Load WorldStereoLight Model

The full WorldStereo pipeline (VNCCS_LoadWorldStereoModel) needs 11–35 GB of transformer. This is the diet version. VNCCS_LoadWorldStereoLightModel loads pre-quantized int4 WorldStereo checkpoints from MIUProject/VNCCS_WorldStereoLight on Hugging Face, so the whole thing actually fits in consumer VRAM. The trade, as ever with aggressive quantization, is quality - but for a pipeline this heavy, "runs at all" is often the win.

The two choices

Required input is model, with exactly two options:

• Camera Light INT4 - the camera model quantized to 4-bit. The sensible default.
• Memory DMD Light INT4 - the 34.9 GB distilled memory model squashed to int4. Still heavy, but a fraction of the float version's footprint.

Optional: offload_mode (default sequential_cpu_offload, plus model_cpu_offload and none) and device (cuda default). No precision menu - the int4-ness is baked into the checkpoint, not a runtime toggle. Each model ships with metadata JSON that tells the loader which model type and precision to fall back on, so the node mostly figures itself out.

How it fits

Same output socket (WORLDSTEREO_MODEL) as the full loader, so it drops into the same WorldStereoGenerate workflows. The changelog notes that recent versions avoid downloading the unused full Wan text-encoder and VAE weights when these single-file Light assets are present - so it's not just smaller to run, it's smaller to download too.

Reasonable expectations: int4 on a generative model this big is genuinely experimental territory. It'll get you keyframe generation on a 12–16 GB card that the full pipeline laughs at, with visible quality cost. If you have the VRAM for the full fp8 loader, use that. If you don't, this is why it exists.

Gotchas

• It still needs the pack's full install - MoGe, PyTorch3D, diffusers, the vendored gsplat_maskgaussian fork. "Light" describes the model, not the setup.
• The first download pulls a multi-GB int4 .safetensors from HF into the cache.
• If generation errors after loading, verify the worldstereo-wrapper dependency - it's the part of WorldStereo's ecosystem that's most often missing and the error message won't tell you.

Install

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

Or ComfyUI Manager → HY-World 2.0.