The study of visual working memory (VWM) has recently centered around a fundamental question: Does VWM have a discrete capacity limit, such that only a few items may be actively maintained at any time, or is VWM a continuous resource that can be flexibly allocated to any number of items with varying degrees of precision? The key signature of discrete theories is that pure guessing will occur on some trials when the number of items to be remembered exceeds the capacity of VWM, whereas continuous models predict that all information in a visual scene can be stored with some precision. However, evidence for guessing has proven difficult to assess, as continuous resource models can predict data patterns that mimic guessing by allowing precision to vary randomly across trials. Here, we evaluated the hypothesis that a major source of trial-to-trial variability in VWM performance is due to variation in perceptual sensitivity across stimuli, an effect that both discrete and continuous models can account for. Participants viewed oriented gratings, and after a brief delay reported the orientation held in memory for a probed grating. When simply modeling the distribution of report errors, a continuous resource model with variable precision provided a better fit to the data than a discrete capacity model. However, precision was markedly higher for vertical and horizontal orientations than for oblique orientations, and this well-known perceptual effect accounted for a substantial proportion of the trial-to-trial variability in VWM performance. After explicitly including this source of variability within both models, the resulting discrete capacity model provided a better fit to the data than the continuous resource model. Our results suggest that the variability in VWM precision arises largely from differences in perceptual sensitivity across stimuli, and taking this variability into account reveals evidence in favor of the discrete capacity model. Meeting abstract presented at VSS 2015.