Blur Faces in a Video, Free nothing stored

Two passes: find every face first, then redraw the clip

The tool reads your clip twice. The first pass decodes a copy no wider than 1280 pixels on its long side and searches it about 15 times per second of video. Each face it already knows is looked for again in a close crop around its last position, where even a modest face fills much of the picture. New faces are looked for one tile at a time, across the whole frame, a 2 × 2 grid and a 3 × 3 grid. The first frame, and the first frame after a cut, get every tile at once.

The second pass draws each frame at the output size and covers every face listed for that moment. A frame that falls between two searches gets the covers of both, which spans the path the face took. If the first pass finds no face at all, the run stops there with a message instead of encoding a clip that would look finished.

When a head turns, a hand passes or a face is found late

A face detector misses faces for a moment when a head turns, a hand passes in front or motion smears the picture. Here a missed face keeps its cover for 1 second, and the cover grows while it waits, in case the face kept moving. When the face is found again inside that second, the cover simply carries on.

The other direction matters too. A face that walks in from the edge is often found a little after it appears. So a newly found face is followed backwards through the last second of searched frames, and its cover starts up to 1 second before the moment it was found. The price of both rules is cosmetic: an oval can stay on screen for about a second after someone has left, or appear a moment before they arrive.

Scene cuts, posters and other face-like shapes

Every searched frame is shrunk to a tiny grey thumbnail and compared with the one before it. A sudden jump in difference counts as a cut: the new shot gets a full search straight away, and no cover is held over or started early across that point. A steady fast pan changes the picture just as much from step to step, so it is not mistaken for a string of cuts.

A face must be found in at least 3 searches, about a fifth of a second of video, and in at least half of the searches across its run before it gets a cover. A pattern that only looks like a face now and then is dropped this way; in testing, a cat on an armchair was. A photo of a person on a wall, a poster or a screen is a real face to the detector, though, and can be covered for a while too.

Distant faces, side profiles and tilted heads

Faces narrower than about 6–7% of the frame width are not found. On a wide shot that means people at the back of a hall, a stand or a street can stay visible, so the cover only protects faces the camera sees reasonably close. A face seen fully from the side, looking down, or half hidden by a hat or a mask is found less often, and leans on the one-second hold.

The covers are upright ovals that reach from the hairline to just under the chin. A head tilted to one side is not followed by the oval’s angle, so its chin or crown can reach past the edge. Cover beyond the face adds room around every face, starting at 25%; raise it for tilted heads, hair and ears. There is no brush or box here to add a face the tool missed.

Blur, mosaic or black oval on a subject in motion

The Effect choices are Blur, Pixelate or Black oval, starting on Blur. Strength scales with each face, so a face far away and a face close to the lens are hidden to the same degree: at the starting value of 60, the blur reaches about a tenth of the face’s width and the mosaic is about seven blocks across. The black oval ignores strength and replaces everything inside it.

Every effect is fully opaque out to the edge of the oval, and only a thin band outside it fades back into the picture. A blur or mosaic still keeps rough colors and shapes inside the oval. When the person in the clip could be at risk, choose the black oval.

Clip length, saved frame size and the sound track

One run takes up to 3 minutes of video on a computer and up to 1 minute on a phone or tablet, because every frame is decoded twice, searched and encoded again in the tab. Files are also checked against a size budget of 4 GB on a computer and 1 GB on a phone. For a longer recording, cut the part you need with the free Video Trimmer, which does not re-encode, then blur that.

The saved video is at most 1920 pixels on its long side on a computer and 1280 on a phone, never larger than the source. It is H.264 in an MP4 where the browser has that encoder, otherwise VP9 or VP8 in a WebM, and a phone clip stored sideways is drawn upright first so every cover lands on the face the player shows. The sound is copied, or converted to AAC for the MP4 when it is in another format. Under the result, Faces covered, Frames processed, Frame and Sound show what the run did.

Replacing a person or sharpening the footage instead

This tool keeps people in the shot and hides their faces. If you would rather show someone else in their place, the paid Viral Character Swapper puts a character from your image into the video and keeps the original motion, and Imagera Recast replaces the people in a short clip with your photos.

Blurring does not improve the picture, and the output stops at 1920 pixels. If the finished clip looks soft, the paid 4K Video Enhancer sharpens and upscales it in the cloud. Blur here first, then enhance the blurred file and check the faces again before posting.

How do you use the video face blur?

1. Open the clip

   Choose a MP4, MOV, WebM or MKV video of up to 3 minutes, or up to 1 minute on a phone.

2. Set the cover

   Pick Blur, Pixelate or Black oval, set Strength, and raise Cover beyond the face to hide hair and ears too.

3. Let both passes run

   Start it and wait while faces are searched first, then every frame is redrawn and encoded.

4. Watch it through

   Play the whole result, pausing where people turn, enter or stand far back, and read Faces covered and Sound.

5. Save your result

   2 free downloads, 8 more after sign-in.

When is the free video face blur enough?

Video Face Blur: questions and answers