A key issue in wavefield separation is to find a domain where the signal and coherent noise are well separated from one another. A new wavefield separation algorithm, called migration filtering, separates data arrivals according to their path of propagation and their actual moveout characteristics. This is accomplished by using forward modeling operators to compute the signal and the coherent noise arrivals. A linearized least-squares inversion scheme yields model estimates for both components; the predicted signal component is constructed by forward modeling the signal model estimate. Synthetic and field data examples demonstrate that migration filtering improves separation of P-wave reflections and surface waves, P-wave reflections and tube waves, P-wave diffractions, and S-wave diffractions. The main benefits of the migration filtering method compared to conventional filtering methods are better wavefield separation capability, the capability of mixing any two conventional transforms for wavefield separation under a general inversion framework, and the capability of mitigating the signal and coherent noise crosstalk by using regularization. The limitations of the method may include more than an order of magnitude increase in computation costs compared to conventional transforms and the difficulty of selecting the proper modeling operators for some wave modes.