TL;DR通过熵正则化的 f - 散度梯度流技术,可以提高生成数据的质量,避免了浪费的样本拒绝,可以应用于各种生成系统,并取得了显著的效果优于 DOT 和 DDLS 方法。
Abstract
deep generative modeling has seen impressive advances in recent years, to the
point where it is now commonplace to see simulated samples (e.g., images) that
closely resemble real-world data. However, generation quality is generally
inconsistent for any given model and can vary dramatic