The paper proposed a normalizing-flow architecture based on a multi-scale decomposition obtained by a wavelet transformation. The benefits of the proposed architecture are faster training (due to improved parallelizability) and super-resolution. These are interesting and novel developments towards scaling up normalizing-flow models. The reviewers had a few concerns, some of which were addressed in the rebuttal (such as training time) while others remain (such as blurriness of samples), and pointed to several directions of improvement and future work. After the rebuttal, all reviewers recommend acceptance. I would strongly advise the authors to take the reviewers' feedback to heart when revising the paper. In particular, it would be good to add the improved comparison of training times to the revised version, and to discuss the issues regarding blurriness of samples, as well as discuss the possible improvements and directions of future exploration that the reviewers suggested.