Tencent introduced Hunyuan3D 1.0. A diffusion model generates several views of the object, and a reconstruction network uses those views to build a 3D asset. For text prompts, the pipeline first generates a reference image with Hunyuan-DiT.
Hunyuan3D
Also known as: Tencent Hunyuan3D · Hunyuan 3D · HY 3D
Plain English
Tencent's Hunyuan3D family generates textured 3D objects from image or text inputs, depending on the version. The 2.1 release includes downloadable weights and training code for generating shape and materials that react to light.
Technical Overview
Hunyuan3D 2.x splits generation into a flow-matching shape diffusion transformer and a paint model that generates textures from multiple views of the mesh. Version 2.1 adds physically based rendering (PBR) maps to the paint stage, which takes either the generated shape or a supplied mesh.
Timeline (5)
2024
Paper2025
ReleaseTencent released inference code and pretrained weights for Hunyuan3D 2.0. Hunyuan3D-DiT synthesizes a mesh from an image, and Hunyuan3D-Paint can texture that mesh or a mesh supplied by the user.
- Release
Tencent released Hunyuan3D 2.1 with model weights and training code under its Hunyuan 3D 2.1 Community License. The new paint model predicts material maps such as metallic and roughness alongside color so lighting can change how the same asset looks.
- Release
Tencent released Hunyuan3D-Omni, a 3.3B image-to-shape model derived from Hunyuan3D 2.1. Its control encoder accepts a point cloud, voxel grid, skeleton, or bounding box to constrain the generated shape.
- Ecosystem
Tencent announced the global launch of its Hunyuan 3D creation engine and enterprise API. The launch highlighted Hunyuan3D 3.0 for generating individual objects and also described separate Hunyuan3D World models for interactive environments.