Recent deep learning-based methods have shown promising results and runtime\nadvantages in deformable image registration. However, analyzing the effects of\nhyperparameters and searching for optimal regularization parameters prove to be\ntoo prohibitive in deep learning-based methods. This is because it involves\ntraining a substantial number of separate models with distinct hyperparameter\nvalues. In this paper, we propose a conditional image registration method and a\nnew self-supervised learning paradigm for deep deformable image registration.\nBy learning the conditional features that are correlated with the\nregularization hyperparameter, we demonstrate that optimal solutions with\narbitrary hyperparameters can be captured by a single deep convolutional neural\nnetwork. In addition, the smoothness of the resulting deformation field can be\nmanipulated with arbitrary strength of smoothness regularization during\ninference. Extensive experiments on a large-scale brain MRI dataset show that\nour proposed method enables the precise control of the smoothness of the\ndeformation field without sacrificing the runtime advantage or registration\naccuracy.\n