I'm mostly just shocked that the model can produce what appear to be entirely consistent models of the same character from different views. I can be convinced, but I was also equally skeptical.

Because that implies a kind of cohesive conceptual understanding of a single object, that I didn't think the current architecture was able to model that. So many image outputs of humans are perfectly fine in local areas, but then they have three arms or a hand where a foot should be or something.

I mean, the chosen character isn't perfectly consistent -- e.g. the side view shows a kind of metal badge on the vest that doesn't exist in the front view, while the back view shows a kind of mid-thigh holster (?) that doesn't appear on the front view, plus the vest on the back view is a few inches shorter. Similarly the collar is popped on the front view, but not the other two.

But nevertheless, it's basically about 99% right. And I simply don't understand how that's possible, unless there author did about ten thousand tries and is showing the only ones that happened to be nearly entirely consistent, somehow just mostly by chance. (Maybe because they're particularly close to the real-life characters used for training?)