Research Editor selected
Elliptic Sampling: Optimal Interpolant Schedules for Gaussians and Beyond
This paper introduces a framework for jointly designing interpolant schedules and discretizations in stochastic interpolant-based generative modeling, using analyticity of complexified ODE trajectories. It yields variance-preserving elliptic schedules parameterized by Jacobi elliptic functions and a sampling algorithm called Zoloflow, which achieves minimax optimality for Gaussian targets and improved complexity for log-concave targets. Empirical tests on non-log-concave distributions and image datasets show practical viability.
Why it matters: Provides a novel theoretical framework with optimal sampling guarantees and practical algorithm, advancing generative modeling efficiency.
generative modelingsampling algorithmsstochastic interpolants