The authors formalize VAP as the conditional velocity variance averaged over the flow interpolant distribution, and prove it equals the posterior action variance rescaled by 1/t^2.
Time-resolved diagnostic separates irreducible velocity ambiguity from reducible training error
Mei et al. introduce Velocity Ambiguity Profiles, which decompose flow-matching loss into a Bayes-risk floor and an excess term, and show that guided time sampling improves CIFAR-100 feature generation at a measurable diversity cost.
Chinese Tech
Yunpeng Mei · Xiaowen Zhu · Chenyu Wang · Chenbo Xin · Hongjie Cao · Jiamin Wang · +2 more
Beijing Institute of Technology · ByteDance
Research Digest··3 min read
Mei et al.
§