The soil and water assessment tool (SWAT) is widely used to quantify the spatial and temporal patterns of sediment loads for watershed-scale management of sediment and nonpoint-source pollutants. However few studies considered the trade-off between flow and sediment objectives during model calibration processes. This study proposes a new multi-objective calibration method that incorporates both flow and sediment observed information into a likelihood function based on the Bayesian inference. For comparison, two likelihood functions, i.e., the Nash–Sutcliffe efficiency coefficient (NSE) approach that assumes model residuals follow the Gaussian distribution, and the BC-GED approach that assumes model residuals after Box–Cox transformation (BC) follow the generalized error distribution (GED), are applied for calibrating the flow and sediment parameters of SWAT with the water balance model and the variable source area concept (SWAT-WB-VSA) in the Baocun watershed, Eastern China. Compared with the single-objective method, the multi-objective approach improves the performance of sediment simulations without significantly impairing the performance of flow simulations, and reduces the uncertainty of flow parameters, especially flow concentration parameters. With the NSE approach, SWAT-WB-VSA captures extreme flood events well, but fails to mimic low values of river discharge and sediment load, possibly because the NSE approach is an informal likelihood function, and puts greater emphasis on high values. By contrast, the BC-GED approach approximates a formal likelihood function, and balances consideration of the high- and low- values. As a result, inferred results of the BC-GED method are more reasonable and consistent with the field survey results and previous related-studies. This method even discriminates the nonerodible characteristic of main channels.