A training objective for generative models that directly maximizes likelihood without requiring variational bounds or adversarial training.