Automatic, but you have the last word
The face finder is YuNet, a compact open detector from the OpenCV model zoo. It returns a box for each face; the page turns each box into an ellipse with the padding you choose, so hair line and jaw are covered too.
Detectors miss faces in profile, in deep shadow or behind sunglasses. Every found face can be switched off, moved or resized, and you can draw extra ellipses for anything it missed. A “Look harder” option lowers the detection threshold for crowded photos.
Block size that actually hides a face
A face pixelated into 20 × 20 tiny blocks is still a face you can recognise. With safety on, each face gets at most ten blocks across its narrow side, whatever the block size slider says — the region list shows when that floor kicks in.
Sharing on WhatsApp or Instagram
Messaging apps recompress photos. Large blocks survive recompression; fine blur and small blocks can smear into something that looks less hidden than it is. Prefer square pixels or hex mosaic at the default floor when the picture will be forwarded.
Questions
How does automatic face detection work here?
When you press “Find faces”, the page downloads the YuNet face detector (224 KB) from our model file server once, caches it in your browser, and runs it on your device with ONNX Runtime Web. The photo itself is never sent anywhere.
It missed a face. What can I do?
Try “Look harder”, which lowers the score threshold, or draw an ellipse over the face yourself with the ellipse tool. Side views and small background faces are the usual misses.
Can I pixelate some faces and keep others?
Yes. Each found face appears in the region list with its own switch. Turn off the ones you want to keep clear.
Is a pixelated face safe from being reversed?
Not from the file you download. The export is a new picture: only the averaged colours are written, the original pixels inside a region are gone, and no layers or edit history are saved. With safety on, each block also gets a small random colour offset and the grid starts at a random position, which defeats the pattern-matching that tools like Depix use on pixelated text.
Does it remove location data?
Yes. The downloaded file is re-encoded from the edited pixels and carries no EXIF block, so GPS coordinates, camera model and capture time are not in it.