We evaluated a dictionary learning (DL) method for seismic-data denoising. The data were divided into smaller patches, and a dictionary of patch-size atoms was learned. The DL method offers a more flexible framework to adaptively construct sparse data representation according to the seismic data themselves. The representation being learned from the data, did not rely on a guess of the data morphology like standard wavelet or curvelet transforms. The method could learn a dictionary and denoise seismic data, whether simultaneously or in two distinctive steps. Empirical study on field data showed promising denoising performance of the presented method in terms of signal-to-noise ratio and weak-feature preservation, in comparison with wavelets, curvelets, anisotropic total variation, and nonlocal total variation.