Face Analyzer Technology

The pipeline behind this face analyzer is intentionally short: load open-source models in the tab, find a face, place 68 landmarks, then score symmetry and a handful of ratios. There is no custom vision API and no GPU farm behind the URL.

Image input

You upload a JPG, PNG, or WebP file, or capture a frame from the camera. The photo stays in memory as a blob URL. We downscale large images before detection so phones do not stall.

Landmark detection

A tiny face detector proposes a box. A 68-point network then estimates the jaw, brows, eyes, nose, and mouth. Those coordinates are the only “understanding” the model has. It does not identify who you are.

Beauty model and report

Left/right pairs feed a symmetry percentage. Length-to-width and feature spans are compared with 1.618 for a golden-ratio fit. Those pieces become a 0–10 face score plus a five-class face shape. Read the full walkthrough if you want the math in slower English.