An Interpretable Evaluation of Entropy-based Novelty of Generative Models