The image never leaves the browser
Decode, sample down, draw back up without smoothing, encode. All of it in the page, none of it over the network. This is the tool where that matters most in practice, because a file you are covering something in is by definition a file with something on it you do not want distributed. Open the network tab and watch it stay silent, or disconnect and redact offline.
How the blocks are made
The region is drawn into a tiny canvas — one pixel per block — so each of those pixels becomes the average of the block it represents. That tiny canvas is then drawn back at full size with image smoothing off, which repeats each pixel across its whole block instead of interpolating. At a 20 px block size, four hundred original pixels collapse into one colour and only that colour is written to the output.
The original values are not stored anywhere in the result. In that narrow sense the operation is irreversible: nothing in the file can be inverted to recover the pixels.
Why irreversible is not the same as safe
An attacker does not need to invert the transform. They need only run it forwards. Take a guess at what was underneath, render it, pixelate it with the same block grid, and compare the averages. If the space of possible answers is small, the matching guess is the answer.
That is why pixelated text is not redacted text. A six-digit number has a million possibilities and a computer will try all of them in moments. A name in a known font from a known list is worse. Published examples of recovering pixelated text go back years and the technique is not sophisticated. The same reasoning applies, more weakly, to anything else drawn from a constrained set.
Faces are a genuinely harder problem, because the space of possible faces is enormous. Heavy pixelation of a face is broadly defensible. Light pixelation of a face is not, because the eye and mouth positions and the overall tonal layout survive the averaging, and people recognise those.
Choosing a block size
The useful rule is to work back from the thing being covered rather than from the image size. Divide the width of what you are hiding by six, and use that or larger. A face 300 px wide wants 50 px blocks. A licence plate 200 px wide wants 35 px or more.
For text, skip the arithmetic and go to the endpoint: set the block size larger than the region itself. The region then reduces to a single block and comes back as one flat colour, which is a solid cover-up and the only thing on this page you should rely on for a document. The result panel tells you when the region has collapsed like that, so you can confirm it did.
Positioning without a drag handle
The region is specified in percentages of the image, which is deliberately awkward and has one real advantage: the same numbers apply to a different photo taken the same way. For a head-and-shoulders portrait the face sits high, so a custom region with the top edge at 10 to 20 percent is usually right, not the 25 percent default.
Ellipse suits faces; rectangle suits plates, document fields and screenshots. To cover two things, run the tool, download, feed the result back in, and set the second region — each pass composites onto the previous one. When there are many faces, pixelating the whole frame is faster and often more appropriate than fifteen passes.
Where the real risk sits
Almost every redaction failure that becomes public is not a cryptographic one. It is the wrong file being attached, the original being the one still in the camera roll, an earlier draft surviving in a shared folder, or a thumbnail cached somewhere that was generated before the edit. The covered file you download here is fine. Give it a name that cannot be confused with the source, put the original somewhere you will not reach for by reflex, and check what you attached before you send it.
Questions people ask
Can it find faces automatically?
No, and that is a considered choice rather than a missing feature. Running a detection model in the browser means downloading several megabytes of weights, and the models still miss profiles, partly occluded faces and small ones in the background. A miss produces a file you believe is covered and is not, which is worse than no automation at all. Looking at the preview and confirming with your own eyes is the check that actually works.
Is a solid black box better than pixelation?
For anything textual, yes, unambiguously. A solid fill contains no information about what was underneath, so there is nothing to attack. Pixelation retains average brightness per block, and averages are enough to reconstruct content drawn from a small set. You can produce a solid fill here by setting the block size larger than the region — the whole area then becomes a single averaged colour.
Can I pixelate a PDF or a screenshot of one?
A screenshot is an image and works normally here. A PDF does not — and worse, covering something in a PDF by drawing a black rectangle over it in a viewer usually leaves the text underneath fully selectable and searchable. If a PDF needs redacting, use a tool with a genuine redaction function that removes the underlying content, or export a page to an image, cover it here, and rebuild the document from images.
Does pixelating make the file bigger or smaller?
Almost always smaller, sometimes dramatically. Flat blocks of uniform colour are trivial for any compressor to describe, so replacing detailed pixels with averages removes most of the information the encoder was working to preserve. A whole-image pixelation at a large block size can produce a file a fraction of the original size.