How to blur part of a video

Choose a cover, follow the detail and check the file you will actually share.

Free toolVideo Blur & Privacy MasksCover faces, names or screens with timed areas. Track a selected object locally, correct its path, then save and review the video.Blur part of a video

Cover every appearance, then watch the saved result

Place a region over the detail, set when it appears and check its full path through the clip. Use an opaque cover for readable private text. A single exposed frame can still be paused.

Find every appearance, including later notifications, reflections and repeated screens.
Choose a cover with a margin. Set its start slightly before the detail appears and its end after it leaves.
Follow movement manually or review an automatic tracker’s output. Both need checking.
Export a new file. Watch that file from beginning to end and listen for spoken information.

Choose solid cover, blur or pixelation

Use a solid cover when you need to replace the selected pixels completely. Blur and pixelation change their appearance, but can retain useful clues.

Effect	Useful for	What to check
Solid cover	Names, account details and readable screen text.	The complete text, its edges and every later appearance must stay inside the rectangle.
Blur	Softening a selected visual detail.	Letter shapes, colours and surrounding context can remain recognisable.
Pixelation	A block-like visual cover for a face or object.	Small blocks may retain facial or text structure. Increasing blocks does not remove context.
Cut the interval	A short segment you can remove entirely.	The detail must not appear elsewhere, and the cut should still make sense.

No visual effect proves someone is anonymous. Clothing, a voice, a badge or a location can identify them. For an active exposed credential, covering a recording does not revoke the credential.

Follow a moving face, plate or screen panel

Manual movement points work when you can inspect the subject’s path. In the video privacy studio, select Move between points, seek to another moment and reposition the rectangle.

A moving face may turn or change size. Add points before a turn, after a turn and wherever a straight path misses it.
A number plate can tilt, shrink or disappear behind something. The rectangular cover must still enclose the full plate.
A screen panel may jump instantly. Jump between points holds each position until the next one, rather than sliding across the screen.

For automatic movement in floi, draw a box around one subject on a clear frame. Choose Follow automatically in Movement, then press Go to area start. Check the box still fits the subject at that frame, then press Track this area. Review its whole path, correct any drift and use Track from here after a correction. Tracking runs locally. It does not automatically discover every face or plate.

Track a whole object when you can hide the surrounding area. In our fictional workspace example, a narrow email-only selection drifted upwards and exposed part of the line after 3.25 seconds. A high model score did not flag it. Use a wider panel cover or manual movement points for narrow screen text.

Other editors offer different tracking controls. Adobe documents masking and tracking in Premiere Pro. YouTube Studio offers face and custom blur for videos on YouTube. Uploading to YouTube is a different privacy decision from opening a local file.

Choose a tracking model from measured results

We compared three browser-side models on 1 November 2026. The public OTB David and Faceocc2 sequences contain 471 and 812 tested frames. Testing used Chrome 156 on an Apple M1 Max with 32 GB memory.

The table measures predicted rectangles against the supplied reference boxes. Overlap uses intersection over union: 1 means identical rectangles and 0 means no overlap. It does not measure how reliably a privacy cover hides a person.

Model and download	David mean overlap	Faceocc2 mean overlap	Processing per frame
NanoTrack v3, 2.6 MB	0.7720	0.7156	12.9 ms
NanoTrack v2, 1.8 MB	0.7491	0.7037	9.5 ms
ViTTrack, 0.7 MB	0.7099	0.5306	7.5 ms

NanoTrack v3 is the default because its average overlap was highest across these two sequences. NanoTrack v2 was faster and did better on our separate synthetic distractor clip. ViTTrack had the smallest download and shortest processing time. No model won every comparison.

Processing time includes decoding, cropping and inference after the model sessions were ready. It excludes download and model preparation. All models also need the shared 11.8 MB runtime on first use. These results describe this implementation and machine. They are not model-wide rankings or phone speed estimates.

Raw box predictions were measured without the editor’s extra cover. The editor adds a 20% margin on each edge by default and preserves the initial box size.
The David video starts at original frame 300 to align its reference annotations. Both public sequences were tested in full. Loss stopping was disabled for this accuracy comparison.
Separate generated clips tested translation, growth, disappearance, a scene cut, exit from the frame and a similar distractor. Obvious disappearance and scene cuts paused all three trackers. A distractor still reduced overlap without reliably causing a warning.
NanoTrack scores remained above 0.96 on the public sequences despite some incorrect boxes. A high score is not proof of coverage. Every automatic path requires your review before saving.

The public clips and annotations come from OpenCV’s copy of the OTB tracking benchmark. These two face sequences and simple generated scenes do not establish accuracy for number plates, screen text, other objects or all lighting. Always inspect your own result.

A worked timing example

Suppose a fictional email is visible from 1.20 to 3.80 seconds in a 30 fps recording. Starting the cover at 1.10 and ending it at 3.90 leaves 0.10 seconds on either side. That is three frames per side for this constant-rate example.

Moment	What to inspect
1.0667 seconds	The last frame before the cover starts. Is the email already visible?
1.10 seconds	The first covered frame. Does the rectangle leave a margin?
Between 1.20 and 3.80 seconds	The entire path. Two correct endpoints do not prove the frames between them are covered.
3.90 seconds	The end is exclusive in floi. The cover is absent on this frame. Has the email gone?

This is a timing calculation, not a universal safety margin. Variable-rate clips do not have evenly spaced frame times. Floi’s Previous frame and Next frame controls read presentation timestamps instead of assuming 30 fps.

What we checked in saved files

In November 2026, we tested a generated three-second, 640 × 360 clip with 72 frames and AAC sound. The machine had an Apple M1 Max and 32 GB memory. FFprobe counted output frames. We sampled the expected cover’s centre in every decoded output frame.

Check	Chrome 156	WebKit 27.2
Cover from 0.50 to 1.50 seconds	72 frames retained. The sampled point was covered in the 24 expected frames.	72 frames retained. The sampled point was covered in the 24 expected frames.
Moving, held and overlapping covers	All 72 expected centre samples covered in each test.	All 72 expected centre samples covered in each test.
Variable frame rate	72 frames across 4.500 seconds.	72 frames across 4.500 seconds.
90° source rotation	360 × 640 picture, with all 72 frames retained.	360 × 640 picture, with all 72 frames retained.
Sound in MP4 / WebM	AAC packets matched the source in MP4. WebM contained converted Opus.	AAC packets matched the source in MP4. WebM contained converted Opus.

These are bounded synthetic checks, not evidence that a real subject stays inside your masks. WebKit is an automated browser-engine test, not a physical iPhone test. Encoder padding made the three-second WebM files 3.020 and 3.055 seconds long respectively. Review your own output.

Check a screen recording beyond the obvious box

Look around the main detail as well as inside its cover. A recording can repeat private information in several places.

Browser address bars, tab titles and autocomplete.
Account menus, avatars, contact lists and notifications.
Terminal history, environment files and request headers.
Captions, spoken names and other audio tracks.

Make one cover per independently timed detail. A wide static cover can be simpler to verify than a tight moving one. You can also crop away an edge or cut out an interval when it is not essential to the story.

Use FFmpeg for a repeatable solid cover

FFmpeg can cover a fixed rectangle for an exact time range. This command assumes an SDR source at its displayed orientation. The box starts 120 pixels from the left and 80 from the top.

ffmpeg -i input.mp4 -map 0:v:0 -map 0:a:0? -vf "drawbox=x=120:y=80:w=420:h=70:color=black:t=fill:enable='gte(t,1.1)*lt(t,3.9)'" -c:v libx264 -crf 20 -c:a aac -movflags +faststart covered.mp4

The drawbox filter replaces that rectangle in each active frame. The expression includes 1.1 seconds and excludes 3.9 seconds. Adjust the coordinates to your clip, then review the result. This example does not track a moving object or mute speech.

Choose the method for the job

A short local clip suits a browser editor. A long recording, complex occlusion or a multi-track project may suit a desktop editor better.

Method	Strength	Tradeoff
Floi browser studio	Timed rectangles, local object tracking, manual correction and review of the encoded file.	You select each subject and review every path. Five-minute and 300 MB caps, browser codec limits.
Automatic tracking editor	Less manual work when faces or objects move.	Missed detections, occlusion and scene cuts still need inspection. Processing and pricing vary.
Desktop editor	Longer projects, more shapes and detailed track control.	Installation and a more involved workflow.
FFmpeg	Repeatable coordinates and timing for fixed details.	Requires commands and a separate review player.

Keep your original, then check the final file after any later compression or editing. A video mask does not certify metadata removal, legal compliance or anonymity.