meshrep — cleaned geometry-only mesh dataset for Mesh-JEPA pretraining

491,429 cleaned, geometry-only GLB meshes (vertices+faces, normalized to a unit cube, ≤20,000 vertices each), packed as tar shards.

subset	source	meshes	shards
objaverse_xl/	Objaverse + Objaverse-XL (github + sketchfab)	339,478	192
trellis/	TRELLIS-500K curation (new OXL assets)	151,951	102

Each tar holds <key>.glb (key = sha256 for github-sourced, uid for sketchfab/objaverse).

Processing

Raw assets → (trimesh for glb / Blender for fbx·obj·etc.) → clean_and_convert: merge duplicate vertices, drop degenerate/duplicate faces, drop F/V≤1, drop non-finite (corrupt) meshes, cap vertices ≤20k, normalize to unit cube + recompute normals. Pipeline: https://github.com/TeCai/meshrep (branch data-pipeline).

Metadata

mesh_metadata.csv: key, source, n_vertices, n_faces (covers the objaverse_xl subset).

License

Derived from Objaverse-XL and TRELLIS-500K; respect their upstream licenses (ODC-By / per-asset for Objaverse-XL; MIT for TRELLIS-500K metadata). Geometry only, no textures/materials.