TileGAN          
September 2020
StyleGAN2, RunwayML, Colab, Machine Learning
1000+ images of Vietnamese cement tiles
~700 images of Portugese azulejos
~700 images of Mexican talavera tiles

The models were trained for around 7000–12000 steps.

Thank you Alba & Maria for letting me use their beautiful collection of azulejos for training!

TileGAN are three Generative Adversarial Networks trained on collections of tile images from Vietnam, Portugal, and Mexico.

TileGAN were based on pre-trained StyleGAN2’s Flickr faces model architecture, using RunwayML and Google Colab for different iterations.

While the tiles were made far apart, they share commonalities in design and motifs. For example, flowers were present across all three tile types, as well as symmetry and tessellated patterns.