SkinTokens-bf16 (Swift-MLX)

Swift-MLX conversion of VAST-AI-Research/SkinTokens — a mesh auto-rigger (UniRig successor): a 3D triangle mesh (GLB) in → the same geometry with a skeleton + per-vertex skin weights (JOINTS_0/WEIGHTS_0) injected out.

⚠️ This is a custom multi-component pipeline, NOT a standard mlx_lm/mlx_vlm/mlx_audio model. It is consumed by the mlx-skintokens-swift package as the MLXEngine meshRig capability (contract 1.19.0, engine ≥ v0.30.0). Loading it standalone with mlx_lm will not work.

Contents

File	What
tokenrig.safetensors	bf16, 672 tensors — Qwen3-0.6B backbone + un-tied rig head + Michelangelo point encoder + embedded SkinVAE. The runtime loads THIS.
skinvae.safetensors	f32, 252 tensors — the standalone pretrained SkinVAE (FSQ [8,8,8,8,8] + chunked decoder); provenance / init.
skeleton_vroid.json	the canonical VRoid (VRM) bone template — the Route-B skinOnly humanoid skeleton order.
config.json	resolved TokenRig + SkinVAE + tokenizer config (reference; the Swift port hard-codes the architecture).

Pipeline

mesh GLB → SamplerMix(54000 pts) → Michelangelo encoder → SkinVAE._encode → TokenRig AR transformer (Qwen3-0.6B on inputs_embeds, grammar-constrained batched beam) → FSQ chunked decoder → per-point skin → cKDTree propagate to original verts → GLB inject. Two modes: auto (generate skeleton + skin) and skinOnly (skin a provided/embedded skeleton — the companion-character VRM path, J-in == J-out).

Licensing

• SkinTokens (the rigging model): MIT (VAST-AI-Research).
• Qwen3-0.6B backbone weights (embedded inside tokenrig.safetensors): Apache-2.0 (Alibaba/Qwen).

Both permissive. Converted for Apple-Silicon MLX inference by the Xocialize MLXEngine effort.