Export WorldStereo Single Model

An exporter for people who are tired of Hugging Face directory sprawl. WorldStereo checkpoints normally live as multi-file diffusers folders; VNCCS_ExportWorldStereoSingleModel repacks one into a single .safetensors file - optionally quantized, optionally with a turbo LoRA fused in, plus a sidecar metadata JSON. It's the tool that makes a model portable: copy one file instead of a tree.

It's flagged experimental, and it's the kind of node you use once per model, then move on. If you never plan to move WorldStereo between machines, you can honestly skip it - but it's also the canonical way to produce the int4 single-file checkpoints that LoadWorldStereoLightModel consumes, so it's load-bearing in the pack's smaller-GPU story.

The inputs that matter

• export_path (required) - absolute path ending in .safetensors, or a directory. It can live outside ComfyUI; it's just a file path. If it's a directory, the node writes the file inside it.
• model_type - which checkpoint to export: worldstereo-camera, worldstereo-memory, worldstereo-memory-dmd, or the two aux models (official-clip-umt5-encoder, official-wan-vae) - handy for exporting the text encoder and VAE standalone, which is what the Light loader uses.
• precision - bf16 (default), fp8, or int4. The tooltip is worth quoting: "int4 is experimental and packs Linear weights to 4-bit with bf16/fp16 fallbacks." Exporting int4 here is exactly how the Light checkpoints get made.
• turbo_lora + lora_strength - optionally fuse a Wan 2.1 acceleration LoRA into the export, so the file ships pre-sped-up.
• device - cpu default, because the tooltip's logic is sound: CPU avoids VRAM pressure but needs lots of RAM; cuda is faster but risky. For a one-shot export, CPU is the safe call.
• overwrite - false by default; the export refuses to clobber an existing file until you flip it. This has bitten exactly once and it's why the flag exists.
• write_metadata_json - true by default; writes the sidecar JSON that the Light loader reads to pick model type and fallback precision.

Output is a single export_info string describing what was written.

Gotchas

• A full export is a big job: loading a 10.9–35 GB transformer, optionally quantizing it, and writing it back out. Budget disk space (int4 is smaller, but bf16/fp8 of the big models are still multi-GB) and patience.
• int4 output is only as good as quanto/torchao's packing - check the exported model actually loads before you delete the source directory.

Install

Same pack as always:

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.