Tencent’s diffusion pipeline builds 3D assets mesh-first, paints later

What it does Hunyuan3D-2 generates high-resolution textured 3D assets through two distinct diffusion stages. A flow-based transformer, Hunyuan3D-DiT, first creates a bare mesh aligned to a condition image. Then Hunyuan3D-Paint synthesizes high-resolution texture maps for that mesh. Because the stages are decoupled, you can also throw hand-crafted geometry at the texture model.

The interesting bit Instead of a monolithic end-to-end model, the project treats shape and texture as separate problems with separate diffusion priors. This avoids the common trap where one bad texture forces a full regeneration. The repo also serves as a model zoo, shipping Turbo and Fast distilled variants, a 0.6B mini model, multiview shape models, and a Blender addon.

Key highlights

• Claims to outperform unnamed open- and closed-source baselines on geometry detail, condition alignment, and texture quality according to the project’s benchmark table.
• Shape generation runs on 6 GB VRAM; full shape-plus-texture generation requires 16 GB.
• The texture model accepts either generated or hand-crafted meshes, making it useful for retexturing existing assets.
• Community and official ecosystem includes ComfyUI wrappers, a Windows portable build, and a Blender addon.
• Supports both image-to-3D and text-to-3D workflows (per repository topics).

Caveats

• The performance table compares against anonymous “Top Close-source Model” entries, so the SOTA claims are hard to verify independently.
• The project has already moved on to v2.1, v2.5, and a separate world-generation model (HunyuanWorld-1.0), meaning this repo may lag behind Tencent’s latest releases.

Verdict A solid bet if you want open-source 3D generation with a clean split between mesh creation and texturing. Look elsewhere if you need a single lightweight, unified model; this is a modular system with a serious GPU appetite.