Water depth and flow velocity are the two river dynamics variables critical for almost every river function, including water resources, navigation, flood generation, sediment transport, and biogeochemical cycling. Traditional approaches for river dynamics modeling are either fast but simple and numerically inaccurate (low-fidelity models) or reliable but complicated and computationally intensive (high-fidelity models). Therefore, it is still challenging to perform a large ensemble of high-fidelity water depth and flow velocity simulations that are needed for planning, design, and forecasting purposes. Recently, a novel downscaling method, the hybrid Low-fidelity, Spatial analysis, and Gaussian Process learning (LSG) model, that can accurately simulate the water depth of flood inundation at the computational cost of a low-fidelity model and preserve the physics of river dynamics, has been developed. But it is unknown if the method can be used for fast and accurate simulations of flow velocity. Here, using the flood simulations of low-resolution (low-fidelity) and high-resolution (high-fidelity) 2-D Overland Flow Model (OFM), we train two LSG downscaling models, one for water depth and one for flow velocity, for the 2017 Houston Harvey flood. The model validation will be performed by comparing the downscaled and high-fidelity OFM simulated water depth and flow velocity for a synthetic future hurricane event based on the Simple Cloud-Resolving E3SM Atmosphere Model (SCREAM) configuration. The validation can confirm that the downscaled and high-fidelity river dynamics solutions are consistent, especially when the downscaled flow velocity is bias-corrected using a machine learning based bias correction method. Our approach will greatly advance storyline high-resolution river dynamics simulations.