Version 3.0 seems to be a bit better at making diverse faces and images, and I feel like some of the stylistic elements are stronger; however, if you don't prompt specific appearances, there's still a good chance elements will show up, similar to the limited training dataset.