Seismic inversion is now the focus of much effort in our industry. Its performance depends critically on the response of the abstract model chosen to mimic the real earth and on the optimization criterion employed to match this response to the actual seismogram. There are two broad categories: travel-time inversion and amplitude inversion. When we invert for travel-times, we seek a subsurface model whose event arrival times best match picked arrivals on the real data. When we invert for amplitudes, we search for a subsurface model whose seismic signature best matches the entire recorded data, and not merely the arrival times. Of the two methods, travel-time inversion is by far the simplest and is currently in widespread use. Amplitude inversion is more general because it makes full use of the recorded data, but implementation is still very costly and the approach has not yet seen extensive practical use. Neither technique can give u s the unique subsurface model-that is, there will always be a multiplicity of subsurface configurations whose responses match the data within a desired tolerance. Physical insight is required to rule out those models devoid of geological significance.