Defer initial root model refresh to the representation that actually use the model data

Full representation already have a lazy loading mechanism, we could make use to that.