ODE Solvers

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

This research paper systematically benchmarks four classical ODE solvers (Euler, Explicit Midpoint, RK4, Dormand-Prince 5(4)) for Flow Matching generative models, implementing them from scratch in PyTorch. It quantitatively compares their efficiency on tasks from 2D distributions to MNIST, showing RK4 at 80 function evaluations achieves sample quality comparable to Euler at 200, and observes Jacobian eigenvalue spectrum stiffening near t=1.