Blur Image Online — Free, Fast & Private
This is a targeted blur, not a filter: you choose what disappears and how completely. In manual mode you drag a rectangle over the detail to hide, resize it with the handles, commit it, and add as many more as the image needs — every region is tracked in the photo's own pixel coordinates, so what you cover in the preview is exactly what is covered in the export. In face mode, an on-device detector finds faces and blurs all of them at once. A single strength slider from a soft 4px to a near-opaque 40px governs both, with a live preview at every step.
Hiding faces before a photo goes public
The most common reason to blur: people who never agreed to be in your post. Bystanders in a street shot, other people's children at a school event, colleagues in an office photo, strangers in the background of a listing. Auto face blur handles the tedious part — it runs BlazeFace, the same compact face detector behind this site's passport tool, finds every face it can, expands each box by about 20% so ears, chins, and hairlines do not leak around the edge, and blurs them all in one pass. Two honest caveats. First, the model is the short-range variant, tuned for faces at selfie distance: small or far-away faces in group shots and wide scenes are routinely missed, so always check the outlined boxes against the photo and add a manual region over anything the detector skipped — auto face blur is a head start, not a guarantee. Second, the detection itself runs entirely on your device; the only thing fetched is the ~0.2 MB model file, downloaded once from Google's model CDN. Your photo is never part of any request.
Plates, addresses, and screenshot details
Manual regions cover everything the face detector was never meant to find: the licence plate in a car-sale photo, the house number behind the subject, a parcel label with your address, the account number in a support screenshot, a name in an email thread you are sharing for its content. Because regions are committed independently, one image can carry several blurs at different places — blur the plate, then the neighbour's window, then the sticker on the laptop — and a misplaced rectangle can be redrawn without disturbing the rest. Metadata deserves the same attention as pixels: a photo that shows no address can still carry GPS coordinates, so strip the EXIF data before the blurred file goes out.
No dark halos: blurring a padded window
A detail most blur tools get wrong: apply a canvas blur to just the selected rectangle and the browser samples past the rectangle's edges into emptiness, darkening the border of the patch — the telltale gray rim around a cheap blur. This tool blurs a padded window instead: it takes the region plus a margin of twice the blur strength on every side, blurs that larger window, and copies only the inner target back into the image. Pixels just outside your rectangle contribute to the blur exactly as they would if the whole image had been blurred, so the patch sits seamlessly in its surroundings with no halo and no visible seam. On browsers without native canvas blur support, a downscale-and-upscale approximation with smoothing on both passes stands in automatically.
The honest part: light blur is not redaction
A gentle blur over text is cosmetic, not secure. Gaussian blur preserves a great deal of the underlying signal, and blurred text — especially at low strength over a known font — has been reconstructed in practice. So treat the slider as a security control, not just a style one: for anything that must not leak, push the strength until the preview genuinely shows nothing, and for passwords, keys, and financial numbers prefer an opaque box over any soft effect. Between the two soft options, blur reads as natural and keeps the photo's tone, while pixelation discards more information per block at equal apparent strength and signals deliberate censorship — the pixelate tool covers that trade-off in depth, including the research on reversing weak redaction. Blur is the right choice when the goal is a person not being identifiable while the scene stays natural; it is the wrong choice for secrets at low strength.
Private by architecture
An image being blurred is an image with something to hide, which makes the usual upload-to-a-server workflow exactly backwards — the unblurred original would arrive on someone else's machine before any hiding happened. Here the entire pipeline is the Canvas API in your own browser: decode, blur, re-encode, download, all local. Manual region mode makes no network request at all and works offline; face mode's single fetch is the small detector model described above, never your image. Export keeps your choice of JPEG, PNG, or WebP with a quality control — PNG is the clean choice for censored screenshots, and if the file needs to shrink afterwards the screenshot compressor finishes the job without smearing the text you kept.