Beyond Log-Concavity and Score Regularity: Improved Convergence Bounds for Score-Based Generative Models in W2-distance
This work develops a new method to find explicit convergence bounds for score-based generative models in W2-distance. It achieves this by relaxing strict requirements like log-concavity and score regularity, instead using weaker assumptions on the data distribution. The framework uses a forward Markov process and a time-reversed process to analyze how these weak properties affect model accuracy.