On the Evaluation of Generative Models in Distributed Learning Tasks