Read this first — the body is harder to read than the face. This is a mirror of two methodologies, not a verdict on a person. It estimates the frame straight from the image — joint positions from 33 pose landmarks, plus the body outline from an on-device segmentation mask — to compute ratios the literature actually uses: waist-to-hip (Singh 1993), a waist-to-height adiposity proxy (Tovée & Cornelissen show body-fat is the single strongest cue for female bodies), the shoulder-to-waist V-taper (Dixson) and leg-to-torso (Sorokowski). But the ceiling is lower than the face: clothing, pose, camera angle and arms-at-your-sides all corrupt the silhouette, a single frontal view has no depth (it can’t see a gut side-on), and true body-fat / muscle / height can’t be measured from pixels — only proxied. Unsupported crops and seated, bent, or side-on poses are refused rather than graded from extrapolated joints. When a trained model (models/body-beauty.onnx) is present it scores the headline — on-device, still no upload — and the geometry becomes the transparent breakdown beneath it; until then the geometry scores both. Everything runs locally; the image never leaves your browser. This is a transparent prototype, not a clinical tool — and unlike the face, nothing credible rates a body’s aesthetics. For the body-fat percentage it can only proxy, a caliper or BIA check at a gym or clinic gives a usable number — or, to track it at home, a body-fat smart scale like the Wyze Scale X (best for trends under consistent conditions, not a precise one-off reading). For best results: face forward, full body in frame, arms slightly away from your sides, fitted clothing.