Through a combination of excited-state-specific quantum chemistry, selected configuration interaction, and variational Monte Carlo, we investigate an approach to predicting vertical excitation energies that achieves high accuracy across a broader range of states than either coupled cluster theory or multi-reference perturbation theory. The approach can be employed in settings beyond the current reach of selected configuration interaction or density matrix renormalization group and retains its high accuracy whether it is treating single excitations, double excitations, or charge transfer excitations. To address the different challenges that arise in these types of states, the method combines excited-state-specific orbital relaxations from complete active space self consistent field theory, an extended active space configuration selection, and improved approaches to optimization and variance-balancing within variational Monte Carlo. In addition to testing in smaller molecules where clear theoretical benchmarks exist, we use the method to offer new benchmark values in triple-zeta-or-better basis sets for a doubly excited state of benzene, a partially-doubly-excited state of uracil, and the widely considered charge transfer state in the tetrafluoroethylene-ethylene system.