Also an unrelated idea: breasts censoring is a bit wonky if you reduce the box size under 100%, I think because it's trying to cover with a single horizontal box something that is not always horizontally aligned.
Could it work better if the model could detect "nipples" instead of (or in addition to) "breasts", so that you could center a smaller square on each target?
I don't want to be a "backseat programmer" here, I admit I know nothing about training AI models, I fully accept my suggestion could be stupid, please believe I'm just trying to be helpful, and sorry if instead I'm being a bother.