How to design loss function to train a model from two-dimension uniform distribution to higher dimensional normal distribution

[zwj]

How to design loss function to train a model from two-dimension uniform distribution to higher dimensional normal distribution?