>>106479827
>I really don't see the point to go vae less.
the paper claims that it gives you faster inference than using a vae
https://arxiv.org/abs/2507.23268