Can we directly sample the latent space of VAE and get a reasonable generation result if we finish training the VAEs？

[hitsz-zuoqi]

@ZENGXH

[ZENGXH]

if we directly sample from a Naïve prior N(0,1), it wont generate a reasonable shape. Since the posterior is in a complex distribution and far from prior, we need to learn the prior to generate reasonable results.

In fact, if you look at the visualization of the latent points, you will have more sense of how far it's from the Naïve prior N(0,1).