We introduce generative distribution embeddings (GDE), a framework that lifts autoencoders to the space of distributions.