PaperlayerView exact revision on arXiv

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

arXiv:2602.03907v1cs.CVcs.AI

TL;DR

Raw 3D repositories contain noisy, non-manifold assets with limited structural granularity, and preparing task-specific data for large-scale training requires substantial computation. HY3D-Bench provides processed, structured assets and synthetic long-tail content; a model trained on its open dataset achieved quality comparable to larger models and outperformed a similarly sized model.

  • Problem

    Raw 3D repositories contain noisy, non-manifold geometry and limited structural granularity that restrict their use in precise geometric understanding, stable generation, and robotic interaction, while large-scale preprocessing is computationally burdensome.

    Source: §1 · p. 1, §2.2 · p. 4

  • Method

    HY3D-Bench curates training-ready 3D assets with watertight meshes, multi-view renderings, and part-level decomposition, and adds AIGC-synthesized content spanning long-tail categories.

    Source: §1 · p. 2

  • Results

    The trained model achieved generation quality comparable to larger Trellis [86] and Hunyuan3D 2.1 [27] models, while outperforming similarly sized Craftsman.

    Source: §5.2 · p. 13

  • Takeaways & Limitations

    The experiments support the authors’ conclusion that their open-sourced dataset is high quality and that data quality plays a crucial role in 3D generation.

    Source: §5.2 · p. 13

    Feedforward 3D generation methods remain constrained by 2D image resolution and limited learning of 3D spatial distributions, making fine-grained, accurate geometry challenging.

    Source: §2.1 · p. 4

Abstract

from arXiv · show

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 is an open-source ecosystem combining data engineering, standardized benchmarks, and scalable AIGC synthesis to support 3D content understanding and creation. It aims to lower research barriers and advance unified understanding and generation of the 3D world.

1 Introduction

  • The asset library provides holistic and part-level assets with watertight meshes, high-fidelity renderings, segmentation results, and 2D masks for part-aware generation and fine-grained perception.A part-merging strategy produces structured assets with consistent part-level decomposition.Source: §1 · p. 2
  • The benchmark covers 400 high-quality objects and provides standard metrics, baselines, fixed experiment settings, training configurations, and pretrained checkpoints for fair comparisons.It addresses fragmentation caused by inconsistent evaluation protocols.Source: §1 · p. 2
  • The AIGC pipeline uses Text-to-Text semantic expansion, Text-to-Image visual synthesis, and Image-to-3D mesh reconstruction to produce long-tail content spanning 20 super-categories, 130 categories, and 1,252 fine-grained sub-categories.It is intended to supplement training data for rare categories and support generation robustness and robotics simulation diversity.Source: §1 · p. 2
  • The ecosystem provides data and infrastructure for downstream 3D generation, perception pre-training, and robotics simulation.Source: §1 · p. 2

2 Related Work

Related work traces 3D generation across four paradigms and identifies both fine-grained geometry limitations and costly dataset preprocessing as barriers that motivate high-quality, training-ready data.

2.1 3D Generation

  • GAN-based methods enable rapid inference but can suffer training instability and mode collapse on diverse, open-domain datasets.The paradigm evolved from direct voxel-grid and point-cloud generation toward NeRF-based synthesis and hybrid tri-plane representations.Source: §2.1 · pp. 2–3
  • SDS methods distill pretrained 2D diffusion priors into 3D representations, but often face Janus artifacts and long optimization times.Later approaches optimize 3D Gaussian parameters to improve generation speed while maintaining visual fidelity.Source: §2.1 · p. 3
  • Feedforward methods amortize generation through multiview reconstruction or single-pass LRM models, yet remain limited in fine-grained geometric accuracy.Their limitations include dependence on 2D image resolution and weak learning of 3D spatial distributions.Source: §2.1 · pp. 3–4
  • Native generation learns 3D representations directly, supporting part-aware and controllable outputs, but depends on high-quality processed data and faces complexity and resolution limits.Related approaches use segmentation-based decomposition or generate explicit part structures; the paper offers processed data for training 3D VAEs and diffusion models.Source: §2.1 · p. 4

2.2 3D Datasets

  • Mainstream models require rendered images, watertight meshes, and SDF fields; preparing millions of models demands substantial GPU and CPU resources.These preprocessing requirements raise the entry barrier and hinder rapid iteration and adoption.Source: §2.2 · p. 4
  • HY3D addresses costly preprocessing with 200k training samples curated from Objaverse and Objaverse-XL, including watertight meshes at 512 resolution.The dataset is tailored for 3D VAEs and diffusion models and aims to reduce computational and technical burdens.Source: §2.2 · pp. 4–5

3 Methods

The method combines a VAE that maps point-cloud shape inputs through latent vectors to reconstructed SDF meshes with image-conditioned diffusion for conditional 3D generation.

VAE

  • The VAE encodes surface-sampled point clouds into latent vectors, decodes them into an SDF field, and extracts an explicit mesh from its iso-surface.The input points include surface normals; the latent set has length L and dimension d.Source: §3 · p. 5

Diffusion

  • Image-conditioned diffusion encodes an image and denoises a shape’s latent representation, with DiT predicting added noise or velocity.For flow matching, training transforms noise sampled from N(0, I) into the data distribution conditioned on image embeddings.Source: §3 · p. 5

4 Hunyuan Objarverse

HY3D-Bench addresses raw Objaverse data’s inconsistent formats and quality, long-tail category gaps, and lack of part structure through full-level and part-level processing plus synthetic asset generation. The resulting collections provide training-ready meshes, multiview data, and broad category coverage.

4 Hunyuan Objarverse

  • Raw Objaverse assets have inconsistent formats and construction, uneven mesh and texture quality, long-tail category imbalance, and mostly lack structured part information.These shortcomings hinder downstream generation, fine-grained understanding, editing, and robotic manipulation.Source: §4 · pp. 5–6

4.1 Existing Enhanced Objaverse Dataset

  • Prior Objaverse enhancements addressed quality filtering [44], orientation normalization [54], and representation diversity [68], but left processing incomplete and training workflows burdensome.HY3D-Bench instead supplies a complete pipeline and training-ready packages, supplemented with product assets.Source: §4.1 · p. 6

4.2 Full-level Data Processing

  • Full-level processing standardizes assets as oriented static meshes, renders orthographic and perspective views, filters geometric, texture, and thin-structure issues, then creates watertight meshes and hybrid-sampled point clouds.Meshes are exported in PLY format; watertight processing uses a UDF grid and graph-cut tetrahedral labeling.Source: §4.2 · p. 7

4.3 Part-level Data Processing

  • Part-level processing splits meshes by topological connectivity, filters and merges components for usable granularity, and produces RGB views, part-ID masks, and watertight whole-object and individual-part meshes.Final samples combine semantically consistent parts with multiview data and support part-aware generation, fine-grained understanding and editing, and robotic manipulation.Source: §4.3 · pp. 8–9

4.4 Synthetic Data Generation

  • Synthetic generation expands product categories into detailed LLM descriptions, uses LoRA-tuned Qwen-Image to create 3D-suitable images, and generates assets with HY3D-3.0.The image tuning targets clean backgrounds, complete centered objects, informative viewpoints, and visible structural and material details.Source: §4.4 · pp. 9–10

4.5 Data Distribution and Visualization

  • The collection contains 252,676 full-level assets, 240,524 part-level samples, and more than 125k synthetic samples.Full-level assets span 19 top-level categories, 74 mid-level subcategories, and 389 fine-grained classifications; part-level samples average 14.13 components (median 11); synthetic data spans 20, 130, and 1,252 categories at those levels, respectively.Source: §4.5 · pp. 10–12
  • Full-level data covers 19/74/389 category levels; part-level component counts cluster across 2–5 (24.63%), 6–10 (24.83%), and 11–20 (27.00%); synthetic data covers 20/130/1,252 category levels.The remaining part-level samples contain 21–50 components.Source: §4.5 · pp. 10–12, Figure 6 · p. 10, Figure 8 · p. 11

5 Evaluation

The evaluation tests the full-level dataset by training Hunyuan3D-2.1-Small and comparing its 3D generation quality with representative open-source methods. It also describes the model’s architecture and progressive training setup.

5.1 Implementation Details

  • Hunyuan3D-2.1-Small reduces the base channel dimension from 2048 to 1536, replaces MoE with a Dense architecture, and contains 832M parameters.Source: §5.1 · p. 12
  • Training progressively increases token count from 512 to 4096; image-mesh consistency is evaluated using ULIP [88] and Uni3D [103].Hunyuan3D-2.1 is the baseline, and the Small model is trained to reduce costs.Source: §5.1 · pp. 12–13

5.2 Experimental Results

  • Despite having far fewer parameters than Trellis and Hunyuan3D 2.1, Hunyuan3D-2.1-Small achieves comparable generation quality and outperforms similarly sized Craftsman.The comparison is reported for image-to-3D generation on the authors’ test dataset.Source: §5.2 · p. 13, Table 2 · p. 13

6 Conclusion

HY3D-Bench is an open-source ecosystem that addresses data-processing bottlenecks for 3D generation through a training-ready object library, structured part decompositions, and synthetic assets. Validation with Hunyuan3D-2.1-Small supports its practical utility, and future work will extend the framework to dynamic assets and broader tasks.

6 Conclusion

  • HY3D-Bench curates 252k high-fidelity 3D objects from large-scale repositories, using a multi-stage pipeline to produce training-ready watertight meshes and multi-view renderings.The repositories include Objaverse and Objaverse-XL.Source: §6 · p. 13
  • The ecosystem introduces 240k structured part-level decompositions to support fine-grained perception, part-aware generation, and controllable 3D editing.Source: §6 · p. 13
  • Its scalable AIGC-driven synthesis pipeline contributes 125k synthetic assets to enrich diversity in long-tail categories and mitigate real-world data distribution gaps.Source: §6 · p. 13
  • Evaluation with Hunyuan3D-2.1-Small confirms the dataset’s practical utility; future work will extend the framework to dynamic assets and broader tasks.The authors aim for HY3D-Bench to catalyze innovation across 3D perception, robotics, and digital content creation.Source: §6 · p. 13

Related papers

Loading 2602.03907v1…