We present a new approach to structural interpretation of 3D seismic data with the objectives of simplifying the task and reducing the interpretation time. The essential element is the stepwise removal of noise, and eventually of small-scale stratigraphic and structural features, to derive more and more simple representations of structural shape. Without noise and small-scale structure, both man and machine (autotrackers) can. arrive at a structural interpretation faster. If the interpreters so wish, they can refine such an initial crude structural interpretation in selected target areas. We discuss a class of filters that removes noise and, if desired, simplifies structural information in 3D seismic data. The gist of these filters is a smoothing operation parallel to the seismic reflections that does not operate beyond reflection terminations (faults). These filters therefore have three ingredients: (1) orientation analysis, (2) edge detection, and (3) edge-preserving oriented smoothing. We discuss one particular implementation of this principle in some detail: a simulated anisotropic diffusion process (low-pass filter) that diffuses the seismic amplitude while the diffusion tensor is computed from the local image structure (so that the diffusion is parallel to the reflections). Examples show the remarkable effects of this operation.