Contrast sensitivity as a function of spatial frequency was determined for 138 neurons in the foveal region of primate striate cortex. The accuracy of three models in describing these functions was assessed by the method of least squares. Models based on difference-of-Gaussians (DOG) functions were shown to be superior to those based on the Gabor function or the second differential of a Gaussian. In the most general case of the DOG models, each subregion of a simple cell’s receptive field was constructed from a single DOG function. All the models are compatible with the classical observation that the receptive fields of simple cells are made up of spatially discrete ‘on’ and ‘off’ regions. Although the DOG-based models have more free parameters, they can account better for the variety of shapes of spatial contrast sensitivity functions observed in cortical cells and, unlike other models, they provide a detailed description of the organization of subregions of the receptive field that is consistent with the physiological constraints imposed by earlier stages in the visual pathway. Despite the fact that the DOG-based models have spatially discrete components, the resulting amplitude spectra in the frequency domain describe complex cells just as well as simple cells. The superiority of the DOG-based models as a primary spatial filter is discussed in relation to popular models of visual processing that use the Gabor function or the second differential of a Gaussian.