Dynamic Guidance for Flow-Based Generation

Abstract: In generative modeling, achieving precise control is essential, yet standard conditional generation methods often rely on static parameters. We show that for flow-based models, this is better framed as a dynamic optimization problem over the path. We introduce a framework that uses fundamental flow principles to derive an adaptive guidance criterion. This allows for the selection of optimal guidance scales during generation. Our results across various flow constructions show that this dynamic alignment improves sample fidelity.