Why does DLow improve diversity when z is also from Gaussian distribution?

[OliverSTH]

Hi @Khrylx ,

Why does DLow improve diversity when z is also from Gaussian distribution?

My understanding is that
CVAE takes $z_1$,$z_2$,... $z_k$ from Gaussian distribution by independent sampling.

However, DLow will use a unique $ϵ$ to map $z_1$, $z_2$,...$z_k$ via the mapping function.

So can I say that the role of network $Q_\gamma$ makes it more diversity than the CVAE method?

Or is there a better explanation?

and

Why does the unique $ϵ$ mapping to different locations in the Latent space can
solve the problem of CVAE's pick the major mode?

Thank you for reply.