Abstract
Flow matching has emerged as a compelling generative modeling approach thatis widely used across domains. To generate data via a flow matching model, anordinary differential equation (ODE) is numerically solved via forwardintegration of the modeled velocity field. To better capture the multi-modalitythat is inherent in typical velocity fields, hierarchical flow matching wasrecently introduced. It uses a hierarchy of ODEs that are numericallyintegrated when generating data. This hierarchy of ODEs captures themulti-modal velocity distribution just like vanilla flow matching is capable ofmodeling a multi-modal data distribution. While this hierarchy enables to modelmulti-modal velocity distributions, the complexity of the modeled distributionremains identical across levels of the hierarchy. In this paper, we study howto gradually adjust the complexity of the distributions across different levelsof the hierarchy via mini-batch couplings. We show the benefits of mini-batchcouplings in hierarchical rectified flow matching via compelling results onsynthetic and imaging data. Code is available athttps://riccizz.github.io/HRF_coupling.