Generative Modeling via Drifting

Mingyang Deng, He Li, Tianhong Li, Yilun Du, Kaiming He · 2026

Read Generative Modeling via Drifting by Mingyang Deng, He Li, Tianhong Li, Yilun Du, Kaiming He (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-2602-04770.txt.

Description

Mingyang Deng and 4 co-authors propose a new way to train image generators in a 2026 paper on drifting models. Generative modeling learns a mapping whose output distribution matches the data distribution. Diffusion and flow models apply that mapping iteratively at generation time, which costs many steps.

Drifting models instead evolve the output distribution during training itself, so generation takes one step. A drifting field governs how samples move, reaching equilibrium when the distributions match. That idea yields a training objective letting the optimizer reshape the distribution directly. On a standard high-resolution image benchmark, their one-step generator achieved state-of-the-art quality scores in both latent and pixel space. The authors hope this opens fresh routes to fast generation.

Catalog record

CREATORMingyang Deng, He Li, Tianhong Li, Yilun Du, Kaiming He
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE85 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-04770

Browse the full library · About the Tower of Babel