Face Landmarks¶

Maps the face with 468 points (eyes, brows, nose, lips, jawline, cheeks). Those points are what the warp nodes use to reshape a face.

Face Landmarks — before / after


In ComfyUI	`Add Node → ComfyUI-JH-PixelPro → face → Face Landmarks (MediaPipe 468)`
Good for	Feeding Face Warp or Face Warp TPS; checking how well the mesh fits a face

When to use it¶

Five points are enough to straighten a face, but not to reshape one. To slim a jaw, lift an eye or copy an expression from another photo, you need a dense map. A landmark is a numbered point on the face, and the same number always sits on the same spot (for example the tip of the nose), so you can compare a face in two photos point by point.

Run the node once on your photo for the "source" points, and once on a reference or edited version for the "destination" points. Feed both into a warp node.

Quick start¶

Connect Load Image to Face Landmarks (MediaPipe 468).
Leave draw_overlay on and look at overlay in a Preview Image to confirm the dots sit on the face.
Connect landmarks to src_landmarks of Face Warp TPS (or Face Warp).
Get dst_landmarks from a second Face Landmarks node run on the target picture.

Settings¶

Setting	Default	What it does	Tip
`max_num_faces`	`1` (1–10)	Most faces to find per image. Missing faces are filled with NaN (no data).	Keep 1 for portraits. The warp nodes use the first face.
`min_detection_confidence`	`0.5` (0.0–1.0)	How sure MediaPipe must be before a face counts.	Lower it if a real face is missed.
`refine_landmarks`	on	Kept for compatibility. The dense 468-point output is always used.	Leave as is.
`draw_overlay`	on	Draws the points as green dots on the `overlay` image.	Turn off if you do not need the preview.

Outputs:

Output	Type	Meaning
`landmarks`	`LANDMARKS`	The 468 points per face, in normalised `[0, 1]` image coordinates. Connect only to nodes that take `LANDMARKS`.
`overlay`	`IMAGE`	Your photo with the points drawn on.

Tips from retouching¶

LANDMARKS is a special type. It does not connect to landmarks_json of Facial Aligner; use a face detect node for that.
Run the node on the same resolution as the image you will warp, or at least the same framing, so points line up.
If a face is not found, its slot holds NaN. Warp nodes then return the image unchanged.
Detection noise is real: two runs on slightly different crops disagree by a few pixels. For this reason Face Warp TPS has a smoothness setting.

Under the hood

The node uses the MediaPipe FaceLandmarker (tasks API) and shares its model file with Face Detect. The result is a (batch, faces, 468, 2) float tensor, NaN-padded when fewer faces are found. The image is converted to 8-bit RGB for MediaPipe. The code lives in core/face_pipeline.py.

Face Warp — Delaunay per-triangle warp.
Face Warp TPS — smooth thin-plate-spline warp.
Face Detect — the 5-point version for alignment.
Face category overview

Example workflow¶

Face pipeline example workflow (Face Landmarks → Face Warp → Face Beauty Blend). Copy workflows/sample_portrait.jpg into ComfyUI/input/ first.