PathGuide Dynamically Optimizes Classifier-Free Guidance for Generative Models

Modern generative models are highly capable of creating complex data, but achieving precise control during inference in conditional generation remains a significant challenge. Classifier-free guidance (CFG) is a primary mechanism for this control, typically treated as a static parameter. However, in flow-based models, the guidance scale directly influences the velocity field and the resulting probability path, suggesting that guidance selection should be a dynamic optimization problem. PathGuide addresses this by reframing scalar CFG selection as an on-policy transport problem. By leveraging the continuity equation, the framework derives a selection criterion with a clear interpretation of path correctness. It proves that if the guided field weakly aligns with the exact conditional field along the generated rollout, the sampler's path will coincide with the target conditional law. This criterion yields a strictly quadratic local objective, allowing for an efficient, closed-form selector for each solver interval. PathGuide can compute optimal guidance scales online during generation or fit them offline into a reusable schedule. Validation on image manifolds and controlled settings demonstrates that this transport-based selector enhances path alignment and sample fidelity, outperforming both fixed and state-of-the-art adaptive guidance methods.