A 3D-generation dataset is most useful when it preserves how each asset was made, not just the finished model. Here, prompts, model identifiers, settings, source assets, previews, and output files are linked at the example level, making the collection useful for examining generation workflows as well as browsing results.
What Sets It Apart
- The 1,342 selected examples are represented by 1,500 downloaded output files, so some examples have multiple outputs.
- GLB dominates the collection (1,271 files), but it also includes Blender, SPZ, PLY, VOX, and FBX files. This offers a view across common 3D delivery formats, though it is not a balanced format benchmark.
- Structured metadata records prompts, model names, settings, file paths, and SHA-256 hashes. That makes provenance and file integrity easier to inspect, while missing input assets remain unavailable remotely.
- Structural checks passed for the outputs, including loading all 137 Blender files with auto-execution disabled. These checks do not establish visual quality; one Blender scene also references two missing textures.
Who It’s For
Great fit if you are comparing image-to-3D or text-to-3D workflows, building asset browsers, or studying how model and settings choices relate to generated outputs. Look elsewhere if you need a large, consistently benchmarked training corpus or verified production-ready meshes: the examples are selected, formats are unevenly represented, and structural validation is not a quality assessment.
