To demonstrate that diffusion-weighted images should be acquired at the instant of maximum blood velocity in kidneys to extract the perfusion fraction (PF) by the bi-exponential intravoxel incoherent motion model. The PF values were measured in Monte-Carlo simulations corresponding to different blood velocities with a constant known PF. The distribution of the measured PF values (PF-distribution) was characterized quantitatively by 3 markers highlighting the deviation of the measurement from the true PF. Diffusion-weighted images of kidneys were acquired in 10 healthy volunteers at the instant of maximal respectively minimal blood velocity in the renal artery (Vmax versus Vmin acquisition). The PF-distributions measured from the Vmax and Vmin acquisitions were compared mutually and with simulated PF-distributions using the 3 markers. A radiologist evaluated the quality of the PF maps. The PF-distributions measured in the simulations were spread around the true PF value, and spreading was reduced as blood velocity increased. A comparison between simulated and in vivo PF-distributions suggests that a similar phenomenon is plausible in vivo. The quality of the PF maps of the Vmax -acquisition was scored higher by the radiologist than those of the Vmin -acquisition in 95% of cases (19 of 20). The PF maps are of better quality when the Vmax -acquisition is used. We show evidence supporting the hypothesis that the variation of PF along the cardiac cycle is due to oscillations between a poor estimation when the blood velocity is low, and a better estimation when blood velocity is higher. © 2018 International Society for Magnetic Resonance in Medicine.