Fast Sampling for Flows and Diffusions with Lazy and Point Mass Stochastic Interpolants

Gabriel Damsholt, Jes Frellsen, Susanne Ditlevsen · 2026

Read Fast Sampling for Flows and Diffusions with Lazy and Point Mass Stochastic Interpolants by Gabriel Damsholt, Jes Frellsen, Susanne Ditlevsen (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-2602-03789.txt.

Description

Gabriel Damsholt and 2 co-authors study the schedules that steer generative sampling in flow and diffusion models. Their 2026 paper works in the stochastic interpolant framework, which unifies the two families. A schedule dictates how a Gaussian starting distribution morphs into the target, and the choice strongly affects sampling speed.

The authors prove how to convert a sample path between arbitrary schedules and diffusion coefficients. They extend the framework to point mass schedules, where the Gaussian start collapses to a single point. For Gaussian data they identify lazy schedule families whose drift vanishes entirely. Applied to a pretrained flow model, their schedule conversion generates images in fewer steps without any retraining.

Catalog record

CREATORGabriel Damsholt, Jes Frellsen, Susanne Ditlevsen
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE100 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-03789

Browse the full library · About the Tower of Babel