Recent studies have demonstrated the ability of magnetic resonance imaging (MRI) to provide anatomically realistic boundary conditions for computational fluid dynamics (CFD) simulations of arterial hemodynamics. To date, however, little is known about the overall reproducibility of such image-based CFD techniques. Towards this end we used serial black blood and cine phase contrast MRI to reconstruct CFD models of the carotid bifurcations of three subjects with early atherosclerosis, each imaged three times at weekly intervals. The lumen geometry was found to be precise on average to within 0.15 mm or 5%, while measured flow and heart rates varied by less than 10%. Spatial patterns of a variety of wall shear stress (WSS) indices were largely preserved among the three repeat models. Time-averaged WSS was reproduced best, on average to within 5 dyn/cm2 or 37%, followed by WSS spatial gradients, angle gradients, and oscillatory shear index. The intrasubject flow rate variations were found to contribute little to the overall WSS variability. Instead, reproducibility was determined largely by the precision of the lumen boundary extraction from the individual MR images, itself shown to be a function of the image quality and proximity to the geometrically complex bifurcation region.