Guide
One sentence carries this page: a low-resolution image gives you less margin for error and is itself evidence that a sharper copy exists somewhere else, so the first job is to get or re-make the better file and the second is to stop relying on your eyes.
Where the low resolution came from, and why that is the first question
Start with provenance, because it decides what your redaction is actually for. Small files arrive by a handful of well-worn routes, and almost all of them leave a larger file behind.
A chat or messaging app re-encoded the picture on the way through — many of them downscale and recompress what they send, and the exact limits differ by app and change over time, so treat the behaviour as likely rather than as a known number. Somebody took a screenshot of a screenshot, on a smaller display, and the second capture inherited the first one’s pixels and then some. The image is a thumbnail or a preview, saved because the full version was behind a login. It is a frame from a video or a screen share, which was already compressed for transport. It is an old scan at a low setting, or a save-for-web export from a design tool. Or you shrank it yourself so it would attach to something.
Each of those implies an upstream file. That matters because it changes the claim your redaction can make. If a sharper unredacted copy has already passed through other hands, covering this copy bounds the future rather than retracting the past — it stops this version spreading further, and that is worth doing, but it does not undo anything. Be straight about that with whoever asked you for the image, because a carefully redacted forward quietly implies containment that may not exist.
If you are the one who made the small copy, the upstream file is on your own machine, and it needs the same care as any unredacted original: it is the file that should not be the one you attach, and the one worth deleting or filing deliberately rather than leaving in a downloads folder.
One route deserves calling out separately, because it is a technique people reach for on purpose: downscaling an image in order to make something unreadable. It does not qualify. Resampling destroys detail everywhere, including the part you wanted to keep, and what it leaves behind in the part you wanted gone is unverifiable — you cannot look at the result and establish that the value is absent rather than merely faint. A cover is checkable. A resample is a hope.
Fewer pixels per character means less margin, not more
There is a persistent intuition that a small image is a safer image, on the grounds that there is less there to read. The arithmetic of drawing covers runs the other way.
A line of interface text on a full-size desktop capture is typically in the region of fifteen to twenty pixels tall. Downscale that capture to a quarter of its width and the same row is four or five pixels tall. Now consider an ordinary slip: your cover lands two pixels higher than you meant. On the original that is a rounding error nobody will ever see. On the small copy it is half the height of the character row, and what is left exposed is the top half of every glyph in the field — which, for digits and for capital letters, is frequently enough to identify them.
The second half of the problem is that you are much less likely to notice. A leftover one or two pixel band in a full-size image reads immediately as a sliver of text poking out from under a box. At low resolution, after compression, the same band looks like exactly what surrounds it: noise. The data is no less present for being camouflaged.
The response is not more care, it is more margin. Treat the field’s visible extent as the minimum extent of the cover and extend past it on all four sides — into the label, into the row above, into the cell border. On a small file the pixels immediately around a value are rarely carrying information anybody needs, and the cost of swallowing them is far lower than the cost of a clipped edge.
What the editor’s own numbers do at small sizes
Three details of the tool on this page behave differently on a small image than on a large one. None of them is documented in the interface, and all three are worth knowing before you draw.
The display is one-to-one, not scaled down. The canvas is created at the image’s exact pixel dimensions and shown inside a scrolling panel with a cap of roughly seventy per cent of the window height, with a max-width that holds it inside the page column. A max-width caps a size; it never enlarges one. So a large photo gets scaled down to fit — which is the usual complaint, that you are judging a huge file on a reduced view — while a three-hundred-pixel-wide screenshot is drawn at three hundred pixels, sitting small in the middle of a much wider column. The view is perfectly honest. It is also too small to aim at.
Browser zoom is the fix, and it is accurate. Pointer positions are converted into image coordinates using the ratio between the canvas’s true pixel width and its currently displayed width, measured from the element itself on every interaction. Zooming the page changes the displayed width, the ratio is recomputed, and your cover still lands where you put it. There is no in-editor zoom control, so page zoom is the intended move: two or three steps in, draw, then zoom back out.
Very small covers are discarded without a word. A rectangle drag under about four pixels in either direction, or an oval under six, is thrown away; the lasso needs at least three points. On a full-size image those thresholds exist to swallow stray clicks and you will never meet them deliberately. On a small image, four pixels can be a whole field. If you draw a tight cover and nothing appears, that is almost certainly what happened — and the fix is the one you wanted anyway, which is to draw it bigger.
Then there is effect strength, which on a small file is the argument for using the solid fill. Blur and Pixelate are both computed from the bounding box of your selection rather than from the shape inside it. Blur copies the box, shrinks it by a factor of ten with a one-pixel floor in each direction, and scales it back up: on a strip only a few pixels tall that floor flattens the whole thing toward a single row of averaged colour, strong in the vertical direction, while still preserving roughly one average per ten pixels of width — so the light-and-dark rhythm of a row of type can persist as a horizontal streak. Pixelate uses square blocks of about a twelfth of the bounding box’s shorter side, with a floor of six pixels, which on a small image means nearly every selection lands on the floor; each block becomes one sampled colour, and a six-character value can be a handful of blocks across. Black Box fills the traced path with a flat dark value and has no strength parameter at all. It is the only method here whose outcome does not depend on how many pixels the region happened to have, which makes it the default on a degraded file rather than the heavy-handed option.
A running order for a small or degraded file
1. Ask where the file came from and whether a better one exists. Request the original from the sender, re-take the screenshot yourself from the live source, or export from the application rather than capturing it. This step is boring and it replaces every other step on this list.
2. If you get the larger file, redact that one. With one eye open about the trade: you now hold an unredacted high-resolution copy you did not have five minutes ago, and it needs the same handling as any original.
3. If you cannot, decide what the cover is for — and say it out loud. Bounding further spread is a reasonable goal. Implying that the value was never out is not.
4. Zoom the browser before you draw anything. Two or three steps. It costs nothing and the coordinate mapping follows it exactly.
5. List the fields before you cover any of them. Counting first gives you something to check against afterwards, which matters more here than usual because the after-the-fact visual check is weaker.
6. Cover with Black Box and overshoot every edge. The pixel you are aiming at is a fraction of a character.
7. Verify on the downloaded file, with the source open beside it. Not on the canvas you were drawing on. See below for what the check looks like when zooming does not help.
8. Send the export, not the file you were sent. Download PNG re-encodes the canvas from scratch, so none of the source file’s tags come across. Expect the export to keep the small pixel dimensions and to look no sharper than the input — it should not, and a file that suddenly looks crisper has been resampled by something.
Verifying a cover when zooming only shows you blocks
The standard check — open the saved file, zoom past one hundred per cent, walk the perimeter of every cover — is described in full on how to verify a photo was actually redacted, and it is the right procedure. It just degrades on a small file, because at four hundred per cent the entire image is blocks, and a two-pixel remnant of a glyph is visually indistinguishable from the compression artefacts beside it. Four things still work.
Check geometry, not legibility. Open the source and the export at the same zoom, side by side, and compare the rectangles: does the cover enclose the field’s extent, with margin, on all four sides? You can answer that question perfectly well about an area whose contents you cannot read.
Count. You listed the fields in step five. Count the covers in the export and reconcile the two numbers. Omission is the failure mode that visual inspection of a degraded image is worst at catching.
Push contrast on the export. Any viewer with levels, curves or an auto-tone control will do. A flat fill cannot be lifted into anything, because there is nothing underneath it. A faint remnant of type will brighten into something you can see — which is precisely what a recipient’s viewer would also do.
Re-open the saved file. The canvas in the page is a live drawing surface, and the thing anybody else opens is the downloaded PNG. Judge that.
What the editor above does and does not do
The tool on this page is region-based: three methods — Black Box, Blur, Pixelate — across three selection shapes, rectangle, oval and freehand lasso. Every operation replaces the pixels inside the region you drew. Undo steps back one cover at a time, Clear restores the untouched image, and Download PNG writes the current canvas out as a timestamped PNG, re-encoded from scratch rather than copied from your source file. Everything happens in your browser on your own machine; nothing is sent anywhere, which is also why there is no stored copy and no sharing link, and why the image, the undo snapshots and the retained original live only in the open tab.
Four limits matter specifically for a degraded file. There is no upscale, resize or crop of any kind — the canvas is set to the source image’s exact pixel dimensions and stays there, so nothing here makes a small image bigger. There is no sharpen, denoise, contrast or enhancement control, so the editor cannot help you read a field you cannot read, and it certainly cannot reconstruct one. There is no zoom control inside the editor; page zoom is the mechanism, and the pointer mapping is written to survive it. And there is no analysis of any kind: the tool cannot locate fields, cannot tell you that a cover clips an edge, and cannot warn you that a value appears twice. Every judgement described above is yours to make. The tool executes covers, and that is all it does.
One pleasant consequence of a small input, for once: the usual warning that a redacted PNG export comes out larger than the file you started with mostly does not apply here. A few hundred pixels of PNG is a small file, and the related problem of an export that is re-compressed again by the platform you send it through is worth a look before you post it anywhere, since that is how the next generation of small copy gets made.
Common mistakes and misconceptions
“It is too small to read, so it is already hidden.” Illegible to you at the size you are looking at it is not the same as absent from the file. Contrast is adjustable at the other end, automated readers do not need crisp type, and a reader who can narrow the value down from context only needs a partial resolve to confirm a guess.
“I will shrink it further to be safe.” Resampling degrades everything, including the content you wanted to keep, and leaves a result you cannot verify. There is also no resize in this editor, which is deliberate.
“The box looked fine on the canvas.” At roughly one image pixel per screen pixel you were judging a four-pixel row by eye, at a size the interface chose rather than one you chose. Zoom the page first, and verify on the export.
“Pixelate is plenty at this size.” Block size falls to its six-pixel floor on nearly every small selection, each block becomes a single colour, and a short value may be only a few blocks wide. It reads as heavy and it is not the strongest thing available to you.
“Asking for the original is overkill.” It is the cheapest fix on the page and it turns guessing back into covering. It is worth a round trip almost every time.
“Redacting this copy settles the matter.” If a sharper copy travelled first, your work bounds what happens next. That is a real and worthwhile outcome, and it is not the same as retraction — so do not let the redacted version be read as one.
“A cover this big looks like I am hiding something.” On a small file a generous cover is proportionate, not suspicious. If the size of the cover is itself the problem, the answer is a sentence of explanation alongside the image, not a tighter box.