Generator pyramid for high-resolution image inpainting

Abstract Inpainting high-resolution images with large holes challenges existing deep learning-based image inpainting methods. We present a novel framework—PyramidFill for inpainting, which explicitly disentangles the task into two sub-tasks: content completion and texture synthesis. PyramidFill attempts to complete of unknown regions in lower-resolution image, synthesize textures higher-resolution progressively. Thus, our model consists pyramid fully convolutional GANs, wherein GAN is responsible completing contents lowest-resolution masked each synthesizing image. Since demand different abilities from generators, we customize architectures GAN. Experiments on multiple datasets including CelebA-HQ, Places2 new natural scenery dataset (NSHQ) resolutions demonstrate that generates higher-quality results than state-of-the-art