AI landscape art is here, courtesy of Nvidia

Charlie Spargo's picture
Nvidia's GauGAN

More intriguing and exciting features from the world of artificial intelligence have been revealed with the unveiling of Nvidia's GauGAN at their GPU Technology Conference in Silicon Valley.

The image creator, which uses generative adversarial networks (GANs), takes simplistic drawings made by the user and turns them into hyperrealistic landscapes. All it takes is one filled box for land and one for the sky, and the system puts out an image almost indistinguishable from reality.

GauGAN named for the French post-impressionist Gauguin, and has been created as a mere example of the potential of Nvidia's neural network platforms for the GTC conference which is taking place right now.

There are just three tools in play - a paint bucket, pen and pencil. However, the processing of the system turns any half-hearted sketch into a stunning natural scene. GauGAN was trained using a million images, and even understands contextual clues - like leafy trees present in a lush summer scene, but empty of leaves when snow is on the ground.

It is also capable of creating variety. Plug in two instances of exactly the same inputs, and the outputs will be different due to the use of randomised numbers.

GANs have been used for many situations. Recently developers created websites capable of generating realistic but ultimately entirely faked human (and cat) faces.