Starts from a flat image.
Face detection
The task of locating a probable face in an image. Detection normally happens before identity comparison.
In the LabTinyFaceDetector analyses the 2D path. MediaPipe FaceLandmarker also performs its own face localisation for the mesh path. Detection does not identify a person.
Bounding box
A rectangle, represented by position and dimensions, around a region that a detector considers a face candidate.
In the LabThe visible scaffold tracks the live camera frame. A Ghostyle composite is analysed separately, so a visible live box does not prove that the modified image produced a face descriptor.
Detection score threshold
The minimum detector score required to report a candidate. Raising it rejects weaker candidates; lowering it accepts more uncertain candidates.
In the LabThe normal TinyFaceDetector threshold is 0.5. A failed composite may be retried at 0.1 for diagnostics. This is separate from the 0.58 identity match threshold.
T-zone
A descriptive pattern of light and dark facial regions around the forehead, nose, eyes, and mouth. It helped explain why high-contrast styling disrupted some early detectors.
Historical contextUseful for discussing CV Dazzle's visual genealogy. It is not a formal input exposed by the current Lab and should not be presented as TinyFaceDetector's explicit rule.
Haar cascade
A classical detector that combines many simple contrast features in a cascade of decisions. It predates modern neural face detectors.
Historical contextEarly CV Dazzle work is often discussed against this family of detectors. Ghostmaxxing currently uses TinyFaceDetector and MediaPipe, not an OpenCV Haar cascade.
False positive
A system reports a face or match when the target condition is absent.
In an experimentRecord whether the error belongs to detection or identity comparison. A permissive threshold can increase one kind of false positive while changing other trade-offs.
False negative
A system fails to report a face or match when the target condition is present.
In an experimentAn intervention may contribute to a false negative, but a single absence can also come from pose, light, blur, timing, or model instability.
Estimated structure.
Face landmarks
Estimated coordinates for facial features such as eyes, brows, nose, mouth, and outline.
In the LabFace-api supplies 68 points for the 2D path. MediaPipe FaceLandmarker supplies 478 points for the mesh and UV path. Point count does not by itself indicate identity accuracy.
Face alignment
The geometric normalisation of a detected face so that identity features can be compared across pose and position.
In the pipelineLandmarks support the face-api recognition pipeline. The Lab does not expose alignment as a user control or an independent confidence score.
Tracking
Maintaining a useful estimate of a face's location or geometry across frames.
In the LabThe live scaffold and mesh update continuously, while the composite identity check runs on a slower interval. A stable overlay and a fresh identity result are not the same event.
Jitter
Visible frame-to-frame movement in estimated points or overlays even when the physical subject appears nearly still.
What to observeOcclusion, poor light, motion and uncertain geometry can increase jitter. It can expose instability that one still screenshot hides.
Smoothing
A temporal filter that reduces rapid changes by combining recent values or coordinates.
Implementation noteSmoothing is a general technique, not a documented guarantee of the current Ghostmaxxing landmark loops. It may improve presentation while also hiding short-lived uncertainty.
Representations and boundaries.
Baseline
A saved reference captured before the intervention and used for later comparison.
In the Lab“Save” stores local face-api and experimental image embeddings under a shared record ID. A baseline is an experimental reference, not verified civil identity.
Embedding
A numeric vector that represents features learned by a model. Nearby vectors are intended to encode some form of similarity defined by the model and its training.
In the projectFace-api produces a dedicated face descriptor. MediaPipe ImageEmbedder produces a general visual embedding. They should not be treated as interchangeable.
Face descriptor
The 128-value identity representation returned by the current face-api recognition model for a detected face.
In the LabThe descriptor is compared with locally saved descriptors. It is biometric-derived data and should not be published or logged casually.
Latent space
An abstract multidimensional space formed by a learned representation. Inputs with features the model treats as similar tend to occupy nearby regions.
Practical limit“Nearby” depends on the model. It does not mean visually or socially identical, and the dimensions are not simple named facial traits.
Euclidean distance
The straight-line distance between two numeric vectors.
In the 2D readoutFace-api distances are compared with 0.58 by default. Lower means closer. The number is not a percentage.
Cosine similarity
A comparison based on the angle between two vectors rather than their straight-line distance.
In the experimental pathImageEmbedder similarities are compared with 0.85 by default. Higher means closer. This direction is opposite to face-api distance.
Match threshold
A boundary that turns a continuous comparison value into a category.
In the LabA face-api distance at or below 0.58 is a match. Changing this value changes the decision rule, not the pixels or descriptor.
What changes, and where.
Ghostyle
A Ghostmaxxing visual intervention implemented as an ES module. It may draw in 2D from face-api landmarks, paint a UV texture for the mesh path, or expose both capabilities.
In the LabA digital Ghostyle is a probe for the local models. A physical makeup test is a separate experiment affected by material, light, pose, camera and movement.
Composite pass
A second analysis performed on an image that combines the camera frame with the active Ghostyle.
In the LabThe composite pass can produce a different box, descriptor or match state from the original camera path. Its slower cadence is intentional.
Adversarial example
An input intentionally modified to cause a model to behave differently while preserving some meaning for a human observer.
Evidence limitA visually unusual result is not automatically adversarial. Establish the target model, intended change, baseline, repeated measurement and relevant conditions.
Adversarial patch
A localised, often optimised pattern designed to influence a model when placed within its view.
Relation to makeupModern patches can be model-specific and need not resemble CV Dazzle. Performance against one pipeline does not transfer automatically to another.
Escaped
A Ghostmaxxing interface label for a local comparison that crossed the current rule while an intervention was active.
Do not overread itIt is not a standard security certification and does not mean anonymous, invisible or safe. State whether the result was detector absence or descriptor distance beyond threshold.
Liveness and presentation-attack detection
Methods intended to distinguish a live authorised presentation from photos, screens, masks or other presentation attacks. Some use motion, depth, texture, sensors or a challenge-response action.
Not implemented hereGhostmaxxing does not currently perform liveness or presentation-attack detection. Expression analysis alone would not establish liveness.