Hunyuan3D V1 - Views2Mesh Pipeline Load

Last loader in the chain, and it has the most personality of the three. While the other pipeline-load nodes handle text-to-image and multi-view diffusion, this one loads the SVRM - the sparse-view reconstruction model that actually turns your six flat views into a 3D mesh. No model, no mesh. It's the difference between "I generated some views" and "I have a .obj file."

What it loads and the use_lite toggle

Under the hood it constructs Tencent's MV23DPredictor, loading the weights from weights/svrm/svrm.safetensors and pinning them to cuda:0. That's a chunky load - it's the single biggest model in the pack's mesh half - and it can take a while the first time.

Its one input, use_lite, mirrors the Image2Views Pipeline Load toggle. Set it true and the mesh stage uses the lightweight path; false uses the full model. The quality difference shows most in fine geometry, and the VRAM savings are real, which matters because the mesh stage is exactly where the pack's documented random torch.OutOfMemoryError hits. The README's advice for that is blunt and practical: just rerun with the same seed and it usually goes through. If it's recurring, use_lite: true here is the first thing to try.

One wiring gotcha worth repeating because it's easy to miss: keep the use_lite setting consistent between this node and the Image2Views Pipeline Load. The two stages use different view orderings depending on which variant they loaded, and mismatched settings is the classic way to end up with a mesh that ignores half your views.

The output

It produces a single pipeline_config - a Hunyuan3D1Views2MeshPipelineConfig - that wires directly into the Views2Mesh node's pipeline_config input. That's it. Load the model, hand the config down, stay out of the way.

Install

Pack-wide install, no exceptions: fresh ComfyUI strongly recommended, Python 3.12, C++ build tools, pytorch3d built from source (the README's own wheel-build recipe is the reliable path), an Open3D py3.12 development wheel, nvdiffrast, Ninja, plus the pack's requirements.txt. The SVRM weights arrive with the main download - huggingface-cli download tencent/Hunyuan3D-1 --local-dir ./weights from inside the pack folder. And keep this pack out of your main ComfyUI install if you can; the README is unambiguous that this dependency stack breaks environments.