Continuous Adversarial MeanFlow Transfer

Training fast generators on new domains with limited data is challenging due to costly multi-step sampling and the lack of common acceleration targets for heterogeneous pretrained models. Additionally, adversarial refinement typically supports instantaneous-velocity flows rather than the finite-interval average velocities predicted by MeanFlow (MF) models. To resolve these issues, the researchers propose MeanFlow-Transfer, a method designed to map heterogeneous source outputs and improve few-step quality through a continuous adversarial framework for better domain adaptation.