One possible approach would be to use the functional transforms API and sample the “random” parameters once and apply them to the input image as well as all targets.
This post gives you an examples and you could also use the reference implementation from here to add more transformations.