Dynsimulation
DynsimulationTechnologies
Face Match & Face Similarity API · 1:1 & 1:N Comparison Engine

Face Match & Similarity
Compare Faces with 99.4% Accuracy in 120ms

Answer “is this the same person?” in the time it takes a page to load. Upload two photos, capture a live selfie, or run a preset pair against the real API — and get back a calibrated score, a verdict against your threshold, and the quality evidence behind both.

Try Live Interactive DemoRequest API Access
99.4%
TAR @ FAR 0.01%
<120ms
Comparison Latency
1:1 & 1:N
Dual Verification Modes
100%
Zero Media Retention
Live Demo API

Live Face Similarity & Comparison Sandbox

This calls the production comparison endpoint, not a canned animation. Try a genuine pair, an impostor pair, and a spoof attempt, and watch where the score lands.

Sample Pairs:
Image 1 (e.g. ID / Passport Photo)
Image 1 (e.g. ID / Passport Photo)
Samples:
Image 2 (e.g. Live Selfie / Capture)
Image 2 (e.g. Live Selfie / Capture)
Samples:
Under the Hood

What Happens Between the Two Photos

Face Match never compares pixels. Each face is converted into a mathematical representation of its structure, and the score you receive is the measured distance between the two.

01

Detect & Score Quality

Every face in both frames is located, then graded on sharpness, pose angle, exposure and occlusion. A photo that is too dark or too rotated to compare fairly is flagged before it can produce a misleading score.

02

Align to Canonical Geometry

Landmarks around the eyes, nose and mouth are used to rotate, scale and crop each face onto the same reference grid, so a tilted passport scan and a handheld selfie are measured on identical terms.

03

Embed Into a Vector

The aligned face becomes a 256-bit embedding — a numeric fingerprint of facial structure, not a picture. Embeddings cannot be reversed into the original photo, which is what makes them safe to store when the source image is not.

04

Compare & Calibrate

The distance between the two embeddings is converted into a calibrated 0–100 score and a verdict against your threshold, returned with per-image quality telemetry so borderline results are explainable.

Two Questions, Two Modes

Verification and Search Are Not the Same Job

Both run on the same embeddings and the same accuracy profile. The difference is what you are asking — confirm a claimed identity, or find one you do not have yet.

1:1 Verification

Is this the same person?

Compare one face against one other face — a document portrait against a live selfie, or a returning user against the photo captured at signup.

  • Photo ID, passport or driving licence against a live capture
  • Returning-user re-authentication before a high-risk action
  • Single request, single score, sub-120ms round trip
1:N Search

Who is this person — and are they already enrolled?

Search one face against an entire gallery to identify it, or to prove nobody has signed up twice under a different name.

  • Duplicate and synthetic-identity detection at onboarding
  • Watchlist and repeat-offender screening
  • Indexed for million-scale galleries, ranked candidates returned
Interpretation Guide

What the Face Similarity Score Means

The score runs from 0 to 100 as a calibrated mathematical distance between extracted 256-bit biometric vector embeddings.

Score BandVerdictBusiness Meaning & Recommended Action
85 – 100Strong MatchSame person with high image quality on both sides. Safe to auto-approve in digital onboarding without manual intervention.
70 – 85MatchSame person, but with slight environmental degradation (e.g. printed ID scan, minor angle rotation, or non-uniform lighting).
40 – 70InconclusiveGenuinely ambiguous due to heavy shadows, sunglasses, or a significant age gap. Route to manual officer review.
0 – 40No MatchDifferent people. Highly reliable rejection; false rejections in this band are under 1 in 10,000.
Applied

Where Face Comparison Earns Its Place

The common thread: a decision that currently waits on a human comparing two photographs, and gets slower and less consistent the more of them there are.

Digital KYC Onboarding

Match the portrait on a submitted ID to a live selfie so an account can be opened in one sitting instead of a branch visit. Strong matches auto-approve; the inconclusive band routes to a human reviewer.

Duplicate Account Prevention

Search each new face against your existing gallery before enrolment. Catches one person farming signup bonuses across many names — the fraud pattern document checks alone cannot see.

Step-Up Authentication

Re-verify a returning user against their enrolment photo before a withdrawal, a password reset or a change of payout details, without sending them back through full re-onboarding.

Workforce & Site Access

Confirm that the person at the gate, on the shift-start app or in the delivery vehicle is the individual assigned to that badge — including contractor and gig-worker fleets that rotate weekly.

Remote Exams & Proctoring

Verify the candidate who started the session is the one who finishes it, with periodic comparisons against the identity captured at check-in.

Fraud & Case Investigation

Compare a disputed photo against known records, or cluster an entire case file by identity, to surface links that manual review across thousands of images would miss.

Honest Limits

How the Score Behaves in the Real World

Benchmark accuracy is measured on clean pairs. Production traffic is not clean, so here is what actually moves the number — and the one thing similarity scoring cannot do on its own.

ConditionEffect on ScoreWhat to Expect
Age gap of 5–10 yearsMinorTypical score drop of a few points. Bone structure dominates the embedding, so most identity documents remain matchable across their full validity period.
Glasses, beard or hairstyle changeMinorHandled directly. The model weights stable facial geometry over surface appearance, so cosmetic change rarely moves a genuine pair below threshold.
Printed or scanned ID portraitModerateHalftone printing and glare cost some detail. Genuine pairs usually land in the 70–85 band rather than 85+, which is why the threshold sits at 70 rather than higher.
Heavy pose angle or partial occlusionModerateFlagged in the quality telemetry before scoring. A capture beyond the usable pose range returns a quality warning instead of a confident but unreliable number.
Presentation attack (printed photo, screen replay, mask)Handled separatelySimilarity alone cannot tell a live face from a photo of one. Pair Face Match with liveness detection so a high score is only trusted when the capture is proven live.

Accuracy figures on this page refer to a 99.4% true accept rate at a 0.01% false accept rate on our internal evaluation set. Your own rates will depend on capture quality, and we will help you measure them on your traffic before you set a production threshold.

Developer First

Integrate Face Comparison in Minutes

One POST with two images. No SDK to adopt, no model to host, no queue to poll — the score comes back on the same request, which is what makes it usable inside a signup flow rather than after it.

  • Accepts multipart image files, base64 strings, or S3/HTTPS URLs
  • Configurable decision threshold per request or enterprise tenant
  • Zero data stored on disk — memory processed & ephemeral
curl -X POST https://api.dynsimulation.com/v1/vision/face-match \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "source_image=@id_passport.jpg" \
  -F "target_image=@live_selfie.jpg" \
  -F "threshold=70"
Response Schema

Every Field You Get Back

A verdict alone is not enough to defend a rejected customer or a passed fraudster. Each response carries the evidence behind the number.

FieldTypeDescription
similarity_scorefloatCalibrated 0–100 similarity between the two faces.
verdictstringMATCH, INCONCLUSIVE or NO_MATCH, resolved against the threshold on the request.
thresholdintThe decision threshold applied, echoed back for audit trails.
faces_detectedobjectFace count per image, so multi-face or empty frames can be rejected explicitly.
qualityobjectPer-image sharpness, pose and exposure indicators behind the score.
execution_time_msintServer-side processing time, excluding network transit.
enginestringModel build identifier, so scores can be tied to a specific version.
Privacy & Deployment

Biometrics You Can Defend to Your Regulator

Face data is the most sensitive category most teams will ever process. The architecture is built so the answer to “where are the faces kept?” is “nowhere”.

Nothing Written to Disk

Images exist only in memory for the duration of the request. No raw photo, crop or thumbnail is persisted, so there is no biometric store to breach or subpoena.

Deploy Where the Data Lives

Use the hosted REST API, or run the same model as a GPU container inside your VPC — including fully air-gapped installs for environments where faces cannot leave the building.

Auditable by Design

Every response carries the threshold, model build and quality telemetry that produced it, which is what a regulator or an internal reviewer asks for when a decision is challenged.

Your Thresholds, Your Policy

Set the decision boundary per request or per tenant. A savings-account opening and a parcel handover do not deserve the same level of certainty, and neither is hard-coded.

Deploy Face Match & Similarity

Request Face Match API Credentials

Tell us what you are verifying and at what volume. You will get sandbox keys, a threshold recommendation calibrated on your own capture conditions, and pricing for hosted, VPC or air-gapped deployment.

Frequently asked questions

What is the difference between 1:1 and 1:N matching?

1:1 compares two specific faces — typically a photo ID or passport against a live selfie — and answers whether they are the same person. 1:N searches one face against an entire gallery to find who it is, or whether they are already enrolled.

How accurate is Face Match?

A stated 99.4% true accept rate at a false accept rate of 0.01%.

How should we read the similarity score?

The API returns a calibrated score from 0 to 100 alongside a match verdict and confidence telemetry, so you can either take the verdict directly or apply your own threshold for the risk level of the transaction.

How fast is a comparison?

Around 120ms for a face comparison.

How large a gallery can 1:N search?

The gallery search is built for million-scale biometric datasets, which is what makes it usable for deduplication as well as identification.

Are the submitted photos stored?

No. Processing is fully ephemeral in memory, with no raw photo written to disk.

Does Face Match detect spoofing on its own?

No, and no similarity model does. A printed photo of the right person will score as the right person, because it is. Pair Face Match with liveness detection so a high score is only trusted when the capture is proven to come from a live face.

Does it still match after glasses, a beard or several years?

Usually yes. The embedding weights stable facial geometry over surface appearance, so cosmetic change and a five to ten year age gap typically cost a few points rather than flipping the verdict. Printed or scanned ID portraits cost more, which is why the default threshold sits at 70.

Can it run on our own infrastructure?

Yes. The same model ships as a GPU container you can run inside your VPC, including fully air-gapped installs for environments where face data cannot leave the premises.

What comes back besides the score?

The verdict, the threshold that was applied, the face count per image, per-image quality telemetry, the processing time and the model build identifier — enough to reconstruct and defend any individual decision later.

Explore the rest of the suite

View all products