PhotoCut Studio

Interactive background-removal tool. One React codebase ships as:

• a desktop app (Tauri + Rust) — full-quality GrabCut engine running natively, offline; and
• a web app (plain browser) — same UI, with its own GrabCut-lite segmenter (k-means color models, hard-constraint trimap, component cleanup, alpha feathering) running fully client-side.

Features

• Cutout workspace — box select, one-click auto cut (border background modeling), keep/remove brushes with live ghost cursor, marching-ants selection, edge feathering, undo, keyboard shortcuts (A 1 2 3 [ ] E ⌘Z), drag & drop and clipboard paste.

• Export — transparent PNG, WebP/JPEG, solid color or image background, copy straight to the clipboard.

• Icon Studio — turns the cutout (or any PNG) into complete app icon sets: iOS AppIcon.appiconset (+ legacy sizes), Android mipmaps + adaptive + Play Store, macOS AppIcon.iconset, a real multi-resolution Windows .ico, favicons + PWA manifest + maskable icon — all zipped in the browser with zero dependencies (own ZIP/ICO writers in src/lib/zip.js / icons.js).

• make_icons.py — the same icon pipeline as a Python CLI (public/make_icons.py, Pillow only; auto-.icns via iconutil on macOS):

  pip install pillow
  python make_icons.py logo.png -o app-icons --padding 0.08 --name "Mi App"
  

This is an original implementation. It is not affiliated with, derived from, or a copy of any existing product. The cutout engine (photocut-core) is written from scratch.

Architecture

photocut/
├─ crates/photocut-core/   Rust GrabCut engine (MIT OR Apache-2.0 deps only)
│   ├─ src/gmm.rs          Gaussian mixture color models
│   ├─ src/maxflow.rs      Boykov–Kolmogorov min-cut (from scratch)
│   ├─ src/grabcut.rs      Iterated graph-cut orchestration
│   └─ src/lib.rs          Public API + compositing + tests
├─ src-tauri/              Tauri desktop shell (Rust backend, commands)
├─ src/                    React frontend
│   ├─ App.jsx             Toolbar, modes, export
│   ├─ components/CanvasEditor.jsx
│   └─ lib/
│       ├─ backend.js      Routes to Tauri (desktop) OR JS engine (web)
│       └─ jsEngine.js     Browser-only fallback segmenter
└─ ...

The single decision point is src/lib/backend.js: it detects Tauri at runtime and calls the Rust commands when present, otherwise runs in-browser. The React UI never changes between the two products.

Browser support (web build)

Evergreen browsers. Minimum tested versions:

Browser	Min	Notes
Chrome / Edge	111+	full experience (worker engine)
Firefox	113+	full experience
Safari	16.4+	full experience
Safari	16.2–16.3	works; engine runs inline on the main thread (no OffscreenCanvas) — brief UI freezes on large photos

Below those versions, color-mix() / backdrop-filter degrade the visuals but the app remains functional. No IE, no legacy Edge.

Prerequisites

• Node.js 18+
• Rust (stable) + Cargo — https://rustup.rs
• Tauri v2 system deps — https://tauri.app/start/prerequisites/ (on Linux: webkit2gtk, on macOS: Xcode CLT, on Windows: WebView2 + MSVC)

Run

Web (no Rust needed):

npm install
npm run dev          # http://localhost:5173

Desktop (Rust required):

npm install
npm run tauri dev

Build for distribution

Web bundle:

npm run build        # outputs static site to dist/ — host anywhere

Desktop installers (.dmg / .msi / .deb / .AppImage):

npm run tauri build  # outputs to src-tauri/target/release/bundle/

Test the engines

Rust (desktop):

cargo test -p photocut-core

Web (smoke tests in your installed Chrome, headless — segmentation accuracy, worker RPC, compositing, icon ZIP, preview panel, 12MP perf benchmark):

npm run test:web

(Each suite is also a plain page under test/*.html if you want to watch it run: npm run dev and open them.)

Workflow in the app

1. Open photo.
2. Box select (1): drag a rectangle around your subject. The engine runs GrabCut and shows the cutout over a checkerboard.
3. Keep brush (2) / Remove brush (3): paint to correct any edges the cut missed. Each stroke re-runs a quick refinement.
4. Export: transparent PNG, a solid color background, or composite over another image.

Selling this

You can sell both products. Notes:

• Tauri is MIT/Apache-2.0 and explicitly permits closed-source, commercial apps. No royalty, no source-disclosure obligation.
• photocut-core dependencies are all permissive (MIT / Apache-2.0): image, nalgebra, thiserror. No GPL/LGPL/copyleft anywhere in the tree. Run cargo tree and cargo deny (recommended) before each release to confirm no copyleft dependency sneaks in via an update.
• GrabCut is a published academic algorithm; implementing it yourself (as here) carries no license burden. Do not pull in OpenCV's GPL-ish modules — not needed.
• The code is MIT-licensed (see LICENSE); keep the copyright in src-tauri/tauri.conf.json consistent with it in src-tauri/tauri.conf.json (copyright, identifier, productName).
• For paid desktop distribution you'll want code signing (Apple Developer ID for macOS notarization; an Authenticode cert for Windows) — configure under bundle.macOS / bundle.windows in tauri.conf.json. Tauri also has a built-in updater plugin for shipping updates to paying customers.
• A common model: free/limited web version for funnel, paid desktop version (one-time or subscription) with batch export, higher resolution, and the full-quality Rust engine. Licensing/activation is app-level logic you add; this scaffold leaves that to you.

Upgrading the web engine (optional)

The browser fallback in jsEngine.js is intentionally simple. For parity with desktop, compile photocut-core to WebAssembly (add a wasm-bindgen wrapper crate and load it from backend.js). Then both products share the exact same engine.

License

PhotoCut's own code (web app, desktop app and the Rust engine) is released under the MIT License.

Third-party components keep their own licenses:

• Bundled open-source dependencies (MIT/Apache-2.0 and others): see the "Third-party licenses" screen and npm run licenses.
• AI models: BiRefNet-lite (MIT weights; trained on DIS5K, whose dataset terms restrict commercial use) and u2netp (Apache-2.0). See public/licenses/ai-models-NOTICE.txt.