Fig. 1. Generation of synthetic texture stimuli.

a Original photographs (first column) are decomposed into V1-like responses (rectified oriented linear filter outputs) differing in orientation and spatial frequency (second column). Response statistics are computed by spatially averaging the products of responses at different orientations, spatial frequencies, and local positions (third column)²³. b Images of Gaussian white noise (top) are iteratively adjusted until their statistics match those of the original images. Initializing with different random seeds yields different samples with identical statistics but differing in detail (columns), and different statistics yield samples from different texture families (rows). c Statistics converge with increasing measurement region width. Median coefficient of variation (standard deviation across samples divided by mean across samples) of example groups of higher-order statistics, as a function of the width of the region over which the statistics are measured. Products of simple cell responses (linear responses; left) and complex cell responses (magnitude responses; right) were measured from regions cropped from images synthesized to be statistically matched across their full extent (512 × 512 pixels). Shaded regions indicate interquartile range across different individual statistics computed from 15 texture families. Source data are provided as a Source Data file.