(Down)Load Hy3D PaintModel

Hunyuan3D is two models wearing a trenchcoat. The shape model builds the geometry; this node loads the paint model, the second half that generates the texture. DownloadAndLoadHy3DPaintModel downloads a diffusers paint pipeline from Tencent's Hunyuan3D-2 repo into ComfyUI/models/diffusers/, loads it as a multiview_pipe, and that pipe is what Hy3DSampleMultiView runs to paint consistent views of your mesh from every angle.

The "multiview" name on the output socket is the clue to how it works. The paint model is a 2.5D-conditioned UNet: you give it the reference image plus the mesh's normal maps and position maps from Hy3DRenderMultiView, and it generates a ring of images of the object from different camera angles - all sharing a consistent style and lighting. Those images then get baked into the UV texture by Hy3DBakeFromMultiview.

Inputs

• model - a dropdown with two choices: hunyuan3d-paint-v2-0 (the full model) and hunyuan3d-paint-v2-0-turbo (the distilled fast variant). Turbo is what it sounds like: notably faster at the cost of some fidelity. Start with the full model for quality; switch to turbo once you're iterating on the rest of the graph.
• compile_args (optional) - a HY3DCOMPILEARGS socket for torch.compile of the UNet and VAE. Requires Triton; torch 2.5.0 recommended per the tooltip. Without it, the pipeline enables model CPU offload instead.

Output: multiview_pipe (HY3DDIFFUSERSPIPE) → Hy3DSampleMultiView. Both paint variants sit in the same download cache, so switching between them doesn't re-download the shared parts.

Wiring it up

The paint stage is the widest part of the workflow and where most of the GPU time goes:

Hy3DGenerateMesh → Hy3DVAEDecode → Hy3DMeshUVWrap → Hy3DRenderMultiView
                                                                 │
DownloadAndLoadHy3DPaintModel → Hy3DSampleMultiView ←───────────┘
                                     ↓
                              Hy3DBakeFromMultiview → ...

Gotchas

The download is the first big moment: hunyuan3d-paint-v2-0 is a multi-gigabyte snapshot from Hugging Face the first time you run it. It lands in ComfyUI/models/diffusers/, so pre-downloading there beats re-fetching. The node deliberately skips the unet/diffusion_pytorch_model.bin and image_encoder files and loads safetensors instead, so a partial or aborted download can leave the folder in a state where the node errors - just delete the folder and let it re-download.

And the pack's standing caveat: Hy3DSampleMultiView (the node this feeds) calls into the compiled custom_rasterizer, which only ships as a prebuilt wheel for Windows 11 / Python 3.12 / CUDA 12.6. On other setups you'll build it yourself or hit No module named 'custom_rasterizer'. That's the biggest single install hurdle in the whole pack, and it's right here at the paint stage.

Install: ComfyUI Manager, search "Hunyuan3DWrapper", or git clone https://github.com/kijai/ComfyUI-Hunyuan3DWrapper into custom_nodes, then pip install -r requirements.txt.