A cyclic solvent injection (CSI) test is conducted in a large 3D physical model (80x40x20 cm(3)). 23.57 l heavy oil (2700 cp@ 21 degrees C) is saturated into this model. The test ran for 30 cycles achieving the recovery factor of 67.1%. Experimental results are compared with those of a cylindrical model (d=15.24 cm, L=30.48 cm) CSI test to investigate dominated driving mechanisms of CSI under different physical models. A linear form of material balance equation (MBE) of CSI is established for the first time to obtain the amount of diluted oil (N-di) after the solvent injected into the model in each cycle, which could be used to evaluate the efficiency of oil dilution (E-sd) and recovery factor of diluted oil (R-d). E-sd is defined as the ratio of the N-di to remaining oil. R-d is defined as the ratio of the cumulative oil production as per cycle to N-di. In the MBE, since non-equilibrium phase behavior is presented in CSI, R-s and B-o measured by conventional PVT test under equilibrium state cannot be directly used in the MBE. In this study, R-s and B-o are first matched through the proposed MBE by using experimental data from large 3D test. With estimated R-s, B-o and other experimental data, N-di of each cycle in large 3D test are obtained by liner regression. At last, to further verify the proposed MBE, the tuned R-s and B-o are used to obtain N-di of the cylindrical model test. Comparison between two tests indicates that the gravity force still dominates the recovery factor rather than the model size. The model shape significantly affects the change of production rate, but not the average production which is mainly determined by the original oil in place (OOIP). The same behavior of minimum production pressure occurred in two tests with the same pressure decline rate. N-di of each cycle in the large 3D test is successfully obtained through the established MBE. Then N-di are used to obtain Esd and R-d. The results show that Esd first increases from 4.75% to 10.70% with the growth of solvent chamber. Then it keeps around 13.78% for main stage. During the very late stage, Esd decreases to 10.11%. Not as Esd dramatically varies, R-d in each cycle is in the range of 28%-34%, which indicates that recovery of CSI has much more potential. A higher recover factor could be expected if Esd and R-d would be both improved. By the tuned R-s and B-o in the large 3D case, N-di of each cycle in the cylindrical model CSI test is matched with very good agreement (average R-2=0.9899), which suggests that the established MBE is reliable to evaluate production performance of CSI.