Wraps the SkinTokens / TokenRig auto-rigger: a static mesh goes in, a rigged .glb (skeleton + skin weights) comes out, with an optional step that relabels the skeleton with Mixamo or UE5 bone names. All glb import/export is pure Python (trimesh + pygltflib) - the upstream project needs Blender for this, and this pack removes that dependency so it runs on a headless GPU server.

Status: feature-complete. Rigging, relabeling, skinned-glb export, texture transfer, the optional voxel post-process, and skin-only-against-an-existing- skeleton all work and are tested. The one open item - animating a rigged output in Kimodo - is waiting on a fix to the Kimodo node itself, not on this pack.

Nodes

| Node | Does | |------|------| | SkinTokens Loader | Loads the TokenRig model (auto-downloaded from HuggingFace on first use) and wraps it for ComfyUI's VRAM manager. Pick a checkpoint from the dropdown - no path to configure. | | SkinTokens Rig | Static mesh → rigged skinned .glb. Accepts a native MESH (from Trellis/remesh) or a File3D (FILE_3D_GLB/GLTF/OBJ/STL, from Load3D). Optional relabel, texture transfer, voxel post-process, skin-only-against-existing-skeleton. | | SkinTokens Relabel | Standalone structural relabel of an already-rigged glb (Mixamo/UE5). Model-free. |

The rigged output is a FILE_3D_GLB, so it drops straight into ComfyUI core's Preview3DAdvanced (which has a showSkeleton toggle) and Save3D - no custom preview node needed.

Install

Via ComfyUI Manager (recommended): search for SkinTokens-NoBlender and install, then restart ComfyUI.

Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/stevelittlefish/ComfyUI-SkinTokens-NoBlender
pip install -r ComfyUI-SkinTokens-NoBlender/requirements.txt

torch is not pinned - ComfyUI provides it. flash-attn is optional (the code falls back to standard attention). The model weights (~14 GB) are not bundled; they download from VAST-AI/SkinTokens to your HuggingFace cache ($HF_HOME) the first time the Loader runs. A gated/private model needs the usual HF auth (huggingface-cli login or $HF_TOKEN).

Usage

Load the example workflow at examples/skintokens_rig.json:

Load3D ─▶ SkinTokens Rig ─▶ Preview3DAdvanced (showSkeleton)
             ▲
SkinTokens Loader

SkinTokens Rig options:

relabel / convention (Mixamo/UE5) / relabel_fingers - rename the generated humanoid bones to a standard convention (structural, topology-driven).
use_transfer (default on) - attach the rig to the original glb, preserving its materials, textures, UVs and scale. Only applies to a File3D input (a native MESH has no source file). Off = write the rigged proxy with a flat default material.
use_postprocess (default on) - refine the skin weights with a voxel-heat heuristic. Slower, but gives cleaner deformation; turn off for a faster pass.
use_skeleton (default off) - skin-only: keep an armature already present in the input glb and only predict skin weights for it, instead of generating a new skeleton.
top_k / top_p / temperature / repetition_penalty / num_beams - generation params (defaults match upstream: 5 / 0.95 / 1.0 / 2.0 / 10).

How it works

Import (skintokens/glb_io.py): trimesh loads the mesh; multi-part scenes are concatenated with correct face offsets. import_skeleton reads an existing armature for the skin-only path.
Inference (skintokens/infer.py): builds the upstream Asset, runs predict_step on the model (moved to the GPU by ComfyUI on demand, evicted when the GPU is needed elsewhere).
Export (skintokens/glb_io.py): writes a skinned glb - translation-only joint nodes, translate(-joint) inverse bind matrices, top-4 JOINTS_0/ WEIGHTS_0 - so the file poses correctly in any glTF viewer.
Relabel (skintokens/relabel.py): topology-driven humanoid recognizer.
Transfer (skintokens/transfer.py): similarity-align the proxy to the original mesh, nearest-neighbour skin, inject skeleton+skin into the target glTF in place so all materials survive.

The upstream torch core is vendored under skintokens/vendor/ (see skintokens/vendor/UPSTREAM.md for the commit and the portability edits).

Development

Local dev is GPU-less and Blender-less; the highest-risk paths (glb export, relabeler, transfer) are covered by pure-Python pytest:

uv venv && source .venv/bin/activate
uv pip install -r requirements-dev.txt   # CPU torch + the runtime stack + pytest
uv pip install -e . --no-deps
python -m pytest -q                        # ~84 passed, ~4 `server` skipped

The .venv is not in git, so a fresh clone (or a machine you haven't set up yet) starts with no dependencies - run the block above before pytest, or it can't even import torch/transformers. No uv? python -m pip install -r requirements-dev.txt inside the activated venv works too.

Tests marked server need the GPU + full weights and are skipped locally; see scripts/server/. The build is spec-driven - see spec/ for the design and spec/TODO.md for the phased task list and validation gates.

Releasing (Comfy Registry)

Publishing is automated. A GitHub Action (.github/workflows/publish.yml) publishes to registry.comfy.org on any push to main that changes pyproject.toml. ComfyUI Manager then picks the new version up from the registry on its own sync cadence.

To cut a release:

Bump version in pyproject.toml (e.g. 0.0.1 -> 0.0.2). This is the trigger - the registry rejects re-publishing an existing version, so nothing publishes until the number changes.
Update the description ([project] description, the one-line text shown in Manager) and/or README.md (the registry page body) if needed.
Commit and push to main. Watch the run under the repo's Actions tab.

The registry API key lives only as the REGISTRY_ACCESS_TOKEN repo secret (never in the repo). Rotate it on registry.comfy.org, then gh secret set REGISTRY_ACCESS_TOKEN - nothing in the workflow changes.

Credits

VAST-AI/SkinTokens - the model and torch reference implementation.
Prior-art ComfyUI node pack: Aero-Ex/ComfyUI-SkinTokens.

License

See LICENSE.