Coarse-grained (CG) molecular dynamics enables the simulation of large biomolecular systems over extended timescales, but recovering atomistic detail from CG representations remains a significant challenge. Here, we present BackMapNet, a generalised and computationally e!cient framework for protein backmapping. The method adopts a staged reconstruction strategy for backbone and side-chain prediction. To enhance structural accuracy and physical realism, BackMapNet integrates geometric priors and