To clarify the relationship between face perception and cortical activation, we manipulated the face recognition performance of 9 subjects by varying the duration (DUR) of stimuli while cortical neuromagnetic responses were recorded. A face image replaced a continuous pixel-noise mask for 17-200 ms, and the subject reported which of the pre-learned faces had been presented. Two cortical responses were clearly stronger to intact than phase-scrambled faces: the temporo-occipital response peaking at 140-200 ms (M170) and a more widely distributed response peaking at 200-500 ms (M300). For the shortest DURs (17-33 ms), face recognition was at chance level and the cortical responses negligible. For DURs of 50-83 ms, the proportion of recognized faces as well as the strength of cortical responses increased steeply. Recognition performance saturated at DURs of around 100 ms, whereas cortical responses continued to increase until the longest DUR of 200 ms. Amplitudes of both M170 and M300 were thus tightly correlated with recognition performance (r=0.98), but comparison of the increment rates as a function of DUR showed the recognition performance to have an even closer similarity to M170 than to M300. In single-trial analysis the variability of response strengths increased in a direct proportion to response amplitude, demonstrating the averaged responses to be composed of graded rather than of all-or-nothing-type single responses.