ANTs provides the option --float to run a registration with single floating point accuracy for computations, but this option is not exposed by the nipype interfaces. It would be nice if it did.