Hunyuan3D

Tencent’s open-source image-to-3D asset model series, generating a textured 3D mesh from a picture.

Hunyuan3D is a series of 3D-asset generation models from Tencent’s Hunyuan team. Version 2.0 was open-sourced in January 2025, working in two steps: a flow-matching-based diffusion Transformer (Hunyuan3D-DiT) generates geometry from an input image, then Hunyuan3D-Paint generates a texture for the mesh. Version 2.1, open-sourced in June 2025, upgraded the texture to PBR materials (physically based rendering materials, including metalness and roughness); the shape model has 3.3B parameters and the texture model has 2B parameters, and the training code was released too. A technical report for version 2.5, published the same month, scaled the shape model up to as much as 10B parameters. In embodied AI, models like this can quickly turn a single photo into an object mesh to expand a simulation asset library — though the generated result still needs mass, friction, and other physical properties added, and a collision shape worked out, before it can go into physics simulation.

ExampleFeeding a photo of a mug to Hunyuan3D-2.1 produces a GLB mesh with a PBR texture; after converting the format, generating a collision shape with convex decomposition, and setting the mass, it can be imported into Isaac Sim as a graspable object.

Also called

Hunyuan3D 2.0, Hunyuan3D-2.1, Hunyuan3D 2.5, Tencent Hunyuan3D

Related

Single-Image 3D Reconstruction · Simulation Assets · Triangle Mesh · Physically Based Rendering · Convex Decomposition · HunyuanWorld (Tencent)

Sources

GitHub: Tencent-Hunyuan/Hunyuan3D-2（含版本时间线） (Chinese)
GitHub: Tencent-Hunyuan/Hunyuan3D-2.1
arXiv 2506.16504: Hunyuan3D 2.5

As of

2025-06