Stable Velocity: A Variance Perspective on Flow Matching
Stable Velocity analyzes flow matching dynamics, revealing a two-regime structure: high variance near prior and low variance near data. It proposes StableVM for unbiased training variance reduction, VA-REPA for selective auxiliary supervision in the low-variance regime, and StableVS for finetuning-free inference acceleration. This framework improves training stability and sampling speed without sacrificing sample quality.