FaceStudio

A private, on-device face editor — retouch, restyle, reshape and re-age portraits entirely in your browser. No uploads, no server, no accounts — your photos never leave your device, and it's free.

→ Live app

FaceStudio is a ground-up rewrite of the original face-average Python/dlib CLI as a WASM-powered static site. It uses Google's MediaPipe Face Landmarker (478-point dense mesh) for detection and a WebGL2 piecewise-affine warp for real-time face manipulation, all client-side.

Features

The flagship is a full face editor; averaging, morphing, replace, enhance, and a "future baby" blend are auxiliary tools built on the same on-device engine.

• Edit — the flagship, FaceApp/Facetune-style but 100% on-device: skin smoothing, teeth whitening, makeup (lips / blush / brows / eye colour), hair recolouring, background bokeh or studio, and vignette via BiSeNet face parsing; shape tools (smile, eye size, nose/face slim, hair volume) as landmark warps; re-aging with a CI-converted FRAN U-Net (MIT) that edits as an identity-preserving delta; one-tap Looks, a hold-to-compare before/after, and share.
• Future Baby — for fun, blend two parents and de-age the result to glimpse a future child (composes the averaging + re-aging engines). Not a genetic prediction.
• Average any number of faces into one, with per-face weight sliders, template-shape mode, Lab-style colour normalisation, and background options.
• Morph between two people with a live blend slider; export the animation as WebM/MP4 or GIF.
• Replace a face in any photo from a multi-image pose bank of one person — the closest-pose source is warped onto the target at native resolution, colour-matched and seam-feathered. No generative models, no quality loss.
• Enhance results with on-device Real-ESRGAN upscalers (photo / anime / general) via ONNX Runtime Web (WebGPU when available).
• Webcam capture, batch drag-and-drop, and a preset gallery to try instantly.
• Installable PWA, offline-capable after first load.

Architecture

• Vite + React + TypeScript + Tailwind.
• ML runs on the main thread via WASM/WebGPU — MediaPipe's loader needs importScripts, unavailable in ES-module workers — while warping runs on the GPU (WebGL2).
• Models are self-hosted (fetched at build time, SHA-checked) and cached by a service worker.
• Deployed to GitHub Pages by CI, which runs vitest unit tests over the pure engine maths and a Playwright smoke test that exercises the real WASM pipeline before every deploy.

Development

npm install
bash scripts/fetch-models.sh   # downloads MediaPipe + ONNX assets into public/models
npm run dev
npm test                       # unit tests (vitest); npm run test:e2e for the Playwright smoke suite

Privacy

Every pixel operation runs on your device — photos are never uploaded, and there are no accounts. The app collects anonymous usage analytics (PostHog, EU region): page views and product events such as which tool was run, its duration and success, exports, and model downloads. No images, filenames, or personal data are ever sent, and analytics is disabled in development and automated (CI) runs.

Legacy

The original Python averaging CLI (dlib, 68 landmarks, .ff cache format) lives on the legacy branch.

Acknowledgements

• Original approach based on Satya Mallick's Average Face tutorial.
• Landmarks by MediaPipe; upscaling by Real-ESRGAN.

License

Copyright © 2018–2026 George Gach. Licensed under GPL-3.0-only — see LICENSE.