Image registration networks aim to estimate a deformation field from an image pair to achieve alignment. To boost registration accuracy, existing methods introduce additional structures to a backbone network or design teacher networks for guiding learning through knowledge distillation. Nevertheless, these additional structures increase computational demands, while the knowledge gap exists between the manually designed teacher network and the student network. Both these registration boosting methods depend on external structures. In this paper, we introduce a novel Collaborative Learning (CL) scheme to boost image registration networks with only themselves inherently. Specifically, we devise correlation collaborative learning to empower the feature correlation among each level, and devise deformation collaborative learning to leverage the fine deformation field to empower the learning of the coarse ones. With this scheme, registration networks can be boosted without the reliance on external structures, and they are exempt from additional computational burdens. Experimental results demonstrate the superiority of our proposed CL scheme compared to other methods for boosting image registration networks.