There’s an obvious trend there of older faces being judged as human and younger faces being judged as AI. A little less obvious, it also looks like “scruffy” male faces may be judged as human while clean shaven males or males with well trimmed beards are judged as AI. I wonder if the study noticed / addressed that at all.
Also the shadows. No shadows (special lightning) is judged as AI generated. Where strong shadows is judged natural.
This is my number one giveaway for AI gen. There’s definitely a style out there which only shows flat, even lighting, with the subject viewed from directly in front. AI Female 44 in the OP is one such example
Also, prettier faces are judged more as AI than less pretty ones.
Intuitively that makes perfect sense to me, since young and attractive humans is what people generate most often
Not only that but there is very likely a huge beauty bias in the training sets.
If you’re looking in the library for books that are at least 100 years old, you’re generally only going to see the ones that people thought were worth preserving for 100 years.
If you’re training your image generation model with stock photographs, you’re generally only going to be giving it images of people who are literally models. Not all models are beautiful, but they’re probably more beautiful on average than the general population.
Yeah, I’m damn sure no AI face collector has got an image of my ugly mug.
Well yeah, look at how bad those real photos are. Two of them are even distorted.
That’s kinda the premise of GAN, it’s generator is going to keep generate more and more “human” faces until it’s discriminator can no longer determine them as generated. It’s literally generating more human humans than real humans (according to itself)
