Do mean hard to be used because of compute required to render it or because of setting up and understanding the training pipeline and related hyperparameters?