Shapes: on-device shape recognition from a single stroke

Takes a single hand-drawn stroke (an ordered list of points) and recognizes it as a clean geometric shape, returning fitted vector geometry ready to snap to. Built for PencilKit-style "smart shapes": draw, pause, and the rough stroke becomes a crisp shape. The model is tiny (about 0.2 MB Core ML, ~1.3 MB LiteRT) and runs in a few milliseconds on device.

✏️ ➜ ▭ · ✏️ ➜ △ · ✏️ ➜ ◯ · ✏️ ➜ ★

Try it

All platforms ship from one repo: Desert-Ant-Labs/shapes (Swift, Kotlin, and JavaScript in a single codebase).

• Live demo: desert-ant-labs/shapes-demo: draw one stroke, get a fitted shape, fully in your browser.
• iOS / macOS / tvOS / visionOS: the Swift SDK (Swift Package Manager) with a one-line PKCanvasView.enableShapeSnapping() and a demo app. It bundles the compiled Core ML model below.
• Android / JVM (Kotlin): Maven Central ai.desertant:shapes with LiteRT (.tflite). The small model is bundled by default; exclude ai.desertant:shapes-tflite-resources to force on-demand download or explicit-directory loading.
• Node / browser (JavaScript / TypeScript): npm i @desert-ant-labs/shapes @litertjs/core for browser builds, or just npm i @desert-ant-labs/shapes for server-side Node. The npm package downloads the model from this repo on first use and caches it (nothing model-sized ships in the tarball); browser inference uses LiteRT.js, and Node uses prebuilt native libraries. Pass directory (Node) or modelBaseUrl (browser) to self-host / run offline.

Files

File	Format	Size	Contents
shapes.tflite	LiteRT / TFLite (fp32)	~1.3 MB	Fixed [1,256,3] features + [1,256] mask window; runs on Android, Linux, Node, and the web (bundled by default in the Kotlin SDK; downloaded on demand by the JavaScript SDK)
shapes.mlmodelc	Compiled Core ML	~0.2 MB	4-bit-palettized classifier, ready to load on Apple platforms (used by the Swift SDK)
shapes_meta.json	JSON	tiny	classes, preprocessing constants, model dims, and snap gates
shapes.safetensors	safetensors	~0.2 MB	packed portable weights (reference)
model.pt	PyTorch checkpoint	~1.5 MB	trained weights (for export / fine-tuning)
config.json	JSON	tiny	class list, preprocessing constants, and per-class snap gates

Older revisions (tag v0.1.0) carry shapes.onnx for SDK versions that predate the LiteRT migration.

How it works

Two stages, the network proposes, geometry verifies:

1. Classify: the stroke is resampled and fed to a compact sequence classifier (Conv1d stem → small Transformer encoder → masked mean-pool → MLP), which predicts the shape type (or none to reject scribbles).
2. Fit + snap: a classical geometric fitter produces clean vector parameters (min-area box, moment/PCA ellipse, max-area triangle, …), then regularizes them (snap to axes, circles, squares, and 15° rotation increments). A fit-residual gate vetoes poor fits so non-shapes stay rejected.

Inputs and outputs

• Input: an ordered list of stroke points in canvas coordinates. Single stroke.
• Output: a shape class plus fitted geometry, or nothing if the stroke is rejected.

Classes

line, rectangle, triangle, ellipse, star, plus none (the reject class: scribbles, partial shapes, and other non-shape strokes). Squares and circles are covered by rectangle and ellipse (snapped when near-regular).

Limitations

• Single stroke only; multi-stroke shapes aren't recognized.
• Tuned for deliberate shapes; very rough or ambiguous strokes are rejected by design.

License

Desert Ant Labs Source-Available License. Free for most apps; a commercial license is required at scale. Full terms are at the link. Licensing: licensing@desertant.com.

Citation

@software{shapes_2026,
  title  = {Shapes: on-device shape recognition from a single stroke},
  author = {Desert Ant Labs},
  year   = {2026},
  url    = {https://huggingface.co/desert-ant-labs/shapes},
}

---

© 2026 Desert Ant Labs · https://desertant.com