Human texture discrimination depends both on spatial-frequency content and on higher-order or multi-point correlations. Spatial- frequency discrimination exhibits a high degree of scale invariance over a range of several octaves, but the scaling behavior of sensitivity to higher-order correlation structure is unknown. We explored the scale dependence of texture discrimination for image ensembles which shared the same power spectrum, but differed in their higher-order correlations. Literally scaling the ensembles so that they occupy larger retinal regions results in discrimination performance that is largely independent of scale over a three octave range. Holding the display size constant and scaling the texture being sampled within the display over the same range produces performance that varies with scale appreciably. The ideal observer performance is computed, and the absolute efficiency is seen to be quite small, on the order of 0.01 to 0.1. As the texture is scaled down, increasing the number of checks within the fixed display size, performance increases while the efficiency decreases. These dependencies remain after the stimulus onset asynchrony is increased from 50 to 500 msec. We created sets of textures which varied both in check number and correlation strength, for which ideal observer performance was equated. For the human observers, efficiency was significantly higher for textures with higher correlation strength, but fewer checks. These results are consistent with a model in which a fixed number of checks is processed in a scale-invariant manner, while the remainder of the display is processed much less efficiently.