The structural tensor method can be used to compute dips and azimuths (i.e., orientation) encased in seismic data. However, this method may produce erratic and uninterpretable orientations when noisy data are encountered. To overcome this difficulty, we incorporate a data-adaptive weighting function to reformulate the gradient structural tensor. In our experiment, the squared instantaneous power is adopted as the weight factor; this can simplify the computation when the instantaneous phase is used as input. The real data examples illustrate that such a weighting function can produce more interpretable and spatially consistent orientations than conventional approaches.