Distributional Diffusion Models with Scoring Rules

Valentin De Bortoli, Alexandre Galashov, J. Swaroop Guntupalli, Guangyao Zhou, Kevin Murphy, Arthur Gretton, Arnaud Doucet · 2025

Read Distributional Diffusion Models with Scoring Rules by Valentin De Bortoli, Alexandre Galashov, J. Swaroop Guntupalli, Guangyao Zhou, Kevin Murphy, Arthur Gretton, Arnaud Doucet (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-2502-02483.txt.

Description

Valentin De Bortoli and six co-authors offer a faster way to run diffusion models, in a paper from twenty twenty-five. These models make new data by slowly removing noise, step by step. Fine steps give good results but cost much time, so many shortcuts have been tried.

Instead of learning only the average clean sample behind a noisy one, their method learns the full spread of likely clean samples. That fuller picture allows much bigger steps during generation with little loss of quality. A scoring rule replaces the usual training loss to make this work. Tests on image making and robot path planning beat standard models when only a few steps are allowed.

Catalog record

CREATORValentin De Bortoli, Alexandre Galashov, J. Swaroop Guntupalli, Guangyao Zhou, Kevin Murphy, Arthur Gretton, Arnaud Doucet
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE141 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-02483

Browse the full library · About the Tower of Babel