Tencent's Hunyuan3D team released HY3D-Bench on February 4, a large-scale open-source collection of cleaned, training-ready 3D assets. The pitch: existing 3D repositories are full of noisy geometry, non-manifold meshes, and inconsistent quality. HY3D-Bench tries to fix that by running everything through a rigorous processing pipeline before release.
The dataset splits into three pieces. A full-level set contains 252K+ watertight meshes sourced from Objaverse and Objaverse-XL, each with multi-view renderings and sampled point clouds. A part-level set offers 240K+ objects with semantic part decomposition, useful for robotic manipulation research and part-aware generation. Then there's a synthetic subset: 125K+ objects generated by Tencent's own HY3D-3.0 model across 1,252 categories, filling gaps in rare object classes that real-world scans don't cover well.
To validate the data, the team trained a baseline model, Hunyuan3D-Shape-v2-1 Small, a 0.8B-parameter DiT, on the full-level subset. Results are in the accompanying paper, though they're self-reported. Independent benchmarks from other teams haven't appeared yet.
The GitHub repo includes download scripts for individual subsets. Total download size isn't officially listed on the dataset card, though community estimates put it north of 20 TB.
The Bottom Line: HY3D-Bench gives 3D generation and robotics researchers a cleaned, structured dataset at a scale that didn't previously exist in open source, if the quality holds up under independent scrutiny.
QUICK FACTS
- 252K+ watertight meshes in full-level subset (sourced from Objaverse / Objaverse-XL)
- 240K+ objects with part-level segmentation
- 125K+ synthetic objects across 1,252 categories
- Baseline model: Hunyuan3D-Shape-v2-1 Small (0.8B DiT)
- Released February 4, 2026; paper submitted February 3
- Benchmark results are company-reported (no independent validation yet)
