I hope they find latent diffusion works for this, without it this will probably be too expensive for private parties to train on big collections.
My understanding was that diffusion models were quite a bit more expensive, but yielded richer latent distributions and better images (for some definition of better).