The paper introduces a novel approach called *non-uniform refinement* to improve the 3D reconstruction quality in single particle cryo-EM, particularly for membrane proteins. Traditional iterative refinement methods often assume spatial uniformity, which can lead to over- and under-regularization in disordered or flexible regions. *Non-uniform refinement* uses a cross-validation approach to derive a regularizer that adaptively adjusts the regularization strength at each voxel, effectively removing noise in disordered regions while retaining signal for aligning particle images. This method is implemented in the *cryoSPARC* software package and has been successfully applied to several membrane proteins, achieving state-of-the-art resolutions and improved 3D map quality. The results demonstrate that *non-uniform refinement* can significantly enhance the resolution and map quality, even for small membrane proteins, making it a valuable tool in structural biology and drug discovery.The paper introduces a novel approach called *non-uniform refinement* to improve the 3D reconstruction quality in single particle cryo-EM, particularly for membrane proteins. Traditional iterative refinement methods often assume spatial uniformity, which can lead to over- and under-regularization in disordered or flexible regions. *Non-uniform refinement* uses a cross-validation approach to derive a regularizer that adaptively adjusts the regularization strength at each voxel, effectively removing noise in disordered regions while retaining signal for aligning particle images. This method is implemented in the *cryoSPARC* software package and has been successfully applied to several membrane proteins, achieving state-of-the-art resolutions and improved 3D map quality. The results demonstrate that *non-uniform refinement* can significantly enhance the resolution and map quality, even for small membrane proteins, making it a valuable tool in structural biology and drug discovery.