Very interesting, and aligns very well with my experience with synthetic data as well. (In that diversity trumps realism by a LARGE margin).