The suggestion of using a classification network as a loss function is brilliant!

I love how we can, in general, elevate the sophistication of ML models by having different models interact and train each other.