Reference plays an important role in analyzing event-related potentials (ERP), because it can change ERP experiment results. Previous work concerning reference, however, has focused on the cognition area, but little is known about emotion area. To test the influence of reference choice on emotional processing, 24 participants performed a passive picture-viewing task involving positive, neutral, and negative pictures while EEG was recorded. Analyses mainly focused on P2 (160 to 230 ms), early posterior negativity (EPN, 220 to 300 ms), and late positive potential (LPP, 400 to 3000 ms) components. We compared three references (average reference; mean mastoids reference; reference electrode standardization technique, REST) on three aspects of these components. e.g., amplitudes, statistic power, and internal consistency. Furthermore, we tested the influence of reference choice on artifact rejection rates of ERP. Self-report behavioral results indicated the emotional pictures differed significantly in valence (ps < .01), and the arousal of emotional pictures were higher than neutral pictures (ps < .01). The ERP results showed that references have distinct influences on amplitudes and statistic power of ERP components as well as artifact rejection rates of ERP. Firstly, the temporal EPN amplitudes were higher under both REST and average reference than those under mean mastoids reference (ps < .05); however, there were quite opposite results for parietal LPP amplitudes (ps < .01). Secondly, the results of Monte Carlo simulations showed that the statistic power of the P2 amplitude difference between negative and neutral conditions increased under mean mastoids reference than those under both REST and average reference. However, the statistic power of EPN’s “positive bias” increased under REST and average reference than those under mean mastoids reference. Thirdly, the artifact rejection rates under mean mastoids reference were significantly higher than those under both average reference and REST (ps < .01). Lastly, the internal consistency of ERPs under three references were relatively good [P2 ( .76 - .87)，EPN ( .87 - .96), and LPP ( .69 - .91)]. In summary, the influence of reference choice on amplitudes was moderated by the biggest active region of specific ERP components. Importantly, high likelihood of the P2 difference between negative and neutral conditions was detected under mean mastoids and average reference, while high likelihood of the EPN’s “positive bias” was detected under both REST and average reference. Additionally, a wider range of threshold to reject the artifact should be used under mean mastoids reference. Reference choice, however, has no effect on the reliability of P2, EPN, and LPP. In summary, in order to avoid distorting results and inappropriate conclusions, future ERP studies are supposed to select reference with caution according to specific components and related work.