The use of well-disentangled representations offers many advantages for
downstream tasks, e.g. an increased sample efficiency, or better
interpretability. However, the quality of disentangled interpretations is often
highly dependent on the choice of dataset-specific hyperparameters, i