Those MS Paint doodles you made in the 90s might have been works of art just waiting to happen, and you never knew it. All they needed was some help from a generative adversarial network (GAN). Nvidia has shown that AI can use a simple representation of a landscape to render a photorealistic vista that doesn’t exist anywhere in the real world.
Nvidia calls its AI-powered landscape generator “GauGAN,” which is a mashup of GAN (generative adversarial network) and Gau (post-impressionist painter Paul Gauguin). The software is dead-simple, but that’s very much the point. It has just three tools: a paint bucket, a pen, and a paintbrush. After selecting your tool, you click on a material type at the bottom of the screen. Material types include things like tree, river, hill, mountain, rock, and sky.
The organization of materials in the sketch tells the software what each part of the doodle is supposed to represent, and it generates a realistic version of it in real time. It even has random numerical elements to ensure two copies of the same doodle will produce different results. Nvidia designed GauGAN to work on a Tensor computing platform powered by an RDX Titan GPU. That card has the necessary power to render the output in real time, but GauGAN should technically work on any platform, even a basic CPU. Although, it would take several seconds to generate images.
Generative adversarial networks are a hot topic in AI research right now because they have the potential to make training networks vastly easier. Rather than feeding a network labeled data until it learns how to process it, GANs consist of two neural networks that compete against each other. One network generates data (in this case, landscapes), and the other decides if it looks real or not. Over time, the networks get better at generating the desired output. In this case, gave the GAN 1 million images from Flickr to train itself.
The results are impressive but not perfect. Nvidia calls the images “photorealistic,” and they are in a way. At first glance, you might think that you’re looking at a real lake or waterfall, but there are telltale glitches and hard edges that don’t look natural. Still, it’s a lot better than most of us can do in MS Paint.
Nvidia hopes to add GauGAN to its AI Playground suite, but it will take a bit more work to get the software ready for public use.
Source: Extreme Tech