P-Guide: Single-Pass CFG Inference for Flow Matching

A new framework called P-Guide has been developed by researchers to provide high-quality classifier-free guidance (CFG) for flow matching in just one inference pass, achieved by adjusting only the initial latent state. When applying a first-order approximation, P-Guide mirrors CFG, directing generation from the prior space without the need for explicit velocity field extrapolation during sampling. This approach takes into account both homoscedastic and heteroscedastic priors, revealing that simultaneous modeling of mean and variance allows for adaptive loss reduction and enhanced robustness against data uncertainty. Experiments indicate that P-Guide cuts inference latency by about 50% while preserving fidelity. The full paper can be found on arXiv, reference 2605.06124.