Hunyuan3D V1 - Image2Views Pipeline Load

Of all the nodes in this pack, this one has the least glamorous job: it loads the multi-view diffusion (MVD) model into memory so the Image2Views node can use it. It's a loader. It has one input. Its entire output is a pipeline_config object that you wire into the next node. But that one input is a decision that matters, so it's worth two minutes of your time.

What use_lite actually does

The toggle is exactly what it sounds like. Set use_lite to true and it loads Tencent's lightweight MVD variant from weights/mvd_lite; set it to false (the default) and it loads the full standard model from weights/mvd_std. The lite model is smaller, uses less VRAM, and generates the six views faster - at a real but modest quality cost. Given that the full Hunyuan3D-1 stack already wants serious VRAM, this is your main lever for squeezing the whole thing onto a smaller card.

Here's the subtle part: the view order changes between the two models. The code reorders the 3×2 view grid differently depending on which pipeline you loaded ([0, 2, 4, 5, 3, 1] for standard, [0..5] for lite) so that the views line up with what the mesh stage expects. You never touch that ordering - it's handled for you - but it's why you shouldn't randomly switch use_lite mid-run and expect identical output. It also means whatever you pick here, the Views2Mesh Pipeline Load's use_lite toggle should match, or the view-to-mesh alignment can get confused.

The rest of it

The node takes a moment to load on first run - it's pulling real weights and pinning them to cuda:0 - and then it's passive. It doesn't generate anything itself. If you're on a low-VRAM setup, use_lite: true here is the single highest-impact setting in the whole image-to-3D chain. If you have the headroom and want the best views, run the standard model.

Install

Same pack-wide reality as everywhere else: not a Manager one-click. This pack needs a fresh (or very careful) ComfyUI, Python 3.12, C++ build tools, pytorch3d from source, an Open3D py3.12 dev wheel, nvdiffrast, Ninja, plus the pack's own requirements.txt. The weights for both the standard and lite MVD models come down as part of the one huggingface-cli download tencent/Hunyuan3D-1 --local-dir ./weights pull - you don't download anything extra for this node. And remember the README's headline warning: keep this pack away from your working ComfyUI install, because the dependency pileup here breaks environments.