Producing robust and hallucination-free generative models is difficult, time-consuming, and expensive.