Over at Nvidia, Tero Karras, Samuli Laine, and Timo Aila published a paper (and accompanying video, viewable above) showing off what their AI is capable of when generating human faces. The results are nothing short of remarkable, and definitely a bit startling when you consider the implications of a computer being able to produce believable photographs of people who don't actually exist. Generating realistic cats, on the other hand, seems to be more of a struggle, but we'll get to that in a minute.
"We propose an alternative generator architecture for generative adversarial networks, borrowing from style transfer literature," the paper (viewable in PDF here (opens in new tab)) reads. "The new architecture leads to an automatically learned, unsupervised separation of high-level attributes (e.g., pose and identity when trained on human faces) and stochastic variation in the generated images (e.g., freckles, hair), and it enables intuitive, scale-specific control of the synthesis."
I'm no smarty pants so I can't tell you what all that means, but the results are downright amazing. For the most part, the human faces the AI generates are completely believable ones, or at the very least don't contain anything to set off our 'uncanny valley' repulsion reactions that can occur when we see something that is almost human but just a teensy bit off.
Both the video and the paper go on to demonstrate the same method applied to images besides human faces: automobiles, bedrooms, and cats. Why those? Possibly because of the huge image libraries available for the AI to study (I'll admit here I didn't actually read the entire paper). The cars look pretty amazing, too, which is impressive since we're so used to seeing makes and models we're familiar with so often, though if you really look through the images you'll definitely spot some weirdness (one has a giant weird half-deflated tire, a few others are sort of non-symmetrical and warped).
I didn't spend much time looking at the bedrooms because pictures of bedrooms are boring and also because holy shit some of the cats the AI generated are seriously messed up.
Cats of Cthulhu! Sure, some look okay at a glance, but others are writhing, misshapen cat-esque horrors. At first it seems odd the AI could nail down human features, hair, freckles, wrinkles, and even eyeglasses so astutely, and then the very same AI could turn around and generate a cat that looks like it's passed through a meat grinder. But then I suppose faces are one thing and entire bodies are something else. If Nvidia's AI tried to generate and entire naked human (and I really hope it does because I'd like to see that) the results might not be any better.
I also like that a few of the cat images produced even have a bit of AI-generated meme text built in. Point a computer at a cat library on the internet, you're gonna get some memes in there, I guess. Maybe we shouldn't be teaching our AI to learn unsupervised. Keep it off Reddit, at least, until it's bit older.
My favorite AI generated cat is The One I'm calling Void Cat:
I think It's cute, don't you? I'm hoping to make It the new mascot of PC Gamer. Yes, if you gaze upon Void Cat too long It will devour your soul and leave you a barren, withered husk, but Void Cat is still adorable in Its own way. I'm not sure if that's its tail or perhaps Its foot (one of thirteen, I'm guessing) covering Its mouth (?) as It deliberates whether or not to spare you from the unknowable, unending abyss.
I'm guessing I will find out soon. We all will find out soon. It won't be long, now. It won't be long.