Skip to main content

Overview

Detects faces in an image and identifies 468 facial landmarks (key points) on each face using MediaPipe’s BlazeFace and FaceMesh models. It also calculates ARKit-52 blendshape coefficients for facial expression analysis. The node can process multiple images in a batch and outputs both the landmark data and bounding boxes for each detected face.

Inputs

ParameterDescriptionData TypeRequiredRange
face_detection_modelThe MediaPipe face detection model to use for landmark detection.FACE_DETECTION_MODELYes
imageThe input image or batch of images to detect faces in.IMAGEYes
detector_variantFace detector range. "short" is tuned for close-up faces (within ~2 m of the camera); "full" covers farther/smaller faces (up to ~5 m) but is slower. "both" runs both detectors and keeps whichever found more faces per frame (~2x detection cost). Default: "short".COMBOYes"short"
"full"
"both"
num_facesMaximum number of faces to return per frame. 0 means no cap (return all detected). Default: 1.INTYes0 to 16
min_confidenceBlazeFace score threshold. Lower values help catch small or occluded faces. Default: 0.5.FLOATNo0.00 to 1.00
missing_frame_fallbackPer-frame behavior when detection fails in a batch. "empty" leaves the frame faceless. "previous" copies the most recent successful detection. "interpolate" lerps landmarks/bbox/blendshapes between bracketing successful frames. Multi-face: pairs faces across frames by greedy bbox-centre NN. Default: "empty".COMBONo"empty"
"previous"
"interpolate"

Outputs

Output NameDescriptionData Type
face_landmarksA structured output containing per-frame face detection results, including 468 facial landmarks, ARKit-52 blendshape coefficients, transformation matrices, and connection sets for mesh visualization.FACE_LANDMARKS
bboxesA list of bounding boxes for each detected face, with coordinates (x, y, width, height), label “face”, and confidence score. One list per input frame.BOUNDING_BOX
This documentation was AI-generated. If you find any errors or have suggestions for improvement, please feel free to contribute! Edit on GitHub

Source fingerprint (SHA-256): f60ed6201288a59d65d62cc98c12f227a353870c36decea8da81a063cfdf2bba