KAUST Researchers Are the First to Generate Images of Realistic and Highly Detailed Texture Maps of Gigapixel in Size Using Deep Neural Networks

Video for SIGGRAPH 2019 Technical Paper TileGAN: Synthesis of Large-Scale Non-Homogeneous Textures

KAUST researchers Anna Fruehstueck, Dr. Ibraheem Alhashim, and Prof. Peter Wonka have developed a novel technique to generate images of realistic and highly detailed texture maps using deep neural networks. The texture images synthesized by their system TileGAN can be of gigapixel size and are created by seamlessly merging smaller texture blocks into a single large image. The underlying neural networks are trained using high-resolution images such as detailed satellite imagery, maps and famous paintings.

Their research will be presented at the upcoming SIGGRAPH 2019 Conference in Los Angeles, CA.

This is a result of TileGAN creating a painting of Jeddah's old town Al Balad in the style of the painter Georges Seurat.

Anna is working on her doctoral dissertation at KAUST under the supervision of Prof. Peter Wonka. Dr. Alhashim is a Postdoctoral Research Fellow working with Prof. Wonka in his research group.