Title: Greed is Good: A Unifying Perspective on Guided Generation

Abstract: Guided generation without the need for training has become a prominent and potent method, empowering users to exert additional control over the generative mechanisms of flow and diffusion models. Currently, two distinct approaches have arisen for addressing this challenge within gradient-based guidance: posterior guidance, which involves projecting the current sample toward the target distribution using the target prediction model, and end-to-end guidance, which relies on backpropagation across the entire ordinary differential equation (ODE) integration process. This study demonstrates that these two seemingly disparate categories can be reconciled by interpreting posterior guidance as a greedy approximation of end-to-end guidance. We investigate the theoretical links between these methodologies and offer a comprehensive theoretical analysis of their performance relative to continuous ideal gradients. Building on these insights, we introduce a technique to interpolate between the two families, thereby allowing for a balance between computational cost and the accuracy of the guidance gradients. Finally, we substantiate our findings through experiments on various inverse image tasks and property-guided molecular generation.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC