# Face match scores and thresholds

Source: https://help.didit.me/liveness-biometrics/face-match-thresholds
Updated: 2026-10-12

Face match returns a similarity score and you decide what happens at each band. Set review and decline thresholds without flooding your queue or missing fraud.

Face match returns a **similarity score**, not a yes or no. You set two
thresholds: below the decline threshold the session is declined, between the two
it goes to review, above the review threshold it passes. Where you put them is a
policy decision, and it is the main lever on both your fraud rate and your review
queue.

## What the score means

[Video: Face match - Didit Academy](https://www.youtube.com/watch?v=7htTWCWsl60&t=376)

[Video: Biometric authentication: re-verify returning users with a selfie - Didit Academy](https://www.youtube.com/watch?v=gJ7BQkNsFjw)

![A verification record in the Didit console with the Liveness and Face match results](https://help.didit.me/screenshots/session-detail-light.webp)

*The score and what it did to the decision are both on the session.*

Face match 1:1 compares the selfie against the portrait on the document and returns a similarity score. It is a measure of how alike two faces are - it is never a certainty about identity, because no biometric comparison is.

That's why the platform gives you two thresholds instead of one verdict.

## The three bands

| Band | What happens |
|---|---|
| Above the review threshold | Passes automatically |
| Between the two thresholds | Session goes to **In review** for a person to decide |
| Below the decline threshold | Session is declined automatically, with `LOW_FACE_MATCH_SIMILARITY` |

Both thresholds are set on the face match step in your workflow, so different workflows can carry different risk appetites - a high-value account opening flow and a low-risk age check don't need the same settings.

## What actually depresses a legitimate score

Before tightening anything, know what moves the score for genuine users:

- **Age of the document photo.** A portrait taken ten years ago is genuinely less similar.
- **Print and scan quality.** Some countries print small, low-contrast portraits; a worn card is worse again.
- **Occlusion.** Glasses, hats, heavy makeup, a beard that wasn't there before.
- **Lighting and camera.** A dim room and a cheap front camera both cost similarity.

None of these is fraud, and all of them are more common in some markets than others. If your users skew toward older documents, a threshold tuned on fresh passports will decline real people.

## Choosing the thresholds

The tradeoff is symmetrical and unavoidable:

- **Tighter thresholds** catch more impostors and send more genuine users to review or decline.
- **Looser thresholds** pass more genuine users and let more impostors through.

A practical way to set them:

1. **Start with the defaults and collect data**

   Run real traffic through the defaults for a few hundred sessions before changing anything. You cannot tune on intuition.

2. **Look at what landed in review, and why**

   Open the review-band sessions. If a reviewer approves nearly all of them, your review threshold is too high - it's generating work with no decisions in it.

3. **Look at the declines**

   If declined sessions are mostly genuine users with old documents, your decline threshold is too high for your population.

4. **Move one threshold at a time**

   Change one, keep the other, and re-measure. Moving both at once tells you nothing about which mattered.

> **Warning:** Do not tune the threshold to make one specific complaint go away. A single
> declined user is not a signal, and loosening a threshold to clear one case
> loosens it for every impostor too. Fix individual cases with a manual review
> decision or a resubmission; change thresholds only on a pattern.

## Widening the review band is not free

A wide review band feels safe - nothing is declined automatically, a human sees everything borderline. In practice it converts a decision problem into a staffing problem: the queue grows, the reviews get less careful, and the effective decision quality drops. See [sessions stuck in review](https://help.didit.me/verification-sessions/session-in-review).

## A stronger check beats a better threshold

If face match is your weak point, the highest-leverage change is usually not the threshold at all:

- **NFC** reads the portrait from the document's chip - the issuer's own image, at full quality - which removes the print-degradation variable.
- **Active liveness** raises the cost of a presentation attack, so a borderline score is less likely to be an attack in the first place.

## Testing your thresholds before going live

Sandbox has scenarios for exactly this: `decline_face_match_low_similarity` for a clear failure and `review_face_match_borderline` for a score inside the review band. Use them to confirm your workflow routes each band where you think it does before real users hit it. See [testing in sandbox](https://help.didit.me/getting-started/sandbox-and-testing).