Can someone explain why restricting the posterior `z` as diagonal Gaussian?

[seekerzz]

Maybe I do not understand this paper throughly, but can someone explain this? The posterior z is modelled as diagonal Gaussian. And in the Zero initialization part, ensures that the posterior distribution as a simple normal distribution. If it is a simple distribution, why a complex prior flow is needed to learn its distribution?

[XuezheMax]

This is the initialization to ensure the posterior distribution as a normal. During training, the posterior distribution will become more and more complex when we update the parameters.