Conditioning Generative Models with Restrictions — LessWrong