ComfyUI-SkinToken

Automatic AI-powered 3D character rigging inside ComfyUI

Features • Installation • Requirements • Nodes • Parameters • Credits

---

Overview

ComfyUI-SkinToken wraps VAST-AI/SkinTokens into ComfyUI custom nodes, enabling fully automatic skeleton generation and skinning weight prediction for arbitrary 3D character meshes. Feed it a TRIMESH input and get back a production-ready rigged GLB/FBX — no manual bone placement required.

Features

• 🦴 Automatic Skeleton Generation — Predicts joint positions, hierarchy, and bone lengths from mesh geometry alone
• 🎨 Skinning Weight Prediction — AI-generated vertex weights with configurable groups per vertex
• 🔄 Transfer Rigging — Transfer a generated rig back onto the original high-detail mesh
• 📐 Bottom Center Origin — Automatically align the skeleton origin to the bottom-center of the mesh bounding box
• 🧹 Auto Mesh Cleanup — Merge vertices by distance and shade smooth by angle before export
• 🏷️ Skeleton Templates — Rename bones to Mixamo or Unreal Engine 5 naming conventions
• 📦 Auto Model Download — Checkpoints are pulled from HuggingFace on first run
• 🖥️ Headless Blender Export — Works without a Blender GUI via background process

Installation

ComfyUI Manager (Recommended)

Search for ComfyUI-SkinToken in ComfyUI Manager and install.

Manual

cd ComfyUI/custom_nodes
git clone https://github.com/Rizzlord/ComfyUI-SkinToken.git
cd ComfyUI-SkinToken
pip install -r requirements.txt

Requirements

Requirement	Details
GPU	CUDA-capable GPU (inference runs on GPU)
Blender	Blender 4.0+ on PATH or set SKINTOKEN_BLENDER_BIN env var
Python	Python 3.10+ (ComfyUI environment)
VRAM	~4 GB for the default checkpoint

Python Dependencies

transformers>=4.57.0
diffusers>=0.35.0
python-box
einops
omegaconf
lightning
addict
fast-simplification
trimesh
open3d
huggingface_hub
numpy>=1.26.0
scipy

Nodes

🔽 SkinToken Download Models

Downloads the required model checkpoints from HuggingFace into ComfyUI/models/skintoken/.

Output	Type	Description
status	STRING	Download result summary
models_dir	STRING	Path to the models directory

🦴 SkinToken Rig

The main rigging node. Takes a TRIMESH input and outputs a rigged mesh.

Output	Type	Description
trimesh	TRIMESH	The rigged mesh
rigged_path	STRING	File path to the exported model
asset	SKINTOKEN_ASSET	Internal asset for chaining
backend	STRING	bpy or blender_headless

Parameters

Parameter	Type	Default	Description
ckpt_name	dropdown	grpo_1400.ckpt	Model checkpoint to use
device	dropdown	auto	Inference device (auto / cuda)
save_file	bool	True	Save output to the output directory
filename_prefix	string	3D/SkinToken_	Output filename prefix
file_format	dropdown	glb	Export format (glb / fbx)
use_transfer	bool	False	Transfer rig onto the original mesh
use_postprocess	bool	False	Apply voxel-based skin postprocessing
group_per_vertex	int	4	Max bone influences per vertex
bottom_center_origin	bool	False	Set skeleton origin to bottom-center of mesh
smooth_angle	float	55.0	Angle threshold for shade smooth
skeleton_template	dropdown	Keep model names	Bone naming convention
top_k	int	5	Top-K sampling for generation
top_p	float	0.95	Nucleus sampling threshold
temperature	float	1.0	Sampling temperature
repetition_penalty	float	2.0	Repetition penalty
num_beams	int	10	Beam search width

Environment Variables

Variable	Description
SKINTOKEN_BLENDER_BIN	Path to the Blender binary (overrides auto-detection)
SKINTOKEN_FORCE_HEADLESS	Set to 1 to always use headless Blender export

Workflow

[Load 3D Model] → [TRIMESH] → [SkinToken Rig] → [Rigged GLB/FBX]

1. Load a 3D mesh as TRIMESH (using any mesh loader node)
2. Connect to SkinToken Rig
3. Configure skeleton template, export format, and generation params
4. Execute — the node predicts skeleton + weights and exports through Blender

Tips

• use_transfer = True produces better results for detailed meshes — it transfers the predicted rig back onto the original geometry preserving all detail
• bottom_center_origin = True is recommended for game engine workflows (Unreal, Unity) where the pivot point should be at the character's feet
• skeleton_template = "Unreal Engine 5" maps bones to UE5 naming for direct Mannequin retargeting
• Lower temperature and higher num_beams produce more deterministic skeletons
• Increase group_per_vertex for smoother deformations at the cost of performance

Credits

• SkinTokens by VAST-AI — the underlying rigging model
• Qwen3-0.6B — language model backbone for token generation

License

This node pack wraps the SkinTokens model. Please refer to the SkinTokens repository for model licensing terms.