Question for the form of latent z

Thanks for your excellent work!
Recently, I am confused by the form of latent z.
In some work, latent z is showed as feature map with size [channel, H, W].
But in multi-modal vaes, I always see z shown as [1, D] vector.
Is it fixed for mmvae+ to use vector? Or I could change the architectures of encoder and decoder to produce a feature map z?

Looking forward to your reply.
