PaperlayerView exact revision on arXiv

Source-linked AI summary

Hunyuan3D Studio: End-to-End AI Pipeline for Game-Ready 3D Asset Generation

Biwen Lei, Yang Li, Xinhai Liu, Shuhui Yang, Lixin Xu, Jingwei Huang, Ruining Tang, Haohan Weng, Jian Liu, Jing Xu, Zhen Zhou, Yiling Zhu, Jiankai Xing, Jiachen Xu, Changfeng Ma, Xinhao Yan, Yunhan Yang, Chunshi Wang, Duoteng Xu, Xueqi Ma, Yuguang Chen, Jing Li, Mingxin Yang, Sheng Zhang, Yifei Feng, Xin Huang, Di Luo, Zebin He, Puhua Jiang, Changrong Hu, Zihan Qin, Shiwei Miao, Haolin Liu, Yunfei Zhao, Zeqiang Lai, Qingxiang Lin, Zibo Zhao, Kunhong Li, Xianghui Yang, Huiwen Shi, Xin Yang, Yuxuan Wang, Zebin Yao, Yihang Lian, Sicong Liu, Xintong Han, Wangchen Qin, Caisheng Ouyang, Jianyin Liu, Tianwen Yuan, Shuai Jiang, Hong Duan, Yanqi Niu, Wencong Lin, Yifu Sun, Shirui Huang, Lin Niu, Gu Gong, Guojian Xiao, Bojian Zheng, Xiang Yuan, Qi Chen, Jie Xiao, Dongyang Zheng, Xiaofeng Yang, Kai Liu, Jianchen Zhu, Lifu Wang, Qinglin Lu, Jie Liu, Liang Dong, Fan Jiang, Ruibin Chen, Lei Wang, Chao Zhang, Jiaxin Lin, Hao Zhang, Zheng Ye, Peng He, Runzhou Wu, Yinhe Wu, Jiayao Du, Jupeng Chen, Xinyue Mao, Dongyuan Guo, Yixuan Tang, Yulin Tsai, Yonghao Tan, Jiaao Yu, Junlin Yu, Keren Zhang, Yifan Li, Peng Chen, Tian Liu, Di Wang, Yuhong Liu, Linus, Jie Jiang, Zhuo Chen, Chunchao Guo

arXiv:2509.12815v1cs.CV

TL;DR

Hunyuan3D Studio addresses the labor-intensive, fragmented process of creating technically ready 3D game assets. It integrates neural modules for generation, segmentation, topology, UVs, and texturing into an end-to-end platform, and the paper reports game-engine-compatible assets with reduced production time and technical barriers.

  • Problem

    3D asset creation requires specialized, costly workflows, while existing generative systems often address isolated stages without simultaneously ensuring visual fidelity and game-engine readiness.

  • Method

    Hunyuan3D Studio integrates modular neural systems for part-level generation, segmentation, topology-aware mesh processing, semantic UVs, and texture synthesis into one asset-creation pipeline.

  • Results

    The authors report that the integrated system produces assets meeting modern game-engine requirements while reducing production time and technical barriers.

  • Takeaways & Limitations

    The platform supports rapid iteration by connecting creative intent to optimized, production-oriented 3D assets within a unified workflow.

Abstract

from arXiv · show

The creation of high-quality 3D assets, a cornerstone of modern game development, has long been characterized by labor-intensive and specialized workflows. This paper presents Hunyuan3D Studio, an end-to-end AI-powered content creation platform designed to revolutionize the game production pipeline by automating and streamlining the generation of game-ready 3D assets. At its core, Hunyuan3D Studio integrates a suite of advanced neural modules (such as Part-level 3D Generation, Polygon Generation, Semantic UV, etc.) into a cohesive and user-friendly system. This unified framework allows for the rapid transformation of a single concept image or textual description into a fully-realized, production-quality 3D model complete with optimized geometry and high-fidelity PBR textures. We demonstrate that assets generated by Hunyuan3D Studio are not only visually compelling but also adhere to the stringent technical requirements of contemporary game engines, significantly reducing iteration time and lowering the barrier to entry for 3D content creation. By providing a seamless bridge from creative intent to technical asset, Hunyuan3D Studio represents a significant leap forward for AI-assisted workflows in game development and interactive media.

1 Introduction

Hunyuan3D Studio addresses the gap between advances in 3D generation and the technical requirements of game-ready assets. It presents an integrated workflow intended to reduce manual effort and support production-quality asset creation.

  • Traditional 3D asset creation is complex, costly, and time-consuming because modeling, UV mapping, texturing, and rigging require specialized software expertise.This workflow can bottleneck game production and limit creative iteration and accessibility.
  • Recent generative systems improve geometry generation from images or text, but often fail to combine visual fidelity with technical readiness for real-time game engines.Existing solutions commonly address isolated pipeline components rather than complete usable assets.
  • Hunyuan3D Studio integrates the full 3D asset workflow to transform high-level creative concepts into game-engine-ready assets with minimal manual intervention.The platform is designed as a comprehensive production environment rather than a standalone geometry generator.
  • The paper details the system architecture and its core modules before discussing their technical implementations and concluding with limitations and future work.

2 Hunyuan3D Studio Pipeline

Hunyuan3D Studio organizes seven technological modules into a sequential but integrated pipeline from concept creation to game-engine export. Shared metadata, parametric control, and reversibility connect the stages while preserving artistic and technical control.

  • The pipeline comprises seven core modules spanning concept design, geometry generation, part decomposition, polygon generation, semantic UV unwrapping, texture synthesis, and animation.Together, these modules address the major stages of game asset creation.
  • Controllable Image Generation (Concept Design): Controllable image generation supports text-to-image and image-to-multi-view synthesis, pose standardization, and style adaptation before 3D modeling.
  • Polygon Generation (PolyGen): PolyGen autoregressively constructs low-polygon assets from geometric surfaces, producing low vertex counts and deformation-aware edge flow for game use.
  • Semantic UV Unwrapping: Semantic UV unwrapping groups surfaces by material and texel-density requirements, while texture synthesis produces PBR sets with nondestructive natural-language editing.
  • Animation Module: An animation module infers joints, bone hierarchies, and vertex weights, and the unified asset graph propagates metadata through reversible downstream updates before Unity or Unreal export.

3 Controllable Image Generation

The controllable image-generation pipeline prepares consistent visual inputs for 3D generation through stylization and pose standardization. It uses pretrained generative models with adapted conditioning and progressive training to preserve character identity and improve fidelity.

  • The pipeline combines image stylization and pose standardization to prepare controllable inputs for subsequent 3D model generation.
  • Image Stylization: Image stylization adapts a Qwen-Image-Edit model with LoRA to preserve subject content while applying predefined game-art styles.For text-only input, the system first generates a reference image and then applies the same stylization process.
  • Pose Standardization: Pose standardization conditions FLUX.1-dev DiT on arbitrary-pose character images to generate standard A-pose views while removing backgrounds and props.
  • Pose Standardization: Progressive training from 512 × 512 to 768 × 768 improves fidelity for facial details and complex clothing textures.Randomized same-character references support generalized pose control and consistent generation across scenarios.

4 High-Fidelity Geometry Generation

The geometry-generation pipeline combines latent shape modeling with image, bounding-box, and multi-view conditioning. These controls are intended to improve alignment, proportions, and high-fidelity character geometry from limited visual input.

  • Hunyuan3D-ShapeVAE encodes point clouds with positions and normals into shape latents, while Hunyuan3D-DiT generates those latents using flow-based diffusion with Mixture-of-Experts layers.
  • Single-image conditioning uses a background-removed 518 × 518 image encoded by DINOv2 and fused with shape latents through cross-attention.
  • Conditional Generation: Bounding-box conditioning encodes height, width, and length into a shape latent that is concatenated with the image condition, with training perturbations encouraging reliance on the box signal.
  • Visualization: The bounding-box signal produces high-quality geometry with appropriate proportions and well-structured forms, including cases where image-only generation fails.
  • Conditional Generation: Generated multi-view images are used as additional conditioning to produce high-fidelity 3D character assets.The multi-view pipeline encodes original and generated views with positional information before forming the final condition.

5 Part-level 3D Generation

The part-level generation pipeline combines automatic 3D segmentation with controllable shape decomposition to produce semantically meaningful, editable, and structurally coherent parts. P3-SAM provides native 3D segmentation, while X-Part uses bounding-box cues and semantic features for controllable decomposition and editing.

  • 5.1 P3-SAM: P3-SAM’s image-to-part pipeline obtains a holistic mesh, detects semantic features and part bounding boxes, then decomposes the mesh with X-Part.The workflow is illustrated in Figure 10.
  • 5.1 P3-SAM: P3-SAM automatically segments complex 3D objects into precise, robust, and semantically coherent part masks using a single positive point prompt.Its architecture uses a feature extractor, three segmentation heads, and an IoU prediction head.
  • 5.1 P3-SAM: A dataset of 3.7 million artist-created meshes with automatically generated part masks enables P3-SAM to achieve scalable and globally coherent 3D part segmentation.
  • 5.2 X-Part: X-Part generates semantically meaningful and structurally coherent parts through a controllable, editable diffusion framework conditioned on object point clouds.
  • 5.2 X-Part: X-Part uses bounding boxes for location and scale control, preserves high-dimensional semantic features, and supports local editing such as merging parts or adjusting their scales.Bounding-box guidance is intended to reduce overfitting compared with finer segmentation cues, while semantic features improve part decomposition.
  • 5.2 X-Part: X-Part achieves state-of-the-art performance in part-level decomposition and generation, evaluated with Chamfer Distance and F-Score on ObjaversePart-Tiny.

6 Polygon Generation with Auto-regressive Models

The polygon-generation pipeline autoregressively converts point-cloud conditions into cleaner mesh topology, then refines localized defects through topology-aware masked preference optimization. Post-training improves completeness and topology quality, while the model generalizes to complex meshes and supports part-aware generation without fine-tuning.

  • 6.1 Mesh Pretrianing with Efficient Tokenization and Architecture: An autoregressive mesh decoder predicts vertices and faces from point clouds to produce clean topology for downstream applications such as UV segmentation and rigging.The architecture combines a point-cloud encoder with an Hourglass Transformer mesh decoder conditioned through cross-attention.
  • 6.1 Mesh Pretrianing with Efficient Tokenization and Architecture: Mesh sequences are tokenized with Blocked and Patchified Tokenization, while truncated training and rolling-cache inference improve efficiency for high-poly meshes.BPT uses spatial block indexing and patch aggregation; training samples fixed-size face slices and inference uses a rolling cache.
  • 6.2 Mesh Post-Training with Topology-Aware Masked DPO: Masked DPO uses quality-aware region masks to refine low-quality mesh areas while preserving satisfactory regions.Regions are high quality only when their faces meet quad-ratio and average-topology thresholds; candidate preferences are built from BER, TS, and HD evaluations.
  • 6.3 Experiments: Post-training substantially improves mesh completeness and topology quality, reducing broken faces and enhancing connectivity.Figure 16 reports gains in completeness and connectivity alongside fewer broken faces; experiments identify post-training as crucial for these improvements.
  • 6.3 Experiments: The model generates more complex meshes with significantly improved topology quality and stability compared with existing polygon-generation methods.The comparison is presented in Figure 17.
  • 6.3 Experiments: Part-aware generation reconstructs meshes from segmented partial point clouds separately without further fine-tuning, making complicated topology easier to generate.The capability is demonstrated in Figure 18.

7 Semantic UV

SeamGPT treats semantic UV unwrapping as autoregressive seam generation, producing artist-style cuts that improve UV quality and editability for production workflows.

  • 7 Semantic UV: SeamGPT generates semantic mesh-cutting seams as an ordered sequence of 3D line segments using an autoregressive model.The method encodes seams as sequential tokens and predicts them in an artist-like decision process.
  • 7 Semantic UV: SeamGPT achieves the best performance across all reported UV-unwrapping metrics and produces semantic cuts across varied surface characteristics.XAtlas over-fragments cuts, FAM misses subtle cuts, and Edge-CLS struggles on smooth regions.
  • 7 Semantic UV: Professional 3D artists rate SeamGPT significantly higher than existing methods for both boundary quality and editability.The study involved 20 professional 3D artists using a 5-point scale.
  • 7.4 Ablation Study: Edge- and vertex-aware point-cloud sampling aligns generated seams with mesh topology and improves model convergence.This alignment helps avoid excessive extra mesh faces by preserving existing vertices and edges.
  • 7.4 Ablation Study: Point-cloud encoding outperforms mesh encoding, whose computational cost scales poorly and whose outputs can misalign with original mesh vertices.The comparison attributes the advantage to more precise shape-position representation.
  • 7.4 Ablation Study: Valid seam granularity typically occurs when the seam-segment-to-vertex ratio R lies between 0.1 and 0.35.Higher values cause over-cutting, lower values insufficient cuts, while controlled autoregressive sampling can produce diverse valid seams.

8 Texture Generation and Editing

Hunyuan3D Studio extends geometry-conditioned texture synthesis into multimodal editing and high-resolution PBR material generation for localized or global asset changes.

  • 8 Texture Generation and Editing: The framework addresses cross-view consistency, geometric alignment, and expansion from RGB textures to photorealistic PBR materials.It lifts a 2D diffusion model into a geometry-conditioned multi-view generator and bakes outputs into high-resolution texture maps.
  • 8 Texture Generation and Editing: The texture system supports text- and image-guided editing, material-based segmentation, and 4K tileable PBR material-ball synthesis.It generates Base Color, Metallic, Roughness, and Normal maps from textual prompts.
  • 8 Texture Generation and Editing: A unified model performs texture synthesis and editing under both textual and visual guidance after training on curated PBR assets and text-texture pairs.The training data includes 80k high-quality 3D assets and 30,000 text-texture pairs.
  • 8 Texture Generation and Editing: Multimodal editing produces diverse material styles on game props and characters, both globally and locally.The reported demonstrations are shown in Figure 24.
  • 8 Texture Generation and Editing: A PBR-VAE compresses multi-domain material data into unified latent representations to support scalable 4K texture synthesis.The system adapts a 3D VAE and fine-tunes a diffusion transformer for material generation.

9 Animation Module

The animation module routes characters through humanoid or general-character branches, combining template-based or autoregressive skeleton generation with topology-aware skinning.

  • 9 Animation Module: The animation system detects each input and routes humanoid characters to a template-based branch or other characters to a general-character branch.This branching accommodates differences in skeletal topology and joint counts.
  • 9 Animation Module: The humanoid branch uses a 22-joint template, motion retargeting, skeletal-and-vertex skinning features, and pose standardization to canonical T-pose.T-pose inputs make motion-retargeting outcomes more reliable and precise.
  • 9 Animation Module: The general-character branch combines autoregressive skeleton generation with geometry topology-aware skinning for varied character structures.The branch is designed for characters with different skeletal topologies and joint counts.
  • 9 Animation Module: The method produces more detailed general-character results with fewer errors and achieves higher overall general-skinning accuracy than existing approaches.Figure 26 compares the rigging and skinning results with UniRig.

10 Conclusion

Hunyuan3D Studio unifies generation, segmentation, retopology, UV unwrapping, texturing, and related asset-processing modules into an end-to-end game-ready workflow.

  • 10 Conclusion: Hunyuan3D Studio transforms single-image or text inputs into game-ready assets with optimized geometry, PBR textures, and engine-compatible topology.Its modular architecture combines geometry generation, component-aware segmentation, PolyGen retopology, and semantic UV unwrapping.
  • 10 Conclusion: The integrated platform is presented as reducing production time and technical barriers, enabling artists and developers to iterate more rapidly on creative work.The conclusion frames this benefit as arising within the unified production workflow.

11 Contributors

The project credits identify sponsors, leaders, and contributors across PolyGen, Part, UV, Texture, Animation, Geometry, Image, Engineering, Data, and Art Designer roles.

  • Jie Jiang, Linus, Yuhong Liu, Di Wang, Tian Liu, and Peng Chen are listed as project sponsors.
  • Chunchao Guo and Zhuo Chen are listed as project leaders.
  • The contributor list spans technical modules including PolyGen, Part, UV, Texture, Animation, Geometry, and Image.
  • Engineering, Data, and Art Designer contributors are also credited.
Loading 2509.12815v1…