Detect Faces (Auraface)

How it fits

Typical backends

Caveats

• Hard constraintmodel must point at a directory holding exactly one ONNX file; zero or several abort startup.
• FallbackRequesting a GPU on a node with none falls back to CPU at startup rather than failing, and an empty input image returns an empty list without running the detector.
• Limitationdet_size is fixed at startup and not effectively mutable at runtime; an ill-matched size silently degrades detection quality.
• I/O contractOutput cardinality varies per frame and an empty list is valid when no face clears threshold; results follow the detector's natural order, not sorted by confidence or position.
• AccuracyEach Landmark confidence is copied from its parent box score — the detector does not estimate per-keypoint confidence — and threshold is captured at startup, so changing it requires a restart.