CryoFIG: Feed-Forward Inference of 3D Gaussians for Ab Initio Cryo-EM Reconstruction

Ab initio cryo-electron microscopy (cryo-EM) reconstruction recovers molecular density from noisy particle images with unknown poses. However, extending learned pose inference to direct density prediction requires combining weak image evidence across views. We introduce CryoFIG (Feed-forward Inference of Gaussians), a shared model for rapid reconstruction that predicts particle geometry and anisotropic 3D Gaussian densities. Its geometry-guided decoder aggregates spatial image features across particles using viewing geometry and known contrast transfer function metadata. Training combines simulated pretraining, shared adaptation on synthetic and experimental particles, and reprojection supervision from disjoint target images. After shared-model training, one checkpoint reconstructs density maps from 32 support particles in approximately seven seconds without an initial volume or inference-time optimization. Across two synthetic benchmarks, CryoFIG achieves finer resolution than the evaluated baselines, as measured by map-to-map Fourier shell correlation. It also achieves competitive reconstruction quality on two experimental benchmarks. These results support learned Gaussian density prediction as a route to rapid cryo-EM reconstruction. Code and pretrained model weights will be released.