Measured
On the production service, 28 October 2026, against the public LFW benchmark (6,000 pairs of news photographs) and its cross-age (CALFW) and cross-pose (CPLFW) variants, at the default threshold. LFW and CALFW photos were first centre-cropped to the subject's face, because the service compares exactly one face per image.
| Measure | Result |
|---|
| Correct decisions, 6,000 LFW pairs | 98.50% |
| Same-person pairs accepted | 97.00% |
| Different-person pairs accepted | 4 of 30,000 (FAR 1.3 × 10⁻⁴) |
| Same-person accepted at FAR ≤ 10⁻³ / ≤ 10⁻⁴ | 97.10% / 96.80% |
| Threshold 48: same-person / different-person accepted | 94.20% / 0 of 30,000 |
| Correct decisions, photos years apart (CALFW) | 94.83% |
| Correct decisions, head strongly turned (CPLFW) | 71.95% |
| Median time through the API | 0.64 s |
The default threshold favours security: the 4 different-person pairs it accepted scored 37 to 40, just above it. A strongly turned head is the weak case: on CPLFW the service declined 43.2% of pairs, mostly because it found no face in one of the photos, and accepted 44.13% of same-person pairs, so ask the user to face the camera and retake. The test sets are news photographs of public figures, not Thai identity documents. Method and limits: the benchmark (PDF).
Limits
- One face per image: a photo that shows a second face is declined, not compared. On uncropped LFW news photographs 16.3% of pairs were declined, nearly all because a second face was visible. Photograph the user alone, or crop to the user's face.
- Built for frontal faces, such as a selfie and an identity-document photo; a strongly turned head is often not found. Poor lighting and occlusion also lower the score.
Data handling
Images are reduced to numeric features for the comparison and are not kept after the response. iApp keeps no copy; consent and notice under PDPA and GDPR remain with you as the data controller.
