From Euler to Dormand–Prince: ODE Solvers for Flow Matching Generative Models

Hao Xiao · 2026

Read From Euler to Dormand–Prince: ODE Solvers for Flow Matching Generative Models by Hao Xiao (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-2605-00836.txt.

Description

Hao Xiao compares four classic solvers for the ordinary differential equations used when sampling from flow matching generative models. Sampling cost is dominated by neural network passes, so a better solver can save many evaluations. Xiao derives Euler, midpoint, classical Runge-Kutta, and Dormand-Prince methods from first principles, codes them from scratch in PyTorch, and benchmarks them from toy distributions to MNIST digits.

Runge-Kutta at eighty evaluations matches Euler at two hundred in sample quality. Two further findings stand out. The learned velocity field stiffens sharply near the end of the path, which explains why the adaptive solver spends steps there. The solver gap also widens for smaller or undertrained models, so solver choice matters most when the model is imperfect.

Catalog record

CREATORHao Xiao
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE20 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.00836

Browse the full library · About the Tower of Babel