Animals actively select visual stimuli by orienting their eyes. In humans, eye positions determine which portions of a visual scene will fall on the fovea and thus be sampled at high spatial resolution. We recorded the eye positions of human subjects while they viewed black and white photographs of real world scenes. We find that this sampling process is not uniform with respect to low-level statistics of the images, such as the power spectrum. The differences reflect a preference for looking at statistical outliers according to local low-level properties. Using a wavelet-based measure of image entropy, we show that this results in an overall increase in the entropy in the foveal visual field. It would be advantageous for the visual system to be adapted to exploit the priors on low level statistics of its natural stimuli. We suggest that eye movements may be part of this adaptation. Priors on the intrinsic statistics of images could be used to orient the eyes toward unexpected signals, and thereby present sensory neurons with a less redundant input ensemble. Visual neurons may then be adapted to the statistics of this ensemble, rather than to the statistics of scenes themselves.