UniRig offers a unified framework for automating 3D model rigging, addressing the time-consuming process of skeleton creation and skinning. It targets 3D artists and animators by providing a single model capable of handling diverse asset types, from humans to animals and objects, significantly streamlining the animation pipeline.

How It Works

UniRig employs a two-stage autoregressive approach powered by large transformer models. First, a GPT-like transformer predicts a topologically valid skeleton hierarchy using a novel Skeleton Tree Tokenization scheme for efficient representation. Second, a Bone-Point Cross Attention mechanism assigns per-vertex skinning weights and predicts bone attributes based on the generated skeleton and input mesh geometry. This unified, end-to-end deep learning approach aims for high accuracy and robustness across various 3D models.

Quick Start & Requirements

• Install: Clone the repository, set up a Python 3.11 virtual environment, and install dependencies via requirements.txt. Key dependencies include PyTorch (>=2.3.1), spconv, torch_scatter, and torch_cluster, requiring specific CUDA versions.
• Model: Pre-trained skeleton prediction checkpoint available on Hugging Face.
• Hardware: CUDA-enabled GPU with at least 8GB VRAM.
• Docs: Project Page, Paper

Highlighted Details

• Unified model for diverse asset categories (humans, animals, objects).
• Automated skeleton generation with topologically valid structures.
• Automated skinning weight prediction.
• Efficient Skeleton Tree Tokenization for compact representation.
• Supports .obj, .fbx, .glb, and .vrm input formats.

Maintenance & Community

The project is developed by Tsinghua University and Tripo. Updates on planned future releases, including datasets and full model checkpoints, will be announced by VAST-AI-Research.

Licensing & Compatibility

The repository is released under a permissive license, allowing for commercial use and integration with closed-source projects.

Limitations & Caveats

Bone attribute prediction and full model checkpoints trained on Rig-XL/VRoid are marked as "Coming Soon." The current skinning prediction performance is noted to degrade significantly if the input skeleton is inaccurate, recommending refinement before skinning.