We develop an image-driven approach to the question of what makes the shape of a woman's body attractive. We constructed a set of 625 images of female bodies by factorially recombining four independent descriptors of shape derived from a principal components analysis of the variation in natural body shape, and had observers rate these images for attractiveness. We then modelled observers' attractiveness ratings with polynomial multiple regression, using the same shape descriptors as explanatory variables. The resulting model agrees well with existing models based on simple anthropometric indices of shape; however, some interesting new findings emerge. There was considerable variation in the shape of bodies that were judged to be equally attractive. Further experiments confirmed that observers could detect these subtle variations in shape suggesting a dissociation between attractiveness judgement and shape discrimination.