Source-linked AI summary
Hunyuan3D 2.5: Towards High-Fidelity 3D Assets Generation with Ultimate Details
Zeqiang Lai, Yunfei Zhao, Haolin Liu, Zibo Zhao, Qingxiang Lin, Huiwen Shi, Xianghui Yang, Mingxin Yang, Shuhui Yang, Yifei Feng, Sheng Zhang, Xin Huang, Di Luo, Fan Yang, Fang Yang, Lifu Wang, Sicong Liu, Yixuan Tang, Yulin Cai, Zebin He, Tian Liu, Yuhong Liu, Jie Jiang, Linus, Jingwei Huang, Chunchao Guo
TL;DR
Existing 3D generation methods struggle with fine-grained shape detail and consistent multiview texturing. Hunyuan3D 2.5 addresses these challenges with a two-stage diffusion pipeline, a scaled shape foundation model, and PBR-oriented multiview material generation. The paper reports stronger shape fidelity, texture realism, and end-to-end results than prior methods, including a 72% image-to-3D user-study win rate.
Problem
Existing models remain limited on complex objects with fine-grained details, while inconsistent multiview images can produce texture artifacts and seams.
Method
Hunyuan3D 2.5 uses a two-stage image-to-3D pipeline with LATTICE for shape generation and a multiview material framework producing albedo, roughness, and metallic maps.
Results
Hunyuan3D 2.5 outperforms current state-of-the-art models in shape detail, surface smoothness, texture consistency, and end-to-end generation.
Takeaways & Limitations
The system provides a model suite for generating realistic, detailed textured 3D assets across diverse real-world images.
Abstract
from arXiv · showhide
In this report, we present Hunyuan3D 2.5, a robust suite of 3D diffusion models aimed at generating high-fidelity and detailed textured 3D assets. Hunyuan3D 2.5 follows two-stages pipeline of its previous version Hunyuan3D 2.0, while demonstrating substantial advancements in both shape and texture generation. In terms of shape generation, we introduce a new shape foundation model -- LATTICE, which is trained with scaled high-quality datasets, model-size, and compute. Our largest model reaches 10B parameters and generates sharp and detailed 3D shape with precise image-3D following while keeping mesh surface clean and smooth, significantly closing the gap between generated and handcrafted 3D shapes. In terms of texture generation, it is upgraded with phyiscal-based rendering (PBR) via a novel multi-view architecture extended from Hunyuan3D 2.0 Paint model. Our extensive evaluation shows that Hunyuan3D 2.5 significantly outperforms previous methods in both shape and end-to-end texture generation.
1 Introduction
3D generation is expanding across industries, but existing methods still struggle with fine-grained shape detail and consistent multiview texturing. These limitations leave high-fidelity, detailed 3D shape generation an open problem.
- 3D generation supports applications ranging from games and embodied AI to film effects and virtual reality.Advances in artificial intelligence have improved automated modeling and texturing efficiency.
- Recent shape diffusion models have advanced 3D generation, but complex objects with fine-grained details remain difficult to generate.The paper identifies high-fidelity and detailed shape generation as an open problem.
- Multiview diffusion methods reduce global consistency problems associated with inpainting and synchronization techniques.
- Highly consistent multiview images remain challenging, leading to artifacts and seams during texture generation.
SHAPE GENERATION
Hunyuan3D 2.5 retains a two-stage shape-and-texture pipeline while introducing scaled shape modeling and PBR-oriented material generation. Extensive evaluations report improvements in shape fidelity, texture realism, and end-to-end visual quality.
- Shape generation: LATTICE is a new shape foundation model trained with larger high-quality datasets, increased model size, and greater computational resources.The model exhibits stable improvement when scaling up.
- Shape generation: The largest model generates sharp, detailed shapes aligned with corresponding images while maintaining clean and smooth surfaces.The authors report that this significantly closes the gap between generated and handcrafted 3D shapes.
- Texture generation: The texture stage extends prior models into a high-fidelity material framework that follows a principled BRDF model.It simultaneously produces multiview albedo, roughness, and metallic maps to describe surface reflection properties and microsurface distributions.
- Pipeline: Hunyuan3D 2.5 retains a two-stage pipeline that generates shape first and texture from that shape second.
- Texture generation: A dual-phase resolution enhancement strategy improves texture-geometry coordination and end-to-end visual quality.
- Evaluation: Extensive quantitative, qualitative, and user-study evaluations across in-the-wild images report that Hunyuan3D 2.5 outperforms open-source and closed-source commercial models.
2 Method
Hunyuan3D 2.5 uses a two-stage image-to-3D pipeline that first generates geometry and then produces detailed materials. Its method combines the LATTICE shape model with multi-view PBR generation, alignment mechanisms, and resolution-enhancement training.
- Overall Pipeline: The image-to-3D pipeline preprocesses an input image, generates an untextured mesh, extracts geometric maps, and synthesizes texture from those outputs.The extracted outputs include normals and UV maps.
- Detailed Shape Generation: LATTICE is a large-scale diffusion model that generates detailed shapes with sharp edges and smooth surfaces from single or four multi-view images.Guidance and step distillation reduce inference time.
- Detailed Shape Generation: The largest scaled shape model produces fine-grained details, including correct fingers, bicycle-wheel patterns, and nested scene objects, approaching handcrafted-design accuracy.These examples are presented in the first row of Figure 4.
- Detailed Shape Generation: Hunyuan3D 2.5 balances sharp edges with smooth, clean surfaces, addressing a difficulty reported for existing models on complex objects.The comparison is illustrated in the second and third rows of Figure 4.
- Realistic Texture Generation: Its material framework generates multi-view PBR maps using mesh-derived normal and CCM conditions, a reference image, and 3D-aware RoPE for cross-view consistency.The framework generates albedo, roughness, and metallic information as material outputs.
- Realistic Texture Generation: Separate albedo, MR, and normal embeddings, shared attention masks, and illumination-invariant consistency training promote channel separation, spatial coherence, and texture-geometry alignment.A two-phase strategy uses 6-view 512×512 training followed by zoom-in training, with inference at up to 768×768 resolution.
3 Evaluation
The evaluation compares Hunyuan3D 2.5 with open- and closed-source baselines for shape generation and textured 3D asset generation. Results report strong shape similarities, improved visual quality, and superior PBR texture generation and user preference.
- Evaluation Setup: Evaluations cover 3D shape generation and textured 3D asset generation using open-source and commercial baselines.The shape baselines include Michelangelo, Craftsman 1.5, Trellis, Hunyuan3D-2, and two commercial models.
- Metrics: ULIP-T, ULIP-I, Uni3D-T, and Uni3D-I measure similarity between generated meshes, input images, and synthesized image prompts.Texture evaluation uses FID, CLIP-FID, LPIPS, CMMD, and CLIP-I.
- Shape Generation: Hunyuan3D 2.5 achieves the best reported ULIP-T, Uni3D-T, and Uni3D-I shape similarities.The authors caution that these metrics do not fully reflect model capabilities and supplement them with visual comparisons.
- Shape Generation: Visual comparisons indicate that Hunyuan3D 2.5 performs better than the other open-source and commercial shape-generation models.Figure 6 provides the visual comparison referenced by the evaluation.
- Texture Generation: For PBR material generation, competing methods struggle to estimate metallic and roughness values and to remove illumination effects from albedo.Figure 7 compares front and back views, complete material maps, and albedo maps.
- User Study: In the image-to-3D user study, Hunyuan3D 2.5 achieves a 72% win rate, 9 times higher than Commercial Model 1.Participants ranked methods on diverse real-world images and the study compared three commercial models.
4 Related Work
Related work spans diffusion-based 3D shape generation and multiview diffusion for texture synthesis. These approaches address shape generation and the global consistency of textured 3D assets.
- Texture Synthesis: Figure 8 presents a user study comparing end-to-end textured results against three commercial models.The figure is an evaluation reference rather than a related-work method.
- Texture Synthesis: Multiview diffusion methods use cross-view attention or related interactions to address global consistency in 3D textures.The related work includes methods such as Zero123++ that spatially concatenate multiview images for interaction modeling.
5 Conclusion
Hunyuan3D 2.5 is presented as a 3D diffusion-model suite for detailed assets, combining advances in shape generation and texture synthesis. Evaluations report improvements in shape fidelity, surface smoothness, texture consistency, and overall asset quality.
- Conclusion: Hunyuan3D 2.5 is an advanced suite of 3D diffusion models for generating high-quality, detailed 3D assets.The conclusion describes the system at the paper level rather than as a single model.
- Conclusion: The reported improvements span shape fidelity, texture realism, shape detail, surface smoothness, and texture consistency.These outcomes are stated as results of the presented system.
- Conclusion: The authors report that Hunyuan3D 2.5 outperforms current state-of-the-art models across the evaluated shape and texture qualities.The conclusion frames the method as a tool for creating realistic and detailed 3D assets.
6 Contributors
The listed contributors are organized by overall contribution, shape generation, texture generation, and additional reported contributors.
- Contributors: Chunchao Guo, Jingwei Huang, and Zeqiang Lai are listed as contributors.
- Shape Generation: The shape-generation contributors are Zeqiang Lai, Yunfei Zhao, Jingwei Huang, Haolin Liu, Zibo Zhao, Qingxiang Lin, Huiwen Shi, and Xianghui Yang.
- Texture Generation: The texture-generation contributors are Mingxin Yang, Shuhui Yang, Yifei Feng, Sheng Zhang, and Xin Huang.
- Contributors: Additional listed contributors are Di Luo, Fan Yang, Fang Yang, Lifu Wang, Sicong Liu, Yixuan Tang, Yulin Cai, Zebin He, Tian Liu, and Yuhong Liu.