Do these models actually learn a 3D representation or do they just learn "something" that is good enough to produce an very convincing impression of 3D ?
Subquestion: if they don't learn 3D, can we say that models learning a 3D representation first will lead to even better productions ?