What are some common challenges you might face when training generative models and how would you address them?