In my previous post I showed a face rater that runs 100% in the browser. A few follow-up notes after shipping it.
Why on-device
Most face-analysis tools upload your photo to a server. That is bad for privacy and slow. I wanted something where the image never leaves the tab.
The stack
- A small vision model quantized to run in WebAssembly
- Browser webcam / file input, no upload
- Local post-processing for facial landmarks
Result
The score is instant, no account, no watermark. You can try it here: FaceGrade
Takeaway
For sensitive inputs like faces, local-first is the obvious choice: lower latency, no data transfer, nothing to store.
