Guiding attention during search using generative target template matching

Template matching, the idea that attention is guided by a specific internal representation of a target, provides a powerful theoretical foundation for understanding visual search. Yet, it remains unclear how such templates operate in naturalistic contexts where target objects may be ambiguous or vary in appearance. We propose that visual search is better explained by generative templates—representations conditionally shaped by features of the current visual input rather than fixed, averaged prototypes. Using deep image synthesis, we generate input-specific target templates and compare their search guidance to conventional fixed templates. Generative templates yield more accurate target localization in naturalistic search arrays and better capture key patterns in human behavior, including search efficiency, fixation probability, and attentional guidance toward structurally similar non-targets. Our input-driven generative template approach offers a flexible alternative to classical template-matching theories, Our input-driven generative template approach offers a flexible alternative to classical template-matching theories, one that can better explain how human categorical search is affected by the variability in visual features both within and between natural object categories.