Face perception is based on the processing and integration of multiple spatial frequency (SF) ranges. However, the temporal dynamics of SF integration to form an early face representation in the human brain is still a matter of debate. To address this issue, we recorded event-related potentials (ERPs) during the presentation of spatial frequency-manipulated facial images. Twenty-six participants performed a gender discrimination task on non-filtered, low-, high-, and band-pass filtered face images, corresponding, respectively, to the full range, spatial frequencies up to 8 cycles/image, above 32 cycles/image, and from 8 to 16 cycles/image. Behaviorally, the task related-performance was more accurate and faster for non-filtered (NF) and mid-range SF (MSF) than for low SF (LSF) and high SF (HSF) stimuli. At both behavioral and electrophysiological levels, response to MSF contained in faces did not differ from the responses to full spectrum non-filtered (NF) facial images. In ERPs, LSF facial images evoked the largest P1 amplitude while HSF facial images evoked the largest N170 amplitude compared with the other three conditions. Since LSFs and HSFs would transmit global and local information respectively, our observations lend further support to the “coarse-to-fine” processing theory of faces. Furthermore, they offer original evidence of the effectiveness and adequacy of the mid-range spatial frequency in face perception. Possible theoretical interpretations of our findings are discussed.