BD MP Face Refine

MediaPipe face masks have crisp, well-placed hulls but chunky boundaries; SAM3 masks have pixel-perfect boundaries but need prompting to know what they're segmenting. BD MP Face Refine is the node that marries them: it matches each MediaPipe feature mask (eye, brow, lips, nose) to the best-overlapping SAM3 segment by IoU, intersects them - SAM3 supplies the pixel-accuracy, MediaPipe prevents bleed - and then computes skin as face_oval − refined features. The result is a set of face feature masks with MediaPipe's placement and SAM3's edges.

How it works

Each feature's MediaPipe hull is dilated by target_expand (4 px) to help find its SAM3 candidate, then matched by IoU against the candidate batch with a min_iou threshold (default 0.05 - deliberately permissive; raise to 0.15+ if the wrong segments keep matching). The matched segment is intersected with the hull, and clip_expand (4) controls how far outside the raw hull the SAM3 segment is allowed to extend. feature_expand (0) dilates the refined masks before they're subtracted from skin, which creates a useful edge buffer. exclusive (on) means each SAM3 candidate can only be matched to one feature - highest IoU wins, in a fixed matching order (lips → eyes → brows → nose).

The two required inputs are candidates (the SAM3 segment batch - each frame is one candidate) and face_oval (MediaPipe's inner face-plate mask, the skin boundary). Wire silhouette_mask (a head outline, e.g. from BD_SAM3MultiPrompt) to clip all outputs to the head boundary, and image to get composited debug/masked outputs with a chosen background (masked_image_bg: transparent/white/black/checker).

Outputs: skin, per-feature masks (left_eye, right_eye, left_brow, right_brow, lips, nose), debug_overlay, match_info, masked_image, masked_skin, and head_mask. That's a full face-segmentation kit in one run.

Installing

Ships in ComfyUI-BrainDead. ComfyUI Manager → search "BrainDead" → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/BizaNator/ComfyUI-BrainDead
cd ComfyUI-BrainDead
pip install -r requirements.txt

Restart. You'll also need a working SAM3 setup upstream (this pack's BD SAM3 Multi-Prompt can provide the candidate batch and auto-downloads the model).

Where people get burned

Two knobs do most of the damage. min_iou too low (default) can match a candidate that's mostly the background if your SAM3 batch is noisy - raise it and watch match_info. And face_oval is the inner face plate while silhouette_mask is the outer head boundary - they're different masks and wiring the wrong one in will either leave hair in your skin or clip the face away. The note on max_target_fill (0.95) matters too: it rejects the "SAM3 returned a full-white not-found result" case that would otherwise swallow your skin mask.