Autofocus

Content-aware focal-point detection for photographs: a Rust engine compiled to WebAssembly (with a native CLI), plus the @modufolio/autofocus npm package that wraps it for the browser. Given an image, it answers one question well — where should a crop keep its focus so the subject survives? — so a face near the frame edge never gets centre-cropped away.

The detector is signals, not a neural net. An adaptive skin model that recalibrates its centres to the skin tones actually present in the image, HOG-based face-likeness, edge and sharpness maps, radial symmetry, and rule-of-thirds priors are fused per block, then a scene classifier (portrait / group / scene / mono) picks the weight preset before the focal point is refined. Deterministic, explainable, fast — and small enough to ship as WASM to every visitor.

Why it exists

• Centre cropping mutilates portraits. The default behaviour of every thumbnail pipeline is to assume the subject is centred. Photographers do not compose that way.
• Saliency-only croppers stop too early. Edge and saturation maps with static skin detection go wrong exactly where photographs matter most. This engine adds an adaptive skin model, face-likeness, eye-band refinement, and scene classification — a portrait, a group shot, and a landscape get different weightings because they are different problems.
• ML croppers are heavyweight and opaque. A model file outweighs this whole engine, and when it picks the wrong subject there is nothing to debug. Here every signal is inspectable (detect_features, debug_blocks), and behaviour is tuned by editing weights you can read.

Layout

• rust/ — the detector crate (autofocus): signal extraction (features, saliency, segments), scene classification (classify, rules, weights), refinement (refine, face_region, zoom), and a native CLI for batch analysis. cdylib + rlib, wasm-bindgen API.
• ui/ — the npm package: a thin loader that fetches the WASM module at runtime and exposes the detection API.

Quickstart

Browser (npm):

import { detectFocus, detectFeatures } from '@modufolio/autofocus'

const { point } = await detectFocus(imgElement)    // → { x, y } in [0,1]
const features  = await detectFeatures(imgElement) // → point + category + image features

The WASM module ships inside the npm package and loads automatically. To self-host it elsewhere instead, set globalThis.AUTOFOCUS_WASM_BASE to the directory URL holding the two wasm/ files, or copy them to the legacy fallback location /assets/wasm/autofocus/:

cd rust && wasm-pack build --target web --out-dir <app>/public/assets/wasm/autofocus

Native CLI:

cd rust && cargo run --release -- path/to/photo.jpg

Reviewing fixture accuracy:

Render every golden fixture with the hand-placed focus point (green ring) and the detected point (orange crosshair) drawn on the photo, a line connecting them, and the distance encoded in the filename so a reverse directory sort surfaces the worst detections:

cd rust && cargo run --release -- --review target/review

Output mirrors the <album>/<file> fixture layout under target/ (already gitignored — the photos are not MIT-licensed, keep renders out of commits).

Evaluating an external dataset:

Point --eval at any directory of photos plus a focuspoints.json mapping each filename to its hand-set point ({"img.jpg": {"x": 0.47, "y": 0.49}}). Prints per-photo distances and the mean; add --review to also render the annotated comparisons. This is how golden sets that cannot be committed are measured:

cd rust && cargo run --release -- --eval path/to/dataset --review target/review-dataset

Rust (as a library):

let (x, y, scene) = autofocus::detect_focus_cli(&rgba, width, height);

How it is used

Born inside the Modufolio portfolio stack, where modufolio/media uses the focal point for content-aware thumbnail crops. The engine has no dependency on any of that — RGBA in, focal point out.

License

MIT — with one carve-out: the fixture photographs under ui/tests/fixtures/ are copyrighted work by the author and are not MIT. They may only be used to run this repository's tests — see ui/tests/fixtures/LICENSE.