Apple's ARKit reference defines the coefficients trackers read

The setup

Apple's ARKit is the iOS framework nearly every iPhone-based VTuber face tracker builds on, and Apple's own developer documentation, not any third-party app, defines the data those trackers actually read. The relevant class is ARFaceAnchor, whose BlendShapeLocation documentation lists the specific facial features ARKit reports on, and the ARKit framework page states the overall system requirement: 'iOS: 11.0.0 -,' meaning support beginning with iOS 11. This is a living-document entry describing the current API, retrieved 16 September 2026, not a single dated release.

What the documents show

Apple's documentation states each blend shape coefficient is 'a floating point number indicating the current position of that feature relative to its neutral configuration, ranging from 0.0 (neutral) to 1.0 (maximum movement).' The BlendShapeLocation page enumerates 52 individually named coefficients grouped under 'Left Eye,' 'Right Eye,' 'Mouth and Jaw,' 'Eyebrows,' and 'Tongue,' each with its own one-line description, such as 'jawOpen,' 'the coefficient describing an opening of the lower jaw.' The ARFaceAnchor page adds the mechanism: this data comes from 'ARFaceTrackingConfiguration,' which 'can track multiple faces simultaneously,' and separately provides a full 'ARFaceGeometry' mesh for the detected face's shape, distinct from the blend-shape coefficients used for animation.

The craft

Apple's own documentation states a developer can use 'as many or as few of the coefficients as you desire,' citing a simple cartoon rig built from only 'jawOpen, eyeBlinkLeft, and eyeBlinkRight' as one end of the range, up to a full professional character rig using the entire coefficient set at the other. This is exactly the design space a VTuber's rigger works within: deciding how many of the 52 named coefficients a given model's blendshapes actually need to be mapped to, a workflow this site documents separately for one specific tracking tool.

Keeping the creator in control

Because ARKit reports named, numeric coefficients rather than raw camera images, an app built on this API can, as a technical matter, transmit only those numbers onward - what any specific third-party tracker actually chooses to send or store is a separate, per-app question this entry does not resolve. Apple's own note on left/right naming, that coefficients are named 'relative to the face' rather than the viewer, is a small but practically important detail for anyone debugging a mirrored rig.

• Does the target VTuber tool map its model's blendshapes to Apple's named coefficients directly, or through an intermediate standard?
• Which subset of the 52 coefficients does the rig actually need for a convincing expression, versus which go unused?
• Does the receiving application document what it does with these coefficients beyond real-time animation?

Apple's documentation is the ground truth for what iPhone-based tracking can report at all; the choices about how much of that 52-coefficient range a given rig actually uses are made downstream, by the tracker app and the rigger.