Preattentive models of early vision have not been supported by the evidence. Instead, an input filtering system, which is dynamically reconfigured so as to optimize performance on the task at hand, is proposed. As a case in point, the authors examined Sagi and Julesz's (1985a) claim that detection tasks are processed preattentively and efficiently (shallow search slopes), whereas discrimination tasks require focal attention and yield inefficient steep slopes. In 5 visual search experiments, efficiency was found to depend not on the nature of the task but on whether the task is single or dual. The second component of a dual task, whether detection or discrimination, is performed inefficiently if it does not fit the configuration of the input system, which had been set optimally for the first component. But, even the second component is processed efficiently if there is enough time to reconfigure the system after processing the first component.