3D生成「ImageNet」来了！腾讯混元开源HY3D-Bench

Core Insights - The article discusses the advancements in 3D generation technology, highlighting the release of the HY3D-Bench dataset by Tencent's Hunyuan team, which addresses key challenges in the field such as data quality, evaluation standards, and long-tail category coverage [3][4]. Dataset Composition - The HY3D-Bench dataset consists of 252,000 high-quality 3D assets, 240,000 component-level structured annotations, and 125,000 AIGC synthetic samples, providing a standardized data foundation for 3D generation research [19][20]. - Early benchmark datasets like ShapeNet had limitations such as imbalanced category coverage and insufficient data volume, which hindered the practical application of 3D generation technology [4]. - The emergence of large-scale datasets like Objaverse has improved the situation, but challenges remain, particularly in the preprocessing of raw 3D data, which requires significant computational resources and expertise [4][6]. Data Processing Pipeline - Tencent's Hunyuan team developed an automated data processing pipeline that filters and processes raw 3D assets into high-quality, training-ready data packages, significantly reducing the technical barriers for researchers [6][8]. - The pipeline includes initial filtering based on polygon count and UV mapping quality, followed by post-processing steps such as watertight processing and multi-view rendering [6][8]. Component-Level Data Processing - The component data processing aims to intelligently decompose static meshes into semantically consistent component sets, facilitating subsequent component-aware generation tasks [8][10]. - This process utilizes topological connectivity analysis to identify physically separated components within 3D assets, enhancing the modularity of 3D generation [8]. AIGC Synthesis - To address the scarcity of long-tail data, the team created a three-step generation pipeline that synthesizes data for embodied intelligent simulation needs [10][12]. - The pipeline includes text expansion using LLMs to generate detailed product descriptions, image generation using text-to-image models, and 3D asset generation using the HY3D-3.0 model [12]. Experimental Results - The lightweight model Hunyuan3D-2.1-Small, trained on the open-source dataset, demonstrates superior generation quality and inference speed compared to traditional methods, achieving a fivefold increase in speed while avoiding common issues like the "Janus Problem" [12][13]. - The dataset's scale includes 252,000 samples for manual modeling, 240,000 samples for component-level data, and 125,000 synthetic samples, providing a robust foundation for 3D generation tasks [13][19]. Future Plans - The team plans to expand the diversity of 3D assets and enhance multi-task adaptability, further exploring the potential of data-driven methods in 3D generation [20].