A few months ago, Nvidia’s AI photo generation technology went viral. The media marveled at the uncanny technological power of the company’s engine, called StyleGAN, which generates photos of people that don’t actually exist.
But while people were busy gawking at how real these machine-generated people looked, they missed the other important part of Nvidia’s experiment: Computer-generated cats.
[Image: courtesy Nvidia]
AIWeirdness’s Janelle Shane highlighted these scary cats in a blog post yesterday. They represent the discards of Nvidia’s mad scientist experiments with its StyleGAN engine, which is capable of generating images of nearly anything made of pixels, including humans, cars, and even bedrooms:
StyleGAN is a generative adversarial network. It’s made up of two algorithms: The first generates cats based on its training on thousands of cat images, while the second evaluates the synthetic images and compares them to the real photos. Then, the second AI gives feedback to the first on its work–until it finally manages to create consistently believable portraits.
Nvidia’s StyleGAN was designed around something called “style transfer.” It doesn’t copy and paste elements of different photos to create a new one. That’s too imperfect and would never look good, according to the scientists who worked on the project. Instead, StyleGAN analyzes three basic things in every photo–which they call styles– and then merges them into something completely new.
The styles are called “coarse,” “middle,” and “fine.” Coarse deals with parameters like the cat’s face, its pose, and the type of hair. The middle is the facial features themselves, like the eyes, mouth, and nose shape. And finally, the fine styles are things like the color of the hair. The scientists describe in their paper how StyleGAN uses this combination of technologies to effectively eliminate noise that is irrelevant for the new synthetic face–for instance, distinguishing a bow on a cat’s head and discarding it as superfluous.
Obviously, things went very wrong during the learning process. This process of generation and critique created plenty of left-over mistakes and failures, which is what we’re seeing in these cats. But it’s all part of the development process–and remember, no cats were harmed in the making of this AI.