Observers estimate a range of social characteristics from images of human faces. An important unifying framework for these judgments is the observation that a low-dimensional social face-space based on perceived valence and dominance captures most of the variance across a wide range of social evaluation judgments. However, it is not known whether or not this low-dimensional space can be used to infer the outcome of new social judgments. Further, the extent to which such social inference may differ across real and computer-generated faces is also unknown. We addess both of these issues by recovering valence/dominance axes from social judgments made to real and artificial faces, then attempt to use these coordinates to predict independent social judgment data obtained from new human observers. We find that above-chance performance can be achieved, though performance appears to be better with artificial faces than real ones.