Abstract
We present InfinityGAN, a method to generate arbitrary-resolution images. Theproblem is associated with several key challenges. First, scaling existingmodels to a high resolution is resource-constrained, both in terms ofcomputation and availability of high-resolution training data. Infinity-GANtrains and infers patch-by-patch seamlessly with low computational resources.Second, large images should be locally and globally consistent, avoidrepetitive patterns, and look realistic. To address these, InfinityGAN takesglobal appearance, local structure and texture into account.With thisformulation, we can generate images with resolution and level of detail notattainable before. Experimental evaluation supports that InfinityGAN generatesimageswith superior global structure compared to baselines at the same timefeaturing parallelizable inference. Finally, we how several applicationsunlocked by our approach, such as fusing styles spatially, multi-modaloutpainting and image inbetweening at arbitrary input and output resolutions