Science

NVIDIA Research's "GauGAN" Is an Incredible, Magic-as-Hell Coloring Book

A virtual art app that's best described as a "magic-as-hell coloring book."

Painting beautiful landscapes could soon become as simple as a few mouse clicks. A new artificial intelligence app, introduced on Monday, has the potential to turn even the crudest of artists into a true Bob Ross.

The magic-as-hell coloring book you can see in the above video was developed by NVIDIA Research, the company behind StyleGAN, the generative adaptive network used to create the creepy one-off site This Person Does Not Exist. This new GAN, dubbed “GauGAN,” after the post-Impressionist painter Paul Gauguin (we love a good tortured French artist pun) used a similar method to train a new A.I. that can turn sketches into hyper-realistic works of art.

NVIDIA has been at the forefront of this type of neural network since GANs were first introduced in 2014 by Ian Goodfellow. GANs are the product of two networks: A generator and a discriminator. The two go head-to-head on millions of images, with the generator “presenting” images to the discriminator, which is then trained using real samples to provide “pixel-by-pixel feedback,” continuously improving on lifelike details.

If this particular strain of A.I. sounds at least a little spooky, you’re not the only one getting the creeps. This Person Does Not Exist creator Phillip Wang built a platform that effortlessly generates very realistic, very fake photos of people, in an attempt to illustrate the deceptive power of GANs.

A diagram of how a GAN works.

Rani Horev / LyrnAI

GauGAN presents a much friendlier side of GANs. The app allows users to create their own segmentation maps — think a green blob that you label “grass,” next to a blue blob that you call “pond” — that are then set upon by a neural network that’s been trained on real landscape images. The pond will suddenly feature glassy water that reflects the grass. The grass, made up of millions of individual sprouts, may even have drops of dew. Change the “grass” label to “snow,” and GauGAN will reconfigure the scene to reflect a winter setting. The result is not just a cut-and-paste job. It’s a unique, synthesized image.

NVIDIA researchers hope GauGAN can help create photo-realistic renderings for architects, urban planners and landscape designers in just a handful of minutes. In the comments on NVIDIA’s introductory blog post, users were already fantasizing about how they could incorporate the virtual art app into their own creative practices.

“Lacking any artistic ability, this would help me visualize elements in my fiction writing,” wrote one user.

“Will be really cool to draw with my daughter,” said another.

Though the app has yet to roll out publicly, Nvidia reportedly debuted GauGAN at their recent GPU Tech Conference booth, allowing participants to test out the GANs’ creative ability in real time. This June, the technology behind GauGAN will be presented at the 2019 CVPR Conference in Long Beach, Calif.

Related Tags