Foundations of the Diffusion AE and the latent space

[jbmaxwell]

Hi, I'm wondering if the Diffusion AE is based on https://diff-ae.github.io, and if so, whether the latent space is likewise divided into two-parts, as in: "Our method can encode any image into a two-part latent code where the first part is semantically meaningful and linear, and the second part captures stochastic details, allowing near-exact reconstruction."

I need a similarity metric and was hoping to use the semantic part, but I'm not sure how that would be accessed—i.e., would it be the first half of the channels or of the features? (This is assuming, of course, that this implementation has a similar latent structure.)