I have two preprints on arXiv exploring an alternative to adversarial training to train generative networks; the first applies it to probabilistic forecasting, while the second is concerned with Likelihood-Free Inference. I feel that is a promising overlooked approach. Happy to hear any feedback! 