Load WorldStereo Model

WorldStereo is the "video" half of this pack - a Wan 2.1-based pipeline that generates new keyframes for a scene so you can extend and explore a reconstruction. VNCCS_LoadWorldStereoModel is the loader that pulls the whole thing together: the transformer checkpoint, the Wan VAE and text encoder, and Microsoft's MoGe depth estimator, all bundled into a single WORLDSTEREO_MODEL socket.

There are no required inputs. The first model_type choice is where your VRAM budget is decided:

• worldstereo-camera - 10.9 GB transformer. Feasible on 16 GB VRAM with offloading. This is the realistic default for most people.
• worldstereo-memory - ~22 GB. You want 24 GB+.
• worldstereo-memory-dmd - 34.9 GB distilled. 40 GB+. This is the one the official WorldGen runs use.

The settings that manage the damage

• precision - fp8 (default) or bf16. The tooltip is blunt: fp8 quantizes the heavy modules via optimum-quanto, bf16 is a half-precision fallback, "never fp32." If loading fails on fp8, bf16 is the fallback that usually works.
• offload_mode - sequential_cpu_offload (default; layer-by-layer, slowest but least VRAM), model_cpu_offload (move whole components between steps; faster, needs more VRAM), none (everything on GPU). This is the single biggest lever for making the 16 GB path actually run.
• turbo_lora - optional Wan 2.1 14B acceleration LoRAs (CausVid 14B rank32 v2, AccVid I2V 480P 14B rank32). Downloaded, loaded, fused, then the pipeline is quantized/offloaded. lora_strength (0–2, default 1) scales it. These exist because raw WorldStereo is slow enough that a turbo LoRA is often the difference between usable and never.
• device - cuda (default) or cpu.

The honest picture

Set expectations now: this is the fiddly, slow, bleeding-edge part of the pack, and the community has the scars to prove it. People report first renders taking tens of minutes to over two hours depending on model and hardware, and WorldStereo's inference code has a habit of refusing to cooperate (missing wrappers, missing depth inputs, "node continues to throw an error"). One careful commenter's verdict after getting it working: "unless you're willing to wait hours ... it may not be worth it." The author himself said the official models were "too heavy for me to even test."

If you're on this page because the workflow needs WorldStereo, budget: a big GPU, fp8, model_cpu_offload or sequential offload on 16 GB, and a turbo LoRA. And know that the loader is only step one - the WorldStereoGenerate node is where the actual inference happens. It's a load of a lot of moving parts, and it's the frontier, not the settled path.

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. PyTorch3D is required for WorldStereo point rendering and must be installed manually (pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable"), and the vendored gsplat_maskgaussian fork from install.py must remain the only gsplat.