Psychophysical thresholds were measured for discriminating small changes in spatial features of naturalistic scenes (morph sequences), for foveal and peripheral vision, and under M-scaling. Sensitivity was greatest for scenes with near natural Fourier amplitude slope, perhaps implying that human vision is optimised for natural scene statistics. A low-level model calculated differences in local contrast between pairs of images within a few spatial frequency channels with bandwidth like neurons in V1. The model was “customised” to each observer’s contrast sensitivity function for sinusoidal gratings, and it could replicate the “U-shaped” relationships between discrimination threshold and spectral slope, and many differences between picture sets and observers. A single-channel model and an ideal-observer analysis both failed to capture the U-shape.