A 3D file is not automatically training-ready
Opening a GLB in a DCC tool does not mean the mesh is scaled, oriented, captioned, or cleared for a training run. A training-ready asset record adds geometry checks, runtime validation, semantic metadata, and per-asset provenance. Appearance alone is not the delivery.
Asset record anatomy
A record typically includes the normalized file, bounding box and up-axis, material slots, category and captions, checksums, and a license row. Optional fields — skeleton, clips, colliders — appear only when the source provides them and the tier requires them.
Validation pipeline
Source ingest, geometry sanity, scale and orientation, materials, metadata, engine load, then provenance close-out. Failed items return for correction instead of being silently dropped into a lake. Target runtime is named; we do not claim universal engine compatibility.
Asset tiers
Tiers have different required fields. Rig, animation, and collider data are not implied for every asset.
Object
Static props with geometry, materials, scale, and category metadata. No rig or animation is implied.
Character
Figure assets that may include a skeleton when the source provides one. Animation clips are recorded only when they exist and are authorized.
Environment
Scene pieces or rooms with spatial bounds and placement metadata. Not every environment is a fully interactive level.
Interactive Asset
Objects with documented interaction points or state fields. Colliders and scripts are included only when validated in the target runtime.
Simulation-Ready Asset
Assets checked for the physics or navigation assumptions of a named engine. This tier is scoped; it is not a default for every file.
Metadata and captions
Names in a filename are not a taxonomy. We attach machine-readable class, parts, and captions scoped to the collection. Caption style follows the model objective — retrieval, generation, or interactive use — rather than a single marketing paragraph per file.
Duplicate and near-duplicate analysis
Exact hashes catch identical exports. Near-duplicates need geometry and texture similarity with an explicit train/eval policy so variants do not leak across the split. Thresholds are program-specific.
Licensing and provenance
Each asset keeps a source and authorization row. Dataset-level footnotes are not a substitute. Training rights do not imply redistribution. Details live on Licensing & Provenance.
Ordinary file versus training-ready record
| Ordinary asset file | Training-ready asset record |
|---|---|
| Opens in one tool | Validated in target runtime |
| Unknown scale | Normalized scale and orientation |
| Visual name only | Semantic metadata and captions |
| Dataset-level license note | Per-asset provenance |
| No duplicate analysis | Exact and near-duplicate checks |
| Manual browsing | Machine-readable manifest |
| Appearance focused | Geometry, usability, and rights checked |