🌍 Load WorldMirror V2 Model

The V2 counterpart to VNCCS_LoadWorldMirrorModel. Same job - fetch Tencent's weights from Hugging Face, hand you a WORLDMIRROR_MODEL - but for tencent/HY-World-2.0, the newer checkpoint, and with one meaningful difference: you get to pick the precision.

V2 is heavier than V1. The node's defaults (device: cuda, precision: bf16) reflect that it's meant to run on the GPU, and the precision menu is where you actually manage your VRAM budget:

• bf16 - recommended. Roughly half the VRAM of float32 with no visible quality loss for this kind of work.
• fp8 - weight-only quantization via torchao, about another half again. Requires torchao and an Ampere+ GPU. If it fits and you're tight on VRAM, this is the one.
• float32 - full precision. You won't want this unless you're debugging or chasing reproducibility; it's a VRAM furnace.

How it fits

Load it, feed the model output into VNCCS_WorldMirrorV2_3D (or the _Clean / _Experimental variants), and reconstruct. It's the same socket type as V1, but don't try to mix loaders and inference nodes - the V1 and V2 reconstruction nodes expect the matching checkpoint and will do unhappy things otherwise.

One important reality check that applies to the whole pack, and the author said it in the release thread: "HY-World 2.0 isn't quite what it claims to be." What shipped is essentially a more advanced WorldMirror - the splat-generation and 3D-model half - not the full character-controlled 3D world from the promo videos. That unreleased half is why this loader exists at all; the reconstruction side is genuinely useful, the rest is roadmap.

Gotchas

• The checkpoint comes down from the HF hub into your HF cache, not ComfyUI/models. First load is a large download; don't panic if it takes a while.
• fp8 needs torchao installed and an Ampere+ GPU. If you pick fp8 on an older card or a missing dependency, the loader fails on import rather than giving you a graceful fallback - bf16 is the safe choice.
• Because this is a CUDA-heavy pack, the install step matters more than the node settings. Via Manager, search HY-World 2.0; manually:

cd ComfyUI/custom_nodes
git clone https://github.com/AHEKOT/ComfyUI_HYWorld2
cd ComfyUI_HYWorld2
pip install -r requirements.txt
python install.py

install.py builds the vendored gsplat_maskgaussian CUDA fork that must be the only gsplat in your environment - don't install upstream from PyPI. Budget for a CUDA Toolkit and, on Windows, MSVC Build Tools. V2 at decent resolution eats a lot of VRAM; the author's tested floor is a 16 GB 5070 Ti, and panorama assembly at high res is where V2 starts to strain a 16 GB card.