English · 한국어
Turn one character illustration into a rigged, animated Spine 4.2 skeleton.
Drop in a PNG. OCS decomposes it into layers, lets you drop dummy parts and place bones, then builds a weighted mesh rig, packs an atlas, and hands back a standalone HTML preview.
flowchart LR
A[image] --> B[layer decomposition]
B --> C[layer cleanup]
C --> D[bone placement]
D --> E[limb split]
E --> F[rig]
F --> G[Spine + preview]
| 1. Upload | 2. Layer cleanup |
|---|---|
| 3. Bone placement | 4. Rig & preview |
Screenshots from a local test project.
Built on see-through for layer decomposition (git submodule, Apache-2.0). spine-animation-ai was read as prior art and is not reused — see NOTICE.md.
- uv, Python 3.12, git
- For layer decomposition: an NVIDIA GPU or Apple Silicon, plus ~20 GB free for model weights
- Everything after decomposition (cleanup, bones, rig, export, preview) also runs on CPU
| NVIDIA | 16 GB is comfortable; 8 GB works with group offload. Blackwell cards need CUDA 12.8 / torch cu128. |
| Apple Silicon | ./scripts/setup_env.sh --with-gpu. Resolution is capped at 768 on MPS. Slower than a discrete GPU. |
| No GPU | The editor still works. Import a PSD or build a demo project. |
Helper scripts: .ps1 for Windows, .sh for macOS / Linux.
git clone --recurse-submodules https://github.com/sleeeppy/OCS.git
cd OCSWindows:
./scripts/setup_env.ps1macOS / Linux:
./scripts/setup_env.sh # editor only, no GPU — under a minute
./scripts/setup_env.sh --with-gpu # + torch and see-throughsetup_env.ps1 fails if CUDA is missing (on Windows that usually means a broken install). setup_env.sh does not — GPU is only required for decomposition.
Optional, for previews that work offline. The Spine Web Player is not committed (redistribution is restricted; you need your own Spine license). Without this step, previews load the runtime from unpkg:
./scripts/fetch_spine_player.sh # macOS / Linux
./scripts/fetch_spine_player.ps1 # Windows./scripts/run_ocs.sh # macOS / Linux
./scripts/run_ocs.ps1 # WindowsThen open http://127.0.0.1:8765/.
Use a transparent-background PNG. OCS can estimate a silhouette from a flat background, but a real cutout is the single biggest quality lever.
The first GPU run downloads ~12 GB of model weights. Later runs take a couple of minutes per image.
- Upload — resolution, steps, and seed are under Advanced. Group offload is on by default (~10 GB VRAM). Leave it on for MPS.
- Layer cleanup — dummy layers are dropped automatically; uncertain ones are flagged. Untick to exclude; auto-dropped layers can be restored.
- Bone placement — drag the joints.
Shift+drag moves the subtree,Xmirrors left↔right,Ssnaps symmetric pairs,0fits the view. - Rig & preview — weighted meshes, atlas,
skeleton.json, and a downloadable HTML preview.
right means the character's right (the viewer's left). Getting this backwards mirrors every animation.
workspace/projects/<id>/export/
skeleton.json Spine 4.2 — weighted meshes + preset animations
skeleton.atlas libgdx atlas
skeleton.png packed texture page
preview.html standalone (assets inlined as data URIs)
idle is exported by default — it works for any pose. walk, wave, jump, and turn_head assume a standing figure; pass animations=[...] to export_skeleton to include them.
Decomposition is the only stage that needs a GPU. If you already have see-through output, or want to iterate on later stages:
.venv/bin/python scripts/import_psd.py path/to/foo.psdNo PSD either — build a synthetic project and open the editor:
.venv/bin/python scripts/make_demo_project.py --all
./scripts/run_ocs.sh.venv/bin/python -m pytest tests -q # macOS / Linux
.venv/Scripts/python.exe -m pytest tests -q # WindowsApache-2.0. See LICENSE and NOTICE.md. Using the Spine Web Player requires your own license from Esoteric Software.