Abstract
Flow-based generative models have been employed for Boltzmann sampling tasks, but their application to high-dimensional systems is hindered by the significant computational cost of obtaining the Jacobian of the flow. We introduce a flow perturbation method that bypasses this bottleneck by injecting stochastic perturbations into the flow, delivering orders-of-magnitude speed-ups. Unlike the Hutchinson estimator, our approach is inherently unbiased in Boltzmann sampling. Notably, this method significantly accelerates Boltzmann sampling of a Chignolin mutant with all atomic Cartesian coordinates explicitly represented, while delivering more accurate results than the Hutchinson estimator.