Learning with gradient inverse consistency and no explicit regularization achieves better diffeomorphic registration performances than vanilla inverse consistency. On a real data experiment, it outperforms state of the art still obtained by optimization based algorithms