Thelocallab/CUDA-wheels

Prebuilt Windows CUDA Wheels for TRELLIS.2 / 3D Generation in ComfyUI

Precompiled Python wheels for the CUDA extensions that TRELLIS.2 needs — so you don't have to set up MSVC, the CUDA toolkit, and Triton on Windows just to run image-to-3D.

If you've ever watched setup.sh --cumesh --o-voxel --flexgemm --nvdiffrast --nvdiffrec fail on Windows, this is for you.

⚠️ Read this before downloading — licences differ per package

These are builds of other people's source code, not original work, and they are not all under the same terms:

• cumesh, flex_gemm, o_voxel — components of microsoft/TRELLIS.2, MIT licensed. Unmodified compilations, redistributed under the same terms.
• nvdiffrast, nvdiffrec_render — NVIDIA research code, governed by the NVIDIA Source Code License (1-Way Commercial), which restricts use to non-commercial research and evaluation. That restriction applies to compiled binaries as much as to source. TRELLIS.2's own README carves these two out as "governed by its own License."

If your use is commercial, do not use the nvdiffrast or nvdiffrec_render wheels here. Contact NVIDIA Research Licensing instead. Mirroring these files does not grant you rights you wouldn't otherwise have, and nothing in this repo modifies NVIDIA's terms — read them yourself before you install.

What's here

Package	Version	Python / Platform	Upstream	Licence
cumesh	1.0	cp311-win · cp312-win · cp312-linux	TRELLIS.2	MIT
flex_gemm	0.0.1	cp311-win · cp312-win · cp312-linux	TRELLIS.2	MIT
o_voxel	0.0.1	cp311-win · cp312-win · cp312-linux	TRELLIS.2	MIT
nvdiffrast	0.4.0	cp311-win · cp312-win · cp312-linux	NVlabs	NVIDIA, non-commercial
nvdiffrec_render	0.0.0	cp311-win · cp312-win · cp312-linux	NVlabs	NVIDIA, non-commercial

What each one does:

• CuMesh — CUDA mesh processing: post-processing, remeshing, simplification, UV unwrapping
• FlexGEMM — Triton-based sparse convolution
• O-Voxel — conversion between textured meshes and the O-Voxel representation
• nvdiffrast — differentiable rasterisation, used to render the generated 3D assets
• nvdiffrec_render — split-sum renderer for PBR materials

Note that cp311 is Windows-only here; cp312 covers Windows and Linux.

Install

Match the wheel to your ComfyUI Python version. Check it first:

python -c "import sys; print(sys.version)"

Then install into the same environment ComfyUI uses — activate your venv first, or use the portable build's embedded Python:

# Windows, Python 3.12 — MIT-licensed components
pip install cumesh-1.0-cp312-cp312-win_amd64.whl
pip install flex_gemm-0.0.1-cp312-cp312-win_amd64.whl
pip install o_voxel-0.0.1-cp312-cp312-win_amd64.whl

For ComfyUI portable, prefix with the embedded interpreter:

python_embeded\python.exe -m pip install <wheel>

Verify:

python -c "import cumesh, flex_gemm, o_voxel; print('ok')"

Troubleshooting

• is not a supported wheel on this platform — Python version mismatch. cp311 wheels require Python 3.11 and cp312 requires 3.12. They are not interchangeable.
• Imports fine in a terminal but ComfyUI can't find them — you installed into the wrong environment. ComfyUI portable does not use your system Python.
• CUDA errors at runtime — these are compiled extensions, so your installed PyTorch CUDA build has to match. Check with python -c "import torch; print(torch.version.cuda)".
• Still stuck — build from upstream source instead. That always works; it's just slow.

Guides

Full walkthroughs, including the ComfyUI node setup these wheels are for:

• Microsoft TRELLIS 2 image-to-3D in ComfyUI
• Hunyuan 3D in ComfyUI

More local-AI guides at locallabdigest.com.

No warranty

Third-party compiled binaries, provided as-is, with no warranty of any kind. If running precompiled CUDA extensions from someone else isn't acceptable in your situation, build from upstream source. All credit for the underlying work belongs to the upstream authors listed above.