Generative Modeling with Flux Matching

Peter Pao-Huang, Xiaojie Qiu, Stefano Ermon · 2026

Read Generative Modeling with Flux Matching by Peter Pao-Huang, Xiaojie Qiu, Stefano Ermon (2026) 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-2605-07319.txt.

Description

Peter Pao-Huang, Xiaojie Qiu, and Stefano Ermon introduce Flux Matching, a generative modeling framework that loosens the rules of score-based methods. Standard training forces the learned vector field to equal the data's score exactly. Flux Matching imposes a weaker condition, admitting endlessly many fields whose stable distribution is still the data distribution.

That freedom lets designers build in structural priors and directed dependencies between variables directly. The method performs strongly on high-dimensional image datasets, according to the authors. They also report gains in sampling speed and interpretability, arguing the vector field itself becomes a design choice rather than a fixed target for generative modeling.

Catalog record

CREATORPeter Pao-Huang, Xiaojie Qiu, Stefano Ermon
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE99 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.07319

Browse the full library · About the Tower of Babel