Explicitly Minimizing the Blur Error of Variational Autoencoders