I won’t say it’s impossible, but your suggestions do seem unlikely to me. We’re talking about unlabeled data, cropped to just the face, and the state of the art

I won’t say it’s impossible, but your suggestions do seem unlikely to me. We’re talking about unlabeled data, cropped to just the face, and the state of the art being found in commercial models in the real world rather than in papers competing to pass standard tests in the literature. (From my direct experience they arguably are pretty close to magical; I think there is just a lot more information in pictures of faces than we thought there was. It’s quite plausible that, since native human machinery was evolved in groups of maximum ~150 distinct faces and was optimized for things like detecting minor emotional cues, our brains are simply throwing a lot of that information out. But that’s not really the point I want to make here; instead, I want to say stuff like “commercial models aren’t trailing research papers; research papers are trailing commercial models”.)