Adaptive Diffusion Guidance via Stochastic Optimal Control

TL;DR

This paper introduces a theoretically grounded, adaptive guidance framework for diffusion models using stochastic optimal control, enabling dynamic guidance strength selection for improved sample quality.

Contribution

It formalizes the guidance strength and classifier confidence relationship and develops a stochastic control approach for adaptive guidance scheduling.

Findings

Provides a formal relationship between guidance strength and classifier confidence.

Develops a stochastic optimal control framework for dynamic guidance in diffusion models.

Establishes a principled method for guidance scheduling that adapts during sampling.

Abstract

Guidance is a cornerstone of modern diffusion models, playing a pivotal role in conditional generation and enhancing the quality of unconditional samples. However, current approaches to guidance scheduling--determining the appropriate guidance weight--are largely heuristic and lack a solid theoretical foundation. This work addresses these limitations on two fronts. First, we provide a theoretical formalization that precisely characterizes the relationship between guidance strength and classifier confidence. Second, building on this insight, we introduce a stochastic optimal control framework that casts guidance scheduling as an adaptive optimization problem. In this formulation, guidance strength is not fixed but dynamically selected based on time, the current sample, and the conditioning class, either independently or in combination. By solving the resulting control problem, we establish a principled foundation for more effective guidance in diffusion models.