MeshGPT

MeshGPT generates compact, artist-style triangle meshes directly — a decoder-only transformer that autoregressively predicts mesh tokens from a learned geometric vocabulary, producing sharp, coherent surfaces instead of dense over-triangulated output. From Matthias Nießner's Visual Computing & AI Lab; CVPR 2024 (preprint Nov 2023). A foundational entry in the group's 3D-generative line that seeds the SpAItial "spatial foundation models" thesis.