That’s pretty neat. It reminds me of how VAEs work: https://en.wikipedia.org/wiki/Variational_autoencoder