Abstract
We present SeamlessGAN, a method capable of automatically generating tileabletexture maps from a single input exemplar. In contrast to most existingmethods, focused solely on solving the synthesis problem, our work tackles bothproblems, synthesis and tileability, simultaneously. Our key idea is to realizethat tiling a latent space within a generative network trained usingadversarial expansion techniques produces outputs with continuity at the seamintersection that can be then be turned into tileable images by cropping thecentral area. Since not every value of the latent space is valid to producehigh-quality outputs, we leverage the discriminator as a perceptual errormetric capable of identifying artifact-free textures during a sampling process.Further, in contrast to previous work on deep texture synthesis, our model isdesigned and optimized to work with multi-layered texture representations,enabling textures composed of multiple maps such as albedo, normals, etc. Weextensively test our design choices for the network architecture, loss functionand sampling parameters. We show qualitatively and quantitatively that ourapproach outperforms previous methods and works for textures of differenttypes.