This study presents a procedure to estimate more accurate direct surface runoff from large rainfall (>25.4 mm)-runoff events. Improved models (M5-M7) are derived by coupling two concepts: (i) initial abstraction as 2% of the total rainfall and (ii) runoff coefficient = degree of saturation. Performance of ten different models including the original SCS-CN method (M1), Mishra and Singh 2002 (M2), Mishra et al. 2006 (M3), Ajmal et al. 2016 (M4), improved models (M5-M7) and their simplified forms (M8-M10) is evaluated using large (7687) number of rainfall events derived from 98 watersheds of USDA-ARS to assess the accuracy of runoff estimation. Quantitatively, it is assessed using seven performance indices, viz., R2, NSE, PBIAS, RMSE, NRMSE, RSR, and MAE; categories; and Ranking and Grading System (RGS). The resulting high values of R2, RSR, RGS, and lowest values of NSE, PBIAS, RMSE, NRMSE, and MAE for the improved models (M5-M7) reveal that improved models performed better than the existing models (M1-M4). Similarly, based on different performance categories, all improved models exhibited superior performance in most of the watersheds than did the existing models. Sensitivity analysis indicated CN to be the most sensitive parameter of the improved model. The proposed model is seen to have overcome the limitations of the original and its previous versions intended for large events and can thus be used for estimating runoff more accurately.