Inverting Variational Autoencoders for Improved Generative Accuracy