Tal Boger¹, Chaz Firestone¹; ¹Johns Hopkins University

Even when completely motionless, some objects look animate (e.g., dogs and elephants) while others don’t (e.g., boots and sofas). A rich literature suggests that the mind automatically encodes this property, reporting striking effects of perceived animacy on visual attention and working memory. However, objects that differ in animacy tend to differ in many lower-level features (e.g., shape), and follow-up work has revealed that such confounds often account for these seemingly high-level effects. Thus, whether animacy itself drives visual processing remains unanswered and even controversial. Here, we take a new approach to this question by exploiting “visual anagrams”: static images whose interpretations change radically with orientation. We used a diffusion model to generate such images in ways that varied animacy — e.g., a dog in one orientation and a boot when rotated, or an elephant in one orientation and a sofa when rotated. Each anagram image contains the exact same pixels in either orientation, such that animacy varies across the two interpretations while nearly all lower-level features remain constant. 9 experiments used this approach to demonstrate that animacy itself guides memory and attention. Experiments 1–2 found that changes to an object in a memory array were more detectable when they altered animacy (e.g., dog→sofa is more detectable than dog→elephant, even when the elephant and sofa are the very same image, just rotated). Experiments 3–6 found that animate targets in a search array were easier to find among inanimate distractors than among other animate distractors (and vice versa). Finally, Experiments 6–9 verified that differences in orientation (one of the only lower-level features uncontrolled by visual anagrams) cannot account for these results; the effects disappeared when using silhouetted, blurred, and pixelated versions of the anagrams. Thus, visual processing extracts animacy itself, over and above its covarying lower-level features.