EMMIVox is a Bayesian inference method for generating accurate structural models from cryo-electron microscopy (cryo-EM) data. It balances experimental information with physico-chemical models of the system and environment, including water, lipids, and ions. EMMIVox explicitly accounts for data correlation and noise, and accurately infers B-factors to produce structural models with excellent fit to the data and high stereochemical quality. It outperforms state-of-the-art refinement techniques by providing both single-structure models and structural ensembles. EMMIVox is implemented in the PLUMED library and is flexible for determining high-quality structural models. It addresses challenges such as data correlation and noise, and improves data fitting and stereochemical quality. EMMIVox is particularly effective in refining coarse-grained models of large protein complexes and determining conformational ensembles from low-resolution cryo-EM maps. It is also useful for structural ensemble refinement, where it can model conformational heterogeneity that is often overlooked by conventional methods. EMMIVox is benchmarked on nine complex biological systems and demonstrates improved quality metrics compared to existing methods. It is implemented in the open-source PLUMED library and aims to set a new standard for single-structure and ensemble refinement by integrating cryo-EM maps with accurate physico-chemical models. EMMIVox provides accurate structural models and ensembles, which can advance understanding of molecular mechanisms underlying biological functions.EMMIVox is a Bayesian inference method for generating accurate structural models from cryo-electron microscopy (cryo-EM) data. It balances experimental information with physico-chemical models of the system and environment, including water, lipids, and ions. EMMIVox explicitly accounts for data correlation and noise, and accurately infers B-factors to produce structural models with excellent fit to the data and high stereochemical quality. It outperforms state-of-the-art refinement techniques by providing both single-structure models and structural ensembles. EMMIVox is implemented in the PLUMED library and is flexible for determining high-quality structural models. It addresses challenges such as data correlation and noise, and improves data fitting and stereochemical quality. EMMIVox is particularly effective in refining coarse-grained models of large protein complexes and determining conformational ensembles from low-resolution cryo-EM maps. It is also useful for structural ensemble refinement, where it can model conformational heterogeneity that is often overlooked by conventional methods. EMMIVox is benchmarked on nine complex biological systems and demonstrates improved quality metrics compared to existing methods. It is implemented in the open-source PLUMED library and aims to set a new standard for single-structure and ensemble refinement by integrating cryo-EM maps with accurate physico-chemical models. EMMIVox provides accurate structural models and ensembles, which can advance understanding of molecular mechanisms underlying biological functions.