variational autoencoder (VAE) has widely been utilized for modeling data
distributions because it is theoretically elegant, easy to train, and has nice
manifold representations. However, when applied to image reconstruction and
synthesis tasks, VAE shows the limitation that the generat