unirig

name	unirig
description	Automatically rig 3D models with UniRig (VAST-AI-Research, SIGGRAPH'25) — predict a skeleton, predict skinning weights, and merge the rig back onto the original mesh. Use when the user wants auto-rigging for .obj/.fbx/.glb/.gltf/.dae/.vrm assets, a skeleton or skin weights for a character or creature, a UniRig environment prepared on a CUDA machine, batch rigging of a model directory, or an honest comparison between UniRig, SkinTokens, Tripo, Mixamo, AccuRig, and Blender Rigify. Triggers on: unirig, auto rig, auto-rigging, 3D rigging, skeleton prediction, skinning weights, rig a character, armature generation, rigged glb, rigged fbx, bone weights.
allowed-tools	Bash Read Write Edit Glob Grep WebFetch
compatibility	Requires Python 3.11 and an NVIDIA CUDA GPU for inference (spconv, flash_attn, torch_scatter/torch_cluster). The routing, dry-run planning, and GLB inspection paths work on any machine.
license	MIT
metadata	{"tags":"unirig, 3d-rigging, auto-rigging, skeleton-prediction, skinning-weights, blender, fbx, glb, vrm, character-animation, game-art, pytorch","version":"1.0","source":"https://github.com/VAST-AI-Research/UniRig"}

unirig — automatic 3D rigging (skeleton → skin → merge)

Keyword: unirig · auto-rigging · skeleton prediction · skinning weights

Only rig assets the user is licensed to modify. UniRig is MIT-licensed, but its checkpoints, the Rig-XL/VRoid/Objaverse data, and the user's own models each carry their own terms.

UniRig is a two-stage autoregressive rigging framework: a GPT-like transformer predicts a topologically valid skeleton from mesh geometry, then a bone-point cross-attention model predicts per-vertex skinning weights. A third step merges the predicted rig back onto the original (full-resolution, textured) asset.

The single most common failure is skipping the merge stage or merging the wrong file — see Step 5. The second most common failure is an environment that silently lacks CUDA extensions.

When to use this skill

• The user wants a static 3D character/creature/prop turned into a rigged asset
• The user needs a skeleton only, or skin weights for a skeleton they already edited
• The user needs UniRig installed and verified on a CUDA machine, or wants to know up front that their machine cannot run it
• The user wants to batch-rig a directory of models
• The user is choosing between UniRig, its successor SkinTokens, or classical/hosted riggers
• The user has a rigged output and wants to verify that joints and skin weights actually landed

Instructions

Step 1: Capture the intake packet and pick the stage

Collect four facts before running anything:

• Asset: file format (.obj, .fbx/.FBX, .dae, .glb, .gltf, .vrm), poly count, single model or directory, textured or not
• Goal: skeleton only, skin only, or a fully rigged deliverable
• Hardware: NVIDIA GPU + CUDA version, VRAM, OS (no CUDA ⇒ no inference; route out)
• Constraint: install budget (UniRig needs spconv, flash_attn, torch_scatter, torch_cluster, bpy), and whether hand-authored control over the skeleton is required

Routing rules:

1. Static mesh, no skeleton yet → ()