We introduce a method using Fourier neural operators to rapidly generate a high-quality starting model for FWI or joint inversion.