BD MP SAM3 Face Segment

BD MP SAM3 Face Segment is the heavy hitter of this pack's face pipeline, and the name tells you exactly how it works: MediaPipe localizes each facial feature, then SAM3 segments that feature with pixel accuracy. MediaPipe gives the location; SAM3 gives the fidelity. The thing it's built to fix - stated right in the node description - is the stylized-brow offset: on anime and stylized renders, the painted eyebrow sits somewhere other than where MediaPipe's landmarks are, and a landmark-only band misses it entirely. This node lets SAM3 grow the brow to what's actually painted.

How it works

For each feature (brows, eyes, lips, nose), the node builds a SAM3 prompt from the MediaPipe localization: a bounding box, positive landmark points, and sibling negatives. So SAM3 knows roughly where to look, what is the feature, and what isn't. Then the SAM3 mask is clipped to a "bleed guard" - a dilated version of the MediaPipe zone - so precision can't wander off into the forehead. The guards are per-feature for a reason: brows need a large guard (~40–45) to grow out to the painted shape, while lips need a small one (~12) or they overflow onto the face.

The best part for setup: no wiring needed. The node auto-loads and auto-downloads the official SAM3 checkpoint in-house (you can wire a comfy-core SAM3 MODEL to override, but you don't have to). It also auto-downloads MediaPipe's model. For a pack whose other segmentation nodes assume you already have SAM3 loaded, this one is refreshingly turnkey.

The inputs that matter

• do_brows / do_eyes / do_lips / do_nose - toggle which features SAM3 refines. Brows is the one you bought this for; the others default on too.
• bleed_guard / brow_bleed_guard / eye_bleed_guard / lips_bleed_guard - the per-feature clipping. If brows read thin, raise brow_bleed_guard; if lips bleed onto the face, lower lips_bleed_guard. These are the tuning knobs you'll actually turn.
• refine_iterations - defaults to 1, and the tooltip's warning is worth quoting: extra passes tend to shrink masks on stylized/non-photo renders. Raise only if SAM3 over-segments.
• edge_refine - optional edge snapping after cleanup: off / guided (fast, edge-aware) / matting (PyMatting, CPU) / vitmatte (deep matting, best on soft/hair edges, auto-downloads the model). On the ROI crop, so it stays cheap.
• remove_background - computes a clean head silhouette in-house (text-grounded SAM3, head prompts minus neck/clothing negatives), optionally with neck_cut using an auto-downloaded SegFormer face parser for a real neck-removal instead of a jawline cut. When on, everything clips to that silhouette and it's emitted as the silhouette output.
• skin_color_filter - refines the skin output by colour (adaptive LAB or fixed HSV): drops non-skin pixels like a dark mustache from the skin mask. drop_facial_hair punches uncolored facial hair out of the clean cutout for repainting.
• drop_eyes / drop_brows / drop_lips / drop_nose - control which features get punched out of the head_cutout_clean image output, leaving holes for BD MP Face Infill to fill. Nose defaults to kept (it's usually skin).

Outputs

You get the full mask family (face_oval, skin, eyes, brows, irises, lips, nose, ears, forehead, hair) matching BD MP Face Mask's layout, plus head_mask, masked_skin, the silhouette, a debug_overlay, and two image outputs - head_cutout (head on transparent/black/white) and head_cutout_clean (features punched out, ready for infill). The rgba output is a zone map: R=lips, G=brows, B=eyes, A=face_oval.

Installing

Standard pack install - ComfyUI Manager search "BrainDead" or clone + pip install -r requirements.txt. Needs mediapipe, and it will fetch the SAM3 checkpoint and VitMatte model on first use. All V3 API.

The catch, honestly: this is a lot of machinery and a few auto-downloads. For a photoreal face you might get 90% of the way with BD MP Face Mask alone. The moment you're on stylized art and the brow or lip masks come back wrong, this node is the upgrade - and it's the reason the pack's FaceWrap and Face Segmentation templates exist.