Enhancing subspace-based inversions through an efficient multi-scaling scheme