Riverine flooding can disrupt local economies for weeks. Appropriate flood risk management strategies hinge on accurate flood projections with adequate uncertainty quantification. Higher spatial resolution flood projections tend to be more accurate, especially for spatially heterogeneous areas like cities. Higher spatial resolution flood projections also tend to be more computationally expensive. At a high enough resolution, they can take so long to run that we cannot tune their inputs and possibly cannot run the model at all. In such situations a logical alternative is to (1) tune the inputs for a lower resolution version of the model and (2) downscale a low resolution model projection at a realistic input setting. To downscale a low resolution projection  is to map it onto a higher resolution grid. Many downscaling approaches exist. There are (1) general probabilistic approaches that can be applied to many types of physics-based models and (2) deterministic downscaling approaches that are developed specifically for flood hazard models. Each type of downscaling approach comes with its own set of benefits but does not share the benefits of the other type. We develop a probabilistic downscaling approach that is specific to flood hazard models, achieving the benefits of both types of downscaling approaches, mainly (1) appropriate uncertainty quantification and (2) improved accuracy from deriving information from physical processes. Compared to a more general probabilistic downscaling approach, our approach shows improved performance in estimating the high resolution projection and quantifying the associated uncertainties.