Diffracted multiples are a persistent problem, especially in the cross-line direction. Methods to remove these multiples typically require dense and complete sampling. I use non-stationary prediction-error filters to interpolate extra shots in the in-line direction and extra receivers in the cross-line direction. The dimensionality of the interpolation does not make a big difference in the in-line direction, but makes a large difference in the cross-line direction. This is shown with several tests on a synthetic dataset.