demonstrate the excellent and stable coloring effect of our proposed model.