Auto Rigger (DL)

Automatically rig 3D models into animation-ready s&box .vmdl files using deep learning, entirely inside the editor and entirely in pure C#: no Blender, no Python, no native DLLs. Skeletons are predicted by real neural networks (full C# ports of the released research models, validated numerically against PyTorch on the official weights).

Using it

1. Install the library into your project and open View -> Auto Rigger (DL).
2. Enable at least one model via Manage Models... (one-time setup; see below).
3. Add Model: drop in .fbx / .glb / .gltf / .obj files. Each file is analyzed and rigged automatically with the best enabled model, or pick a specific model from the dropdown ("Auto" selects for you per mesh).
4. Preview: once a row finishes rigging, its eye button opens a preview - the textured model rendered see-through with the skeleton visible inside it (red joints, blue bones, the traditional convention). Scroll to zoom.
5. Export: writes the skinned .fbx (import it in Blender if you like) plus the generated .vmdl, compiled in place with its textures. Compiler errors (if any) surface on the row.

Models

Model	What it is	Status
RigNet (SIGGRAPH 2020)	Graph networks predicting joints, hierarchy, and skin weights.	Available
UniRig (VAST-AI 2025)	350M-parameter autoregressive skeleton transformer conditioned on a point-cloud shape encoder.	Available
SkinTokens / TokenRig (VAST-AI 2026)	UniRig's successor: skeleton and skin tokens generated as one sequence on a Qwen3 backbone, with native FSQ skin-VAE weights.	Available
MagicArticulate (ByteDance Seed3D 2025)	Autoregressive bone-pair transformer trained on Articulation-XL.	Available
Puppeteer (ByteDance Seed3D 2025)	Skeleton transformer with parent indices carried in the token sequence, plus a native skinning transformer (PartField features) for neural skin weights.	Available
RigAnything (Adobe 2025)	Autoregressive diffusion skeleton generation with neural skin weights. Noncommercial research license.	Available
Anymate (2025)	Three-stage neural rigging: joint prediction, connectivity, and neural skin weights (Apache-2.0).	Available

Transformer models run deterministically (greedy decode) on CPU; expect one to several minutes per mesh depending on the model. SkinTokens, Puppeteer (with its skinning checkpoint installed) and RigAnything predict their own NEURAL skin weights; the other models use the library's geodesic voxel skinner.

Managing models

Manage Models... shows every supported model with its license, download size, and a hardware verdict for your machine. Models your machine cannot run are disabled automatically. Download fetches weights directly from the authors' official distribution with progress and resume; nothing is redistributed by this library, and an ATTRIBUTION.md is written next to the weights. You can also load your own RigNet-format checkpoint zip.

Notes

• Engine-agnostic core in Code/ (whitelist-safe), editor UI in Editor/.
• Development harness, PyTorch golden tests, and gate scripts live in dev/ (not part of the shipped library).
• Attribution:
  • RigNet: Xu, Zhou, Kalogerakis, Landreth, Singh, SIGGRAPH 2020 (GPL-3.0 or commercial license from UMass).
  • UniRig: VAST-AI Research, One Model to Rig Them All (2025).
  • SkinTokens: VAST-AI Research (2026).
  • MagicArticulate: ByteDance Seed3D (2025), code Apache-2.0.
  • Puppeteer: ByteDance Seed3D (2025).
  • RigAnything: Liu et al., Adobe (SIGGRAPH TOG 2025). Adobe Research License: noncommercial research use only.
  • Anymate: Deng et al. (2025), Apache-2.0.
• All model weights are downloaded by the user from the authors' official sources; this library ships no weights.

Cloud rigging (vast.ai)

For machines that cannot run a model locally, Cloud... rents a GPU on vast.ai for a single rig. Renting is never automatic: the picker lists offers with hourly rates and a worst-case cost, and nothing is charged until you click rent. The instance created for the job is recorded in a local ledger and destroyed with verification the moment results are back. Only that one instance is ever touched; pre-existing instances on your account are never enumerated or cancelled.

Roadmap

• Broader corpus validation and per-model quality presets.