Source-linked AI summary
HY3D-Bench: Generation of 3D Assets
Team Hunyuan3D, :, Bowen Zhang, Chunchao Guo, Dongyuan Guo, Haolin Liu, Hongyu Yan, Huiwen Shi, Jiaao Yu, Jiachen Xu, Jingwei Huang, Kunhong Li, Lifu Wang, Linus, Penghao Wang, Qingxiang Lin, Ruining Tang, Xianghui Yang, Yang Li, Yirui Guan, Yunfei Zhao, Yunhan Yang, Zeqiang Lai, Zhihao Liang, Zibo Zhao
TL;DR
3D generation is constrained by noisy, structurally incomplete raw assets and costly preprocessing. HY3D-Bench addresses this with training-ready geometry, structured part decomposition, and scalable synthetic data, and its dataset supports comparable generation quality despite fewer parameters than larger baselines. The paper positions this ecosystem as a unified resource for 3D generation and related downstream research.
Problem
Raw 3D repositories contain noise, non-manifold geometry, insufficient structural granularity, and costly preprocessing requirements that limit their use for 3D research.
Method
HY3D-Bench combines curated training-ready assets, structured part-level decomposition, standardized data resources, and AIGC synthesis for long-tail categories.
Results
Despite significantly fewer parameters than Trellis and Hunyuan3D 2.1, the trained model achieves comparable generation quality and outperforms similarly sized Craftsman.
Takeaways & Limitations
The ecosystem provides unified resources for 3D generation, fine-grained perception, controllable editing, robotics simulation, and digital content research.
Takeaways & Limitations
Feedforward generation remains constrained by 2D image resolution and limited learning of 3D spatial distributions, hindering fine-grained accurate geometries.
Abstract
from arXiv · showhide
While recent advances in neural representations and generative models have revolutionized 3D content creation, the field remains constrained by significant data processing bottlenecks. To address this, we introduce HY3D-Bench, an open-source ecosystem designed to establish a unified, high-quality foundation for 3D generation. Our contributions are threefold: (1) We curate a library of 250k high-fidelity 3D objects distilled from large-scale repositories, employing a rigorous pipeline to deliver training-ready artifacts, including watertight meshes and multi-view renderings; (2) We introduce structured part-level decomposition, providing the granularity essential for fine-grained perception and controllable editing; and (3) We bridge real-world distribution gaps via a scalable AIGC synthesis pipeline, contributing 125k synthetic assets to enhance diversity in long-tail categories. Validated empirically through the training of Hunyuan3D-2.1-Small, HY3D-Bench democratizes access to robust data resources, aiming to catalyze innovation across 3D perception, robotics, and digital content creation.
1 Introduction
HY3D-Bench establishes an open-source foundation for 3D content research by combining structured, training-ready assets, standardized evaluation, and scalable synthetic data generation.
- HY3D-Bench provides a high-quality, structured, and reproducible ecosystem supporting both 3D content understanding and creation.
- Its asset library includes watertight meshes, high-fidelity renderings, part decompositions, masks, and part-level meshes for fine-grained perception and part-aware generation.
- The benchmark standardizes evaluation through 400 high-quality objects, unified metrics, baselines, fixed settings, training configurations, and model checkpoints.
- The AIGC pipeline uses text-to-text, text-to-image, and image-to-3D stages to synthesize long-tail assets across 20 super-categories, 130 categories, and 1,252 fine-grained sub-categories.
- The ecosystem aims to lower barriers to 3D research and support generation, perception pre-training, robotics simulation, and broader generalization.
2 Related Work
3D generation has progressed from adversarial and optimization-based methods toward feedforward and native approaches, but dataset quality and preprocessing remain central constraints.
- The field evolved from manual modeling and scanning toward four paradigms: GAN-based, SDS-based, feedforward-based, and native 3D generation.
- GAN-based generation: GAN-based methods established high-fidelity synthesis, while explicit voxel and point-cloud representations faced cubic memory complexity and limited resolution.
- SDS-based generation: SDS-based methods address scarce annotated 3D data by optimizing differentiable 3D representations using pretrained 2D diffusion priors.
- Feedforward generation: Feedforward methods improve inference efficiency through direct prompt-to-3D mappings, including multiview synthesis followed by reconstruction and single-pass LRM approaches.
- Feedforward generation: Feedforward methods remain constrained by 2D image resolution and limited understanding of 3D spatial distributions, hindering fine-grained geometry generation.
- Native generation: Native methods learn 3D representations directly and can improve geometric consistency, but computational complexity and resolution still limit quality.
- 3D datasets: Benchmark datasets have historically suffered from limited categories, simple geometry, and small quantities, while large-scale preprocessing imposes substantial computational and expertise requirements.
- 3D datasets: HY3D-Bench addresses these dataset bottlenecks by providing 200k preprocessed samples with diverse categories, complex geometry, and watertight meshes at resolution 512.
3 Methods
The method uses a VAE to compress point-cloud geometry into latent vectors and reconstruct an SDF field, while diffusion models learn conditional generation from image embeddings.
- VAE: Given a surface point cloud with coordinates and normals, the VAE extracts features and resamples them into a latent vector set.
- VAE: A decoder reconstructs the signed distance function field from the latent representation, after which iso-surface extraction produces an explicit mesh.
- Diffusion: The 3D diffusion model encodes an input image into embeddings and uses DiT layers to predict noise or velocity during conditional generation.
4 Hunyuan Objarverse
HY3D-Bench addresses raw 3D datasets’ quality, structural, processing, and distribution gaps through standardized full-level and part-level assets, plus category-diverse synthetic data. Its resulting ecosystem provides training-ready geometry and representations spanning broad object categories and component complexities.
- Motivation: Raw 3D repositories contain format inconsistencies, geometric defects, severe category imbalance, and insufficient part structure for precise understanding, generation, and robotic interaction.These limitations include non-manifold geometry, scarce long-tail categories, and holistic meshes lacking hierarchical decomposition or assembly descriptions.
- Data Processing: HY3D-Bench processes Objaverse assets through cleaning, standardization, filtering, and post-processing to produce high-quality static meshes and structured part-level representations.The full-level workflow aligns orientations and formats, while part-level processing creates semantically consistent components and watertight holistic and part meshes.
- Full-level Data: 252,676 full-level assets are curated from Objaverse and Objaverse-XL as training-ready data, spanning 19 top-level categories, 74 mid-level subcategories, and 389 fine-grained classifications.The collection undergoes the complete processing pipeline before downstream use.
- Part-level Data: 240,524 part-level samples have a mean of 14.13 components and a median of 11, with 27.00% containing 11-20 components.The distribution also includes simpler 2-5-component and 6-10-component objects, plus highly complex objects containing 21-50 components.
- Synthetic Data: More than 125k synthetic samples cover 20 top-level categories, 130 mid-level subcategories, and 1,252 fine-grained product classifications.The synthetic category system spans everyday goods, consumer electronics, and professional industrial products.
5 Evaluation
The evaluation trains and assesses Hunyuan3D-2.1-Small using a progressively increasing token resolution and compares it with representative open-source 3D generation methods. Despite having fewer parameters than larger baselines, the model achieves comparable quality to them and surpasses similarly sized Craftsman when trained on HY3D-Bench.
- Evaluation Setup: ULIP and Uni3D measure consistency between images and generated meshes during evaluation.The experiments use Hunyuan3D-2.1 as the baseline while training a lightweight Small model.
- Model Architecture Adjustments: Hunyuan3D-2.1-Small is derived from Hunyuan3D-2.1 with reduced channel dimensions and a simplified dense architecture to lower training costs.The base channel dimension decreases from 2048 to 1536, and the Mixture of Experts structure is removed.
- Progressive Training Strategy: The training strategy progressively increases token resolution from 512 to 4096 tokens to improve representation fidelity.The detailed training configurations are provided in Table 1.
- Experimental Results: Hunyuan3D-2.1-Small achieves comparable generation quality to Trellis and Hunyuan3D 2.1 despite having significantly fewer parameters.The comparison includes Michelangelo, Craftsman, Trellis, and Hunyuan3D 2.1; quantitative results are reported in Table 2.
- Experimental Results: Hunyuan3D-2.1-Small outperforms the similarly sized Craftsman when trained on the open-sourced dataset.The authors use this result to support the quality of the dataset and its usefulness as a training resource.
6 Conclusion
HY3D-Bench addresses data processing bottlenecks in 3D generation through a high-fidelity asset library, structured part-level decomposition, and scalable AIGC synthesis. Its empirical validation with Hunyuan3D-2.1-Small supports the practical utility of these resources for 3D research.
- 6 Conclusion: HY3D-Bench provides a high-fidelity library of 252k 3D objects with watertight meshes and multi-view renderings.The objects are derived from large-scale repositories through a rigorous multi-stage pipeline.
- 6 Conclusion: The ecosystem includes 240k structured part-level decompositions for fine-grained perception, part-aware generation, and controllable 3D editing.
- 6 Conclusion: A scalable AIGC-driven synthesis pipeline contributes 125k synthetic assets to enrich diversity in long-tail categories.
- 6 Conclusion: Empirical validation using Hunyuan3D-2.1-Small confirms the practical utility of the dataset.
- 6 Conclusion: The authors aim for HY3D-Bench to catalyze innovation across 3D perception, robotics, and digital content creation.
7 Contributors
The paper lists its authors and other contributors alphabetically by first name.
- Authors: The authors are listed alphabetically by first name.
- Authors: The listed authors include Bowen Zhang, Chunchao Guo, Dongyuan Guo, Haolin Liu, Hongyu Yan, Huiwen Shi, Jiaao Yu, and Jiachen Xu.
- Authors: The author list continues with Jingwei Huang, Kunhong Li, Lifu Wang, Linus Penghao Wang, Qingxiang Lin, Ruining Tang, Xianghui Yang, Yang Li, Yunfei Zhao, Yunhan Yang, Zeqiang Lai, Zhihao Liang, and Zibo Zhao.
- Other Contributors: Other contributors are also listed alphabetically by first name.
- Other Contributors: The listed other contributors include Chao Zhang, Edwarrd Wang, Hao Zhang, Jiaxin Lin, Peng He, Yirui Guan, Yonghao Tan, and Zheng Ye.