The ever increasing computer power, together with the improved accuracy of atomistic force fields, enables researchers to investigate biological systems at the molecular level achieving remarkable sizes and detail. However, the computational resources required to cover biologically relevant length and time scales are unavailable but to a few groups worldwide; furthermore, it is often the case that the physical and biological questions one tackles by means of computer simulations and in silico models remain unanswered even having access to cutting-edge technology, when multiple runs and/or fast simulation times are required. One of the most effective solutions to these limitations is offered by simplified, or coarse-grained, models of large biomolecules, capable of providing insightful information on the system's properties at a lower computational cost with respect to all-atom models. Generally, coarse-grained models employ a uniform mapping, that is, a given group of atoms (e.g. an amino acid or a nucleic acid basis) is mapped onto the same effective interaction site irrespectively of its specific environment, properties, and role. Consequently, relevant fine-grained detail is removed or poorly embedded in these models, with detrimental consequences for the realistic and accurate representation of the system. An alternative to this intuitive yet limiting strategy is the usage of non-uniform mappings in which the resolution of the system can vary depending on local as well as global features. In this talk I will make the case for this kind of approaches, presenting and discussing algorithmic procedures developed to identify the representation of a given system which optimally describes it, distributing the atomistic detail and the crude coarse-grained simplification so as to balance efficiency and accuracy.