PartNeXt-Mat: Learning Pipeline-Authored Material Priors on Structured 3D Assets

We introduce PartNeXt-Mat, a material-prior layer aligned with PartNeXt geometry, part masks, and hierarchy: 333,152 part assignments in 23,221 objects across 50 categories. Evidence fusion assigns a material, traceable grade resolution attaches applicable density and elastic priors, and verification-aware routing records solver caveats. A frozen 4,054-object visual benchmark studies part prediction and fixed-query completion. On its original split, category/name lookup reaches 82.9% accuracy, exposing near-total exact-key overlap. In an exploratory protocol withholding exact category/name keys from supervision, a fused visual-and-semantic model reaches 63.9% accuracy and 31.0% Macro-F1, compared with 57.8%/7.3% for a global prior, 54.5%/21.6% for train-only name/category backoff, and 57.8%/23.4% for a size-matched semantic MLP. The fused-minus-matched F1 gain is 7.60 points (key-cluster 95% interval 0.96–12.72). In name-disjoint completion, three observed labels raise set-attention accuracy by 1.67 points (0.70–2.63), although a support-aware table remains competitive. Numerical targets contain only 26 complete property tuples. The resource supports learning and completing authored, provenance-recorded part priors; it does not establish asset-specific measured mechanics.