Browser-Based Anti-Spoof Engine
Face Liveness That Runs Entirely in the Browser
Passive, Client-Side, Zero Backend Required
Pixl Face Liveness Check is a client-side face liveness & anti-spoof engine — the web counterpart to Pixl's native Android detector. Every check runs on-device, in the page, with no camera frame ever leaving the browser.
At a glance
9 Gated
Pipeline Stages
468 3D
Facial Landmarks
16 Frames
per Verdict (~2s)
0 Frames
Sent Off-Device
Works Fully
Offline
On-Device Analyzer
Capturing 16 Frames
Stage: Landmark Mapping
Hold Still…
On-Device Verdict
"frames_evaluated": 16,
"real_vote_share": "93.7%",
"spoof_detected": false,
"frames_sent_off_device": 0
Privacy-First
No camera frame ever leaves the browser — every check, model and score is computed on the visitor's own device.
No App Required
Runs in any modern browser tab that already asks a user to face a camera — no native install and no live backend to stand up.
One Family, Two
Platforms
Shares its detection logic and model assets with Pixl's native Android FaceSpoofDetector — this is the web port of that same pipeline.
Browser-Based Detection Engine
What is Face Liveness Check?
Pixl Face Liveness Check is a browser-based face liveness & anti-spoof engine. A capture only reaches a verdict after passing nine sequential gated checks — each one able to reject the frame before liveness scoring even runs. Everything happens client-side, inside the page.
Because every check runs on-device, teams get a privacy guarantee that a server-side check
simply can't offer:
Get Started
- 468 3D facial landmarks tracked per frame for consistent measurement.
- A dual-model anti-spoof ensemble scored at two independent crop scales.
- 16 frames captured over roughly two seconds, majority-voted before a verdict.
- Every dependency bundled with the page — no CDN calls at runtime, works fully offline.
Capabilities
Key Benefits
What "liveness detection" actually means for this engine — a real, gated feature list, not a single score.
On-Device Privacy
No camera frame or captured image is ever transmitted off the device — the whole engine runs inside the browser tab.
468-Point Facial Mapping
3D landmarks plotted across the face form the coordinate basis for distance, pose, brightness and glare checks.
Dual-Model Anti-Spoof Ensemble
The face is cropped at two independent scales and scored by two dedicated anti-spoof models, combined into one signal.
Multi-Frame Temporal Voting
A single passing frame is never a verdict — 16 frames are captured over ~2 seconds and majority-voted first.
Anti-Injection Camera Screening
Screens the device label against known virtual/software cameras before a face is ever evaluated.
Works Fully Offline
Every dependency ships bundled with the page — no CDN calls at runtime, so it keeps working once loaded.
Process Flow
The Verification Pipeline
A capture only reaches a verdict after passing nine sequential gated checks — this sequence is the actual feature list.
01
Anti-Injection
Camera Acquisition
Requests the camera with progressively looser constraints and screens the device label against known virtual/software cameras.
02
MediaPipe · BlazeFace
Face Detection
Locates a single face in frame. Frames with zero faces or more than one face are rejected outright.
03
MediaPipe FaceLandmarker
Landmark Mapping
468 3D landmarks are plotted across the face — the coordinate basis every later measurement reads from.
04
Geometry Check
Distance Gate
Estimates camera-to-face distance from interocular spacing, holding the user inside a fixed working range.
05
Pose Check
Frontal-Pose Gate
Confirms eyes are level and the nose is centered within a tight tolerance, ported from Pixl's Android detector.
06
Luma Check
Lighting Gate
Samples brightness across forehead, cheeks and chin, rejecting underlit or one-sided lighting.
07
Glare Check
Eyeglass-Glare Gate
Flags specular highlights and colored-lens reflections that would otherwise blind the texture signal.
08
Dual-Model Ensemble
Liveness Ensemble
Cropped at 2.7× and 4.0× the face box and scored by two dedicated anti-spoof models, combined into one signal.
09
Temporal Voting
Multi-Frame Capture & Verdict
16 frames are captured over roughly two seconds and majority-voted before the user sees a result.
Verdict Logic
How the Accuracy Model Works
There's no single "accuracy score" — a pass is a temporal vote across a whole capture window, gated by two thresholds at once.
Capture
window: 16 frames, ~130ms apart (≈2.0s total)
Per-frame
signal: combined output of 2 models, 2 crop scales
Majority
requirement: ≥ 60% of frames vote "real"
Confidence
requirement: ≥ 85% averaged real-class confidence
PASS only when both conditions hold at once — otherwise the check returns to "Hold still…"
Concrete Proof Point
In testing, holding a printed/screen photo up to the camera scored 54.9% average real-confidence — well under the 85% bar, and correctly rejected. A spoof attempt that would slip past a naive "which class is highest" check gets caught by the confidence bar.
Note
The 85% / 60% thresholds are engineering-set operating points tuned against known failure cases — not a certified accuracy, FAR or FRR figure from a labeled benchmark dataset. This is how the check works, not a percentage we publish as a benchmark claim.
Under the Hood
SDK Profile: What Ships Today
Useful for scoping what claims are accurate about the product's current form factor.
Browser Support
Applications
Where It Fits
Confirmed fit today, plus how digital teams are framing it for specific industries.
Confirmed by the Product
Itself
Any web property that already asks a user to
face a camera in-browser — no native app install required
Contexts that need an on-device privacy
guarantee ("nothing leaves the browser")
Deployments that need to keep working with no
live backend / offline
Products already on Pixl's native Android SDK,
wanting a matching web step
Well-Suited Use Cases
Digital onboarding / KYC for
fintech and neobanks
Remote proctoring & exam
identity checks
Access control for web
portals (HR, healthcare, member sites)
Browser-based attendance /
check-in kiosks
Age-gated content and
age-verification flows
FAQs
Frequently Asked Questions
Answers to common queries about Pixl Face Liveness Check.
Passive liveness detection verifies a real person is in front of the camera without asking them to blink, turn their head or follow any on-screen prompt. Pixl Face Liveness Check runs a nine-stage gated pipeline and a dual-model anti-spoof ensemble across a short capture window to reach a verdict passively.
No. Pixl Face Liveness Check runs entirely client-side in the browser tab. No camera frame or captured image is ever transmitted off the device, and every model dependency is bundled with the page so it keeps working fully offline once loaded.
Chrome, Edge, Firefox and Safari on desktop; Chrome, Samsung Internet, Firefox and embedded WebViews on Android; and Safari on iOS. A secure context (HTTPS or localhost) is required for camera access.
Camera acquisition screens device labels against known virtual/software cameras before a face is ever evaluated, and the dual-model anti-spoof ensemble is tuned to catch printed-photo and screen-replay attacks, alongside known virtual-camera spoofing.
Yes. Pixl Face Liveness Check shares its detection logic and model assets with Pixl's native Android FaceSpoofDetector — this is the web port of that same pipeline, for any web property that needs a matching in-browser step.
See It On Your Own Site
Try Pixl Face Liveness Check
Add a browser-based, privacy-first liveness step to your onboarding, proctoring or access-control flow — no live backend required.
Nothing Leaves the Browser
No camera frame or captured image is ever transmitted off-device.
Works Fully Offline
Every dependency is bundled with the page — no runtime CDN calls.
~2 Second Verdict
16 frames captured and majority-voted before the user sees a result.
One Family, Two Platforms
Same detection pipeline as Pixl's native Android SDK.
Contact Form