GauGAN: Nvidia's AI That Turns Drawings into Landscapes

Nvidia has made a massive bet on artificial intelligence. One of its most spectacular advances is the creation of fake portraits of people who don't exist, and the company doesn't hesitate to share other projects whenever it gets the chance. At its latest GPU Technology Conference, the green giant from Santa Clara introduced GauGAN, a software powered by generative adversarial networks that can create entire landscapes with just a few solid color strokes in the style of Paint.

AI and Art

Their relationship is much closer than the public imagines, and with each new generation, their abilities become more advanced and complex. We still remember well Google's initial attempts, the strange interpretation of nudes, and the first international recognition with the work "El Hijo del Carnicero". However, there is an important difference between creating art and assisting the user in creating art. That is the direction Nvidia is taking with its new GauGAN platform. The company is in the middle of GTC 2019 (which will last until tomorrow), and it took the opportunity to share the state of this particular tool.

How GauGAN Works

In relaxed terms, GauGAN is an image editor not far from Paint. Its interface features a total of three tools: pencil, brush, and paint bucket. But its real power lies at the bottom. There, each color has a special description, and when painting with it, the generative adversarial networks take care of the conversion. Precision is optional: a thick, general stroke is enough for GauGAN to create rivers, trees, rock walls, waterfalls, wood, flowers, sand, snow, grass, mountains, and much more. Another impressive aspect of GauGAN is that changes apply almost in real time, thanks to a hardware structure based on RDX Titan cards. With some code modifications, GauGAN could work on conventional graphics cards and even processors, as long as you accept a slight loss in performance.

Training and Future Plans

Nvidia used one million Flickr images under Creative Commons to train GauGAN, and besides creating landscapes from scratch, it can also synthesize hundreds of thousands of objects. For example, if the user specifies that the season is winter, leaves will disappear from branches. Nvidia's current plan is for GauGAN to become part of its future AI Playground, but no commercial launch is expected in the short term. Still, the possibility of a public demo has not been completely ruled out. Obviously, GauGAN's results are not perfect, and Nvidia recognizes that someone could use this technology for malicious purposes, but it would be a mistake to deny its enormous creative potential.

Source: TechCrunch