Scalable Normalizing Flows for Visual Generation

Industrial Problems Seminar

Jiatao Gu
Apple

Abstract

Normalizing flows offer a principled framework for generative modeling with exact likelihood estimation, efficient sampling, and structured latent representations. However, scaling flow-based models to high-dimensional visual generation has historically been challenging, especially compared with the recent success of diffusion and autoregressive models. In this talk, I will discuss recent progress on making normalizing flows more scalable and effective for visual generation, with a focus on modern architectures and training strategies for images and videos. In particular, I will present STARFlow, a family of scalable autoregressive normalizing flow models that bridge desirable properties of likelihood-based modeling and high-quality visual synthesis. I will discuss how these models relate to diffusion and autoregressive approaches, what advantages they offer in efficiency and modeling flexibility, and why flow-based methods remain a promising direction for large-scale visual generative modeling.