"The foundation model behind Hunyuan3D 2.5 and 3.0."
Democratize High-Fidelity 3D Generation at Scale
Zeqiang Lai1,2★, Yunfei Zhao2★,
Zibo Zhao2, Haolin Liu2
Qingxiang Lin2, Jingwei Huang2,
Chunchao Guo2†, Xiangyu Yue1†
1MMLab, CUHK · 2Tencent Hunyuan
★ Equal contribution † Corresponding authors
"The foundation model behind Hunyuan3D 2.5 and 3.0."
VoxSet is a novel 3D representation that unifies voxel regularity with the flexibility of set-based models. It encodes geometry in a structured sparse format, enabling both efficient scaling and precise local detail.
This hybrid design allows the system to benefit from predictable scaling laws and consistent test-time improvements, bridging the gap between generated and handcrafted 3D assets.
Our generation pipeline adopts a two-stage strategy. The first stage generates a coarse structure with any off-the-shelf 3D Generator such as Hunyuan3D-2, and the second stage generate the detailed geometry via LATTICE.
Each latent in VoxSet is anchored to a 3D voxel grid, allowing direct positional embedding into the diffusion transformer for stronger spatial guidance and better model scaling.
Powered by Hunyuan3D-2
Powered by LATTICE
Explore the key features that make LATTICE stand out in 3D generation
The model achieves a level of accuracy approaching that of handcrafted designs, such as the correct number of fingers, the detailed bicycle wheel pattern, and even a bowl within a large scene.
Tiny and subtle decorations
Complex details
Correct number of fingers
Explore how training and test-time scaling affect mesh quality.
Browse, filter, and interact with generated 3D assets by Hunyuan3D 2.5.
Category
Polygons
Textures
Materials