This paper proposes a new method for training multimodal generative models based on Jensen-Shannon Divergence. At the beginning, it received 5666 scores; after rebuttal, the scores have been increased to 6666, which are 4 weak accept recommendations. All the reviewers agree that this paper is well organized and the main techniques are also clearly presented. Therefore, the AC recommends accepting the paper. The reviewers also gave a lot of more detailed comments. The authors are encouraged to use these comments to further improve the paper.