Hy3D Bench

Released in February 2026 by Tencent, HY3D-Bench is a large-scale collection of 3D datasets. It comprises over 252,000 watertight meshes with multi-view renderings, 240,000 objects with part-level decompositions, and 125,000 AI-synthesized objects across 1,252 categories. The dataset is hosted on Hugging Face and was last updated in April 2026.

Use Cases

• Train 3D generative models based on the collection of over 252,000 watertight meshes.
• Develop part segmentation algorithms based on the dataset of 240,000 objects with fine-grained part decompositions.
• Benchmark synthetic data generation techniques based on the 125,000 AI-synthesized objects.
• Perform multi-view 3D reconstruction based on the provided renderings and sampled points.

Strengths

• Large scale with over 252,000 watertight meshes.
• Includes fine-grained part decomposition for 240,000 objects.
• Contains a diverse synthetic subset of 125,000 objects across 1,252 categories.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for individual subsets is unknown, which may limit suitability assessment.

Provenance

Source

Tencent

Collection Method

Collection includes real-world objects and AI-synthesized objects, as described.

Time Range

Dataset released in February 2026.

Freshness

Last updated 2026-04-03 11:50:49; freshness should be verified.

Geography

null