This paper generated a large amount of discussion. On the positive side, the reviewers noted that the paper shows large benefits compared to baseline approaches in terms of producing scale-consistent outputs and significantly reduced training time. The reviewers appreciated the clarification on the phrase "efficiency" as explained in the rebuttal, as well as other clarifications about priors. For the final revision, we strongly recommend that the authors: - Describe in more detail why their method leads to faster training times, compared to the previous methods; this was a very confusing point to some of the reviewers - Explain the positive benefits of reduced training time, e.g. lowering the computational resources needed for training and thereby facilitating reuse of the method - Compute inference efficiency, i.e. runtime of the trained model at test time, compared to baseline approaches