Fig. 1. Schematic illustration of Blush regularization and slices of example volumes.

a, Training procedure, showing two passes for both half-maps and recycling of the denoiser output (in pink), with calculation of a mean squared error (L2) loss. b, Iterative reconstruction with spectral trailing. Each half-map is reconstructed separately. At each iteration, the FSC is used to estimate a cut-off frequency (ρ), which is subsequently used to low-pass filter the denoiser output. The final output does not pass through the denoiser but is subjected to a Wiener filter, similar to baseline reconstruction. c, Denoiser U-net architecture, consisting of five consecutive encoder blocks and a convolution block, followed by five consecutive decoder blocks. SiLU stands for sigmoid linear unit; Norm for batch normalization. d,e, Slices through maps before (left) and after (right) a single application of the denoiser to the final iteration of the reconstruction for PfCRT (d) and the spliceosome (e). f,g, Slices through maps of baseline reconstruction (left) and after Blush regularization (right) of the FIA (f) and the Aca2–RNA complex (g). Scale bars, 30 Å.