HY3D-Bench provides a comprehensive suite of high-quality 3D datasets designed to overcome the limitations of existing repositories, such as noise and lack of structural detail. Targeting researchers in computer vision, generative modeling, and robotics, it offers clean, structured, and diverse 3D content crucial for training robust models and establishing benchmarks. The benefit lies in enabling more reliable and advanced research in 3D asset generation and understanding.

How It Works

HY3D-Bench comprises three distinct, complementary datasets. The Full Dataset offers over 252K watertight, normalized, and cleaned meshes with multi-view renderings, ideal for training generative models and general 3D benchmarks. The Part Dataset provides over 240K objects with consistent part-level segmentation and individual part assets, facilitating part-aware generation and fine-grained geometric perception. The Synthetic Dataset leverages a Text-to-3D pipeline (LLM → Diffusion → Image-to-3D) to generate over 125K AI-synthesized objects across 1,252 categories, specifically addressing long-tail distributions for enhanced model robustness and data augmentation. This multi-faceted approach ensures high-quality, structured data addressing various research needs.

Quick Start & Requirements

Dataset download is facilitated via the Hugging Face CLI: hf download tencent/HY3D-Bench --repo-type dataset --local-dir "your/local/path". Specific subsets can be included using the --include flag. Significant disk space is required, with the Full-level dataset alone approaching ~11 TB. Detailed usage instructions are available in baseline/README.md and Part_README.md.

Highlighted Details

• Three high-quality 3D datasets: Full (252K+ watertight meshes), Part (240K+ objects with decomposition), and Synthetic (125K+ AI-synthesized objects).
• All meshes are guaranteed watertight, normalized, and cleaned, ensuring training-ready quality.
• Synthetic dataset covers 1,252 categories via a Text-to-3D pipeline, emphasizing long-tail object distributions.
• Includes a baseline model, Hunyuan3D-Shape-v2-1 Small (0.8B parameters), trained on the Full-level data.

Maintenance & Community

Community engagement is encouraged via WeChat and Discord groups. The project acknowledges contributions from the "Team Hunyuan3D" and lists numerous authors in its associated BibTeX entries. Links to Xiaohongshu and X (formerly Twitter) are also provided.

Licensing & Compatibility

The provided README does not explicitly state the software license for the HY3D-Bench datasets or associated tools. This lack of clear licensing information presents a significant caveat for adoption, particularly concerning commercial use or integration into closed-source projects.

Limitations & Caveats

The primary limitation is the absence of a clearly defined license, hindering assessment of usage rights and compatibility. While the datasets are presented as high-quality, the README does not detail specific performance benchmarks or comparative analyses against other datasets, nor does it outline potential limitations in terms of data diversity beyond the stated long-tail focus of the synthetic set.