Setup lighting (better camera you can move)
- The theory wants "an evenly lit face, a stable camera position, and a simple background", but that's is true if you want to sistematically study the effect of every dot.
- Allow camera access. The models are served from this site and may take time to initialise.
- Only one face in frame.
- Keep pose, distance, light, and camera as similar as possible if you're comparing states.
- Once a while move the camera, your angle, and see how it affects the detection.
Baseline means the saved face descriptor used as the comparison reference. It stays in the browser on this device. See baseline.
From camera to comparison.
Wait for the camera and models
The live stage should replace the loading message. Before a face is saved, the readout asks you to save one. This is not yet a match result.
Select Save
Keep your face steady and use Save in the bottom bar. The saved-face count should become one. The same face is now the active local baseline.
Read the unmodified state
Leave the Camera view active. The number beside MATCH is a descriptor distance. In the current 2D pipeline, smaller means closer to the saved baseline.
Inspect landmarks
Switch between 2D points and 3D mesh. These views show geometry estimated from the frame. They are not identity scores.
Choose a Ghostyle
Open Settings, review the current threshold, and select a Ghostyle. The overlay is rendered on the face and the composite image is analysed separately.
Compare the result
Watch the distance and status over several updates. Record the Ghostyle, threshold, camera, light, pose, and any instability. Do not keep only the highest or lowest transient value.
Geometry is not identity.
Original frame plus active Ghostyle
The cleanest view for following the visible intervention and the top readout.
68 face-api landmarks
Points used to locate facial regions and anchor compatible 2D Ghostyles.
478 MediaPipe landmarks
A dense normalised face mesh used by the 3D path and UV-capable interventions.
Distance against a baseline
A separate comparison output. Read it with the current threshold and state.
Change one condition at a time.
A Ghostyle is a modular visual intervention. Some Ghostyles draw from the 68-point geometry, some use the 3D mesh or UV texture, and command Ghostyles such as Face Brush add their own controls.
The Lab keeps tracking the original camera image while it also builds a composited image containing the Ghostyle. That distinction explains why a face box can remain visible while the result of the modified-image pass changes.
Local until you choose otherwise.
- Saved baselines and their thumbnails persist in browser storage on this device.
- Delete all in Settings removes the saved face records from this browser.
- Record creates a short clip and places it in the browser workflow for review.
- An upload is a separate action. It requires review and explicit consent before the clip is sent to the configured server endpoint.
- Copying or downloading an image exports data from the page to your device or clipboard.
Common problems and what to check.
| What you see | Likely meaning | What to check |
|---|---|---|
| Loading message remains | Models or camera did not become ready. | Local model files, browser console, camera permission, and secure origin. |
| No face in frame | The detector produced no usable face for that pass. | Light, angle, occlusion, distance, and whether the video frame is valid. |
| Save your face | No active baseline exists. | Select Save while one readable face is present. |
| Distance changes rapidly | Frame-by-frame predictions or input conditions are unstable. | Hold position, improve light, and observe a series rather than one value. |
| 2D and 3D disagree | The two pipelines use different representations and decision directions. | Record both outputs and treat the result as partial or unresolved. |
Do not read “Escaped” as “safe”. It means the current comparison crossed the configured local threshold, or that the modified-image detector did not return a face. It says nothing by itself about another model, camera, deployment, or future frame.