Shape Guides Visual Pretense

People often imagine everyday objects are something else. A turned over bottle becomes a car, a teapot becomes a swan. Such pretense is common in play, pedagogy, and narratives. The relationship between a real and pretend object is flexible, but not arbitrary. In this work, we used a behavioral and computational approach that compares people and performant multi-modal vision models to study the features that guide the construction of visual pretense. In four studies (N = 716 in total), we show that people have systematic preferences in visual pretense, and that these preferences are better accounted for by spatial and physical alignment (specifically shape similarity), over surface feature similarity (such as color). We also found that people systematically align the subpart structure of real and pretend objects. We further show that people's visual pretense preferences are not accounted for by current common approaches to multi-modal vision models, likely due to their reliance on surface features rather than spatial and physical ones.