skin-tokens.cpp: SkinTokens auto-rigging ported to CPU without PyTorch

The skin-tokens.cpp port to C++23/GGML moves the SkinTokens/TokenRig auto-rigger from VAST-AI-Research, which previously required an NVIDIA GPU with 14 GB of memory and PyTorch, to CPU and Vulkan.

Compiled by Sergey KostenchukPublished 2026-09-01Updated 2026-09-18

2026-09-01 Coding NVIDIA

🦴 skin-tokens.cpp: auto-rigging on CPU without PyTorch

A port of the SkinTokens/TokenRig auto-rigger from VAST-AI-Research to C++23/GGML has been released by localai-org on GitHub: the model generates a skeleton and skinning weights from a static GLB mesh or binds a mesh to an existing skeleton. Weights are in GGUF on Hugging Face.

🌍 The original required an NVIDIA GPU with 14 GB of memory and PyTorch. The port allows embedding auto-rigging into C++ tools and game pipelines, and the Trellis2 + Kimodo combination closes the “mesh → skeleton → motion” loop on regular hardware. The claimed closeness to the PyTorch reference is inference equivalence, not rigging quality.

👤 GGUF weights are on Hugging Face, the CLI is built with CMake and runs on CPU: rig, skin, web demo in the browser, C API skintokens.h. Code is Apache-2.0, the distribution terms for the GGUF weights are not yet confirmed.

Source 1: https://github.com/localai-org/skin-tokens.cpp Source 2: https://github.com/VAST-AI-Research/SkinTokens

Sources