We simulate transport of a solute through three-dimensional images of different rock samples, with resolutions of a few microns, representing geological media of increasing pore-scale complexity: a sandpack, a Berea sandstone, and a Portland limestone. We predict the propagators (concentration as a function of distance) measured on similar cores in nuclear magnetic resonance experiments and the dispersion coefficient as a function of Péclet number and time. The behavior is explained using continuous time random walks with a truncated power-law distribution of travel times: transport is qualitatively different for the complex limestone compared to the sandstone or sandpack, with long tailing, an almost immobile peak concentration, and a very slow approach to asymptotic dispersion.