SkinTokens: A Learned Compact Representation
for Unified Autoregressive Rigging

SkinTokens is a learned, compact, and discrete representation for skinning weights. Built on this representation, TokenRig is a unified autoregressive framework that models the entire rig, i.e., skeleton and skinning weights, as a single token sequence. Given an input 3D mesh, it generates a complete skeleton hierarchy and skin weights that can be directly imported into standard 3D pipelines for character animation and simulation.

SkinTokens is the successor to UniRig (SIGGRAPH '25). While UniRig uses separate stages for skeleton prediction and skinning, SkinTokens unifies both into a single autoregressive sequence via learned discrete skin tokens, yielding 98%–133% improvement in skinning accuracy and 17%–22% improvement in bone prediction over state-of-the-art baselines.

🔮 Overview

TokenRig takes a single 3D mesh as input and autoregressively produces a fully rigged asset — a coherent skeleton hierarchy plus dense per-vertex skinning weights — in a single unified sequence. Method in three stages:

1. Learn SkinTokens — An FSQ-CVAE compresses sparse skinning weights into a compact discrete vocabulary.
2. Unified Autoregressive Modeling — A Qwen3-0.6B-based Transformer generates the full rig (skeleton followed by SkinTokens) as one interleaved sequence.
3. RL Refinement via GRPO — Tailored geometric and semantic rewards (volumetric joint coverage, bone-mesh containment, skinning sparsity, deformation smoothness) refine the model for out-of-distribution assets.

Qualitative comparison of skinning prediction. TokenRig produces clean, locally coherent influence maps that closely match the ground truth, while baselines suffer from "bleeding" artifacts across disconnected mesh parts.

See the project page for the full teaser video and additional qualitative comparisons (skeleton generation and impact of GRPO).

📦 Installation

Prerequisites

• Hardware: An NVIDIA GPU with at least 14 GB of memory is required for inference.
• Software:
  • Python >= 3.11
  • CUDA Toolkit >= 12.1
  • uv is recommended for managing dependencies.

Installation Steps

1. Clone the repo:

   git clone https://github.com/VAST-AI-Research/SkinTokens.git
   cd SkinTokens

2. Create a virtual environment and install PyTorch:

   uv venv --python 3.11
   source .venv/bin/activate
   uv pip install torch==2.7.0 torchvision==0.22.0 torchaudio==2.7.0 --index-url https://download.pytorch.org/whl/cu128

   [!NOTE] Adjust the CUDA version in the PyTorch index URL to match your driver. See PyTorch Get Started for other CUDA versions.

3. Install dependencies:

   uv pip install -r requirements.txt

4. Install flash-attn:

   uv pip install flash-attn --no-build-isolation

5. Download pretrained models:

   python download.py --model

   This downloads the TokenRig and FSQ-CVAE checkpoints to experiments/, and the Qwen3-0.6B config to models/.

🤖 Pretrained Models

We provide the following pretrained models on Hugging Face:

Model	Description	Download
articulation_xl_quantization_256_token_4	TokenRig autoregressive rigging model, trained on ArticulationXL 2.0 + VRoid Hub + ModelsResource and refined with GRPO (recommended)	Download
skin_vae_2_10_32768	FSQ-CVAE (SkinTokens) — skin-weight tokenizer used to encode and decode skinning weights	Download

💡 Usage

Hugging Face Space Demo

The easiest way to try TokenRig without any local setup is the hosted Hugging Face Space — upload a mesh and get a rigged result in the browser.

Gradio Demo (local)

python demo.py

Then open http://127.0.0.1:1024 in your browser.

Command Line

# Rig a single model
python demo.py --input examples/giraffe.glb --output results/giraffe.glb

# Rig with original texture and scale preserved
python demo.py --input examples/giraffe.glb --output results/giraffe.glb --use_transfer

# Skin a model using its existing skeleton
python demo.py --input examples/giraffe_skeleton.glb --output results/giraffe.glb --use_skeleton --use_transfer

# Batch process a directory
python demo.py --input examples/ --output results/ --use_transfer

Generation Parameters

Parameter	Default	Description
--top_k	5	Top-k sampling
--top_p	0.95	Top-p (nucleus) sampling
--temperature	1.0	Sampling temperature
--repetition_penalty	2.0	Repetition penalty
--num_beams	10	Number of beams for beam search
--use_skeleton	False	Use existing skeleton (generate skin only)
--use_transfer	False	Transfer original texture and scale
--use_postprocess	False	Apply voxel-based skin postprocessing

Troubleshooting

• Server fails to start: Make sure http_proxy / https_proxy environment variables are unset or correctly configured.
• Blender export issues: Remove the glTF_not_exported node when importing results into Blender.

🙏 Acknowledgements

• UniRig — the predecessor to this work.
• Qwen3 — the LLM architecture used by the TokenRig autoregressive backbone.
• Michelangelo — 3D shape encoder.
• 3DShape2VecSet — shape-representation backbone used by the FSQ-CVAE.
• FSQ — Finite Scalar Quantization, the discretization scheme behind SkinTokens.
• GRPO (DeepSeekMath) — the policy-optimization method used for RL refinement.
• Tripo — the 3D generative studio from Tripo, a broader context for this line of work.

We sincerely appreciate the contributions of these excellent projects and their authors. We believe open source helps accelerate research, lower barriers to innovation, and make progress more accessible to the broader community.

License

This project is licensed under the MIT License.

📜 Citation

If you find this work helpful, please consider citing our paper:

@article{zhang2026skintokens,
  title   = {SkinTokens: A Learned Compact Representation for Unified Autoregressive Rigging},
  author  = {Zhang, Jia-Peng and Pu, Cheng-Feng and Guo, Meng-Hao and Cao, Yan-Pei and Hu, Shi-Min},
  journal = {arXiv preprint arXiv:2602.04805},
  year    = {2026}
}