AI Face Landmarker

What it maps

Upload a portrait and the page places 478 points across the face — jaw, brows, eyelids, nose, lips, and the irises — then connects them into a fine mesh. The result is drawn over the photo and can be downloaded as face-landmarks.png.

This is a geometry tool, not a recognition tool. It does not say who the person is, how they feel, or how old they are. It says where the facial features are, in pixel coordinates, on this one image.

The model and its settings

The one-face limit is a configuration choice, not a model limitation. If two people are in the frame, the page maps the first face the model returns and ignores the rest. That makes the output predictable for portraits and unsuitable for group analysis.

The blendshape and matrix outputs are worth mentioning because they exist in the result object: the model does compute expression coefficients and a 3D transform. The interface simply does not surface them. There is no expression readout, no head-pose value and no export of the numbers.

What the overlay means

Each landmark is drawn as a small green dot at the corresponding pixel, and the tesselation connections are stroked between them, producing the familiar wireframe. Because the dot radius is fixed at one and a half pixels and the line width at one, the mesh stays legible on a large photo and becomes dense on a small one.

The coordinates are normalised by the model and multiplied by the canvas dimensions, so the overlay lines up with the image at its native resolution. The downloaded PNG therefore has the same pixel dimensions as the source, with the mesh burned in.

The side panel reports 478 points and a short explanation. Nothing else is listed — no per-point table, no confidence values, no angles.

Step by step

1. Click the upload area and choose a photo. Drag and drop is not supported.
2. Wait for the runtime and model to download on first use, then for the mapping pass.
3. Review the mesh. A well-aligned mesh follows the jaw and brow closely; a misaligned one usually indicates a turned head or heavy occlusion.
4. Download the annotated PNG if you need it.

The Load AI Model button preloads the runtime so that the first photo maps immediately. After a successful load it shows "Model Ready" and stops accepting clicks.

Where the points are useful

• Face alignment. Landmark positions give the rotation and scale needed to straighten a portrait before cropping.
• Augmented reality prototypes. Glasses, masks and filters need anchors at the eyes, nose and jaw; this mesh supplies them.
• Animation reference. The mesh can be used as a starting point for rigging a stylised character.
• Photo retouching guidance. Seeing where the eye line and lip contour fall helps plan edits.
• Research and teaching. A visual demonstration of how a landmark model decomposes a face.

For measurement — distance between pupils, face width in millimetres — the tool is not enough on its own. Pixel distances need a known scale, and a single 2D photo does not provide one.

Where it goes wrong

Extreme angles are the main failure. A profile view, a face tilted far back, or a head partially outside the frame gives the model less to anchor on, and the mesh drifts or collapses toward the visible half.

Occlusion is the second. A hand, a mask, a microphone or hair across the face leaves gaps that the model fills by inference, which can bend the mesh in ways that look plausible but are wrong. Sunglasses are usually handled, because the eye region is inferred from the surrounding structure, but heavy shadow reduces accuracy.

Low resolution hurts: below roughly a hundred pixels across the face, the points become noisy. Very large images are fine, but the face still needs to occupy a reasonable portion of the frame. Cartoons, statues and paintings may produce a mesh at the wrong scale or no mesh at all.

Reading a mesh

The density of the mesh is a useful signal. On a well-lit, front-facing portrait the points sit tightly along the jaw and brow, the eye contours close cleanly, and the lip outline follows the mouth. When the head is turned, one side of the face compresses and the mesh crowds there, while the far side spreads out. When the model is guessing, points drift onto hair, background or the edge of a hand. Learning to spot that drift matters more than the raw count, because the count is always 478 whether the fit is good or bad.

What it does not do

No identity, no recognition, no age, no gender, no emotion labels, no liveness detection, no deepfake detection, no multi-face output, no video tracking, no 3D model export and no coordinate download. The mesh is a drawing. If you need the raw numbers, this page will not give them to you.

It is also not a medical or anthropometric instrument. Landmark positions are model estimates, not measurements, and should not be used for diagnosis, identification or any decision about a person.

Privacy

The photo never leaves the browser. It is decoded locally, passed to the WebAssembly model, and drawn onto a canvas; no request carries the image and nothing is stored. The page downloads the runtime and model from public CDNs on first use, which reveals that the tool was opened but not what was processed.

That said, facial geometry is sensitive data. If you are handling images of other people, local processing solves the transmission question but not the consent question.

Limits

One face per image, one image per run, no batch, no webcam, no video, no confidence threshold, no expression readout and no numeric export. Annotated PNG only. Very large photos take longer to decode, and the mesh density can make small images look cluttered. A modern browser with WebAssembly and network access on first load is required.

Common mistakes

• Uploading a group photo. Only the first face is mapped.
• Expecting emotion or identity. Neither is part of the model's output here.
• Measuring in pixels and calling it millimetres. Without a scale reference, the numbers are relative.
• Trusting a mesh on a profile shot. The points drift when half the face is hidden.
• Looking for a coordinate export. The only output is the annotated image.
• Using it on children's photos without consent. Local processing does not settle that question.

Frequently Asked Questions