Indeed, neural operators can do zero-shot super-resolution as long as they are trained on data of sufficient resolution. There is a tradeoff between training costs and resulting fidelity. If we can get rich training data, then we can do faithful operator learning.

→ View original post on X — @animaanandkumar