Characterizing materials to atomic resolution and first-principles structure-property prediction are two pillars for accelerating functional materials discovery. However, we are still lacking a rapid, noise-robust framework to extract multilevel atomic structural motifs from complex materials to complement, inform, and guide our first-principles models. Here, we present a machine learning framework that rapidly extracts a hierarchy of complex structural motifs from atomically resolved images. We demonstrate how such motif hierarchies can rapidly reconstruct specimens with various defects. Abstracting complex specimens with simplified motifs enabled us to discover a previously unidentified structure in a Mo─V─Te─Nb polyoxometalate (POM) and quantify the relative disorder in a twisted bilayer MoS2. In addition, these motif hierarchies provide statistically grounded clues about the favored and frustrated pathways during self-assembly. The motifs and their hierarchies in our framework coarse-grain disorder in a manner that allows us to understand a much broader range of multiscale samples with functional imperfections and nontrivial topological phases.