Detect on request
Only the categories you select run, and the source remains untouched after detection.
Free local privacy tool
Privacy Blur: blur or pixelate only the regions you approve
Request face and licence-plate detection when useful, add manual regions for everything else, and correct every box before Apply. Faces can fetch ~15 MB; plates add ~81 MB plus any uncached shared ONNX Runtime assets. Requests begin only after Detect and are cached; manual regions need no detector download.
Your image stays on this device. Nothing starts until you choose a file.
Result proof
A privacy result you can audit before export
01
Input
Image + reviewed regions
02
Local pass
Blur or pixelation
03
Result
Separate PNG
How to verify it: Count the applied regions, inspect every edge and reflection at 100%, and confirm that no unreviewed area changed. An empty detection must leave the source untouched.
Human-reviewed privacy
Detection proposes regions; it never hides pixels by itself
Faces, licence plates, and manual regions stay visible and editable until Apply. A successful no-detection result remains empty instead of triggering a heuristic that invents boxes.
3 region types
Face · plate · manual
Exact correction
Move, resize, remove, or enter numeric bounds for every proposed privacy region.
Two obscuring styles
Choose adjustable blur or block pixelation, then commit the reviewed set as one history operation.
Know before you process: Automatic detectors can miss small, occluded, rotated, or stylised subjects. Review the complete frame and add manual regions when needed.
Detection is evidence, not permission
A detector can miss a subject or draw an imprecise box. That is why this workflow stops after detection: the regions remain ordinary editable geometry until you explicitly apply an obscuring style. Successful no-detection results stay empty rather than being replaced by guessed skin or contrast boxes.
What people use this for
Street and event photos
Review faces in the crowd and add boxes for screens, badges, or bystanders the detector misses.
Vehicle listings
Obscure licence plates while keeping the vehicle and surrounding scene sharp.
Documents and screenshots
Add exact manual regions over names, addresses, account details, or interface identifiers.
Tracked video privacy
Continue in the Video Editor to review tracked regions before an audio-preserving render.
Limits worth knowing
- Small, occluded, rotated, reflected, illustrated, or unusually lit faces and plates may be missed.
- Weak blur can retain identifying structure; inspect at the actual delivery resolution and choose sufficient strength.
- Privacy also depends on audio, context, metadata, shadows, clothing, tattoos, and information outside the selected region.
- Tracking can drift in fast motion or after occlusion, so video regions must be reviewed across time.
How the local workflow runs
- Load the source and choose whether to request face detection, plate detection, or neither.
- Press Detect if wanted; faces can fetch ~15 MB and plates add ~81 MB plus any uncached shared ONNX Runtime assets.
- Move, resize, remove, or numerically edit every proposal, then add any missing manual regions.
- Choose blur or pixelation and set an appropriate strength.
- Apply the reviewed set once, inspect the actual result, and download the separate PNG.
Frequently asked questions
Does detection blur anything automatically?
No. Detect creates editable region proposals only. Move, resize, remove, or add boxes, choose blur or pixelation, and press Apply when the reviewed set is correct.
What downloads when I press Detect?
Faces can fetch approximately 15 MB for MediaPipe WASM, JavaScript, and the face task. Licence plates add an approximately 81 MB pinned model. Licence-plate detection may also fetch uncached shared ONNX Runtime assets. The browser caches these assets; manual regions require no detector download.
What happens when no face or plate is found?
The tool reports that nothing was detected and leaves the image unchanged. A successful empty model result does not trigger a heuristic that invents privacy boxes.
Can I blur something other than a face or plate?
Yes. Draw a manual region or enter exact x, y, width, and height values. Manual boxes use the same review and Apply workflow.
Should I use blur or pixelation?
Pixelation makes the treatment obvious and can be stronger at high block size. Blur fits some editorial styles, but weak blur may leave identifying structure. Increase strength and inspect at 100%.
Does this work on video?
Yes in the Video Editor, where reviewed regions are tracked over time and audio is preserved. This standalone page handles still images; both use the same detect-review-apply rule.
Does Privacy Blur guarantee anonymity?
No. Clothing, tattoos, reflections, context, metadata, voices, and missed regions can still identify someone. Review every frame or image and strip metadata when appropriate.
Related tools
Object Remove & Replace
Reconstruct a reviewed region instead of visibly obscuring it.
Metadata Remover
Inspect and remove EXIF, GPS, and other identifying metadata.
Smart Crop
Reframe after privacy regions are safely applied.
Video Editor
Track reviewed privacy regions through a clip while retaining audio.
Ready to try it?
Detection never applies a change, and manual boxes need no detector download.
Review privacy regions →