Shape, a self-supervised 3D foundation model for CAD geometry, achieves 0.729 reconstruction R2 and 98.1% top-1 retrieval on held-out industrial meshes.

Researchers introduced Shape, a 10.9M-parameter foundation model that converts surface meshes into dense per-token embeddings for industrial CAD workflows. The model was pretrained on 61,052 CAD meshes from Thingi10K, MFCAD, and Fusion360.

Shape combines a structured 3D latent grid, a multi-scale geometry-aware tokenizer (MAGNO) with cross-attention, and a transformer processor using grouped-query attention and RMSNorm. It uses masked-token reconstruction and multi-resolution contrastive consistency for pretraining.

On a held-out split of 2,983 meshes, Shape achieves reconstruction R2 = 0.729 and 98.1% top-1 retrieval under the Wang-Isola protocol, with near-zero reconstruction train/val gap. A 2×2 ablation shows per-dimension normalization is critical: without it, performance drops to R2 < 0.14 and top-1 < 88%; with it, both losses succeed at R2 > 0.70 and top-1 > 96%.

Code, embeddings, and an interactive demo are released at the project URL.