TileGAN: a New Approach to Generating Large-Scale Textures

Let’s have a look at another important paper from SIGGRAPH that is dedicated to Synthesis of Large-Scale Non-Homogeneous Textures.

Creating large-scale textures of high quality is an extremely challenging task. Existing solutions have limitations like repetition and artifacts. The paper discusses TileGAN — a texture synthesis framework that is capable of generating large-scale textures using generative adversarial networks.

“We tackle the problem of texture synthesis in the setting where many input images are given and a large-scale output is required. We build on recent generative adversarial networks and propose two extensions in this paper,” says the abstract. The research proposes an algorithm to combine outputs of GANs trained on a smaller resolution to produce a large-scale plausible texture map with virtually no boundary artifacts. One of the best things here is that you get a user interface for artistic control.

TileGAN lets users interactively generate and paint textures using latent brushes. The team trained a network using progressive growing on GAns on texture data. The researchers state that the generated texture tiles have very high quality and rich details. The key here is careful tiling of intermediate latent tiles.

You can learn more and find the full paper here. You can also find the code on GitHub.