The goal of this paper is to propose a unified framework for nonlinear refinement of optical flow. The first model is a two-phase refinement process where an initial estimate obtained in the first phase is subsequently refined in the second phase using additional constraints. We study the mathematical well-posedness of this formulation using an evolutionary-PDE approach. The second model is a single-phase improvement process involving an anisotropic regularization of the curl of the flow. For implementation we use the first-order primal-dual Chambolle-Pock algorithm. We observe that the results obtained by both methods are comparable in nature. We perform several numerical experiments and empirically demonstrate that by using a two-phase refinement process, a faster convergence rate of the order O(1/N) is achieved than the single-phase process which has a higher convergence rate of the order O(1/N^2).