The theoretical understanding of eyewitness identifications made from a police lineup has long been guided by the distinction between absolute and relative decision strategies. In addition, the accuracy of identifications associated with different eyewitness memory procedures has long been evaluated using measures like the diagnosticity ratio (the correct identification rate divided by the false identification rate). Framed in terms of signal-detection theory, both the absolute/relative distinction and the diagnosticity ratio are mainly relevant to response bias while remaining silent about the key issue of diagnostic accuracy, or discriminability (i.e., the ability to tell the difference between innocent and guilty suspects in a lineup). Here, we propose a signal-detection-based model of eyewitness identification, one that encourages the use of (and helps to conceptualize) receiver operating characteristic (ROC) analysis to measure discriminability. Recent ROC analyses indicate that the simultaneous presentation of faces in a lineup yields higher discriminability than the presentation of faces in isolation, and we propose a diagnostic feature-detection hypothesis to account for that result. According to this hypothesis, the simultaneous presentation of faces allows the eyewitness to appreciate that certain facial features (viz., those that are shared by everyone in the lineup) are non-diagnostic of guilt. To the extent that those non-diagnostic features are discounted in favor of potentially more diagnostic features, the ability to discriminate innocent from guilty suspects will be enhanced.