TTP_Hunyuan3DNode

Tencent's Hunyuan3D is one of the few image-to-3D models you can actually run on your own GPU without an API key, and this node wraps the entire pipeline in a single box: feed it an image, get back a mesh. It's the whole reason the Comfyui_Hunyuan3D pack exists - the other two nodes in the pack exist to serve it. SquareImage preps your input to a square, and GifImageViewerNode shows you what it made.

Everything runs locally. No cloud, no key, no Tencent account. The trade is that it's a genuinely heavy pipeline - the first run pulls ~10GB of weights from Hugging Face - and slow enough that you'll want to preview with low steps before committing to a full render.

What it actually does

One node, four stages, all visible in the source:

1. Background removal - rembg's u2net cutout, the same salient-object model used in a thousand other ComfyUI workflows.
2. Multi-view generation - a diffusion model hallucinates six consistent views of your subject (front, sides, back).
3. Mesh reconstruction - an SVRM model turns those views into an actual 3D mesh with vertex colors.
4. Optional extras - texture mapping and a pytorch3d turntable GIF.

If the weights/svrm/svrm.safetensors file is missing, the node auto-downloads the whole tencent/Hunyuan3D-1 repo into a weights folder inside the pack on first run. Convenient - and also why your first execution takes a while. Expect the multi-view UNet to be around 5GB and the mesh model around 1GB.

Inputs that matter

• image - any ComfyUI IMAGE tensor. The model wants roughly square input around 1024×1024; run it through the pack's own SquareImage node first or the mesh quality quietly drops.
• step - 50 by default. More steps means better views at a linear time cost. For previews, 20–30 is plenty.
• max_number_of_faces - 90,000 by default, though the author's shipped workflow uses 200,000. Higher = denser mesh, slower. Most people won't need more than the default.
• do_texture_mapping - off by default, and that's the right call. It's the slow, fiddly stage, and it's the one the community's practical advice keeps saying to skip.
• do_render_gif - renders that 120-frame turntable GIF. Great for showing off, real time cost.
• use_lite - swaps the ~5GB multi-view UNet for the ~3.5GB lite version. Slightly worse view consistency, meaningfully less VRAM. If your card is struggling, this is the switch to flip.

Outputs

• rmbg_image and multiview_image - the cutout and the six-view sheet, so you can inspect each stage with a plain PreviewImage node.
• obj_file_path and glb_file_path - strings, not images. This is where people get stuck: wire them into ComfyUI-3D-Pack's [Comfy3D] Preview 3DMesh node to view inside ComfyUI, or just open them in Windows 3D Viewer, Blender, or MeshLab.
• gif_file_path - the turntable, only when do_render_gif is on. Wire it into GifImageViewerNode.

Every run lands in ComfyUI/output/3D_output/<number>/ with img_nobg.png, views.gif, mesh_with_colors.obj, and mesh.glb. The pack keeps 30 numbered folders round-robin and deletes the oldest when the set fills, so copy out anything you want to keep.

Installing

ComfyUI Manager: search Comfyui_Hunyuan3D and install. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/TTPlanetPig/Comfyui_Hunyuan3D

On the portable Windows build, run the pack's installer from its own folder:

cd ComfyUI/custom_nodes/Comfyui_Hunyuan3D
..\..\..\python_embeded\python.exe install.py

That script pulls the ugly part: pytorch3d and nvdiffrast compiled from source, plus rembg, open3d, trimesh, xatlas, and libigl. If the build fails on Windows, the README blames a Python-version mismatch and points you at MrForExample's Comfy3D_Pre_Builds for prebuilt binaries to drop into python_embeded.

Gotchas

• Viewing is the trap. The node runs fine without ComfyUI-3D-Pack, but seeing the mesh inside ComfyUI requires it - and that pack is notoriously the hardest custom node to install, needing a CUDA compiler and Visual Studio build tools on Windows. The README's own advice: skip it and open the output folder.
• It's slow and it isn't magic. Community reports run to minutes per generation even on decent cards, and complex shapes come out blobby. Great for a quick concept mesh, not a production asset.
• License. Hunyuan ships under Tencent's community license - commercial use is fine under 100M monthly active users, but the grant excludes the EU, UK, and South Korea entirely.