chibifire/interactor-skintokens-viewer

SkinTokens viewer for the desktop, with auto-rig

Open a .glb, .gltf, .vrm or OpenUSD (.usd, .usda, .usdc, .usdz) model, then press R to rig a .glb or OpenUSD model with SkinTokens on your own GPU. No internet connection is necessary: each package contains the viewer, the rig program and the weights.

Package	System
skintokens-viewer-macos-arm64.dmg	macOS on Apple silicon: the program SkinTokens Viewer.app
skintokens-viewer-windows-x86_64.zip	Windows, x86-64
skintokens-viewer-linux-x86_64.tar	Linux, x86-64

SHA256SUMS has the checksums.

What each package contains

• skintokens-viewer (.exe on Windows): the Godot program.
• skintokens-viewer.pck: the viewer, from V-Sekai-fire/entities-model-explorer.
• The godot-sandbox library, which runs the RISC-V guest programs in the pck: the rig program skintokens.elf, the OpenUSD reader usd.elf (OpenUSD and oneTBB), the walk mode mm.elf, the eye-close shapes headfit.elf and the animation program motion.elf.
• The 72 walk-mode clips in the pck: 51 clips made with Kimodo and 21 motion-capture clips from the O3DE MotionMatching gem.
• w/: the three weight files from chibifire/SkinTokens-GGUF-BF16 at 912d4d1, a format change of VAST-AI/SkinTokens.
• LICENSE.md: each part, its licence and the full licence text. NOTICE: the attribution notices. CITATION.cff: how to cite the package and its parts. LICENSE-model-explorer.txt: the viewer's licence.

Keep the files together in one folder.

The macOS .dmg file contains SkinTokens Viewer.app, with the program, the pck, the godot-sandbox library, the weights and the licence files inside it.

Use

1. On Windows and Linux, extract the package and start skintokens-viewer. On macOS, open the .dmg file, drag SkinTokens Viewer.app to Applications, and start SkinTokens Viewer.app.
2. Drop a .glb or OpenUSD file on the window. The viewer changes an OpenUSD file to a .glb in its user folder, and rigs that file.
3. Press R. After one to two minutes the viewer opens the rigged model. The program saves it beside the source file as <name>.rigged.glb, and shows that path for 10 seconds. If the source folder is read-only, the program saves the file in its user folder: ~/Library/Application Support/Godot/app_userdata/ModelExplorer on macOS, %APPDATA%\Godot\app_userdata\ModelExplorer on Windows and ~/.local/share/godot/app_userdata/ModelExplorer on Linux. A second rig of the same file replaces the earlier result.

Walk mode

The viewer also has the walk mode (M). Its page and its own downloads, without weights, are on chibifire/walk-mode.

First start

On macOS, the program is not notarized. Before the first start, enter xattr -dr com.apple.quarantine "/Applications/SkinTokens Viewer.app" in Terminal. On Windows, SmartScreen can stop the first start: select "More info", then "Run anyway".

Tests

• macOS, Apple M2 Pro (Metal): three test characters rigged in this package, each byte-identical to the reference rig.
• macOS .dmg file, Apple M2 Pro: the rig of a test character from SkinTokens Viewer.app is identical to the rig from the macOS .zip file of this release. The program also started with open.
• Windows, NVIDIA RTX 4090 (Vulkan): one test character rigged in this package, with the same joints as the reference rig; skin weights differ by at most 0.000001.
• Linux: not tested with auto-rig or walk mode.

License

The packages hold parts with different licences. LICENSE.md lists each part with its full licence text, and each package contains a copy of it, of CITATION.cff and of NOTICE:

• The rig program skintokens.elf (interactor-skintokens-guest): all rights reserved, licensed only to K. S. Ernest (iFire) Lee. A free and open-source licence is not yet decided, so the packages as a whole do not have one.
• MIT: the viewer, Godot, ggml, ggml-rd, godot-vrm, interactor-usd-guest, interactor-motion-guest, interactor-motion-matching-guest (with godot-motion-matching), interactor-headfit, nlohmann/json, the SkinTokens code and the SkinTokens weights.
• Apache-2.0 OR MIT, used here under MIT: the O3DE MotionMatching gem clips.
• The 51 Kimodo walk-mode clips are the output of Kimodo-SOMA-RP-v1.1 (NVIDIA Open Model License). No person wrote them. The packages do not contain the Kimodo weights.
• Apache-2.0: skin-tokens-ggml, kimodo-ggml, motion-bricks-ggml, Kimodo and oneTBB.
• BSD-3-Clause: godot-sandbox and libriscv.
• Tomorrow Open Source Technology License 1.0: OpenUSD.
• The pck holds 18 prompt embeddings for the animation program (used by the Kimodo motion viewer). An LLM2Vec encoder based on Meta Llama 3 made them ahead of time; they are the output of a program and no person wrote them. The packages do not contain the encoder. Built with Meta Llama 3 (Meta Llama 3 Community License).

Credit for the model goes to the SkinTokens authors (VAST-AI).