This paper evaluates four featural models of stimulus similarity using data collected for a set of 16 nations. Algorithms are developed for finding stimulus representations, and the important issue of balancing data-fit against model complexity is addressed by using the Geometric Complexity Criterion. Although the data clearly incorporate both common and distinctive features, Tversky’s (1977) Contrast Model seems unable to express these regularities in an appropriate manner. However, we show that a new version of the Contrast Model that treats each feature as either being common or distinctive is better able to capture the essential aspects of the similarity judgments.