A new hybrid empirical downscaling technique is presented and applied to assess 21st century projections of maximum and minimum daily surface air temperatures (Tmax, Tmin) over the Midwestern USA. Our approach uses multiple linear regression to downscale the seasonal variations of the mean and standard deviation of daily Tmax and Tmin and the lag-0 and lag-1 correlations between daily Tmax and Tmin based on GCM simulation of the large-scale climate. These downscaled parameters are then used as inputs to a stochastic weather generator to produce time series of the daily Tmax and Tmin at 26 surface stations, in three time periods (1990–2001, 2020–2029, and 2050–2059) based on output from two coupled GCMs (HadCM3 and CGCM2). The new technique is demonstrated to exhibit better agreement with surface observations than a transfer-function approach, particularly with respect to temperature variability. Relative to 1990–2001 values, downscaled temperature projections for 2020–2029 indicate increases that range (across stations) from 0.0 K to 1.7 K (Tmax) and 0.0 K to 1.5 K (Tmin), while increases for 2050–2059 relative to 1990–2001 range from 1.4 K to 2.4 K (Tmax) and 0.8 to 2.2K (Tmin). Although the differences between GCMs demonstrate the continuing uncertainty of GCM-based regional climate downscaling, the inclusion of weather-generator parameters represents an advancement in downscaling methodology.