Flow matching loss varies across time steps due to both fundamental ambiguity and data scarcity—VAP separates these, letting you identify where better data helps versus where the problem is inherently hard, improving efficiency on real generative tasks.
This paper introduces Velocity Ambiguity Profile (VAP), a diagnostic tool that breaks down flow matching training loss into two components: an irreducible Bayes-risk floor (intrinsic ambiguity) and a reducible data-limited error.