Does the amount of information stored in visual short-term memory (VSTM) reflect the specific demands of the ongoing task or does the system routinely make full use of its capacity? Eye movement behavior in multi-modal (i.e., visual-manual) tasks suggests a preference for a ‘just in time’ processing strategy that minimizes the use of memory. In the present study, a scene comparison paradigm was introduced to determine whether the preference holds when the task is primarily visual and when more complex naturalistic scenes are used as stimuli. Participants made same or different judgments in response to simultaneously presented pairs of scenes that were identical or differed by one object. In Experiment 1, differences were created by deleting an object, replacing it with an item from the same basic-level category (token-substitution) or replacing it with an item from a different basic-level category (type-substitution). The number of objects fixated during each glance to a scene and the number of fixations intervening between glances to corresponding objects suggest that frequently one object at a time is maintained in VSTM. In Experiment 2, type- and token-substitution conditions were administered in blocks in order to test the hypothesis that the preferred strategy would depend on whether detection could be based solely on conceptual information versus when visual information was needed. There were no differences between substitution conditions, suggesting either that participants relied on visual information even when identity information would have been sufficient, or that the minimal memory preference extends to the use of identity codes. A third experiment using arrays of objects and the objects’ names was conducted to explore this latter possibility. Overall, the results suggest a strong general bias toward minimal use of VSTM in complex visual tasks.