5FA
CAPTURE STUDY · PROTOTYPE

BIOMETRIC RECOVERY RESEARCH

Can a camera capture
enough detail?

Compare Cloudflare vision models on guided face, eye, and palm captures using your MacBook Air or iPhone 15 camera.

Capture-quality study only. Models can comment on framing, focus, and lighting. They do not match identity, verify liveness, or authenticate you. Frames are sent to Cloudflare Workers AI for inference and are not saved by this prototype.

STUDY SETUP

Record conditions

Metadata only

Repeat captures at a consistent camera position, distance, and light. Use a new session code when conditions change. Avoid names or account identifiers.

01 / CAMERA

Guided capture

Camera off
◉
Start the camera to begin
On-device landmarksNot started
Live face and hand measurements will appear here. Raw video stays in this browser until you explicitly analyze a frame with a selected model.

Use bright, even light. For palms, move close enough to show crease detail while keeping the whole palm in frame. Landmark measurements are normalized image ratios, not millimeters. This model locates hand joints; it does not read palm creases or measure pupil diameter. The MacBook camera is the baseline; try the iPhone rear camera as a second capture device.

02 / MODEL COMPARISON

Vision results

Waiting for frame
Calls: 0 · saved trials: 0

Still-image models receive the same captured frame in parallel batches. Clef Omni can analyze a short live video clip; this tests video-level capture quality, not identity or liveness.

No live video clip analyzed yet.
Captured frame preview

Loading the Cloudflare vision model registry…

0 trial records in this session

Checking Groq analyzer configuration…

When you click Analyze, saved trial summaries, aggregate on-device landmark ratios, human labels, and model outputs are sent to Groq. No video clips or images are sent. The analyzer provides exploratory commentary; human labels and predefined metrics remain the evaluation basis.

No study analysis generated.

WHAT THIS TEST CAN TELL US

Camera suitability

Resolution, brightness, contrast, sharpness proxy, and human-rated usability by device and target.

Model usefulness

Latency, structured output completion, agreement with human labels, and recapture guidance by model.

What it cannot tell us

Identity match accuracy or spoof resistance. Those need separate biometric and attack evaluations.