Non-uniform refinement improves single particle cryo-EM reconstruction by addressing spatial variability in 3D density maps. This method uses cross-validation to automatically regularize 3D maps during iterative refinement, accounting for structural differences and yielding higher resolution and better 3D map quality. Traditional methods assume uniformity, but non-uniform refinement adapts to local variations, removing noise from disordered regions while preserving useful signal for alignment. It has been implemented in cryoSPARC and has been successfully used in structural studies of membrane proteins. Non-uniform refinement is reliable and automatic, requiring no parameter changes between datasets and no manual masks or labels. It improves resolution and map quality for small membrane proteins, which is important for structural biology and drug discovery. The algorithm uses a cross-validation regularization framework, allowing for adaptive regularizers that can find better 3D structures with improved latent pose estimates. The regularization parameters are optimized using cross-validation, leading to better signal-to-noise ratios and improved resolution. Non-uniform refinement has been tested on various membrane proteins, including STRA6-CaM, PfCRT, and NaV1.7, showing significant improvements in resolution and map quality. It also improves B-factors, indicating more efficient use of signal from the same data. Non-uniform refinement is particularly effective in handling disordered regions, such as detergent micelles and lipid nanodiscs, and has been shown to reduce over-fitting and improve alignment accuracy. The method is implemented in cryoSPARC and has been used in several notable structural studies.Non-uniform refinement improves single particle cryo-EM reconstruction by addressing spatial variability in 3D density maps. This method uses cross-validation to automatically regularize 3D maps during iterative refinement, accounting for structural differences and yielding higher resolution and better 3D map quality. Traditional methods assume uniformity, but non-uniform refinement adapts to local variations, removing noise from disordered regions while preserving useful signal for alignment. It has been implemented in cryoSPARC and has been successfully used in structural studies of membrane proteins. Non-uniform refinement is reliable and automatic, requiring no parameter changes between datasets and no manual masks or labels. It improves resolution and map quality for small membrane proteins, which is important for structural biology and drug discovery. The algorithm uses a cross-validation regularization framework, allowing for adaptive regularizers that can find better 3D structures with improved latent pose estimates. The regularization parameters are optimized using cross-validation, leading to better signal-to-noise ratios and improved resolution. Non-uniform refinement has been tested on various membrane proteins, including STRA6-CaM, PfCRT, and NaV1.7, showing significant improvements in resolution and map quality. It also improves B-factors, indicating more efficient use of signal from the same data. Non-uniform refinement is particularly effective in handling disordered regions, such as detergent micelles and lipid nanodiscs, and has been shown to reduce over-fitting and improve alignment accuracy. The method is implemented in cryoSPARC and has been used in several notable structural studies.