Diverse non-homogeneous texture synthesis from a single exemplar

Capturing non-local, long range features present in non-homogeneous textures is difficult to achieve with existing techniques. We introduce a new training method and architecture for single-exemplar texture synthesis that combines a Generative Adversarial Network (GAN) and a Variational Autoencoder (VAE). In the proposed architecture, the combined networks share information during training via structurally identical, independent blocks, facilitating highly diverse texture variations from a single image exemplar. Supporting this training method, we also include a similarity loss term that further encourages diverse output while also improving the overall quality. Using our approach, it is possible to produce diverse results over the entire sample size taken from a single model that can be trained in approximately 15 min. We show that our approach obtains superior performance when compared to SOTA texture synthesis methods and single image GAN methods using standard diversity and quality metrics.