As machine-learning advances are giving rise to a vibrant artificial-intelligence-generated art scene, developers have begun to create versions of these tools easy enough for even the most code-challenged creatives to use.
The latest addition to this growing array is a platform called Runway ML, which hosts dozens of AI-based creative functions ranging from utilitarian tasks to experimental works-in-progress and whimsical novelties.
The hub draws together a range of recent breakthroughs that have transformed the creative capacity of AI, including research group OpenAI’s text generator, GPT-2, new methods of image fabrication and an Nvidia web tool that turns doodles into photorealistic landscapes.
“We are on the verge of a new creative revolution,” Runway ML’s mission statement says. “Recent advances in machine learning and artificial intelligence research are producing radical changes in the way digital content is made, understood and processed, unfastening previously unimagined ways of creating.”
The key innovation driving much of this progress is a machine-learning model called a Generative Adversarial Network (GAN), in which a neural network generates images by honing them until another neural network can no longer tell the difference between the fake images and the real ones of the data set. The setup was first formalized in 2014 by Ian Goodfellow, now Apple’s head of machine learning.
A milestone in the nascent field came last fall when famed auction house Christie’s sold a painting created with a GAN for nearly half a million dollars.
On Runway ML, users can generate portraits like that pricey artwork—as well as photorealistic faces, landscapes and a few other types of images—through an application called StyleGAN, an open-source tool created by Nvidia researchers last December. A similar tool from researchers at Google’s DeepMind division called BigGAN offers 1,000 categories of image generation, though many of them tend to produce images that are a bit more garbled.
StyleGAN can create portraits similar to the one that Christie’s auction house sold as well as realistic human faces.
On the more experimental side is a tool from Runway ML creator Cristóbal Valenzuela called AttnGAN, which creates images based on text descriptions of a scene. The results are rarely photorealistic depictions of the given prompt, but the tool usually makes for entertainingly abstract takes on the text nonetheless.
Runway ML tool AttnGAN turns text to images.
Two different tools on the platform will stylize images in the mold of famous artists and genres, while a third transplants the perceived style of one image onto another. Another set of programs perform simple photo editing tasks, like colorizing black-and-white images, removing specific objects from a frame and other discrete transformations similar to tools from companies like Adobe and IBM.
The hub also offers apps for video editors, such as real-time skeletal tracking, facial feature detection and an easy way to turn human actors into 3D graphic models for animation purposes.
The video section of Runway ML’s toolkit.
On the text-based front, there is an image-to-text-description translator and a version of the copy generator GPT-2, which spits out an article when prompted with a chunk of text. OpenAI deemed the system too dangerous to release in full earlier this year for fear that it would be used to mass-produce fake news.
In fact, much of the other functionality offered within Runway ML is based on technology that some experts worry has the potential for misuse in so-called deepfakes, or AI-fabricated footage that can be indistinguishable from the real thing. But the chances of doing so with the toolkit’s simplified versions of these programs seem slim, especially when the most advanced forms of this type of AI are still in early stages.
Agencies, brands and designers have already begun to experiment with some of the creative capabilities of generative AI. AKQA recently created its own sport with a neural network, complete with a GAN-generated logo; French designer Philippe Starck used Autodesk’s software to create the first commercial designed through AI—a chair for homeware brand Kartell; and Goodby Silverstein and Partners used deepfake tech to create an interactive persona of Salvador Dalí for his namesake museum.