Blur faces automatically, without uploading the photo
Most “blur face” apps send your picture to a server to find the faces. This one doesn’t. When you open a photo, a face detector runs inside your browser: the UltraFace model (MIT license, about 1.2 MB) on ONNX Runtime WebAssembly. The model and the runtime are fetched from this site the first time you use the tool (about 15 MB in all before compression, then cached by the browser); your photo stays on your phone or computer the whole time.
The model looks at the image 320 × 240 pixels at a time, so the tool scans your photo in several passes: once as a whole for big faces, then in overlapping tiles for the small faces of a group photo or a crowd. Each face it finds gets a box on the preview and is hidden straight away. The count appears in the side panel, for example “3 faces found”.
Detection is good but not perfect. Faces seen from the side, partly covered, very small or in the dark can be missed, and once in a while a pattern is taken for a face. Always look at the preview before you share the result.
Fix the result: skip faces, add the missed ones
- Keep a face visible: click its box. The outline turns dashed and the face is shown untouched. Click again to hide it again.
- Cover a missed face, a name tag or a license plate: drag on the preview to draw a box. It works with a mouse, a pen or a finger. Drawn boxes have a ✕ button to remove them, and “Clear drawn areas” removes them all.
- Shape: an ellipse follows the shape of a head; a rectangle covers more of the hair and background.
Blur, pixelate or black out: which is safest?
| Effect | Looks | How safe |
|---|---|---|
| Blur | Soft, natural | Safe at high strength; a light blur can leave a face recognizable |
| Pixelate | Square blocks | Safe with few, large blocks; fine blocks can be partly reconstructed |
| Solid black | Black shape | Nothing of the face is left in the file |
Strength is measured against each face’s size, so a small face at the back of a group gets the same protection as a close-up. At the default of 85%, a pixelated face is about seven blocks wide. Researchers have shown that software can sometimes recognize faces hidden with a light blur or small blocks (for example McPherson, Shokri and Shmatikov, “Defeating Image Obfuscation with Deep Learning”, 2016), which is why the default is strong. For anything sensitive — a child, a witness, someone at risk — use solid black. For a whole picture or any other area, see Blur Image and Pixelate Image. For more ways to do it, read how to blur a face in a photo.
What you download
The censored photo is saved in the format you opened when the browser can write it: JPG stays JPG (quality 92%), PNG stays PNG and WebP stays WebP (PNG in browsers that can’t save WebP, such as Safari). HEIC and AVIF photos are saved as JPG, GIF and BMP as PNG. The full resolution is kept; only the preview is shown smaller.
Because the file is rebuilt from pixels, hidden metadata such as GPS location, date and camera model is not carried over. To remove metadata without changing the picture, use Remove EXIF Data, or see the guide on how to remove EXIF data from photos.
Privacy
Detection and blurring both happen on your device, with the Canvas API and WebAssembly. Nothing is uploaded, the page keeps working offline once loaded, and the site’s security policy only lets it talk to its own server.