SkinTokens

From the project README

SkinTokens mesh auto-rigging extension for Modly. It wraps the upstream VAST-AI SkinTokens / TokenRig runtime and exposes a Modly process node that converts an input mesh into a rigged .glb workflow output.

This repository contains the Modly extension shell only: manifest, setup entrypoint, runtime adapter, readiness checks, and tests. SkinTokens model weights are not committed to this repository.

What this extension provides Install / Repair through setup.py with JSONL progress events for the Modly UI. A Modly process node named Rig Mesh with SkinTokens. Deterministic SkinTokens source staging under the extension runtime directory. SkinTokens checkpoints and required config files stored under Modly's normal models/ storage. Runtime progress stages for setup, readiness, SkinTokens inference, GLB export, and output validation. Fail-closed GPU compatibility checks for the current FlashAttention-based upstream runtime.

Requirements Modly with GitHub/process extension support. Python >= 3.11 provided by the Modly setup payload. NVIDIA CUDA GPU. NVIDIA Ampere / RTX 30-series or newer is required (sm80+). RTX 20-series, GTX, and older GPUs are not supported because upstream SkinTokens uses FlashAttention in the TokenRig and SkinVAE paths. At least 14 GB VRAM is recommended by the upstream runtime profile. CUDA >= 12.1; this extension prepares the PyTorch cu128 lane by default. Internet access to download SkinTokens source/assets from GitHub and Hugging Face. Compatible flash-attn, bpy/Blender, and native Python dependencies for the target platform.

If the host GPU is older than Ampere and Modly passes GPU metadata, setup fails early with gpu-too-old before heavy installs/downloads. If GPU metadata is not available during setup, runtime performs the same check before generation.