Mosaic, blur or a solid block: which to use when redacting
What each of the three actually does to the pixels, which of them destroys information and which only looks like it does, what the recovery research really shows, and the decision rule that follows from it.
Three ways to hide something in an image, and they are not interchangeable. A mosaic averages blocks of pixels. A Gaussian blur mixes each pixel with its neighbours. A solid block throws the area away and paints one colour. All three look like redaction to a person glancing at the result; only one of them is redaction in the sense of "the information is gone".
The short answer is at the top so you can leave: use a solid block for anything with a fixed format or a small set of possible values — codes, card numbers, plates, dates of birth, names from a short list. Use a coarse mosaic when you want the reader to still perceive the structure of what you covered, and only for content nobody is motivated to recover. Use blur for aesthetics, never for safety.
What each one does to the pixels
Mosaic, also called pixelation
The image is divided into a grid and every pixel in a cell is replaced by the cell’s average colour. How much survives depends entirely on the cell size. A grid of four pixels over ten-pixel-tall text barely disturbs the letter shapes. A grid of forty pixels over the same text leaves a row of coloured squares carrying only the rough darkness of each region. This single parameter is the whole story, and it is the parameter people never look at, because the tool ships with a default and the default is usually tuned to look nice rather than to destroy information.
Gaussian blur
Every output pixel is a weighted mixture of the input pixels around it. That is a linear operation, and linear operations are in principle undoable. Real blurs are not perfectly invertible because the result is quantised to eight bits per channel and clipped, so some information genuinely is lost — but far less than the smooth, satisfying appearance of a blur suggests. Blur is the technique that most overstates its own effectiveness to the eye, which is exactly the wrong property for a safety tool.
A solid block
The pixels under it are gone. There is no parameter to get wrong, no residual signal, nothing to deconvolve or brute-force. It is ugly and it is correct, and the only way a solid block fails is if it was never flattened into the exported image.
What the recovery research actually shows
There is a real literature here and it is worth stating precisely, because it gets flattened into slogans in both directions. Deep-learning work has shown that pixelated and otherwise obfuscated images can be re-identified at rates far above chance when the set of possible answers is known and limited — recognising which of a known set of people a pixelated face belongs to, for instance. Separately, practical tools recover pixelated text by brute force: render a candidate string in the likely font at the likely size, apply the same pixelation with the same grid alignment, and compare the result to the redacted region. When the string is short and its format is fixed, this searches a tiny space and wins.
Now the conditions, which are what make it a useful finding rather than a scare story. Those attacks generally need a guessable candidate space, a font and rendering the attacker can reproduce, and a grid coarse enough to have averaged but fine enough to leave a distinguishing signal. Take those away — a coarse mosaic, unknown text, no fixed format, an unknown rendering pipeline, a flattened export — and there is no practical recovery. So it is false to say that a mosaic is always reversible, and it is equally false to treat a mosaic as equivalent to a block. Both statements are wrong in ways that get people into trouble.
The reasonable conclusion is a decision rule rather than a claim about images: if you would be materially harmed by being wrong, choose the option whose safety does not depend on any assumption about the attacker. That is the solid block, every time. Reserve the mosaic for the case where preserving structure is worth something and the content is not worth attacking.
- Fixed format and a small candidate space — a six-digit code, a licence plate, a card suffix, a date of birth: solid block. The value is guessable, so anything that leaves residual signal is a hint.
- Free text you would not want quoted — a name, an address, a message: solid block if it matters, a coarse mosaic if it merely needs to be unreadable at a glance in a screenshot nobody will attack.
- Something whose shape is part of the point — a face you want to show was present, a chart whose outline matters, a UI region you are pointing at: mosaic, because a block deletes the information that there was anything there.
- A blur, at any strength: use it for aesthetics, not for secrets. It is a linear operation and it is the one of the three that most reliably looks safer than it is.
When a mosaic is the right answer
It is not never. A mosaic communicates something a block cannot: that there was something there, roughly how much of it, and roughly what kind of thing it was. A chat log with mosaicked names still reads as a conversation between several people. A dashboard with mosaicked figures still reads as a dashboard. Black bars turn both into an unreadable page of censorship, and readers respond by distrusting the whole image.
So the useful pattern in a single screenshot is often mixed: a coarse mosaic over the ambient material you are hiding for tidiness, and a hard block over the two or three items that would actually cost you something. Most tools make this awkward because they force one mode per image. It should be a per-stroke decision, because it is a per-item decision.
The mistake that beats all three
None of this matters if the cover is still a separate object when the file leaves your hands. A rectangle drawn over text in a PDF hides nothing — select the text underneath and copy it. A layered image file carries the layer. An editor with non-destructive edits keeps the original so that a revert button can exist.
The habit that prevents all of it: export a new flat file, open that file, and inspect it. If it is a PDF, try selecting text across the redacted area. If it is an image, brighten it hard in any editor and look again — that alone catches every "I lowered the opacity so it looks nicer" mistake, which is more common than any of the sophisticated attacks described above.
Doing it in Marka
Marka treats redaction as a brush with three modes: mosaic, blur and a solid colour, with a strength slider and three brush widths. Because it is a brush and the mode is a property of the stroke, the mixed approach above is the natural way to work — mosaic over the ambient material, then switch to solid and drag over the two things that matter. The mosaic mode is free; blur and solid are part of the one-time Pro purchase. On Mac there is also an on-device scan built on Apple’s Vision framework that proposes candidates to cover and covers only what you tick; it never acts on its own and it will miss things, and it is not present on iPhone or iPad. Everything is processed on the device, with nothing uploaded, which is the only arrangement under which a redaction tool makes sense. The interface is in Simplified Chinese in this version, which is worth knowing before installing.
Frequently asked questions
Is a mosaic reversible?
Not unconditionally, and anyone who tells you it always is has overstated a real result. Pixelated short text in a known format and a known font can be recovered by brute force — render candidates, pixelate them the same way, compare. A coarse mosaic over unpredictable content, flattened into an export, is not recoverable in any practical sense. The safe rule is to not have to know which case you are in: use a solid block when it matters.
Which is safer, mosaic or blur?
A coarse mosaic, by a wide margin. Blur mixes pixels rather than discarding them and is close to a linear operation, so it retains more of the original signal than its smooth appearance suggests. Blur is the weakest of the three and the one that most flatters itself to the eye.
How coarse does a mosaic need to be?
Coarse enough that a single cell covers a substantial part of a character — as a working rule, cells at least as large as the text is tall, so that no letter shape survives as a distinct pattern. If you can still see where the words break, the grid is too fine. Increase it until the covered region carries no structure at all, then look at it again at full size.
Why do professionals use black bars if a mosaic looks better?
Because a black bar has no parameter to get wrong and no residual signal, so its safety does not depend on any assumption about who is looking or what tools they have. A mosaic is a judgement call about an attacker; a block is not a judgement call at all. That is worth more than looking good.
Does re-saving as JPEG add any protection?
No. Treat compression as noise that an attacker also has to deal with, not as a defence. It removes a little high-frequency detail and it is not designed to destroy information — that is the opposite of what a codec is for. Re-encoding a weak redaction does not turn it into a strong one.
Do it in Marka
Marka is a screenshot markup app for iPhone, iPad and Mac. Brush a mosaic over what should not be seen, point an arrow at what should, number the steps, crop it to the shape you are posting in, and send. Everything is processed on the device — no upload, no account, no analytics, and a 2.4 MB download. Opening, cropping, the mosaic brush and export are free; Pro is a one-time purchase that covers all three platforms.