Nvidia is showing off the power of a cutting-edge field of artificial intelligence with a new tool that turns rudimentary computer doodles into images that could pass for real photographs.
The software, which the Silicon Valley chip company plans to eventually integrate into its forthcoming AI Playground web studio, draws on a data set of around 1 million photos to flesh out a photo-quality landscape based on the rough shapes and colors of an MS Paint-caliber sketch.
Nvidia envisions the tool and the other deep-learning-based photo editing widgets in the AI Playground suite appealing to creative professionals like architects, game developers and content creators.
“It’s much easier to brainstorm designs with simple sketches,” said Bryan Catanzaro, vice president of applied deep learning research at Nvidia, in a statement. “It’s like a coloring book picture that describes where a tree is, where the sun is, where the sky is. And then the neural network is able to fill in all of the detail and texture, and the reflections, shadows and colors, based on what it has learned about real images.”
The software uses a generative adversarial network, or GAN, in which two neural networks—AI machines with computational nodes that mimic the neurons of a brain—are trained with vast data sets of images until one is able to generate an image that the other is unable to distinguish from the rest of the real photos. While the basic concept has existed for years, the name was first codified in a 2014 paper by Google senior staff research scientist Ian Goodfellow.
Such programs have made headlines recently for generating the first piece of AI-made art to sell at an auction house for nearly half a million dollars as well as their more dystopian potential to create hyper-realistic fake news images and videos called “deepfakes.” Fabrications like these could eventually include, for instance, AI-generated incriminating footage on which a public figure’s likeness can be superimposed and then be spread online.
But their more pedestrian application in design work of any kind could have big implications for creative fields, especially when combined with generative audio, language and text tools. Still, given the slightly off appearance that still marks most of a GAN network’s output to actual human eyes, don’t expect them to replace designers or artists any time soon.