What your similarity score means

How ranking works

ArcFace produces a face representation for each usable photo, and the service normalizes it. Multiple photos are combined into quality-weighted templates, and cosine similarity ranks only the relatives you uploaded. “Closest match” means the highest visual similarity in that uploaded group.

How the display score works

The raw model comparison is converted to a bounded 0–100 entertainment score. If an approved calibration artifact is configured, the mapping follows consented human resemblance ratings; otherwise the interface labels the configured mapping as experimental.

• 0–24: Limited visual resemblance
• 25–44: Some visible similarities
• 45–64: Moderate visual resemblance
• 65–79: Strong visual resemblance
• 80–100: Very strong visual resemblance

A 46% example

A 46% score falls in the moderate visual resemblance band: the model found several noticeable shared visual traits alongside clear differences. It does not mean a 46% chance of being related.

Score and confidence are different

The similarity score describes the visual comparison. Result confidence summarizes photo suitability, the lead over the second-ranked relative, and consistency when multiple usable photo pairs are available; a fixed stability baseline is used when only one pair is available. It is never identity or biological confidence.

What feature details describe

Eye spacing, nose position, and mouth proportions use detected landmark positions, while face shape uses face-box proportions. These detector-geometry cues are aggregated across usable photo pairs. They are not explanations of ArcFace decisions and do not describe eye color, identity, genetics, or inherited traits.

Why a score can change

Age, expression, lighting, camera quality, pose, crop, blur, and model or detector updates can affect the result. Clear photos with similar angles usually produce a more consistent comparison.