A paradoxical finding from recent studies of face perception is that observers are error-prone and inconsistent when judging the identity of unfamiliar faces, but nevertheless reasonably consistent when judging traits. Our aim is to understand this difference. Using everyday ambient images of faces, we show that visual image statistics can predict observers' consensual impressions of trustworthiness, attractiveness and dominance, which represent key dimensions of evaluation in leading theoretical accounts of trait judgement. In Study 1, image statistics derived from ambient images of multiple face identities were able to account for 51% of the variance in consensual impressions of entirely novel ambient images. Shape properties were more effective predictors than surface properties, but a combination of both achieved the best results. In Study 2 and Study 3, statistics derived from multiple images of a particular face achieved the best generalisation to new images of that face, but there was nonetheless significant generalisation between images of the faces of different individuals. Hence, whereas idiosyncratic variability across different images of the same face is sufficient to cause substantial problems in judging the identities of unfamiliar faces, there are consistencies between faces which are sufficient to support (to some extent) consensual trait judgements. Furthermore, much of this consistency can be captured in simple operational models based on image statistics.