Install HYWorld2 Dependencies

ComfyUI_HYWorld2 is one of the most dependency-heavy packs in the 3D corner, and this node is the author's answer to that: a single node you queue like any other workflow step, and it runs pip install on the pack's requirements.txt, builds the vendored gsplat fork, and tries to install PyTorch3D. It's an output node (marked OUTPUT_NODE), so it only runs when it's actually in your graph - which is by design: you drag it in, hit Queue, watch the log.

It takes exactly one input, install, which is locked to RUN_DEPENDENCY_INSTALL - there's no choice to make. The output is a single install_log string, and the real progress appears in the ComfyUI console.

What it actually installs

Under the hood it runs the same things the README tells you to run manually:

pip install -r requirements.txt
python scripts/build_gsplat.py   # builds the vendored gsplat_maskgaussian CUDA fork

The heavy hitters in requirements.txt are the 3D stack - open3d, trimesh, plyfile, ninja, pybind11, plus MoGe (pinned from git+https://github.com/microsoft/MoGe@…), diffusers, optimum-quanto, bitsandbytes, and a cupy-cuda13x for the WorldGen trajectory pipeline. PyTorch3D is attempted separately because it can't go in the plain requirements list. If you already ran the manual install, this node's job is mostly done and it's a no-op safety net.

The honest gotchas

• It only works if the prerequisites for a compile are in place. Building gsplat_maskgaussian needs a CUDA Toolkit matching your PyTorch's CUDA version, ninja, and on Windows the MSVC C++ Build Tools (Desktop development with C++ workload). No toolkit → the build fails and the node reports it in the log.
• The pip step can take a long time and may look stalled. flash-attn in particular is a notorious from-source compile (the README suggests pip install flash-attn --no-build-isolation separately); the node doesn't hide that, but budget for it.
• Do not install upstream gsplat from PyPI in the same environment. This pack needs its vendored fork to be the only package named gsplat - the fork adds distloss and gauss_masks rasterization arguments the native trainer and worldgen require. Mixing them in is the single most common cause of mysterious failures in this pack.
• If your ComfyUI Python is the embedded Windows one, the README points at scripts\pipinstall.bat as the safer path for exactly this reason - pip inside the embedded runtime can be finicky.

Verifying success

The pack ships a CUDA smoke test that calls gsplat.rendering.rasterization with distloss=True and gauss_masks - if that succeeds, your fork is healthy. Otherwise, the install node's log is the first place to look.

Install the pack itself via ComfyUI Manager (search HY-World 2.0) or git clone https://github.com/AHEKOT/ComfyUI_HYWorld2 into custom_nodes, then restart before queuing this node.