Original title: Guided Flows for Generative Modeling and Decision Making
Authors: Qinqing Zheng, Matt Le, Neta Shaul, Yaron Lipman, Aditya Grover, Ricky T. Q. Chen
In this article, the focus is on using guidance without classifiers to enhance conditional generative models, especially Flow Matching models. Unlike diffusion models, which have benefited from this guidance, Flow Matching methods like Continuous Normalizing Flows (CNFs) haven’t explored this avenue. The study investigates the impact of Guided Flows across various tasks such as conditional image generation, speech synthesis, and offline reinforcement learning, a novel application for flow models. It’s revealed that Guided Flows notably enhance sample quality in image generation and zero-shot text-to-speech synthesis. Remarkably, these improvements come with significantly reduced computational requirements, without compromising the overall performance of the agent. This research marks a pioneering step in employing guidance techniques within Flow Matching models, showcasing their potential for various applications while optimizing computational resources.
Original article: https://arxiv.org/abs/2311.13443