Fast, robust, accurate content-aware diffing.
- Robust: can diff 99.95% of code changes
- Content-aware: 98.8% of changes diffed using content-specific algorithms
| Content | Fast: under 100ms | Accurate: perfect | Accurate: near-perfect |
|---|---|---|---|
| Code | 97% | 67% | 90% |
| Pictures | 96% | 69% | 91% |
| Cursors | 100% | 85% | 100% |
| Fonts | 95% | 82% | 100% |
| Archives | 88% | 87% | 87% |
| Message catalogs | 100% | 83% | 83% |
| PDFs | 57% | 100% | 100% |
Speed measured on a T14s, Intel Core i7-1370P with 32 GB RAM (see Fast). PDFs have fewer than ten judged test cases (see Accurate).
See it in the browser: twenty real code
changes, recorded from the command-line tool and compared side by side in Unix diff and in
OmniDiff, and content examples: an
OpenDocument file, a compiled translation catalog, and pictures, still and animated.
Various light and dark themes are available:
Files are diffed based on type. A PDF diagram with one box renamed, and a web font with twenty icons redrawn:
cargo install --locked omnidiff
This command builds OmniDiff from source, with the dependency versions it was tested with. You
need a C compiler on PATH and a Rust toolchain, rustc 1.92 or later. The build compiles every tree-sitter grammar from C.
The first cargo install takes a few minutes, because of this and the lto = "fat" release
profile.
Prebuilt binaries for Linux, macOS (Intel and Apple Silicon), and Windows are attached to every GitHub release.
macOS and Linux:
brew install ivankovic/omnidiff/omnidiff
.deb packages for amd64 and arm64 are available in a signed apt repository:
sudo install -d -m 0755 /etc/apt/keyrings
curl -fsSL https://ivankovic.github.io/omnidiff/apt/omnidiff-archive-keyring.gpg \
| sudo tee /etc/apt/keyrings/omnidiff-archive-keyring.gpg > /dev/null
echo "deb [signed-by=/etc/apt/keyrings/omnidiff-archive-keyring.gpg] \
https://ivankovic.github.io/omnidiff/apt stable main" \
| sudo tee /etc/apt/sources.list.d/omnidiff.list > /dev/null
sudo apt update && sudo apt install omnidiffOn NixOS, or anywhere with Nix installed:
nix run github:ivankovic/omnidiff
If flakes are not enabled in your Nix configuration, add
--extra-experimental-features 'nix-command flakes' after nix.
Recipes for Arch and Gentoo live in packaging/: the PKGBUILD builds locally with
makepkg -si, and the Gentoo ebuild is ready for an overlay. The two distros are overwhelmed and
their repositories open and close to new packages sporadically.
VS Code does not yet let an extension replace its built-in diff view. The API is implemented upstream but unreleased; the extension will adopt it when it ships.
- VS Code - omnidiff-vscode. Search for
OmniDiff in the Extensions view, or
code --install-extension ivankovic.omnidiff. Also on Open VSX for VSCodium, Cursor and Windsurf. - Neovim - omnidiff.nvim.
omnidiff # open an empty viewer; press o to pick each file
omnidiff BEFORE AFTER # open directly into the diff of two files
Press ? in the viewer for the full list of keybindings.
The TUI uses 24-bit color when the terminal advertises it with COLORTERM=truecolor, and the
nearest 256 colors otherwise (macOS Terminal.app, or most terminals over ssh, which does not
forward COLORTERM). If your terminal supports 24-bit color but does not advertise it, run
export COLORTERM=truecolor.
For reviewing changes in a browser, see codereview.
Note: codereview is at v0.0.0; use it at your own risk.
omnidiff --headless BEFORE AFTER, or its synonym --batch, prints the diff as plain text, with
optional color, instead of opening the TUI. Use this for scripts or CI. Headless mode also starts
automatically whenever stdout is not a terminal, for example when piped into less or redirected to
a file. Because of this, omnidiff BEFORE AFTER | less works without the flag.
OmniDiff collapses long runs of unchanged lines. It keeps 3 lines of context on each side of a
change (override with --context N), the same convention as diff -u. OmniDiff also prefixes each
hunk with the nearest enclosing function, class, or struct line, when that line is not otherwise
visible. This shows the location of a change deep inside a large file.
Colors are on by default; pass --color never, or set NO_COLOR=1, to disable ANSI colors, for
example when you redirect output to a file - --color always forces them even under NO_COLOR.
OmniDiff exits 0 on success and 2 on error. For scripting, pass --exit-code to additionally
get 1 when the files differ. This follows the diff(1) convention. --exit-code is opt-in rather
than the default for the same reason git diff exits 0 even when files differ: OmniDiff's usual
non-interactive callers are version control systems driving it as a display tool, and they read a
non-zero exit as "the tool failed" - jj warns on every file, and git difftool with
difftool.trustExitCode=true aborts the whole diff. (The 7-argument GIT_EXTERNAL_DIFF form stays
at 0 even with --exit-code, since git treats a non-zero exit there as fatal.)
omnidiff --mode json BEFORE AFTER prints the diff as one JSON object. Its fields are described
in src/tui/json_output.schema.json, a JSON Schema.
OmniDiff is a git difftool backend. Run the interactive setup wizard, which asks
whether to configure it globally or for the current repository only:
omnidiff git configure
git config difftool.omnidiff.cmd 'omnidiff "$LOCAL" "$REMOTE"'
git difftool --tool=omnidiff
Run git config diff.tool omnidiff to make plain git difftool use OmniDiff by default, without
needing --tool. If you do not want git to ask "view diff ... [Y/n]?" before every file, run git config difftool.prompt false.
git difftool opens the interactive TUI. git diff and git log -p never do. git diff
pipes its output through git's pager, and a full-screen TUI cannot draw onto a pipe, so OmniDiff
always falls back to plain text there regardless of terminal or GIT_EXTERNAL_DIFF config (see
"Headless / batch mode" above). If you want the interactive viewer from git, use git difftool,
not git diff. If git difftool still doesn't open interactively over SSH, reconnect with
ssh -t — the session needs an allocated pseudo-terminal; tmux panes always have one.
OmniDiff also works directly with git diff and git log -p, through GIT_EXTERNAL_DIFF. This
path needs no difftool config:
GIT_EXTERNAL_DIFF=omnidiff git diff
Pictures, archives, fonts and the rest of the supported content get a short
report of what changed in them. Any other binary file gets a one-line Binary file <path> differs
notice instead of a diff, the same stand-in git and diff(1) print for them. Neither ever blocks
the rest of a git diff: an external diff that exits non-zero makes git abandon the entire run,
so OmniDiff reports an unshowable file as a successful diff of nothing rather than as a failure.
jj does not read git's difftool/diff.external settings, even in a colocated repo, so it needs
its own configuration. Run the setup wizard:
omnidiff jj configure
jj config set --user merge-tools.omnidiff.program omnidiff
jj config set --user merge-tools.omnidiff.diff-args '["$left","$right"]'
jj config set --user merge-tools.omnidiff.diff-invocation-mode file-by-file
That registers jj diff --tool omnidiff. To make it the default for plain jj diff as well:
jj config set --user ui.diff-formatter omnidiff
Use --repo in place of --user to configure the current repository only.
diff-invocation-mode = "file-by-file" is required. jj's default hands a diff tool two
directory trees; OmniDiff diffs two files, so without this setting every invocation fails. With
it, jj passes one changed file pair at a time, keeping each file's real path and extension, so
language detection works exactly as it does under git.
jj diff runs its formatter under a pager, so OmniDiff renders in its non-interactive text mode
there - the same output git diff gets. jj has no equivalent of git difftool's interactive
per-file viewer (its terminal-attached hook, ui.diff-editor, is for jj diffedit/jj split,
which edit the right-hand side and read it back - not something a read-only viewer should claim to
do), so for the full-screen TUI on a jj repo, run omnidiff BEFORE AFTER directly.
The language is detected from the file extension. A file with an unknown extension is diffed as plain text, line by line, so nothing is refused.
25 languages are parsed with a tree-sitter grammar and diffed structurally:
| Language | File extensions |
|---|---|
| C | .c, .h |
| C# | .cs |
| C++ | .cc, .cpp, .cxx, .hpp, .hh, .hxx |
| CSS | .css, .scss |
| Go | .go |
| HTML | .html, .htm |
| Java | .java |
| JavaScript | .js, .mjs, .cjs, .jsx |
| JSON | .json |
| Kotlin | .kt |
| Lua | .lua |
| Perl | .pl, .pm |
| PHP | .php |
| Python | .py, .pyi, .pyw |
| R | .r |
| Ruby | .rb |
| Rust | .rs |
| Scala | .scala |
| Shell (Bash) | .bash, .sh |
| Swift | .swift |
| TSX | .tsx |
| TypeScript | .ts, .mts, .cts |
| Vimscript | .vim |
| XML | .xml, .xht, .xhtml, .svg, .glif, .plist, .ui, .qrc, .glade, .xib, .storyboard, .xaml, .vcxproj, .filters, .csproj, .fsproj, .vbproj, .props, .targets, .nuspec, .resx, .wxs, .iml, .manifest, .config, .policy, .xsd, .xsl, .xslt, .xlf, .xliff, .kml, .gpx, .rss, .atom, .graphml, .dae |
| YAML | .yaml, .yml |
Recognised by extension but diffed as plain text, since no grammar is compiled in: Bazel (.bazel), Dart (.dart), Emacs Lisp (.el), Markdown (.md, .markdown), Protocol Buffers (.proto), SQL (.sql).
Files that are not text are recognised by their bytes, never by their names, and diffed by what they hold:
| Content | Formats | Diffed |
|---|---|---|
| Pictures | PNG, JPEG, GIF, WebP, BMP, ICO, TIFF | pixel by pixel, as the eye compares them: how much changed and where; animations frame by frame |
| Archives | zip (and so jar, Office documents, EPUB), tar, gzip, xz, bzip2 | file by file, each by what it is, archives inside archives included; Java class files as a listing of their fields and methods |
| Fonts | TrueType, OpenType, WOFF, WOFF2, EOT | glyph by glyph, each drawn, and the font's names and metrics |
| Cursors | Windows .cur and .ani, X cursors, Hyprland .hlc |
size by size and frame by frame |
| Message catalogs | gettext .mo, Qt .qm |
message by message |
| Documents | page by page, each page drawn | |
| Text in other encodings | UTF-16 and UTF-32 with a byte order mark | as text, like any other file |
Only the members that changed are opened, so a font with two edited glyphs shows two glyphs. Each change is also named by how much of it a reader would see: invisible (only the bytes differ), imperceptible, artifacts (what compression or resampling leaves behind), edited, redrawn or replaced - for text, whitespace, formatting and rewritten stand in for the second, third and fifth.
The TUI shows pictures as pictures: in the terminal's own graphics where it speaks the kitty,
sixel or iTerm2 protocol, and in Unicode half blocks everywhere else. b switches to half blocks
and back, and OMNIDIFF_GRAPHICS=halfblocks (or kitty, sixel, iterm2) overrides the
detection; inside tmux, the terminal in front of the pane decides. An archive, a font or any other
container lists its changed members beside the one selected, g shows every changed glyph or
picture at once, and Enter opens a text member in the full diff view.
Anything else binary gets the one-line Binary file <path> differs notice.
Copyright (C) 2026 Marko Ivankovic
This program is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, version 3 of the License.
See the LICENSE file for the full text of the License.
A commercial license is available as a monthly subscription through GitHub Sponsors. It covers internal use of omnidiff by your organisation without the source-disclosure obligations of the AGPL. The terms are in LICENSE-COMMERCIAL. Pick the tier that names the commercial license as a benefit. For invoicing or other arrangements, contact me at marko@ivankovic.me.
OmniDiff's goal is:
- A median diff in 100ms or less.
- A 99th-percentile diff in 1000ms or less.
Both are met. Over the 2,223 fixtures in src/test/data/diffs/, measured 2026-10-17 on a laptop,
an Intel Core i7-1370P (6 performance and 8 efficiency cores, 20 threads) with 32 GB RAM: p50
4.2ms, p90 39.4ms, p99 199ms, slowest 644ms. 100ms is the 96.6th percentile — 75 fixtures
take longer than that, and none takes longer than a second.
Times belong to the machine that measured them, so each machine keeps its own (see
research/data/machines.csv): the Intel Xeon E3-1275 v5 server that measured the research
figures takes about 2.5 times as long per fixture. Runtime is reported, never gated: make check-quality prints this distribution against the machine's own committed baseline on every
push, and warns when the runtime is more than twice the baseline's.
Other content is timed over pairs sampled from the same corpus, decoding included - unpacking an archive, drawing a font's glyphs or a PDF's pages - measured 2026-10-17 on the same laptop:
| Content | Pairs | p50 | p90 | p99 | Under 100ms |
|---|---|---|---|---|---|
| Pictures | 230 | 1.8ms | 41ms | 526ms | 96.1% |
| Cursors | 22 | 0.6ms | 1.3ms | 15ms | 100% |
| Fonts | 100 | 6.4ms | 57ms | 826ms | 95% |
| Archives | 100 | 2.6ms | 156ms | 1,938ms | 88% |
| Message catalogs | 81 | 0.4ms | 1.9ms | 4.7ms | 100% |
| PDFs | 100 | 69ms | 310ms | 985ms | 57% |
Archives miss the 99th-percentile goal: the slowest hold thousands of changed members. PDFs meet
both goals, but only 57% of them take under 100ms, since a PDF draws every changed page. make benchmark-content measures this and writes research/data/performance/content_speed/<machine>.csv.
OmniDiff's goal is to process 100% of all commits.
The full test dataset holds the git commit history of about 7,500 open-source git repositories,
as available on the main branch. This list of repositories comes from the Gentoo Linux
distribution. Find it in list_of_repositories.csv.
Measured over every modified code file in the most recent 50 commits of each of those
repositories - 442,530 readable before/after pairs in 25 languages, diffed with no size cap under
a 120-second budget and a 6 GB memory cap per process - 442,322 (99.95%) completed
successfully. 96 pairs ran past the budget and 112 past the memory cap; the latter are twenty
files, nineteen of them generated or embedded - tree-sitter parser tables, codegen, minified
bundles, a PNG as a C array - and one commit of a 40,000-line single-header C++ library. Given
24 GB and 300 s, 11 of the 28 files that first hit the cap complete. The run, its harness and every pair that did not complete are documented in
research/data/performance/PROVENANCE.md.
Beyond code, over every file changed in the same window - 2,018,010 changes in 7,352
repositories, pictures, fonts and archives included - 98.8% are diffed as text or by their
content (98.2% counting each repository at most 1,000 times, so that a few giant ones do not
decide it; 6,481 repositories entirely). What is left is mostly text in legacy encodings, compiled
objects, audio, and EOT fonts compressed with MicroType Express. The census is documented in
research/data/corpus_stats/PROVENANCE.md.
OmniDiff must match a human's own reading of a change, measured against the hand-authored
ground-truth mappings in src/test/data/diffs/:
- 90% of test cases with zero mismatched bytes.
- 99% of test cases with at most 1% of bytes mismatched.
Neither is met yet. 1,141 of the 2,223 fixtures carry a hand-painted ground truth - every byte of
both files labelled with what a human says happened to it - and OmniDiff's own highlighting is
compared against it byte by byte, under each of its two highlighting presets (--full, which keeps
brackets, separators and leading whitespace, and --minimal, which drops them):
| preset | zero mismatched bytes | at most 1% mismatched | mismatched bytes, whole corpus |
|---|---|---|---|
--full |
701 (61.4%) | 981 (86.0%) | 91,169 of 57.0M (0.16%) |
--minimal |
769 (67.4%) | 1,028 (90.1%) | 47,088 of 57.0M (0.08%) |
The whole-corpus rate is far below 1% because most of each file is unchanged and nobody gets that
wrong; the per-test-case numbers are the ones that count, since a reader meets the mistakes one
diff at a time. Most of what is left is not in the matching: rendering the human's own node
mapping still disagrees with the painting on 80% of the --full bytes and 67% of the --minimal
ones, so the highlighting rules own the gap more than the matcher does.
make update-painting-attribution measures this and writes one row per fixture and preset to
research/data/quality/painting_attribution.csv, which the table is counted from;
make check-painting-attribution fails if any fixture gets worse.
Other content is judged rather than painted: a human names how much each change is (invisible, imperceptible, artifacts, edited, redrawn or replaced, see supported content) and OmniDiff's verdict is compared with theirs. A test case is perfect when every verdict matches, and near-perfect when every one is at most one level off:
| Content | Judged | Perfect | Near-perfect |
|---|---|---|---|
| Pictures | 130 | 90 (69%) | 118 (91%) |
| Cursors | 13 | 11 (85%) | 13 (100%) |
| Fonts | 11 | 9 (82%) | 11 (100%) |
| Archives | 23 | 20 (87%) | 20 (87%) |
| Message catalogs | 12 | 10 (83%) | 10 (83%) |
| PDFs | 4 | 4 (100%) | 4 (100%) |
make benchmark-content counts these too.
This project uses substantial AI assistance, currently Claude Code. Most commits disclose this
with a Co-Authored-By trailer and a link to the session that produced them. This project does not
hide that fact. This project does not treat AI assistance as a lesser way to write software.
Contributions are not accepted at this time, to keep the development speed high (see CONTRIBUTING.md). When that changes, AI-assisted contributions will be as welcome as any other, disclosed the same way: whoever submits the work is responsible for understanding it and standing behind it.
See CONTRIBUTING.md for the technology overview, code-quality and testing
expectations, project structure, and what CI checks on every push and PR. AGENTS.md has
additional AI-agent-specific conventions.