Improving alignment of flow and diffusion models during inference
Steepest Guidance: A Practical and Principled Approach to Inference-Time Alignment of Flow and Diffusion-based Models
Machine Learning
Summary
Generating content with flow and diffusion models needs a way to sync their outputs well, but this is hard to do in practice. The authors propose a new method called Steepest Guidance, which treats this syncing as a step-by-step optimization problem, focusing on small, local improvements. Their approach is based on a solid mathematical principle and they show through experiments that it works effectively. This helps make generative models more flexible and reliable.
What this means in practice
- •For machine learning engineers: Improve generative model flexibility and control by better aligning flow and diffusion models during inference.
- •For computer graphics developers: Enhance image or video synthesis pipelines by applying a principled method to align generative models for higher output quality.
Authors
Shokichi Takakura, Akifumi Wachi, Rei Higuchi, Kohei Miyaguchi, Taiji Suzuki
Abstract
Inference-time alignment of flow and diffusion-based models is critical for achieving flexible generative modeling. Theoretically, Doob's $h$-transform provides an elegant solution to this problem, and most existing methods are based on this principle. However, in practice, estimating the optimal guidance derived from Doob's $h$-transform at inference time is challenging. To deal with this issue, we regard inference-time alignment as a sequential optimization problem in the space of probability measures and propose a novel framework called *Steepest Guidance*, based on the principle of maximizing local improvement in the objective. We provide a theoretical analysis of the proposed method and demonstrate its effectiveness through extensive experiments.