Structures will be mispositioned across prestack migration gathers in the presence of inaccuracies in the velocity model. Stacking these misaligned gathers runs the risk of destroying important structures in the stacked migration image. To mitigate this problem, we propose a trim statics inspired by the non-local means algorithm originally developed for image denoising. This method differs from the conventional one in two fundamental respects. First, the trim statics are computed by comparing image patches instead of individual image traces. Second, no global pilot trace is needed because only two migration images at a time participate in trim statics and are stacked into one image. A multitude of migration images are stacked recursively in this two-to-one fashion. Tests with a Gulf of Mexico dataset show a noticeable improvement in the feature coherency of the stacked migration image.