Hy3DGenerateMeshMultiView

Single-image 3D is a guessing game with extra steps: the shape model has to invent everything it can't see, which is why the backs of objects come out soft and blobby. Hy3DGenerateMeshMultiView is the upgrade for when you actually have more than one photo. Instead of one image, you feed it front, left, right, and back views, and the shape model conditions on all of them - the result is dramatically better geometry on the sides and back, with far less hallucination.

It's the same HY3DMODEL pipeline and the same latent output as Hy3DGenerateMesh, just with the conditioning spread across four views. This is the node behind kijai's second example workflow (hy3d_multiview_example_02.json), which is the multi-camera path people use when they're scanning a real object or a character sheet with a turn-around.

Inputs

• pipeline - the HY3DMODEL from Hy3DModelLoader.
• front, left, right, back - optional IMAGE sockets. Here's the clever bit: leave one unconnected (or connect a black/zero image) and the model falls back to single-image mode for that direction. So you can feed 4 views, 2 views, or 1 view - it degrades gracefully rather than erroring. In practice: 4 views for real objects you photographed; 1-2 for concept art where only the front is defined.
• guidance_scale (default 5.5), steps (default 30), seed (default 0) - same meaning as the single-image node.
• scheduler (optional) - FlowMatchEuler (default) or ConsistencyFlowMatchEuler for fewer steps.

Outputs are three: latents (HY3DLATENT → Hy3DVAEDecode), plus image and mask - the RGB views and their alpha masks the node assembled, which are handy for previewing exactly what the model saw.

When it's worth it

The honest cost/benefit: setting up four aligned views of a real object (photographing front/back/left/right with consistent framing) is real work, and the payoff is mostly in the occluded regions. For a prop you're going to 3D-print, the back being slightly speculative rarely matters. For something you'll display in the round, or where the back is actually visible and characterful, the multi-view path is a big step up and worth the setup.

One practical tip from the example workflow: each view gets background-removed and resized to a common resolution before it reaches this node, using the standard ImageRemoveBackground/ImageResize stack. Mismatched backgrounds or wildly different scales between views will confuse the model, so normalize them.

Install

Ships with the pack (ComfyUI Manager, search "Hunyuan3DWrapper", or clone into custom_nodes), needs the converted shape model in ComfyUI/models/diffusion_models/, and benefits from the same "new ComfyUI" requirement for the Preview3D nodes in the surrounding graph. No extra node-specific dependencies beyond the pack's requirements.txt.