Abstract. For landscapes with riparian buffers, we develop and analyze models predicting landscape discharge based on material release by an uphill source area, the spatial distribution of riparian buffer along a stream, and retention within the buffer. We model the buffer as a grid of cells, and each cell transmits a fixed fraction of the materials it receives. We consider the effects of variation in buffer width and buffer continuity, quantify the relative contributions of source elimination and buffer retention to total discharge reduction, and develop statistical relationships to simplify and generalize the models. Width variability reduces total buffer retention, increases the width needed to meet a management goal, and changes the importance of buffer retention relative to source elimination. Variablewidth buffers are less efficient than uniform-width buffers because transport through areas of below-average buffer width (particularly gaps) dominates landscape discharge, especially for narrow buffers of highly retentive cells. Uniform-width models overestimate retention, so width variability should be considered when testing for buffer effects or designing buffers for water quality management. Adding riparian buffer to a landscape can decrease material discharge by increasing buffer retention and by eliminating source areas. Source elimination is more important in unretentive or wide buffers, while buffer retention dominates in narrow, retentive buffers. We summarize model results with simpler statistical relationships. For unretentive buffers, average width is the best predictor of landscape discharge, while the frequency of gaps was best for narrow, retentive buffers. Together, both predictors explain Ͼ90% of the variance in average landscape transmission for any value of buffer retentiveness. We relate our results to ecological theory, landscape-scale buffer effects, buffer management, and water quality models. We recommend more empirical studies of buffer width variability and its effects on material discharge. Landscape models should represent width variability and the nonlinear interactions between buffers and source areas.