We measured the upper threshold for directional motion discrimination (Dmax) in two-frame random binary luminance patterns (random dot kinematograms) in which either one or both frames was spatially low-pass filtered by convolution with a Gaussian filter. When both frames were low-pass filtered, Dmax increased as a function of the standard deviation of the Gaussian blurring function, in agreement with previous findings. However, when only one of the two frames was blurred, Dmax showed little change with blurring space constants below about 20 min arc, and at larger space constants motion discrimination became impossible. We take this as evidence against the proposal that Dmax is preferentially determined by motion signals from high spatial frequencies; and as evidence for the alternative that Dmax depends upon the mean spatial interval between features in the pattern after a single stage of spatial frequency pre-filtering. The breakdown in motion discrimination for space constants above about 20 min arc can be predicted from the computed effects of blurring upon the correlation between features (zero-bounded regions) in the broad-band and spatially filtered patterns. At values of blur where motion discrimination began to collapse there was a temporal order asymmetry; discrimination was easier when the low-pass pattern preceded the broadband pattern than when the broadband pattern appeared first. We propose that the temporally sustained high spatial frequency signal in the broadband pattern is delayed relative to the more transient low frequency signal; or alternatively, that the inhibitory surround of the spatial prefilter is switched in after a delay relative to the excitatory centre. The processing-delay interpretation was tested and confirmed in a second experiment by manipulating the frame duration.