Direction-discrimination performance was measured for two-frame random-dot kinematograms in which one or both frames were spatial frequency filtered with a one octave band-pass filter and the centre frequency of this filter was varied in the range 0.75-9 c/deg independently for each frame. When both frames were filtered so that they contained common (overlapping) spatial frequencies direction discrimination was extremely good but it deteriorated rapidly as the degree of spectral overlap between the two frames decreased. These results are consistent with previous findings that suggest that the mechanisms that mediate the initial stages of motion detection are narrowly tuned for spatial frequency and cannot combine information conveyed at disparate frequencies in order to compute an unambiguous estimate of the direction of local motion. However, when only one of the frames was band-pass filtered and the other was unfiltered (broadband), the correct direction of stimulus motion could be discriminated reliably for a broad range of filter centre frequencies. Performance was best when the centre frequency of the filtered frame was at medium spatial frequencies and tended to deteriorate as the centre frequency approached either extreme of the spatial frequency range examined. This basic pattern of results may be attributed to the visual system's differential sensitivity to the Fourier components present in the unfiltered frame.