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 Core ML model is about 0.2 MB and runs in a few milliseconds on device.

✏️ ➜ ▭ · ✏️ ➜ △ · ✏️ ➜ ◯ · ✏️ ➜ ★

Try it

• Live demo: desert-ant-labs/shapes-demo: draw one stroke, get a fitted shape, fully in your browser.
• iOS / macOS / visionOS: shapes-swift: Swift package with a one-line PKCanvasView.enableShapeSnapping() and a demo app.
• Android / Kotlin / JVM: shapes-kotlin: the Kotlin SDK.
• JavaScript / TypeScript: shapes-js: the npm package (Node + browser).

Files

File	Format	Size	Contents
shapes.mlmodelc	Compiled Core ML	~0.2 MB	4-bit-palettized classifier, ready to load on Apple platforms
shapes.safetensors	safetensors	~0.2 MB	packed 4-bit palettized portable weights for JS/Kotlin
shapes_meta.json	JSON	tiny	classes, preprocessing constants, model dims, and snap gates for JS/Kotlin
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

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.ai.

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.ai