To improve seismic inversion, we propose to use prior information from empirical estimates of spatial rock property distributions obtained from outcrop training images. Using photographic images of chalk outcrops, we derive empirical distributions of spatial rock property gradients and estimates of spatial correlations. We use this information as a priori information in probabilistic, seismic inversion. Our results are realistic, high-resolution posterior samples of subsurface models, whose variability can be used as an estimate of model uncertainties.