Objaverse++: 100K+ Curated 3D Objects with Quality Scores

Objaverse++ contains between 100,000 and 1,000,000 curated 3D objects filtered from the original Objaverse collection for high-fidelity modeling. Developed by Chendi Lin et al. and updated in late 2025, the dataset includes automated quality annotations for geometric structure and texture.

Use Cases

• Filtering training sets for text-to-3d models using the geometric structure scores
• Benchmarking image-to-3d reconstruction using the curated 3D object meshes
• Analyzing 3D asset distributions based on the texture quality annotations

Strengths

• Includes quality scores for geometric structure and texture
• Scale of 100,000 to 1,000,000 curated 3D objects
• Utilizes a specialized classifier to clean noise from the original Objaverse source

Limitations

• Quality scores are generated by an automated classifier which may introduce systematic labeling bias
• Specific 3D file formats and metadata schema are not explicitly detailed in the summary

Provenance

Source

Arxiv 2504.07334

Collection Method

Automated classification and curation of the existing Objaverse dataset

Freshness

Last updated December 2025