I mean, I just typed "/imagine asian warrior nun --v 4" into Discord, thinking of Beatrice from the recent Netflix series, and three of the four results show what I'm talking about: https://i.imgur.com/499QCn6.png
I only see "stuff on the face" in the third image, which I guess is an improvement. I'm not sure I'd call hands "fixed" based on this image alone, but they're better.
With classifiers the issue is that if you place sufficient objects in space that co-occur the model will believe that it is a class with said objects, eg a face, but the problem is the relative positioning of them plus all angles of rotation.
I think geometric deep learning has a solution for the rotation ia rotation invariant models, but I haven’t gone through that book yet.