AI Face Detector

What it finds

Upload a photo and the page draws a green rectangle around every face the model can see, then reports how many it found. The annotated photo can be downloaded as detected-faces.png; if the count is zero, the download button never appears.

That is the whole feature set. No names, no identities, no emotions, no ages, no landmarks and no matching against a database. It is a detector: it locates faces, and nothing more.

The model, and why "short range" matters

BlazeFace short-range is designed for faces that are large in frame — selfie distance, video-call framing, a portrait filling most of the picture. It is fast and reliable in that setting. It is not a crowd-counting model, and it is not designed to pick out small faces in the background of a wide shot.

This single design choice explains most of the disappointing results people get. A group photo taken from across a room, a stadium shot, or a street scene with distant pedestrians will under-count, sometimes dramatically. A head-and-shoulders portrait will be found immediately.

What the overlay shows

The canvas is set to the image's natural size, so the export keeps full resolution. Detected faces get a two-pixel green rectangle. There are no labels, no confidence percentages and no keypoints drawn on the image, which keeps the output clean enough to reuse as a visual aid.

The side panel shows a large count — "3 faces detected" — and nothing else. Because the threshold is the model's default rather than something you can adjust, borderline faces either appear or do not, with no slider to explore.

Using the page

1. Click the upload area and pick an image. Drag and drop is not supported.
2. The first image triggers the runtime and model download; the panel reports each step.
3. Detection runs automatically after loading, with no button to press.
4. Check the count and the boxes. If a face is missing, the cause is usually size or angle rather than a bug.
5. Press Download to save the annotated image.

The Load AI Model button preloads everything if you want to avoid waiting on the first photo. Once loaded, it reports "Model Ready" and further images are near-instant.

Detection is not recognition

It is worth being precise about the difference, because the two are often confused in conversation and in law.

• Detection answers "is there a face here, and where?" — what this page does.
• Recognition answers "whose face is this?" — requires a database and is not part of this tool.
• Analysis answers "what is this person feeling or how old are they?" — also absent, and unreliable in general.

Nothing is stored, compared or transmitted. The output is a rectangle on a picture.

Good uses

• Counting people in a photo for captions or alt text, where approximate numbers are acceptable.
• Checking framing before publishing a portrait — a box that clips the chin tells you the crop is wrong.
• Planning a blur or redaction pass. The boxes show where to apply it, though the tool does not blur for you.
• Testing a pipeline. Quickly confirm that a set of images contains detectable faces before running heavier processing.
• Teaching. A clear visual demonstration of what a face detector actually sees.

Where it fails

Small faces are the big one. So are profile views, faces turned sharply away, heavy shadow, masks covering most of the face, and faces cut off by the frame edge. Sunglasses are usually fine; a full-face covering is not.

Low-resolution images lose the detail the model needs. Very large images are fine in principle, but a face occupying fifty pixels in a 6000-pixel-wide photo is effectively a small face. Cropping to the subject before uploading is the simplest fix.

Cartoons, statues, mannequins and posters can produce false positives; the model looks for facial patterns, not for living people. Conversely, an unusual angle on a real person can produce a false negative. Neither outcome is an error in the usual sense — it is the model doing its job at the edge of its training.

Counting people in practice

If a number matters, treat the count as a lower bound and verify it by eye. Crop the frame to the group and run it again: a tighter crop makes each face larger and often reveals one or two the wide shot missed. Repeat until the count stops changing. For anything formal — event attendance, capacity checks, editorial captions — a manual count of the boxes is still the reliable method, and the overlay makes that quick because every detected face is already marked.

Consent and handling

Because processing is local, there is no upload to worry about. That does not remove the human questions. Photographs of identifiable people are personal data in many jurisdictions, and using a detector to count or locate faces in someone else's image can still require a lawful basis. The tool gives you boxes; it does not give you permission.

If you are redacting, remember that the downloaded PNG still shows the faces under green rectangles. It is an aid for planning, not a redacted output.

Privacy

Everything happens in the browser: the file is read locally, decoded to a bitmap, and passed to a WebAssembly model. No image, box or count is sent to a server, and nothing is retained after the page is closed. The only network traffic is downloading the library and model weights from public CDNs on first use.

Limits

One image per run, no webcam, no video, no batch, no confidence threshold control, no keypoints, no landmarks, no recognition, no emotion or age estimation, no export of coordinates and no blurring. The annotated PNG is the only output. Large images take longer to decode, and the page requires a modern browser with WebAssembly plus network access on first load.

Common mistakes

• Using a group photo from a distance. The short-range model is built for nearby, large faces.
• Expecting identities. Detection only; there is no recognition.
• Looking for confidence scores. The page shows boxes and a count, not per-face numbers.
• Assuming a missed face is a bug. Profile, shadow, occlusion and size all cause misses.
• Treating the annotated file as redacted. The faces are still visible underneath.
• Forgetting consent. Local processing does not change the ethics of handling someone's photo.

Frequently Asked Questions