Only managed to identify 3/10 faces correctly.
2/10. Are there any giveaways that a face is AI generated?
Faces are very well done these days, you can go to https://thispersondoesnotexist.com/ and keep refreshing images. Eventuall you’ll notice little things, like earrings not matching or sunglasses being 2 different halves melded together, hats and hair can look weird, etc
But most of them appear very human
That’s wild. Some of them are really good, but I see what you mean once you keep refreshing the page. Glasses are a good tell once you look closely. Earlobes also appear to be difficult to perfect as well.
I went back to the NYT article and tried again with my goldfish memory using that deduction and got 6/10.
Also, these are all emulating professional camera conditions. In a more standard phone photo scenario, the lower quality can lead to all sorts of weird looking illusions in real photos that can make them look ai generated. I was playing with some AI photo editing and saw some things that looked off which I thought was the generation messing up, but after checking the original photo again I saw that the weird stuff was actually in the original photo due to weird shadows or motion blur.
For me it was the eyes. Pretty much all of the real faces had realistic reflections in the eyes. The AI ones didn’t.
I got 8/10. The blond AI and the dude with glasses were the two I missed.
I know you’re joking, but the opposite is likely the future of game graphics: First a conventional render, perhaps even with path-tracing (although that might not be necessary), then AI on top:
I got 6/10. For me the thing I was looking for most was images that seemed to have wrinkles and things that seemed slightly blurred out. I don’t know if this is because it’s trained on a dataset where real people frequently do this (which is why it’s only a 6/10, because many pictures of real people have this feature as well), or if it’s because it’s generating detail and that’s just how it looks in current versions.
I was mostly looking for details to indicate the person was real. Got 8/10. Someone had an odd tooth, someone else’s eyebrow hair looked like they had just woken up. There was one picture I immediately knew was AI, because it just seemed off. Another had some strange wrinkle texture on the neck that just seemed unnatural. One of the 2 I got wrong because I thought the eye wrinkles were too much for the rest of the face. Turned out to be a real person.
I’ve done a lot of stuff with generative models called GANs, like StyleGAN (2?) which I believe these pictures to be from. My main focus is the “hair bubble effect”. This works best for people with longer hair, which is why I had 30% wrong in this test. Basically, by starting at images generated by these models for a long time, I started noticing that it is bad at creating the few loose hairs that stand out from the main pack. These plucks of hair often seem to go around some invisible “bubble” or weirdly flow together with the background. So my main point of focus is often the transition between hair and background, or just the hair in general, since that’s where it’s most likely to mess up. But the images picked here were also intentionally picked to be the most confusing according to the rest of the article, so it’s not that weird that these are hard to classify. Some of the real ones looked extremely AI to me, and it was only after the first false positive that I got a lot more careful with labeling some as “AI” than I normally would.
Example, the strands of hair here (though admittedly the effect is not very convincing here):
I surprisingly got 6/10. The AI ones that I guessed had some faint lines on the skin near the chin that seemed kinda blurry.
Well since it’s a bunch of real faces combined it’s not that difficult to make a realistic face.
6/10. Of the ones I got wrong 3 of them seemed heavily photoshopped.
I got the same score, though I got the dark haired woman. I don’t remember the one I missed instead.
I feel like trying to identify unedited images and AI isn’t as hard, but when they’re edited it gives many of the same details that AI generated ones have. Mostly it’s the smoothing out of details. This could be just that the AI was trained on a dataset pulled from images available online (which are almost all edited to some extent) or just because the way it generates it makes details not quite clear. Probably a little of both.
