Vi3W
Skip to content

Guides

UniRig explained: open-source AI rigging, from skeleton to skin weights

UniRig explained: VAST's MIT-licensed AI rigging model, skeleton and skinning prediction, supported formats, GPU needs and the Blender workflow.

2 min read

A cartoon adventurer in a T-pose, half textured and half a glowing blue skeleton wireframe
Illustration. UniRig predicts the skeleton and the skin weights for a mesh like this.

Quick answer

What is UniRig?

UniRig is an open-source AI rigging model from VAST, the team behind Tripo, presented at SIGGRAPH 2025 under the MIT licence. It predicts a skeleton for a 3D model, then per-vertex skin weights, and merges them into a rigged FBX or GLB. It handles humans, animals and other shapes on a CUDA GPU with 8 GB or more.

What UniRig does

From its GitHub README, UniRig works in two stages:

  1. Skeleton prediction: a GPT-like transformer predicts a valid bone hierarchy using what the authors call Skeleton Tree Tokenization.
  2. Skinning weight prediction: a Bone-Point Cross Attention mechanism predicts per-vertex weights (and bone attributes, such as for physics) from the skeleton and the mesh.

Both stages are available in the current release, with a checkpoint trained on the Articulation-XL2.0 dataset and hosted on Hugging Face. Unlike Mixamo, it isn't limited to two-legged humanoids.

Requirements

  • Python 3.11 (per the README's setup)
  • A CUDA GPU with at least 8 GB VRAM for generation
  • Inputs: .obj, .fbx, .glb and .vrm

The three commands

The README's example uses a giraffe:

  1. Skeleton: bash launch/inference/generate_skeleton.sh --input examples/giraffe.glb --output results/giraffe_skeleton.fbx (change --seed for variations)
  2. Skin: bash launch/inference/generate_skin.sh --input examples/skeleton/giraffe.fbx --output results/giraffe_skin.fbx
  3. Merge: bash launch/inference/merge.sh --source results/giraffe_skeleton.fbx --target examples/giraffe.glb --output results/giraffe_rigged.glb

Batch versions take --input_dir and --output_dir.

A puppet linked to floating control rings
Illustration. AI places bones and weights; animators still add controls.

Using UniRig output in Blender

  1. File ‣ Import ‣ glTF 2.0 (or FBX) for the merged result.
  2. Select the armature, Ctrl+Tab into Pose Mode, and rotate major bones to test deformation.
  3. Fix weights in Weight Paint mode where joints collapse.
  4. Add IK and controls for animation; UniRig gives deform bones, not an animator-friendly control rig.

Checking rigs: Blender rig not moving mesh and AI rigging for Blender.

Where it fits

  • Creatures, animals and props that Mixamo can't rig.
  • Batch rigging many generated characters.
  • Research and pipelines where you can tune the setup.

For humanoids that only need stock motion, Mixamo is still quicker: Mixamo with an AI character.

Start with a clean mesh

Rigging, AI or not, works best on clean, watertight topology. Vi3W generates characters and creatures from a prompt, photo or sketch and retopologises them to clean quads, exported as GLB or FBX for UniRig or Blender. Vi3W doesn't rig itself. Try Vi3W.

Last reviewed