Face Similarity Scores Explained

Two scores, two ways to read the comparison

The search represents facial features as an embedding: a numerical vector. Vector similarity is the cosine similarity between the uploaded face and an indexed face. Higher values mean closer representations. The mathematical cosine range is −1 to 1; the useful range depends on the model and the images.

Similarity maps the raw value between configured lower and upper thresholds onto a scale from 0% to 100%, clipping values outside that interval. It makes the display easier to scan, but the mapping is not a calibrated identity probability. Compare scores within this search rather than with percentages from a different service.

Why several results can all show 100%

Every raw score at or above the configured upper threshold reaches the same displayed maximum. In the example below, cards with different vector values still show 100%. That is a ceiling in the display, not a claim of certainty.

For a purely illustrative configuration with a lower threshold of 0.45 and an upper threshold of 0.72, a raw score of 0.60 maps to roughly 56%. Values of 0.83 and 0.95 both map to 100%. Your active configuration may differ; these example values are not universal identity cutoffs.

● ● ●CS Match / Face SearchProduct example

What lowering Min similarity can change

Min similarity applies to the rescaled main-match score. A setting of 0.55 means a minimum displayed similarity of 55%; it does not mean raw cosine 0.55. Lowering the filter can admit additional candidates, but it does not make a weak photo clearer or add missing images to the index.

Max results, Per domain, candidate retrieval and the indexed content also limit the gallery. If the current results already fill a limit, a lower threshold can leave the visible cards almost unchanged. Similar faces is a separate lower-scoring exploration band with its own selection and diversity rules, rather than an unlimited continuation of the main list.

Read the numbers alongside the evidence

Pose, lighting, occlusion, age differences and image quality can change the comparison. Start with a clear photo, open a promising card, inspect its Domain and Variants views, and visit the source page before drawing a conclusion. A score cannot establish a name, relationship or context on its own.

• Compare the original image and the result at a useful size.
• Look at the raw vector score when percentages tie.
• Check another clear photo if a result seems inconsistent.
• Treat Similar faces as visual leads that may show unrelated people.