BD MP Face Mask

BD MP Face Mask extracts per-region face masks from a single image using MediaPipe Face Landmarker landmarks - eyes, brows, lips, nose, ears, forehead, hair, the whole face, as separate masks. The headline numbers are the whole pitch: deterministic, CPU-only, roughly 5 ms per frame, no SAM3 prompts needed. If you need face region masks and you're currently typing "left eyebrow, right eyebrow" into a segmentation model, this node is dramatically cheaper - and it's a natural first stage for the pack's face pipeline.

The masks are drawn from the 468/478 landmark points rather than painted with a big blurry oval: eyebrows are enveloped from their contour, eyes use eroded eyelid regions, lips follow the organic mouth contour. That's the "shared tight drawing" the pack brags about, and it's what makes these masks usable for things like skin repainting where precision matters.

The outputs (there are a lot)

The node produces the full family: face_oval (jaw to hairline, no neck), skin (the oval minus eyes/brows/lips - your primary painting region), individual and combined left_eye/right_eye/eyes, left_brow/right_brow/brows, left_iris/right_iris/irises, lips, nose, left_ear/right_ear/ears, forehead, and hair. Note the left/right convention: "left eye" is the subject's left, which appears on the right side of a front-facing image. Plus status and bbox_json.

The inputs that matter

• head_mask - the one you'll reach for most. MediaPipe's face oval can miss bald heads, full chin edges, and unusual head shapes. Wire an external head silhouette (e.g. from SAM3) and it becomes the skin base instead of the oval; MediaPipe still runs for landmarks, it just stops deciding where the head ends.
• feature_expand - grows eye/brow/lip regions before they're subtracted from skin. Default 4; 3–6 px covers lashes and lip borders. Too low and skin bleeds into features; too high and you carve skin away.
• face_expand - expands the oval itself. +4 catches jaw-edge hair strands. Negative contracts.
• subtract_nose - exclude the nose region from skin too, if that's what your pipeline needs.
• iris_expand - fills out the 4-point iris ring into the coloured disc (3–6 px reads as a natural circle).
• bbox_feature + bbox_frame - emits the bounding box of a chosen feature as bbox_json, for feeding box-prompted SAM3 (e.g. SAM3_Detect).

The honest limits

These are geometric masks, not semantic ones. They know where an eyebrow is; they don't know whether the pixel at that location is actually eyebrow-colored. That's why the pack layers BD MP SAM3 Face Segment on top when it needs pixel-accurate results - and if you have a stylized brow that's offset from where MediaPipe's landmarks sit, the pure landmark version will be off. For the fast, deterministic, no-GPU baseline, though, this is the node. MediaPipe is also the license-clean play here (Apache 2.0), so it's the one that survives shipping.

Installing

Same as the pack - ComfyUI Manager search "BrainDead", or clone + pip install -r requirements.txt + restart. Requires the mediapipe package; the face_landmarker.task model auto-downloads on first run. All V3 API, so keep ComfyUI current.

Reach for this as the free first pass, then decide how much precision your actual output needs before pulling in SAM3.