Two high resolution quantitative precipitation forecasts with different levels of realism are evaluated. Classical scores (bias, correlation and scores based on contingency tables) confirm that the two forecasts do not have the same quality. A multi-scale extension of these scores has then been made to produce a validation for hydrological purposes. Rainfall fields are integrated over surfaces of various scales. For better simulation, scores indicate an increase in the quality of the simulated precipitation for larger surfaces (typically more than 100 km2): the localisation errors are reduced by the aggregation. This helps to determine the usefulness of such forecasts for hydrological purposes.