Face Avatar
Become a live cartoon. A real face-tracking model runs entirely on your device, reads 478 points and 52 expressions off your face, and a character mirrors your every blink, smile, and head turn in real time.
🔒 100% private — your camera never leaves your deviceHow it works
When you press start, the page asks for your camera and downloads a small face-tracking model called MediaPipe Face Landmarker, about 3.7 MB, one time. After that it runs on your own machine. Every frame it locates 478 landmark points on your face and estimates 52 expression values — how open your jaw is, how much each eye is closed, whether your brows are up or furrowed, which way you are looking. Those numbers are wired straight into a hand-drawn vector character, so the avatar smiles when you smile and blinks when you blink, with no perceptible lag.
How is this running in my browser?
There is no server doing the work. The model is compiled to WebAssembly and runs the heavy math on your GPU through WebGPU when it is available, falling back to your CPU otherwise. Because the camera frames are processed locally and thrown away immediately, nothing is uploaded and nothing is stored. The only thing that ever leaves your device is a clip you choose to download yourself. It even keeps working with your internet disconnected once the model is cached.
Things to try
Raise one eyebrow, then both. Wink — each eye is tracked separately. Open your mouth wide and watch the jaw drop, then purse your lips. Turn your head side to side and the whole character follows. Switch the character and the backdrop, drop the smoothing to zero for snappy puppet motion or raise it for a calmer look, and hit record to save a short clip of your avatar talking.
Is my camera uploaded anywhere?
No. The face-tracking model runs entirely in your browser. Your camera feed is processed on your device and discarded each frame. Nothing is uploaded, and nothing is saved unless you press record.
How does a real model run in a web page?
It downloads MediaPipe Face Landmarker once (about 3.7 MB) and runs it on your CPU or GPU through WebAssembly and WebGPU. It finds 478 face points and 52 expression values every frame, which drive the avatar.
Do I need to install anything?
No app, no plugin, no sign-up. Open the page, allow the camera, and it works. After the first load it runs offline.
Can I use it on a Zoom call or stream?
You can record a clip or save a snapshot here. For a live call, use OBS window capture or screen sharing of the avatar, since a web page cannot register a system virtual camera by itself.
It says my browser is not supported.
This needs camera access and a modern browser. Use a recent Chrome, Edge, Brave, or Safari on desktop or Android. iOS works in Safari. If the camera is blocked, check the site permissions in your address bar.
I build something like this every day.
Senior full-stack engineer, available for senior or contract work, fully remote. See the rest of the lab or get in touch.