This study introduces DiffScale, a diffusion model with classifier-free guidance, to enhance wind speed predictions by downscaling subseasonal to seasonal (S2S) forecasts. DiffScale efficiently super-resolves spatial information across continuous downscaling factors and lead times, leveraging weather variables and regional priors to conditionally sample high-resolution forecasts. Unlike traditional methods, it directly estimates the density of target S2S forecasts without auto-regressing over lead time. Synthetic experiments using ECMWF S2S forecasts and ERA5 reanalysis data demonstrate significant improvements in wind speed prediction quality through continuous downscaling and bias correction.