🌐 Load WorldMirror Model

Every WorldMirror workflow in this pack starts here. VNCCS_LoadWorldMirrorModel grabs Tencent's WorldMirror V1 weights from Hugging Face (tencent/HunyuanWorld-Mirror), loads them, and hands you a WORLDMIRROR_MODEL handle to plug into VNCCS_WorldMirror3D. It's the least glamorous node in the pack and the one you can't skip.

There's genuinely nothing to configure before the first run - no required inputs at all. First load downloads ~5 GB into the Hugging Face cache, so budget a coffee break. After that it's instant.

The knobs you can ignore

Everything here is optional, and most of it you can leave alone:

• device - cpu by default, which is safe for loading but slow for inference. Set cuda if your VRAM allows; you'll see VNCCS_WorldMirror3D offer its own offload scheme anyway.
• sampling_strategy - uniform or conservative. This is how the model picks Gaussian splat sampling points. Uniform is the default; conservative trades coverage for fewer, higher-confidence points. If your output looks noisy, conservative is a cheap experiment.
• enable_conf_filter + conf_threshold_percent - filters the lowest-confidence points right at load time. Default off because VNCCS_WorldMirror3D has its own confidence_percentile knob that does the same job later, but setting it here bakes the filter into the model's behavior for every run.

Output is a single model socket of type WORLDMIRROR_MODEL, which only the reconstruction nodes in this pack accept - you can't mix it with a V2 model or anything from outside the pack.

Why you'd ever touch this instead of a full loader

Honest answer: you wouldn't, much. This node exists to keep the model-loading step separate from inference so you can re-run reconstruction with different settings without re-downloading or re-loading the checkpoint. The two settings that genuinely matter are the sampling strategy and confidence filter, because they change the character of the output at the source rather than as a post-process.

One real gotcha: the model lives in the Hugging Face cache, not in your ComfyUI models folder. If you're the type who likes everything inside ComfyUI/models/checkpoints, that'll annoy you - and if you're offline after a first clean install, the loader fails until you've fetched the weights once.

Install

Search HY-World 2.0 in ComfyUI Manager, or:

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

The install.py step is not optional - it builds the vendored gsplat_maskgaussian CUDA fork that replaces upstream gsplat, and it's the most common source of "why won't this pack import." Don't install gsplat from PyPI separately; the pack needs its own fork installed under that name. On Windows, that means MSVC C++ Build Tools and a CUDA Toolkit matching your PyTorch. Author's word: it runs on a 16 GB 5070 Ti, and to set expectations correctly, the whole thing is "not very stable yet" - but the loader is the boring, dependable part.