Title: Stabilizing Diffusion Steering: A Framework for Preventing Generative Path Collapse

Abstract

Inference-time steering offers a method to adapt pretrained diffusion and flow models to novel tasks without the need for retraining. This approach typically employs ratio-of-densities mechanisms to reweight time-indexed marginals using fixed exponents. However, we identify a critical failure mode termed "Marginal Path Collapse," where the intermediate density resulting from such compositions becomes non-normalizable, even when the start and end points are valid. This instability often emerges when combining heterogeneous experts that have been trained with incompatible noise schedules or when utilizing negative exponents and partial supports.

To resolve these issues, this work introduces two key contributions: (i) a rigorous Path Existence Criterion that delineates the conditions under which composed intermediate densities remain mathematically well-defined, and (ii) Adaptive Path Correction with Exponents (ACE). ACE extends the capabilities of Feynman-Kac steering by accommodating time-varying exponents. Our theoretical analysis demonstrates that ACE regulates the quantile radius of intermediate distributions, thereby offering a mathematical explanation for the path stabilization observed in empirical results.

We evaluate ACE on a complex drug design task involving flexible-pose scaffold decoration, which requires the integration of de-novo, conformer, and protein-conditioned experts. In this scenario, ACE successfully prevents path collapse and achieves performance significantly superior to baselines relying on constant exponents. Additionally, ACE enhances attribute success rates in compositional image generation, positioning it as a robust and general framework for compositional sampling.

Project Page: https://ziseoklee.github.io/projects/ACE/

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC