IIRC there was a NIPS 2016 paper that showed that the blurry images are because of the variational approximation used in the generator/decoder network. They rep

IIRC there was a NIPS 2016 paper that showed that the blurry images are because of the variational approximation used in the generator/decoder network. They replaced this by doing MCMC inference on the encoder network, and showed that you get sharper images, but much less efficiently. In general, variational inference will always tend to model the mean or mode of the data.