Towards High-Order Mean Flow Generative Models: Feasibility, Expressivity, and Provably Efficient Criteria

Yang Cao and Yubin Chen and Zhao Song and Jiahao Zhang · 2025

Read Towards High-Order Mean Flow Generative Models: Feasibility, Expressivity, and Provably Efficient Criteria by Yang Cao and Yubin Chen and Zhao Song and Jiahao Zhang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-07102.txt.

Description

Image-making models that flow noise into pictures can now work in one step, by chasing an average speed instead of an instant one. Yang Cao and three co-authors push that trick one order up in 2025: chase the average acceleration too. Three proofs carry the paper.

First, the average push obeys a sameness law like its first-order kin, so one-step sampling stays sound and the training loss stays tame. Second, the sampling act fits in a shallow circuit class, a known mark of true cheapness. Third, fast approximations give a buildable path. The aim is richer births at the same one-step cost. Theory leads here, with the bench tests as the stated next step.

Catalog record

CREATORYang Cao and Yubin Chen and Zhao Song and Jiahao Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE102 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-07102

Browse the full library · About the Tower of Babel