Find faces runs the YuNet face detection model from OpenCV inside your browser; it is downloaded once, about 1 MB, only when you press the button. Blur is decorative: a 2016 study recognized 57.75% of faces blurred the way YouTube did it, so the default solid block is the mode to use when a person must stay unidentified.
Why cover faces before sharing
Most of the time a face in a photo is harmless. Sometimes it is not. The person in the background of your vacation shot never agreed to be posted. A coworker at an event may not want their employer's name tied to a party picture. People at a protest, a support meeting or a shelter can face real consequences if they are recognized. A witness in a video of an incident, a customer in a shop photo, or a stranger in a viral clip did not choose to be part of it.
Face recognition is now cheap and widely available, which changes the math. A face that a human would never place can be matched by software against public profile photos. Covering the faces you do not need is a small step that removes that possibility from your picture.
How to cover faces
- Pick the photo using the file button, or drag it into the box at the top.
- Press Find faces. The first time, your browser downloads the face model and its runtime from this site (about 1 MB; the size is shown next to the button before anything starts). After that, detection runs on your device.
- Look at every box the model drew. Drag in any face it missed (turned heads, sunglasses, small faces at the back), move boxes that sit too tight, and delete boxes on faces you want to keep visible.
- Leave Cover with on Solid block for anyone who must stay unidentified. Pick Blur only for a softer look where recognition does not matter.
- Press Export and save the new copy from the download link or the share button.
How the automatic finder works, and where it fails
The finder is YuNet, a compact face detection model published in the OpenCV model zoo under the MIT License. It runs through TensorFlow.js (its WebAssembly backend) on your processor, so it does not need graphics acceleration and still works in browsers where WebGL is switched off. YuNet was trained on faces roughly 10 to 300 pixels across in the image it is given, which is a good fit for group photos and street scenes, but very tiny faces at the back of a large crowd can still slip through.
Expect misses in these situations and check for them yourself:
- faces in profile, looking down, or turned mostly away;
- faces partly hidden by hands, hats, masks, helmets, hair or face paint;
- tiny faces at the back of a crowd or in a wide landscape shot;
- motion blur, deep shadow or strong backlight;
- faces that are not "live": photos on a wall, a poster, a TV screen, a reflection in a window or a mirror;
- drawings, statues and pets, which may be boxed by mistake.
The finder only proposes boxes. You decide what gets covered, and the export uses exactly the boxes on screen when you press the button.
A face is not the only thing that identifies someone
Covering the face and leaving everything else can still give a person away. While the photo is open, look for:
- name badges, lanyards and ID cards;
- distinctive tattoos, birthmarks or hair;
- team jerseys with names and numbers, school or company logos;
- a car with its plate visible, a house number, a street sign;
- other photos that show the same person uncovered.
All of these can be boxed the same way. For plates, see cover a license plate.
Blur versus solid: what the research says
Blur spreads each pixel into its neighbors. It looks much more natural than a black box, which is why video sites and news outlets like it. It also leaves a lot of information in place.
In 2016, researchers at Cornell Tech and the University of Texas showed that an ordinary image-recognition network, trained on blurred examples, picked the right person in 57.75% of faces blurred with YouTube's face-blur tool. The set had 40 people, so chance alone would score 2.5%. Distortion effects fare no better: in 2007 a suspect's face hidden behind a digital swirl in photos posted online was reconstructed by a German police unit, and the restored picture led to his identification within days of a public appeal.
Your photo will rarely face that kind of effort. But you cannot know in advance who will look at it, so the page defaults to a solid block, which keeps none of the original pixels.
What you get back
Export builds a new image from the canvas, so:
- covered faces contain only the black fill or the blur, not the original pixels;
- the hidden data that cameras and phones write into photos does not come along: EXIF (camera, date), GPS position, XMP and IPTC fields, maker notes, thumbnails and Content Credentials;
- the picture keeps its original size. Finding faces and drawing the preview use a scaled-down copy no wider or taller than roughly 2,048 pixels, and your boxes are stretched back to full size for the export.
If a photo is too big for the browser to encode in one piece, you get a copy whose longer side is 4,096 pixels, plus a message explaining why.
Supported photos and devices
- JPEG, PNG and WebP open in all current browsers on computers and phones.
- HEIC opens only where the browser can display it on its own, which in practice means Safari. Elsewhere you are asked to convert to JPG first.
- One photo is edited at a time; you can choose several and work through them.
- Face finding needs a browser with WebAssembly, which every current browser has. On an older phone the first run may take a few seconds while the model loads.
Where your photo goes
Nowhere. The photo is read from your device into this tab, the face model runs in the tab, and the finished copy is created in the tab's memory for you to save. The model arrives from this site on your first press of Find faces; the photo is not part of that request or any other. Closing the tab clears everything the page made.
How the pieces fit together
The face finder is TensorFlow.js with its WebAssembly backend running the YuNet model file, both served from this site at fixed versions. The editor keeps box coordinates relative to the original photo and draws the final image when you export. Open your browser's developer tools and the Network panel lists the model download and nothing carrying your photo; the methodology page explains how to watch for that yourself.
FAQ
Is blurring a face enough to hide who it is?
Not reliably. In a 2016 test, a recognition network matched 57.75% of faces blurred with YouTube's tool, against 2.5% by chance. Blur is fine for looks; use Solid block when a person must not be identified.
Why did the tool miss a face?
The model works best on faces pointed roughly toward the camera. Profiles, hats, masks, very small faces and motion blur cause misses. Draw a box over any face the finder skipped before you export.
Does the face model upload my photo?
No. The model file is downloaded to your browser, and the detection runs there. Your photo is not sent with that download or afterwards.
How big is the download?
About 1 MB, shown next to the Find faces button before you press it. It covers the TensorFlow.js WebAssembly engine and the YuNet model, fetched once from this site.
Can I cover only some of the faces?
Yes. Delete the boxes on faces you want to keep, and add boxes anywhere the finder missed. Only the boxes on screen are applied.
Does the exported photo keep the location and camera details?
No. The new file is encoded from scratch, so EXIF, GPS, XMP, IPTC and Content Credentials are not carried over.
Can I do this on a phone?
Yes. Open the page in your phone's browser, tap Find faces, adjust the boxes with your finger, and use the share button to save the result.
Sources
Checked 2026-10-11.
- Defeating Image Obfuscation with Deep Learning (arXiv 1609.00408, 2016): 57.75% recognition of YouTube-blurred faces on a 40-person set.
- OpenCV Zoo: YuNet face detection: model description, the 10 to 300 pixel face range, MIT License.
- ABC News, October 19, 2007: a digitally swirled face reconstructed by a German police unit.
Last updated: