BigColor: Colorization Using a Generative Color Prior for Natural Images

For realistic and vivid colorization, generative priors have recently been exploited. However, such often fail for in-the-wild complex images due to their limited representation space. In this paper, we propose BigColor, a novel colorization approach that provides diverse with structures. While previous are trained synthesize both image structures colors, learn color prior focus on synthesis given the spatial structure of an image. way, reduce burden synthesizing from expand its space cover images. To end, BigGAN-inspired encoder-generator network uses feature map instead spatially-flattened BigGAN latent code, resulting in enlarged BigColor enables robust inputs single forward pass, supports arbitrary input resolutions, multi-modal results. We demonstrate significantly outperforms existing methods especially