You may have heard recently about a new AI algorithm designed to extrapolate what a face looks like from really poor-quality, low-resolution sources, called Face Depixelizer. The tool is pretty remarkable in how incredibly naturalistic the resulting faces that it generates look, though if we’re honest what you’ve likely heard about it has more to do with how weird the results are when it’s fed video game characters as input, or how it seems to have a bit of a racial bias. I just wanted to run some cars through it, so that’s what I did. And oh boy, did it get weird.
Given a low-resolution input image, model generates high-resolution images that are perceptually realistic and downscale correctly.
— Bomze (@tg_bomze) June 19, 2020
Of course, this tool was never meant for cars, but I’ve always anthropomorphised the crap out of the “faces” of cars, and I have a lot of fiercely-held opinions about car-face anthropomorphization. I was really curious to see how the AI handled a selection of car faces, and I know you are too, so first I’m going to show you the faces it generated, and let’s see if you can guess what cars I used as inputs.
Ready? Here we go:
Now, if you know what cars I have, that topshot likely gave away two of these answers, but I think you’ll still be pretty challenged. As I was doing these, I noticed some interesting quirks of the AI.
First, the overall car shape, when over a blank background, seems to get filled in with hair, which, if we’re honest, is a pretty malleable part of a head, shape-wise.
Sometimes parts of the cars became background blurs behind the head, but not that often. The most important factor I found when it came to producing a head that wouldn’t make an average person shit their pants with fear when seen has to do with the quality of the input image; specifically, the crappier, the better.
I tested this with the same image of a mid-1960s Volkswagen Type 2 bus, in 64 pixel, 32 pixel, and 16 pixel sizes. The less information, the more the AI could extrapolate and assume, and those results made a more “normal”-looking image.
See for yourself:
That 64 pixel one looks like a Frankenstein’d corpse of J. Jonah Jameson from the Spider-Man comics, gender-swapped and severely beaten. The 32 pixel one is a bit more normal, but still pretty strange, but that 16 pixel humanised bus could be any number of mums of friends I had growing up.
OK, look again at those faces up there. Make some guesses. Look at the overall shape of the head for a hint. Squint. Imagine. Ready to see how you did?
Here’s the full set of cars and faces I used, along with pixel sizes for each input image:
That’s got the full gamut from terrifying monsters to people I’d be ok taking a road trip with. Many of them might look better with a good haircut and re-colouring, but I think there’s at least four to six faces there (Beetle, Civic, Rolls-Royce, Sprite, and maybe the Porsches) that feel convincing.
The others, uh, less so.
Here’s what I think the best and worst results are:
For the best — as in the most convincing — I think this early ‘70s Beetle has to be it. Not only does the face seem completely plausible, but she even seems to match the Beetle’s time and place of origin. That looks like a sweet German lady who could have been born in the early ‘70s, doesn’t it? Nice to meet you, Inge.
Notice that her hair is the same shape as the Beetle, and look at how the tires became the shadowed areas under her ears. I think the seat headrests became her eyes, and the headlights became earrings? Windshield becomes forehead, bumper mouth, I think? Amazing.
For the worst, it has to be our poor early ‘60s BMC Mini here. Oh god, look at that monster. The headlights seem to have transformed into graying temple hair, the grille into a strong jaw and nose, eyes are hidden with a band of red cosmetics like that replicant woman in Blade Runner, the head shape has corners, oh no no no.
I think this was an example of too much detail input, which only leads to terror.
I love crap like this — it’s fascinating to see just what the algorithms do with this sort of information, and if I’m polluting their dataset with car pictures instead of faces, then, well, that’s my little way of fighting the Rise of the Machines.
Want to see who the person lurking in your cars are? Go here and try for yourself!