SkinTokens

SkinTokens is an AI rigging project for 3D artists and researchers preparing meshes for character animation or simulation. Its TokenRig model generates a skeleton hierarchy and the skin weights that control how the mesh moves with its bones. Local inference requires an NVIDIA GPU with at least 14 GB of memory.

• Complete rig generation: Turn a single 3D mesh into an asset with both a skeleton and per-vertex skin weights for import into standard 3D workflows.
• Joint skeleton and skin prediction: TokenRig predicts both parts of the rig in one sequence, rather than using the separate prediction stages of its predecessor, UniRig.
• Local browser interface: Use the model through a Gradio interface running on your machine, with pretrained models available from Hugging Face.

TokenRig uses a Qwen3-0.6B-based model and reinforcement learning refinement aimed at joint coverage, keeping bones within the mesh and producing smoother deformation on unfamiliar assets. SkinTokens represents the skin weights as compact tokens that the model can generate alongside the skeleton.

The Python project is open source under the MIT license and requires CUDA for local inference. A separate hosted Hugging Face Space lets you upload a mesh and receive a rigged result in the browser; that route sends your mesh to the hosted service.