A data-driven regularization method called Blush regularization improves cryo-EM structure determination by using denoising convolutional neural networks (CNNs) to enhance image alignment and reduce noise. This approach outperforms traditional methods, particularly for low signal-to-noise ratio data. The method was tested on various biological macromolecules, including a 40 kDa protein-nucleic acid complex previously considered intractable. Blush regularization improves reconstruction quality by incorporating prior knowledge about biological structures, reducing overfitting, and enhancing resolution. It was applied to datasets with small or complex structures, such as a dimeric G-protein-coupled receptor (Ste2), a spliceosome complex, and an amyloid filament (FIA). The method also improved the resolution of a 40 kDa Aca2-RNA complex. Blush regularization was implemented in the open-source software RELION-5 and demonstrated effectiveness in 3D classification, multi-body refinement, and autorefinement. The approach reduces the size barrier for cryo-EM structure determination by improving the accuracy of reconstructions for smaller or more complex biological macromolecules. The study highlights the potential of deep learning to enhance cryo-EM data analysis and expand the applicability of the technique for a wide range of biological structures.A data-driven regularization method called Blush regularization improves cryo-EM structure determination by using denoising convolutional neural networks (CNNs) to enhance image alignment and reduce noise. This approach outperforms traditional methods, particularly for low signal-to-noise ratio data. The method was tested on various biological macromolecules, including a 40 kDa protein-nucleic acid complex previously considered intractable. Blush regularization improves reconstruction quality by incorporating prior knowledge about biological structures, reducing overfitting, and enhancing resolution. It was applied to datasets with small or complex structures, such as a dimeric G-protein-coupled receptor (Ste2), a spliceosome complex, and an amyloid filament (FIA). The method also improved the resolution of a 40 kDa Aca2-RNA complex. Blush regularization was implemented in the open-source software RELION-5 and demonstrated effectiveness in 3D classification, multi-body refinement, and autorefinement. The approach reduces the size barrier for cryo-EM structure determination by improving the accuracy of reconstructions for smaller or more complex biological macromolecules. The study highlights the potential of deep learning to enhance cryo-EM data analysis and expand the applicability of the technique for a wide range of biological structures.