I can see another use for time-series generative models; to create the right shape of noisy data.

To improve the robustness of machine learning models, artificial noise is often introduced into actual training data. Something like Timestral could simplify the process of generating it.