Text diff checker with word-level highlighting

Most free diff tools compare whole lines only: change one word and the entire line goes red on one side and green on the other, leaving you to find the actual edit yourself. This page runs a second comparison inside each pair of changed lines, so “Run the test suite on Node 22.” against “…Node 24.” marks two characters and leaves the rest plain. Character-level would go further and mark the “2” and the “4” alone — which is right for a typo like “colour” against “color”, and wrong for anything larger, because a rewritten sentence turns into dozens of one-character fragments scattered across the line. Words are the unit people read, so words are the unit used here, with punctuation and the gaps between them treated as their own tokens: “config.json” splits into “config”, “.” and “json”, because the dot is often exactly where the change is.

Almost always something invisible. Trailing whitespace — a space or a tab sitting after the last visible character — is the usual culprit, and it shows up here as a highlighted blank at the end of the line; switch on “Ignore leading and trailing whitespace” to confirm that is all it was. Next most common is a character that looks like one you know and is not: a non-breaking space instead of a space, a curly quote instead of a straight one, an en dash instead of a hyphen, or a zero-width character carried in from a web page or a word processor. Those all survive copy and paste and none of them are visible. A tab against four spaces is the same story and reads identically at most indent widths. If you want the text cleaned rather than merely ignored, the Text Cleaner tool on this site strips them; line endings are the one cause this page cannot show you, for the reason in the answer below.

It is the hunk header, and it says where this block of changes belongs. The part after the minus is the original file: start at line 1, and this hunk covers 5 of its lines. The part after the plus is the new file: start at line 1, covering 5 lines. Both counts include the unchanged context lines printed around the change, which is why they are usually larger than the number of “-” and “+” lines. That exact header is what the sample above exports. Two more rules matter when you read one: a count of 1 is written without the comma (“@@ -3 +3,2 @@”), and an empty range — a hunk that only adds lines to a file that had none — names the line before the insertion point, which is why inserting into an empty file produces “-0,0”.

The limits are 5,000 lines and 400,000 characters per side, and beyond 1,200 single-line edits between them the tool switches to a coarser answer. Those are not arbitrary. Finding the smallest possible set of edits costs roughly the size of the input multiplied by the number of differences, so two ten-thousand-line files that differ everywhere is on the order of a hundred million operations — the tab stops responding, and you lose the text you pasted with no explanation of why. Refusing with a sentence costs you less than that does. When only the edit limit is passed, nothing is thrown away: the lines the two share at the start and the end are still matched, and the region between them is reported as one block that differs, with a note saying so.

Twice the number of identical lines, divided by the total number of lines on both sides. The sample above has 5 lines each side and 3 identical lines, so it scores 2 × 3 ÷ 10 = 60%. A line that changed by a single word counts as zero here, not as a near-match, which makes the figure deliberately stricter than what your eye reports — it answers “how much of this is untouched”, not “how similar do these feel”. Two empty inputs score 100%, because nothing about them differs.

Because those options make lines count as equal when they are not. A patch is a set of instructions for turning the left text into the right text byte for byte, so if the tool decided “Total” and “total” were the same line, the patch would leave that line out — and applying it would produce a file that is not the right-hand text. It would look correct in the preview and fail on a real repository, possibly after half of it had already been applied. So the export is only offered for an exact comparison. The on-screen result is still perfectly useful with the options on; it just is not a patch.

A newline at the end of a file is the end of the last line, not an extra blank line, so it is never shown as a row — a file that ends properly would otherwise appear to carry a phantom empty line at the bottom. It is still a real difference: when one side ends with a newline and the other does not, the page says so in a note above the result, and the exported patch records it the way diff does, with “\ No newline at end of file” written under the affected line. Windows line endings are a different matter, and the honest answer is that this page cannot see them. Every browser normalises carriage returns to line feeds as text enters a text box, so a file written with CRLF and the same file written with LF arrive here identical — pasting CRLF text above and getting no difference at all is that behaviour, not a bug in the comparison. That is also why there is no “ignore line endings” switch: there would be nothing left for it to ignore. Check line endings with “file”, with “git diff --stat”, or in your editor’s status bar.

Myers’ diff algorithm — the O(ND) greedy variant described in his 1986 paper, written out in this page rather than pulled from a library, since a diff package would cost more of the page weight than the whole tool. It finds a minimum edit script: no shorter set of insertions and deletions turns one text into the other. What it does not guarantee is that the minimum is unique. Where several alignments have the same cost — most commonly around repeated or blank lines — one tool may show a block as moved down and another as moved up, and both are correct. If two diffs of the same pair disagree, check that the number of added and removed lines matches; that part will agree.

No. The comparison, the word-level highlighting and the patch are all produced by JavaScript this page has already loaded, there is no upload of any kind, and nothing is written to localStorage — closing the tab leaves nothing behind. That matters more here than on most pages: the things people compare are contracts, config files, credentials and drafts that have not been published. The page does load Google’s ad script, as every page on this site does, but it is never given what you paste.