Hyunwoo Gu¹, Justin L. Gardner¹; ¹Stanford University

Classical attentional capture studies using simplified search arrays have shown that goal- and salience-driven systems can be placed in conflict, making search less efficient. However, unlike artificial arrays, salience and target-relevant features in natural scenes show complex statistical relationships that vary between concordance and conflict. This raises a key question about which system exerts greater influence on gaze selection, and how their influences evolve under conflict in natural scenes. We analyzed published natural visual search datasets using saliency models and vision-language model template matching, finding that human fixation choices are biased by salience features. Furthermore, search templates directly estimated from gaze patterns recapitulated this bias, showing a tendency to be similar to salience templates derived from free-viewing. Using deep stimulus synthesis, we further demonstrated that this salience bias can be used to control gaze selection patterns during visual search. Consistent with classical attentional capture, we found that mismatch between salience and target features lowers search performance. However, capture was not static; we observed that salience bias was strongest during initial fixations and attenuated as search progressed. To further characterize these dynamics, we fit an observer model that chose fixations from either salience or target maps at each step. We found that the initial probability for salience was higher but the probability of remaining in the target mode increased, indicating a tendency to switch from salience to target drive as search progressed. Finally, analyses of natural stimuli revealed that salience features alone yield above-chance search performance. This stands in contrast to classical search arrays, which are often designed to make salient distractors task-irrelevant, thereby motivating observers to suppress them. Thus, in natural scenes, salience serves as an informative signal that helps constrain search, justifying the initial bias despite the potential for capture.