There is actually a specific approach of this concept for generating synthetic data for training datasets called UDAPDR[0].

It or something like it could likely be applied to any form of generation including what you are describing.

[0] - https://github.com/primeqa/primeqa/tree/4ae1b456dbe9f75276fe...