MaRS: A Fast Sampler for Mean Reverting Diffusion Based on ODE and SDE Solvers

Ao Li 1,2 Wei Fang 3,4∗ Hongbo Zhao 1,2∗ Le Lu 3 Ge Yang 1,2 Minfeng Xu 3,4† 1 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 2 Institute of Automation, Chinese Academy of Sciences (CASIA) 3 DAMO Academy, Alibaba Group 4 Hupan Laboratory, Hangzhou, China · 2025

Read MaRS: A Fast Sampler for Mean Reverting Diffusion Based on ODE and SDE Solvers by Ao Li 1,2 Wei Fang 3,4∗ Hongbo Zhao 1,2∗ Le Lu 3 Ge Yang 1,2 Minfeng Xu 3,4† 1 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 2 Institute of Automation, Chinese Academy of Sciences (CASIA) 3 DAMO Academy, Alibaba Group 4 Hupan Laboratory, Hangzhou, China (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-2502-07856.txt.

Description

Ao Li and co-authors speed up in this 2025 paper a diffusion model family prized for guided image repair. Mean-reverting diffusion builds the task's condition into its equations, but sampling it took hundreds of steps. Their sampler, MaRS, solves the reverse equations and the matching flow in semi-analytic form, splitting each step into a closed part and a learned integral.

No retraining is needed, and the recipe accepts all standard model output styles. On ten restoration tasks, from deblurring to inpainting, quality held while speed rose ten to twenty fold. The authors pitch the method as making controllable generation practical. Photo apps gain the most from the speed.

Catalog record

CREATORAo Li 1,2 Wei Fang 3,4∗ Hongbo Zhao 1,2∗ Le Lu 3 Ge Yang 1,2 Minfeng Xu 3,4† 1 School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing, China 2 Institute of Automation, Chinese Academy of Sciences (CASIA) 3 DAMO Academy, Alibaba Group 4 Hupan Laboratory, Hangzhou, China
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE80 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-07856

Browse the full library · About the Tower of Babel