Prevent model from predicting the same examples for different inputs

I have a ANN model, that predicts a fixed length curve. The problem is, those curves are really similar to each other, for example these two:

To compare those curves, I use RMSD between their points. The problem is, that the model for most of the parameters, learns a mean-like curve representation, that has low RMSD for most of the curves, so the model predicts the same curve for different inputs.

Is there any loss function/metric to penalize for predicting the same outputs for different inputs ?