AI BriefingFlow Matching That Follows the Mean
Key point
Proposed a new Flow Matching method that steers the generative flow using a reference set.
Details
Leveraging the fact that in deterministic interpolants the velocity field is determined by the conditional endpoint mean, the paper proposes a way to steer the generative flow simply by changing the reference set.
Reference-Mean Guidance (RMG) is a training-free approach that computes a closed-form endpoint-mean correction from a reference bank and applies it to the frozen FLUX.2-klein (4B) model. This controls color, identity, style, and structure while leaving the prompt, seed, and weights unchanged.
Semi-Parametric Guidance (SPG) absorbs the same idea into the model itself by adding an explicit mean anchor and a learned residual refiner. Experiments showed that on AFHQv2, it matches the quality of an unconditional DiT-B/4, while still allowing the reference set to be swapped at inference time.
The key point is a new control interface that adapts a generative model using only example data, without any parameter updates.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.