Thank you for the reference, it looks very promising. I've heard of ways to smooth the latent space through Lipschitz regularisation, but then got disappointed again when I read "ah well it's just layer normalisation". So many things in ML come in a different appearance and actually mean the same thing once you implement them.