Image Outpainting and Harmonization using Generative Adversarial\n Networks

Abstract

Although the inherently ambiguous task of predicting what resides beyond all\nfour edges of an image has rarely been explored before, we demonstrate that\nGANs hold powerful potential in producing reasonable extrapolations. Two\noutpainting methods are proposed that aim to instigate this line of research:\nthe first approach uses a context encoder inspired by common inpainting\narchitectures and paradigms, while the second approach adds an extra\npost-processing step using a single-image generative model. This way, the\nhallucinated details are integrated with the style of the original image, in an\nattempt to further boost the quality of the result and possibly allow for\narbitrary output resolutions to be supported.\n