Blur Faces in Photos

About the Blur Faces in Photos

Add one photo or a whole folder from an event, and every face is found automatically and covered — blurred, pixelated, hidden behind an emoji or a solid box, as an oval or a rectangle. Detection runs on your device with Google’s MediaPipe face detector and two BlazeFace models: one for faces in ordinary photos (down to about 5% of the picture and 5 metres away, by its model card) and one for close-up selfies. Group photos are also scanned in overlapping tiles for small faces, and turned 90°, 180° and 270° for sideways and upside-down ones.

No detector is perfect, so you stay in control: review each photo, drag over anything that was missed to cover it, move or resize any box, or remove a wrong one. Adjust the extra area around faces and the strength, then download one photo or all of them as a ZIP. Every saved photo is built from new pixels, so it carries no EXIF data — no GPS location, camera or date. Nothing is uploaded.

Common uses

• Blurring bystanders in street photography and travel photos before posting.
• Hiding children’s faces in school, sports or party photos for a newsletter or social media.
• Anonymising people in photos for a dataset, a report or a public complaint.
• Covering faces of patients, clients or colleagues in workplace photos.

Which effect hides a face best?

• Solid box replaces the face with one colour; nothing of it remains in the saved file.
• Emoji covers the face with an emoji drawn over a strong blur, so nothing recognisable shows around its edges either.
• Pixelate averages the face into blocks — from 17 blocks across at strength 1 to 4 at strength 10.
• Blur applies a Gaussian blur whose radius grows with the face (σ from about 4% to 15% of its size).

Blur and pixelation hide a face from people, but researchers have shown that recognition software can sometimes still identify blurred and pixelated faces (McPherson, Shokri and Shmatikov, “Defeating Image Obfuscation with Deep Learning”, arXiv:1609.00408, 2016). When identities really must not be recoverable, use a solid box or an emoji with some extra area around the face.

How faces are found, and what can be missed

The photo is checked several times with the BlazeFace detectors (Bazarevsky et al., arXiv:1907.05047; models under the Apache 2.0 licence): whole, in overlapping tiles for small faces, and turned for sideways ones. Overlapping finds of the same face are merged, keeping the most confident. According to the models’ cards, faces should be roughly upright (tilted by at most about 45°), turned no more than sideways, and at least 70% inside the picture.

So faces can be missed when they are very small, seen from behind or in profile, heavily shadowed, blurred by motion, or partly covered — and now and then something that is not a face is marked. Always look over every photo before you share it. “Find more faces” under Detection lowers the confidence needed, which finds more faces but also more false matches.

Detection runs on your device’s processor: for these small models that is faster than the graphics processor (MediaPipe lists 2.94 ms on the CPU and 7.41 ms on the GPU of a Pixel 6 for the short-range model) and works the same in every browser.

Limitations

• Faces that are very small, seen from behind or the side, covered, or blurred by motion can be missed — check every photo and draw over anything left.
• Up to 100 photos of up to 80 MB each at a time; very large photos are saved at a reduced size within browser canvas limits, with a note.
• The emoji’s look depends on your device’s emoji font.
• Saved photos have no EXIF metadata at all, including the date taken and copyright fields; keep the originals if you need them.

Privacy

Everything happens in your browser. What you enter or open here is not uploaded or stored by MySmartCoPilot. The first time, the face detection engine (about 11 MB) and two face models (1.3 MB) are downloaded from this site; your photos are never sent anywhere.