SemiPainter: Learning to Draw Semi-realistic Paintings from the Manga Line Drawings and Flat Shadow

Keyue Fan, Shiguang Liu, Wenhuan Lu

Abstract: Semi-realistic paintings are widely used in game concept art, film posters and animation, etc. However, none of existing work can reproduce this type of art. We propose the SemiPainter to generate semi-realistic paintings from line drawings based on a deep learning network. We divide the complex work into two parts, generating the shadow from a line drawing and a simple flat shadow image at first, then generating flat color from line drawings. We merge these two paintings and get the final result. We use two networks with U-Net structure to realize these two stages of work. In order to ensure the global consistency of picture illumination, we add the lighting direction to the shadow network. The color of each part of the painting is uncertain, so color hints from users is provided when coloring, so that the network can realize controllable colorization. To the best of our knowledge, this is the first framework for reproducing semi-realistic paintings.