MediaPipe Face Mesh Eyes Only

Sometimes the most useful node in a pack is the one that does one tiny thing well. Runware_mediapipe_face_mesh_eyes_only detects faces in an image, runs MediaPipe's face mesh landmarking, and returns a mask isolating just the eyes - with controls for how tight, blurred, and padded that mask is. It's a preprocessing utility (runware:35@9, task type imageMasking), and it exists so you don't have to install MediaPipe, wire up face-landmark nodes, and hand-tune the geometry yourself.

Where does this slot into a real workflow? An eyes-only mask is the classic input for eye-focused inpainting or img2img - regenerating just the eyes (the part of a face everyone looks at first) without touching skin, hair, or mouth. It's also handy for eye color edits, gaze adjustments, and any pipeline where you want a precise face-region mask and the local MediaPipe stack is more overhead than you want. Because it runs on Runware's cloud, there's nothing to install beyond the pack itself.

Inputs and controls

• image (required) - the face image, as an IMAGE socket.
• settings.confidence - detection threshold, 0–1, default 0.5. Raise it to reject weak detections.
• settings.maxDetections - 1–20, default 6. More faces in frame = consider raising it; fewer, tighter = lower it so the strongest faces win.
• settings.maskBlur - 0–100, default 5. Blur radius on the mask edges; a few pixels stops hard, crunchy boundaries when you inpaint.
• settings.maskPadding - 0–200, default 10. Extend (or with a negative, shrink) the masked area. The tooltip calls out the negative range - padding isn't just "more," it's also "tighter."

The image output is a mask in IMAGE form - the classic white-on-black segmentation output you'd feed into an inpainting node or a mask-aware generation.

What's worth knowing

• It's a mask, not an edit. The output is the segmentation, not a "fixed eyes" result. The node's job ends where the eyes-only region is isolated; what you do with that region is the next node's job.
• This is the package-deal version of a local task. If you already run Impact Pack + face-detection locally, this duplicates capability. If you don't, it's a one-node replacement for a whole dependency tree.
• Toggles and tooltips are your friends - confidence, maxDetections, blur, and padding are all settings under the pack's standard pattern, and the tooltips spell out each one's behavior precisely.
• Face size matters. A tiny face in a big group shot will be under the confidence threshold; crop or raise... no, lower the threshold, and watch maxDetections.

It's not glamorous, and that's the compliment. When you need an eyes-only mask without building a MediaPipe pipeline from scratch, this is the node you'll quietly reach for every time.