A 3D asset is usually a finished mesh: useful to render, but difficult for a model to edit or reason about part by part. This collection changes the learning target from the mesh alone to a compact program that rebuilds it. That makes the geometry, dimensions, parts, and—in articulated objects—joints explicit, creating a bridge between visual examples and editable 3D code.
What Sets It Apart
- Each record pairs a Python program with a render and an OpenUSD export, so researchers can compare the generated shape with its visual and scene representations.
- The programs use a shared modeling language with named parts and parameterized dimensions. This gives models a more structured target than unconstrained 3D output and makes generated objects easier to inspect and modify.
- Its 66,887 Objaverse examples include 14 articulated objects with 57 joints. The train/validation split separates Objaverse assets by Sketchfab author, reducing the chance that the same creator’s style appears on both sides.
- Quality filtering uses a surface F-score threshold, but that score comes from the reconstruction pipeline itself rather than an independent evaluation. It is useful as metadata, not a substitute for benchmarking.
Who Should Use It
A good fit for researchers studying text- or image-conditioned 3D generation, program synthesis, or models that need editable object representations. Look elsewhere for broad coverage of organic forms or a general-purpose 3D asset archive: this release is mostly hard-surface objects, and small source details may be omitted. Reuse also requires checking each record’s source license and preserving its attribution where applicable.
