The paper has been very well received by the reviewers for its novelty and the strong experimental evaluation. The ability to handle large scale inputs and the high compression ratios reported make this work a great contribution for the community. I would encourage the authors to make the final effort in further improving their work by addressing remaining concerns of the reviewers, one being discussion on limitations of the approach for instance. Eventual further comparisons with GAN based approaches would also be desired but not a required change.