The submission is an empirical work which shows that a robust classifier can be used for various image synthesis tasks. The empirical study is solid, extensive, and conclusive. Both quantitative and qualitative results suggest that the hypothesis is correct and robust classifiers solve various image synthesis tasks with a simple likelihood maximization. The submission is interesting and would be a good addition to the conference. One of the issues raised by the reviewers is additional experiments specifically trying the proposed methodology with a non-robust network. Including it in the camera-ready version would be good for the sake of completeness. The paper has a potentially large impact on the community. Synthesizing images requires learning the data distribution of natural images. And, the paper suggests that the robust networks do a better job in learning this prior. This result might have a large impact on the community since the learned data distribution can be used in various downstream tasks.