What it is
UniRig is a framework for automatic 3D rigging. It predicts a topologically valid skeleton hierarchy from a 3D mesh, then predicts per-vertex skinning weights from the skeleton and mesh geometry. The initial release provides code and a model checkpoint trained on Articulation-XL2.0 for skeleton and skinning prediction, while bone attribute prediction and full Rig-XL/VRoid checkpoints are listed as planned releases.[1]
What it can do in game development
- UniRig generates a skeleton for a 3D model in .obj, .fbx, .glb or .vrm format.[1]
- UniRig predicts per-vertex skinning weights from a mesh and a skeleton, and the README recommends refining the skeleton before this step for better results.[1]
- UniRig merges the predicted skinning result with the original 3D model to produce a rigged asset.[1]
How to install and run (as the README describes)
- The UniRig repository is cloned from GitHub.
[1]git clone https://github.com/VAST-AI-Research/UniRig - A conda environment named UniRig is created with Python 3.11.
[1]conda create -n UniRig python=3.11 - The UniRig conda environment is activated.
[1]conda activate UniRig - PyTorch and torchvision are installed with pip.
[1]python -m pip install torch torchvision - The project requirements are installed from requirements.txt.
[1]python -m pip install -r requirements.txt - spconv is installed for the CUDA version in use.
[1]python -m pip install spconv-{you-cuda-version}
What you need
- Python 3.11 is required.[1]
- PyTorch 2.3.1 or newer is the version listed as tested.[1]
- Generation requires a CUDA-enabled GPU with at least 8GB VRAM.[1]
- Training the skin model takes at least 60GB of memory on a single GPU, even with batch_size=2 in the data config.[1]
Limits to know
- Bone attribute prediction for physics-based secondary motion is listed as coming soon.[1]
- The README states that skinning results may degrade significantly if the skeleton is inaccurate, for example when tail bones or wing bones are missing, and it recommends refining the skeleton before skinning.[1]
- Merging a skeleton file produces no skinning, and the README says to merge the predicted skinning result for a skinned asset.[1]
Licence
MIT。 Read the licence file for what it allows.
Sources
- VAST-AI-Research/UniRig/README.mdindexed 2026-11-14
