It wouldn’t be the first “impossible” thing to eventually be implemented.
Natural neural networks can do it. We have proof for that; geometry-correct realistic art made by humans exists after all. So it’s very likely not impossible to do this with artificial neural networks.
Whether it’s possible to do in an end-user GPU I don’t know. Probably eventually. May take a year, a decade, a century… who knows. The future is hard to predict.
But I don’t see why it would be impossible. We are just not there yet.