Proteins and other biomolecules form dynamic macromolecular machines tightly orchestrated to move, bind, and perform chemistry. Uncovering their structural heterogeneity is key to understanding mechanism and function. Imaging tools such as cryo-electron microscopy (cryo-EM) and cryo-electron tomography (cryo-ET) can access the intrinsic heterogeneity of these complexes present in vitro or in their native cellular contexts. To address the outstanding computational challenges, here, we design a novel deep generative model for heterogeneous 3D reconstruction from 2D cryo-EM imaging data. Our model, DRGN-AI, can operate ab initio, that is, reconstruct without any structural priors or input poses and thus enables structure determination of highly dynamic complexes or mixtures of unknown composition. Using DRGN-AI, we reveal new conformational states in large datasets, reconstruct previously unresolved motions from unfiltered datasets, and demonstrate ab initio reconstruction of biomolecular complexes inside cells. With this expressive and scalable model for structure determination, we hope to unlock the full potential of cryo-EM and cryo-ET as a high-throughput tool for structural biology and discovery.