- Need to make it clearer we're computing log-likelihoods (or the KL bound on these, rather) - Need to fix the latent resampling to harmonize with the method used in the transformer inferer and allow it to work in 3D