Psychophysical studies have shown that the human face recognition system utilizes only a limited band of the spatial frequencies (sfs) present in facial images. To find cortical responses that would show a similar tuning, we recorded whole-scalp neuromagnetic responses to facial images that contained narrow-band spatial noise (10 noise bands with central sfs from 2 to 45 c/image). The signal-to-noise ratio was 0.74 across all sfs and made recognition impossible around the 16 c/image condition, but easy in the low and high sf conditions. The stimuli (7×7 deg2, mean luminance 130 cd/m2, RMS contrast 0.174) appeared abruptly on an average gray background for 0.5 s once every 2.5 s, and stimuli with different noise sfs were presented in random order. The subject had to respond to an image of a target person. In the six subjects, the strongest modulation by noise sf occurred in the 150–200 ms temporo-occipital responses: The largest signals were elicited by the noiseless and lowest-sf noise stimuli. The signals decreased as a function of increasing noise sf until they disappeared at the medium sfs, and increased again at the highest frequencies. For different subjects, the U-shaped tuning curve had a minimum at 5.6–23 c/image, and a bandwidth of 2–3 octaves. The sustained 250–650 ms response in the same cortical area was modulated less. The 80–110 ms midoccipital responses showed a different pattern, being weak for the lowest noise sfs, then increasing as a function of sf, and attenuating at the highest noise sfs. In conclusion, these results show that the 150–200 ms temporo-occipital response resembles behavioral face recognition in its sensitivity to noise sf. Thus it could reflect processes that are critical for the recognition of faces.