The innovation is that it doesn't predict image patches (like older autoregressive image models) but somehow does some sort of "next scale" or "next resolution" prediction.

It still predicts image patches, left to right and top to bottom. The main difference is that you start with patches at a low resolution.