In this paper, a multi-scale direct learning control (DLC) scheme is proposed. DLC is defined as the direct generation of the desired control profile from existing control inputs without any repeated learning. DLC schemes developed hitherto only consider trajectories with a single-scale relation in that any two trajectories are spatially proportional to each other with a single scale for all coordinates (dimensions). In comparison, the new scheme is capable of handling trajectories with multiple scales, i.e., each spatial coordinate of a trajectory is proportional to that of another trajectory associated with a distinct scale. It is shown that the new scheme is applicable to linear time-varying high-order systems. Simulation results demonstrate its effectiveness.