A rating with context. Measurements with definitions.

How the PSL rating works

After consent, a resized JPEG is sent through our server to Replicate’s google/gemini-3-flash model. It evaluates the visible features in one photograph and returns a subjective 1–8 rating in half-point increments. The full report includes eight feature readings, visible strengths, practical next steps and photo limitations. Welcome credits reveal only the cover score; a credit pack purchase unlocks the full explanation. A paid guest report includes the full result without an account.

The score summarizes a subjective visual impression. The eight feature readings explain particular observations; they are not objective measurements or a claimed mathematical formula for the overall score. No attractiveness dataset, calibration study or accuracy benchmark validates this rating.

Different photos and model runs can produce different scores. Unclear or unsuitable photos may receive no score. We do not calculate the score from the ratios below, assign insulting tier labels or recommend cosmetic procedures.

How the analysis runs

Your image is decoded locally. The analysis canvas is reduced to a maximum side length of 1,280 pixels while preserving its aspect ratio. The model detects up to two faces so the tool can reject photos with more than one detected face.

A successful result needs one detected face. Normalized landmark coordinates are converted into image-pixel coordinates before distances are calculated. This prevents a non-square photo from distorting the ratios merely because of its aspect ratio.

Eye spacing

The distance between the inner eye corners is divided by the average of the left and right visible eye widths. A result of 1.00 means those distances are equal in the photograph. It is not a target or an attractiveness threshold.

Face width-to-height

The distance between lateral face landmarks is divided by the distance from the upper forehead landmark to the chin landmark. These are the model’s visible landmarks, not clinical bone measurements or a precise hairline measurement.

Mouth-to-nose width

The distance between mouth corners is divided by the distance between selected lateral nose landmarks. Expression, occlusion and landmark placement can change this ratio.

Visible eye-width difference and eye-line tilt

Eye-width difference is the absolute difference between the two measured eye widths, divided by their average, expressed as a percentage. It describes this image and can increase with head rotation or partial occlusion.

Eye-line tilt is the angle of the line between outer eye corners relative to the image horizontal. It helps you notice a tilted photo. Neither measurement is a clinical assessment of facial symmetry.

What this tool cannot establish

These measurements do not establish an objective PSL rating, attractiveness percentile, medical condition or personal trait. We do not infer identity, ethnicity, gender or age. The model can make errors and may behave differently across lighting, faces and devices.

The number of landmarks is not a confidence score. We do not claim benchmark accuracy or a validated relationship between these ratios and perceived attractiveness. Read results as approximate descriptions of a single photograph.

Software and privacy

The implementation uses Google’s MediaPipe Tasks Vision face landmark detector. The runtime and model are served from this website. Landmark results are processed in browser memory. The separate AI rating sends a resized photo through our API to Replicate after consent. Raw photos are not stored in our database. Generated report text is stored in Supabase so the account or guest owner can return to the result. With your consent, a temporary photo copy stays in this browser during sign-in or checkout; the pending record expires after 30 minutes for account flows or 7 days for guest orders and is removed after successful generation or photo replacement.

Downloaded reports contain the text and results available to you, not the photo. Model files may be cached by the browser. The original face illustration on this website is a diagram, not evidence of an actual detection result.