gizmobench

Image Blur

Choose a picture, set a radius in pixels and blur the whole thing, or drag a box over one face, one number plate or one line of a screenshot and blur only that. The blur is a separable Gaussian: the radius is the standard deviation, the kernel is cut off at three times it and normalised so the weights sum to one, and it is run across and then down, which is the same result as the two dimensional convolution and a great deal less work. Pixels inside the box sample their real neighbours from outside it, so the edge of the box is blurred exactly as a whole image pass would blur it rather than against a repeated edge, and every pixel outside the box is copied into the new file byte for byte. Where a picture has transparency the convolution runs on premultiplied colour, so a cut-out keeps its own colour along the edge instead of picking up a dark rim from the black that usually sits under a transparent pixel. The page prints the assumptions next to the result: the radius, the number of taps, the edge rule, how many pixels are blurred and, in Selection mode, how many are copied unchanged. Nothing is uploaded, there is no account, and the file that comes out is the original width and height.

originalnothing yet

The picture appears here.

blurrednothing yet

The blurred copy appears here.

or drop one on the stage, or paste a screenshot

Never uploadedNo accountBlurred in this tabFull resolution out
Radius
Mode
Size
no picture yet
Format

Radius is the standard deviation of the Gaussian, in pixels, from 0 to 100, over pictures up to 30 megapixels. Drag a box on the left to blur one part, or switch Mode to Whole.

  • Whole imagea hard edge, 0 against 255, radius 1
    becomes 77 against 178, with the step spread over 6 pixels
  • Dragged regionone rectangle, 4 by 4, radius 1.5
    16 pixels blurred, the other 84 identical to the original
  • Radius 0no change
    every byte of the original returned

Those three lines are printed by running the same code your picture goes through, not typed in by hand.

Your picture stays on this device. A separable Gaussian blur computed on this device: each output pixel is the weighted mean of its neighbours, rounded to a whole channel value, with weights that sum to one, so a flat area comes back exactly as flat as it was and a radius of 0 returns the picture untouched. The file you save holds only the blurred pixels, never the originals, but blurring is a visual effect and not a security control: a small radius over text can still be read or recovered, so paint over anything that has to stay secret. Nothing is uploaded. There is no upload, no account and no watermark, and no image byte is handed to analytics either: the only thing the page records is that the tool was used.

Common questions

How do I blur only one part of an image?
Switch Mode to Selection and a box is already waiting over the middle of the picture. Drag a new box anywhere on the picture, drag the box itself to move it, or type the four numbers under the readout: Left, Top, Width and Height, in the picture's own pixels. Only the pixels inside that box are replaced. The blur inside it still samples the real pixels around it, so the edge of the box is blurred exactly as a whole image pass would blur it, and everything outside the box is copied into the file without a single byte changing. Switch Mode back to Whole and the same radius is applied to the entire picture.
What does the blur radius actually mean?
It is the standard deviation of the Gaussian, in pixels, and it runs from 0 to 100. A radius of 0 returns the picture untouched rather than an error, which is the honest answer to asking for no blur at all. The kernel is cut off at three times the radius, so at radius 12 each pixel is mixed with its neighbours up to 36 pixels away, across 73 taps, and the page prints that tap count beside the result. The radius is measured in the picture's own pixels, not in the pixels of the preview on screen, so a large photograph needs a larger radius than a small one to look equally soft.
Can a blur be undone, and is it safe for hiding private details?
Treat it as a visual effect, not as a security control. The file you download holds only the blurred pixels, so there is no hidden layer sitting underneath them, but that is not the same as being safe. A blur is a linear operation, and it can be attacked by rendering candidate strings, blurring them the same way and matching the result, which is how researchers have read back text that was blurred at a small radius. If what you are covering has to stay secret, such as a password, an account number or an address, paint a solid shape over it in an image editor instead, so the pixels that come out carry nothing from the pixels that went in.
Why do blurred cut-outs get a dark edge in other tools, and not here?
Because a transparent pixel still has a colour stored under it, and it is usually black. A blur that averages the raw bytes pulls that black into every pixel near the edge of a cut-out, which shows up as a grey or dark rim once the picture is placed on a light background. This tool blurs premultiplied colour instead: each pixel's colour is weighted by how much of it is actually there, the blurred colour is divided back out by the blurred alpha, and the result is that a white shape on a transparent background stays white right to the edge while its alpha fades. The tests hold that exactly: every visible pixel of that fixture comes out white.
Does transparency survive, and should I save a PNG or a JPEG?
In PNG, yes: the alpha channel is blurred along with the colour, so a soft edge stays soft and a transparent corner stays transparent. JPEG has no transparency at all, so choosing JPEG fills transparent areas with white before the file is written, and the page says so before you save. PNG is also the better choice for screenshots of text and for anything you plan to edit further, because it stores every pixel exactly. JPEG is smaller for a photograph, and is written at high quality. Either way the download is the full picture at its original pixel size, named after your file with the radius in it, so a radius 12 job on photo.jpg arrives as photo-blurred-12px.png.
Is my picture uploaded anywhere?
No. There is no server in this. The file is read in the page, decoded by the browser's own image decoder, blurred by code running in this tab and handed back as a download or put on your clipboard as a PNG. No image byte is sent to an endpoint, to analytics or to error reporting. The only thing written to your browser's storage is the setting row: the radius, the mode and the format you last used. No picture is ever stored, and if storage is blocked the tool still works.
Is there a limit on the size of the picture?
Nothing is uploaded, so the bytes on disk are nobody's business, but the pixels have an honest ceiling because the arithmetic runs on your own processor. Pictures up to 30 megapixels are read, which covers anything a phone or a full frame camera produces. The work a Gaussian costs is the area it covers times the radius, so the two are capped together at 360 megapixels times radius: a 30 megapixel photograph takes radius 12, a 3 megapixel one takes the full 100, and a small selection inside a large picture costs only what the selection costs. Past either limit the page says which number was too large and what to change, rather than freezing the tab.
Can I use it without a mouse?
Yes. The radius is a number field and a slider, both labelled and both reachable by tab, and typing something that is not a number produces a message saying so rather than a silently corrected value. Tab to the box drawn over the picture and the arrow keys move it, with shift and an arrow resizing it; each press moves one screen pixel, or one source pixel when that is the larger. The four number fields set the box exactly. The picture, the mode, the box and the processing assumptions are all described in words on the page, so a screen reader gets the same facts the two canvases show.

A separable Gaussian blur computed on this device: each output pixel is the weighted mean of its neighbours, rounded to a whole channel value, with weights that sum to one, so a flat area comes back exactly as flat as it was and a radius of 0 returns the picture untouched. The file you save holds only the blurred pixels, never the originals, but blurring is a visual effect and not a security control: a small radius over text can still be read or recovered, so paint over anything that has to stay secret. Nothing is uploaded.