Detect Faces (Ultralytics YOLO)

How it fits

Typical backends

Caveats

• Hard constraintThe model weights slot must hold exactly one Ultralytics YOLO face `.pt` checkpoint; zero or multiple `.pt` files fail to load, and a checkpoint with a different head shape produces nonsense boxes.
• I/O contractFive-point landmarks require a keypoint-capable YOLO-face checkpoint; a detection-only checkpoint pairs every box with an empty keypoint list on the landmark arms.
• I/O contractThe ([BoundingBox], [[Landmark]]) arm emits two index-aligned lists: box i describes the same face as landmark-list i, the same faces the paired arm carries.
• FallbackRequesting a GPU when none is visible falls back to CPU, and an empty Image returns an empty result without invoking the model.
• Parameter interactionthreshold filters per-box confidence before non-maximum suppression, so lowering it surfaces weak faces but multiplies the overlap that iou then has to merge.
• LimitationAll settings are captured once at startup; switching the model weights, device, or image_size at runtime has no effect and requires a redeploy.