Use the camera or a local video.
The camera is the default. Open Source to select a local video; the file remains on this device and can be paused, sought, and replayed. Changing source invalidates the current baseline.
Source → Brightness changes only how bright the video looks on screen. The model reads the original video pixels, so this control cannot improve lighting or change a measurement.
Baseline, application, measurement.
Set the baseline
Use a neutral frontal pose and stable light, then press Set baseline. Hold still and follow the centre target while five accepted samples are distributed across at least two seconds.
Start one makeup step
Press Start makeup step before applying one change. The timeline marks the event and shades the application interval.
Return to comparable conditions
Finish the change, then return to the baseline pose, framing, and room light. Leaving the frame during application produces a marked no-face interval without an invented distance line.
Measure the step
Press Measure this step and hold still. Eight comparable frames over at least three seconds become one temporal snapshot.
Read and repeat
Record the distance, baseline noise floor, and step temporal spread together. Use Start next step for the next cumulative change.
Different pose and light are omitted, not normalised.
Baseline capture checks confidence, scale, framing, 2D roll/yaw, and face-crop exposure. After makeup starts, confidence and face-crop exposure may legitimately change, so they no longer reject a frame. Scale, framing, roll, and yaw still apply, followed by a comparison with the baseline's averaged 2D conditions and ambient luminance outside the face box.
The gate reduces obvious confounds; it cannot prove that makeup caused the remaining distance. It does not estimate pitch and cannot separate every intervention from every shadow.
Read the trace and the 128-value fingerprints.
| View | What it shows | Do not infer |
|---|---|---|
| Distance timeline | Smoothed live distance, gaps, phase backgrounds, event lines, measured noise floor, and the cited 0.60 dlib reference. | That the space between the two reference lines has a calibrated meaning. |
| Equalizer | Absolute change for each of 128 descriptor coordinates. | That one row corresponds to a facial region. |
| Radial distribution | All 128 deltas sorted into 16 ranked RMS bands, with the live reading as an outline and the latest measured step as a filled shape. | That a spoke or direction identifies a makeup line, cheek, or eyebrow. |
| Step capture | The sharpest accepted frame from the measurement window, watermarked with local result metadata. | That one model-local result transfers to a deployed system. |
Use video as a repeatable fixture.
Local-video event labels include media time. Pause or seek to inspect the recording, then mark application and measurement moments. Starting either interval resumes a paused video so the temporal collector does not repeatedly inspect one frozen frame.
Local does not mean insignificant.
- Camera baselines persist in localStorage; local-video baselines exist only for the current page session.
- Manual, sustained-spike, and declared-step JPEG captures use browser IndexedDB.
- The selected video is neither uploaded nor persisted by the page.
- Delete baseline does not erase the gallery. Delete captures separately when they are no longer needed.