Neural synthesis of tileable textures

Abstract

According to various embodiments, an artificial intelligence framework capable of neural synthesis of tileable textures is provided. Using a non-tileable texture input, such as single or multi-layer texture stacks, high-quality tileable textures can be generated that match the appearance of the input samples, while increasing their spatial resolution. Embodiments are provided that leverage a latent space approach in a generative network for synthesizing seamlessly tileable textures, which can maintain semantic consistency in boundaries between tiles. A tileability metric is provided as feedback to improve and optimize the tileability of the output texture, for example using a sampling algorithm that can generate high-resolution, artifact-free tileable textures. In embodiments, a convolutional discriminator is provided for detecting artifacts in the synthesized textures by locally estimating the quality of the synthesized maps. The convolutional discriminator can also provide the feedback-based approach for optimizing an input selection process.