Reviewers appreciated the empirical results in this paper most of all, and the fact that they come from relatively small but well-motivated technical improvements to a well-known recent approach for unsupervised keypoint discovery. These two key changes, combined with the idea of using the discovered keypoints for long term video prediction is the main contribution in this work. Initial experiment design and evaluation in the submission was somewhat sloppy and relied on subjective scores of a small sample of video generations, but the author response to reviews now shows strong quantitative results across datasets, which are compelling. I learn marginally towards accept on this submission.