Abstract: A snow initialization scheme based on anomaly rescaling has been introduced into IAP AGCM4 (Institute of Atmospheric Physics Atmospheric General Circulation Model 4) to obtain realistic initial snow conditions over the Qinghai-Xizang Plateau. Two suites of seasonal ensemble hindcasts with and without snow initialization were implemented, and the predictability skills of large-scale atmospheric circulations and precipitation over eastern China were estimated. The results show that the potential predictability of spring snow depth over the Qinghai-Xizang Plateau increased greatly when more realistic snow initial anomalies were applied, especially in the western part of Qinghai-Xizang Plateau. Furthermore, the results indicate that the potential predictability for the large-scale atmospheric circulation at middle-high latitude Eurasia was enhanced. The predictive skill also showed some improvement for the East Asian summer monsoon. These results are favorable for the improved prediction of rainfall anomalies in China. Hence, the results of the present study indicate that seasonal prediction for summer rainfall anomalies over eastern China is improved by using more realistic initial snow states over the Qinghai-Xizang Plateau, especially in the years with pronounced snow anomalies. Large improvement can be found for the Yangtze River valley with an anomaly correction coefficient from 0.02 without snow initialization to 0.11 with snow initialization for the 36-year average. Finally, the effectiveness of this snow initialization was validated through a case study. The results of the case study show that the atmospheric circulation responds to the improved snow anomalies over the Qinghai-Xizang Plateau with snow initialization and then impacts the predictability of rainfall anomalies over eastern China. Thus, it is suggested that snow initialization based on anomaly rescaling is effective for improving the predictive skill of rainfall anomalies over China, and this snow initialization scheme can be used in real-time predictions.