Burton and Bruce's (1992) model of face naming predicts a "fan effect", in which naming of famous people about whom many descriptive properties are known should be slower than naming of celebrities about whom few properties are known. An experiment is reported that showed that, contrary to this prediction, knowledge of many descriptive properties facilitated face-naming latency. An alternative architecture for an interactive activation model is proposed in which descriptive properties are represented in separate pools of units for each domain of information and in which names are represented by a separate pool of lexical output units. Computer simulations showed that this model could simulate the previously available empirical data as effectively as Burton and Bruce's (1992) original model. However, the proposed model could also simulate the effect of the number of known descriptive properties upon face-naming latency observed in the experiment reported. The new architecture also has the advantage of being more compatible with current models of speech production, and it allows preserved access to unique semantic properties in the context of impaired face naming as reported in the neuropsychological literature.