Measured cheek-to-cheek using facial landmark points. Typical error ±2–3 mm on a 130–155 mm face.

Estimated from iris centre positions. With your prescription PD as calibration, error drops to ±0.5–1 mm.

Inner nose landmark span. Consistent across scans; ±1–2 mm variation depending on lighting and head tilt.

Derived from temple-to-temple span using nose landmark as the symmetric centre. ±2–3 mm expected range.

Why 3 captures?

Every scan averages 3 auto-captured frames taken 600 ms apart. This smooths out micro head-movements and sub-pixel iris jitter, giving you a stable, reproducible reading every time.

8-frame rolling buffer — before any capture, MediaPipe processes 8 consecutive video frames and checks that all 5 alignment gates (face size, yaw angle, nose position, camera distance, no mask) are passing consistently. This ensures your head is truly still, not just still for one frame.

3 snapshots at 600 ms intervals — once stable for 1.6 seconds, the system captures 3 frames spaced 600 ms apart. Each capture sends the image to our AI server for full landmark analysis and mm conversion.

Averaged result — all measurement values (face width, PD, nose bridge, etc.) from the 3 captures are averaged together. Outlier frames (e.g. from a tiny blink) are naturally diluted. The final result is statistically more stable than any single capture.

The result: smaller error, better fit

In testing, single-frame PD estimation showed a standard deviation of ±1.8 mm. With 3-frame averaging under the stable-gate condition, this dropped to ±0.7 mm — a 60% improvement in precision without any additional hardware.

Cheek-to-cheek horizontal distance. Used to determine frame size category (S / M / L) and whether a frame will sit flush.

Forehead top to chin bottom. Helps distinguish Oblong from Square or Oval face shapes and informs lens height suitability.

Centre-to-centre distance between your pupils. Critical for centring prescription lenses in any frame. Most important for lens cutting.

The gap between the inner corners of both eyes. Closely matches the bridge width you need in a frame — prevents pinching or gaps.

Measured across the nose at the inner-eye level. Helps match frames with narrow or wide nose bridges to ensure the frame sits evenly.

Temple-to-temple span from nose centre outward, doubled. This is the total front width of an eyeframe that will sit flush without extending past your face.

Face Shape (7th output)

Using the width-to-height ratio plus forehead and jaw landmark positions, the AI classifies your face as Oval, Round, Square, Heart, Oblong, or Diamond — and maps it to the best-fitting frame styles in our catalogue.

Oval

Balanced proportions — slightly wider forehead, gentle chin. The most versatile face shape. Almost all frame styles work well.

Round

Width and height are nearly equal with soft curves. Angular or rectangular frames add definition and make the face appear longer.

Square

Strong jaw and forehead of similar width. Round or oval frames soften the angles. Avoid boxy frames that echo the face shape.

Heart

Wider forehead tapering to a narrow chin. Bottom-heavy or rimless frames balance the forehead. Avoid top-heavy browline frames.

Oblong

Longer than wide with a straight cheekline. Taller frames (large lens height) or wide frames add width and balance the length.

Diamond

Narrow forehead and jaw with prominent cheekbones. Cat-eye or oval frames with detailing at the top complement the cheekbone width.

Iris centre as anchor — the midpoint between your left and right iris centres (in pixels) is the horizontal anchor. The vertical anchor places the iris at 45% from the top of the frame image — matching how most frame product photos are framed.

Scale from mm to pixels — the frame's total front width (lens + lens + bridge in mm, from the product listing) is converted to pixels using your mm-per-pixel scale factor captured during the scan. This makes the frame appear exactly as large as it would sit on your face.

Multiply blend removes white backgrounds — frame product photos have white backgrounds. The canvas uses globalCompositeOperation: multiply to mathematically remove white while keeping the frame colours — no green-screen needed.

Opacity slider — adjust from 30–100% to see how opaque or see-through the frame appears. Lower opacity helps verify bridge and temple alignment on your face.

Best results: use frames that have product images specifically from a front-facing angle. The View Product button on each try-on frame takes you directly to that frame's listing — where you can check exact mm dimensions before ordering.