Significance Solving crystal structures of large biological macromolecules in the absence of experimental phase information, especially at low resolution, is a tedious problem prone to mistakes and overfitting. Due to errors stemming from poorly interpretable parts of the electron density map, human experience and intuition are imperative for building a correct atomistic model. For this reason, tools for automatic structure determination used nowadays require constant human supervision. Here we present a method that can overcome the difficulties; it greatly reduces or even eliminates human involvement in the solution process by working with ensembles of possible solutions. This approach can find solution of a higher quality than can humans and can solve difficult cases not amenable to other methods.