HY3D-Bench

HY3D-Bench is a large-scale collection of 3D model datasets released by Tencent's Hunyuan research team, designed to give AI researchers high-quality data for training and testing 3D generation and computer vision models. Existing 3D datasets often have noisy or broken geometry that makes them hard to use directly for training — HY3D-Bench addresses this by providing cleaned, "watertight" meshes (3D shapes with no holes or gaps) along with rendered images and structured metadata.

The release includes three separate datasets totaling over 600,000 objects and roughly 22 terabytes of data. The Full-level dataset contains over 252,000 complete 3D objects with multi-view photo renders and point cloud samples, suitable for training 3D generation models. The Part-level dataset contains over 240,000 objects broken down into labeled individual parts (useful for robotics grasping tasks or fine-grained shape analysis). The Synthetic dataset contains over 125,000 AI-generated objects across 1,252 categories, generated through an automated pipeline that uses language models to expand text descriptions, image diffusion models to create visuals, and then image-to-3D reconstruction to produce the final meshes — covering rare categories that are hard to find in real-world scans. A baseline 0.8-billion-parameter 3D shape generation model trained on the full dataset is also released on Hugging Face. The data is available for download via the Hugging Face datasets hub.