Good paper that employs generative models for optimizing black-box objectives, where Bayesian optimization might otherwise be used. The insight that the generative model need only be locally accurate to provide a descent direction for optimization is interesting. Overall, the reviewers agree that the paper has merit, and that the proposed method is simple and powerful. The reviewers made several suggestions for improvement. I would encourage the authors to take to heart the reviewers' suggestions when preparing the camera-ready, and to add the additional details (e.g. baselines, wall-clock times) that the reviewers requested.