Skip to content

Releases

Vision OCR 1.16.0 — Accurate mode reads difficult pages with stronger on-device readers, chosen by kind of page

Choose a tag to compare

@charlesapetersen charlesapetersen released this 16 Oct 23:17

Accurate mode: stronger on-device readers, chosen by kind of page. Settings has a new Accurate mode box, off by default and best left off for everyday scans. Switched on, it reads difficult print with an extra on-device reader, newspapers above all, and fits that reader's words onto the lines Vision finds, so the text layer still sits on the words. A table beside it shows how accurate, how fast, how much memory and how large a download each reader is, and a slider sets the most memory a reader may use. Readers are downloaded only after a confirmation that names the total (about 1.3 GB for Falcon-OCR with the shared runtime and layout model, 0.9 GB for PaddleOCR-VL). Every file is checked before it is installed, and nothing is downloaded while the mode is off. "Measure on this Mac" times each downloaded reader on a few pages, so the speed column shows estimates for your Mac. Accuracy is not measured there.

Four kinds of page. Choose pages sets a reader for each kind: Newspapers, Photographed pages, Handwritten pages and Other documents. The app sorts each page into a kind by itself. It spots a camera's photo of a page from the camera's tags or from how the page was lit, and handwriting from the shape of the ink in its lines. Sideways and upside-down pages are read upright.

Handwriting. A new reader for handwritten pages, LightOnOCR-3 (a 5.2 GB download that needs the memory limit set to 8 GB or more), missed about 96 words in 1,000 on 19 photographed letters and notes, a small sample. On the same pages Vision missed 561 and Falcon-OCR 409. Words read from a hand mostly cannot be placed on Vision's lines, so by default a handwritten page is followed by a page of the reader's text, which can be selected, copied and searched. A setting turns this off.

Faster batches. On a Mac with 16 GB of memory or more, two pages can be read by Falcon-OCR or PaddleOCR-VL at once, and the longest pages go first. In a ten-page test this finished about 18% sooner with PaddleOCR-VL and 13% sooner with Falcon-OCR.

Smaller photos of pages. A dropped photo or image of a page is now compressed like a scanned page. A phone photo of a typed letter went from 830 KB to about 87 KB with the same text. A Photographed pages slider sets how much of the paper behind the words is kept, and a camera's photo now gets a Letter-sized page.

Cleaner type and better selection on scans. Pale typewriting, and typing on grey or yellowed paper, comes out as solid as the scan shows it. Black type no longer comes out near-bold. Faint pencil and dark marks that are not words, such as rules, stamps and ink handwriting, stay legible. Lines that were missing from the text layer, or garbled at their ends, now read as printed. A drag across the top half of a line no longer copies the line above, and a drag down one column of a magazine page stays in that column. On newspapers with a rule printed between the columns, a line read on into the next column is split into its two columns (C28, C41, C52, C53, C54, C55, C56).

Plainer settings. The settings now explain themselves without terms like "invisible text layer" or "JBIG2".

The full list is in CHANGELOG.md.

Model files

Model files Pre-release
Pre-release

Choose a tag to compare

@charlesapetersen charlesapetersen released this 14 Oct 16:48

Runtime and DocLayout-YOLO layout model (Apache-2.0 weights) for the accurate mode. Not an app release. See NOTICE.txt for licences.

Vision OCR 1.15.0 — more of each scanned page is selectable, columns copy in order, and files come out smaller

Choose a tag to compare

@charlesapetersen charlesapetersen released this 29 Sep 21:02

More of each scanned page can be selected. A block of text missed beside a column that was read is now recognised. Lines that recognition had fused into one garbled line come out as the real lines, and lines that stopped partway across the page are whole. Footnotes in small type and table columns that were skipped on the first read are read again and can be found. Scans cut into narrow strips under a vendor's header, as ProQuest newspaper pages are, are now recognised, and newspaper pages pasted up from strips are read one strip at a time at the strips' own resolution (C33, C51, C43, C39).

A drag down a column stays in that column. On pages set in two or three columns a selection runs down one column before moving to the next, lines read straight across the columns are split, words broken at the end of a column join up with the right half, and a headline or byline keeps its own height and width. Text printed sideways can be selected along its printed line. Copying no longer repeats the tail of a word hyphenated at a line end. On scans of two-page spreads the selectable text sits on its words again (C34, C40, C52, C36, C44, C42).

Numbers and typewritten pages read better. Each line of a page rebuilt in black and white is also read from the grey scan and the better reading is kept. Numbers in scanned tables that were garbled ("ถ.120") come out right, and two table cells read as one are split. Such pages take up to twice as long to recognise (C50).

Colour and type quality. Red headings on scanned text pages keep their colour. Light coloured type on a page rebuilt in black and white, such as a pink notice, now comes out black and can be selected and searched. Text on colour pages is dark and solid again, which was a regression in 1.14.0, and it is sharp in Preview (C32, C49, C31, C38).

Smaller files. Scans that already arrive as black-and-white JBIG2, including JSTOR and NBER downloads and ProQuest theses, keep their own compression. On 51 test documents the output went from 258.8 MB to 197.5 MB. Scanned pages whose own images are already smaller than the app's rebuild, such as ProQuest's layered pages and Internet Archive book scans, keep those images, so ProQuest pages that came out 1.5 to 2.2 times their original size now come out at about the original's size. Scans that carry an outline and a born-digital cover page are no longer several times larger than they need to be, and every searchable copy is written more compactly (C37, C47, C35).

The searchable copy keeps more of the original. It keeps the printed page numbers, so Preview's page field and Go To Page use them, and it keeps the title and author. When highlights, underlines and notes are not carried over, the result now says how many were left and of which kinds (C48, C45).

Fewer crashes and less noise. English pages no longer carry stray Arabic letters that recognition read from smudges. On one newspaper page such a line made Preview crash when a click landed at the end of a cartoon's signature (C46).

The full list is in CHANGELOG.md.

Vision OCR 1.14.0 — skipped blocks of text are now read, and spot colour is kept

Choose a tag to compare

@charlesapetersen charlesapetersen released this 26 Sep 04:27

Scanned pages printed with a spot colour keep it. A page with a red rule, a coloured banner or a
colour photograph on a page of type was published in grey, because only pages with a lot of colour
overall kept theirs. Automatic now also keeps colour when enough of the sheet itself is coloured,
ignoring a coloured scanner border. Across the test library 100 pages keep their colour this way, at
about 7% more bytes each.

Scanned pages where whole blocks of text could not be selected are now read in full. Vision, given a
whole page at once, sometimes skips entire paragraphs of clean type; on a six-page 1951 journal scan about
half of each page had no text layer. When a page comes back with inked areas that have no words over them,
it is now read again in overlapping strips and the missing lines are added, and a line the first reading
found is not added a second time. On that document the selectable words went from 2,075 to 3,526. Pages
read fully the first time cost almost nothing extra; pages that needed the second reading took about three
seconds longer each, and so may a page with a picture on it (BUGS.md C30).

Short born-digital pages keep their exact text, and a low-resolution scan with an old text layer is read
again instead of kept as it was.
A part title, half-title or nearly blank page with only a few words of
real text used to be turned into a picture and re-read. It is now kept as it is, provided nothing on it
could be a picture: no image, pattern, note or outlined lettering. A short page that does carry one of
those is still re-read, and the run report now names it. The other half of the fix matters more. The test
for "this page is already a picture" missed scans under 900 pixels wide and scans whose image is written
inline. When such a scan carried an earlier OCR layer, it was taken for a born-digital page, kept with its
old text, and never read by the app. Pages whose text is mostly invisible are now always re-read (BUGS.md C29).

A mixed PDF — born-digital pages among scanned ones — now keeps the compact page compression for its
scanned pages.
Until now, one page kept as it is turned that compression off for the whole file, and the
layered-image saving with it, so a mixed file came out much larger: measured on a ten-page book with a
born-digital first page, 1.4 MB instead of 0.4 MB, and nine tenths of that difference was the layering rather
than the compression itself. The kept page is now put back into the compressed document instead of the
document being rebuilt the old way. Two files still take the old, larger route, deliberately: one that has
bookmarks, because the step that puts the page back cannot carry a bookmark tree across, and one with a
highlight or a note of your own on a born-digital page, because that page arrives already carrying the mark
and copying it again would put two there. Both come out correct, only larger. Measured over the 42 such
documents in the test corpus: 16 have bookmarks and 4 have a mark of their own on such a page, so 22 are
refused for neither reason (BUGS.md C29).

A page that already has real text of its own is now kept as it is, instead of being turned into a picture
and re-read.
This is the fix for the report described in the next entry. Making a mixed PDF searchable used
to rebuild every page as an image, so a born-digital cover sheet or title page came out carrying OCR of a
picture of its own words — American Sociological Association became AMFAKAN FOCAX ONCAL ASSOXUTION, and a
search that worked on the original failed on the copy. Such pages are now copied through untouched: their
text, their fonts and their colour images arrive exactly as they were, and the report says which pages were
kept as well as which were not; what is measured is the page's text, character for character, not its fonts
or its image streams. Scanned pages are rebuilt as before — re-reading those is what the app is for — with
one caveat that is a limitation rather than a promise: the test for "already a picture" needs the page's
image to be at least 900 pixels wide and drawn as a normal XObject, so a very low-resolution scan, or one
whose image is written inline, can be mistaken for a born-digital page and kept with its old OCR instead of
being re-read. Two more limits (the first of which the entry above now removes for most files): a mixed file is written
with the older, larger page compression and loses the layered-image saving with it, so it can come out
noticeably bigger; and a short born-digital page — under
about two lines of text — is still rebuilt, because two lines is the bar the app uses to tell a real page of
text from a plate or a part title (BUGS.md C29). (The first caveat, for scans with an old invisible text
layer, and the short-page limit are both removed by the short-page entry above.)

The run report now tells you when a page that already had real text of its own was turned into a picture
and re-read.
Some PDFs are a mixture: a library or a publisher puts born-digital pages — a download cover
sheet, a title page, sometimes whole chapters — in front of or among scanned ones. Making such a file
searchable rebuilds every page as an image and recognises it, so on those pages exact text is replaced by
OCR of a picture of that text, and a phrase you could previously search for may no longer be found. That
happened silently: the page count is right and the file has a text layer, so nothing looked wrong. The
report now names those pages, with the complete count first. Measured over the 233-document test corpus,
392 pages in 42 documents are of this shape, and on 38 of the 42 nothing else in the app mentions it at all
— including one 300-page book of which 167 pages are born digital (BUGS.md C29). This is a report and not
yet a fix: those pages are still rebuilt — the entry above is the fix, and it landed in the same unreleased
batch, so on a page of two lines or more this note now names pages that were kept rather than pages that
were lost. Pages that only ever carried OCR text are deliberately not
reported — re-reading those is what the app is for.

A page with unrecognised words on it is no longer treated as a page of nothing but text. When every
mark on a page is text the recogniser boxed, both tone layers are stored at an eighth and a sixteenth of
the page's resolution — which is the right thing to do, and was being done to pages that also carried
type the recogniser had missed. Words present in your scan then survived only in that shrunk
background, illegible, and in no text layer either. The layering decision now also asks whether the ink
outside the recognised words is shaped like lines of type, at the page's own type scale, and keeps the
resolution when it is (BUGS.md C28). Measured over the corpus: about 18 pages in 16,987 stop being
shrunk, for roughly 2.7 MB more output in total, and 13 of the 16 pages measured losing content keep it.
No page loses anything it kept before — the change can only ever store a page at more resolution than
it did, never less. Three pages that lose only a hand-made mark — a pencilled annotation, an inked
margin number — are still not covered, and one of those is invisible to this signal for a reason
recorded in the register.

"Smallest files" no longer promises that nothing is lost from your photographs. The setting's
description used to end "though nothing is lost from them", and that is measurably untrue: on a
scanned page whose small print falls outside the words the recogniser boxed, a third of the
resolution can be the difference between legible and not (BUGS.md C28 — a printed correlation matrix
whose last column reads at half resolution and does not read at a third). The clause is gone and
nothing replaced it; the rest of the description is unchanged, and nothing about what the setting
does has changed.

The run report now names the pages whose non-text ink was stored at an eighth. On a page whose
ink is all recognised text, both tone layers are shrunk hard — and anything the recogniser missed
on such a page lives only in that shrunk background. Measured over the whole population
(BUGS.md C28): of the 73 corpus pages the shipped bar still treats this way, sixteen lose content.
Until now nothing said which pages those were, on any run. A batch that publishes such a page now
writes one line into its run report naming the pages and how much of each page's ink fell outside
the recognised words — for example scan.pdf: 3 page(s) were read as holding nothing but recognised text, so the rest of the page was stored at an eighth — anything the recogniser missed on them is degraded: p4 (2.1% of its ink); p6 (0.0% of its ink); p7 (0.0% of its ink)

Every such page is named, with no threshold on that fraction: the same campaign showed that the
pages which lose content and the pages which do not cannot be told apart by it, and one page
measured losing a pencilled annotation reads 0.0%. The conversion still succeeds and the file is
still whole — this reports the degradation, it does not yet prevent it.


Requires macOS 13 (Ventura) or later, on an Apple Silicon or Intel Mac. Drag Vision OCR onto Applications. The app is not notarized, so the first launch needs System Settings ▸ Privacy & Security ▸ Open Anyway. Control-click ▸ Open no longer works on current macOS.

Vision OCR 1.13.1 — lines of prose the recogniser missed are no longer destroyed

Choose a tag to compare

@charlesapetersen charlesapetersen released this 20 Aug 16:44

A page the app judged to be nothing but type could lose whatever the recogniser did not box. On such a page both tone layers are stored at a fraction of their resolution — a large saving, and safe as long as "nothing but type" is true. What decides it is how much of the page's ink Vision found no words for, and that threshold was too permissive. Pages carrying a figure, a table's rules, a hand-drawn mark — and, as it turned out, mostly words — were being called all-text and then destroyed.

Measured over the whole test library before the number moved: 16 pages of 2,129 sampled stop being shrunk, across 10 documents of 233. Thirteen of the sixteen were rendered at the old and new settings and read by eye — eleven lose something and eight lose content outright: thirteen values out of a statistical table on one page, seven lines of running prose on another, two lines of 1939 typescript on a third. The price is about +0.55% of total output size.

The three pages this was first reported on were re-rendered and looked at. The line drawings that had been erased, or left as a ghost, are whole and legible again — where before they were smudges.

What this does not do. Moving one number rescues the pages that sat just above it. 73 of the sampled pages are still stored this way, in 22 documents; eight of those were rendered and read, and four of the eight still lose content — two lines of prose each on two pages, figures out of a statistical table on a third, a handwritten signature on a 1939 letter on the fourth. Sorted by the quantity this threshold tests, the pages that lose content and the pages that do not are interleaved, so no value of this number separates them. The real fix is to the 1-bit stencil, which today keeps only what the recogniser found words for; that work is open and is sequenced for 1.14.0, along with a separate defect where spot colour is discarded.

Photo detail = Maximum was exempt from this shrink before the change and still is.


Requires macOS 13 (Ventura) or later, on an Apple Silicon or Intel Mac. Drag Vision OCR onto Applications. The app is not notarized, so the first launch needs System Settings ▸ Privacy & Security ▸ Open Anyway — Control-click ▸ Open no longer works on current macOS.

Vision OCR 1.13.0 — pale drawings and photographs come out intact

Choose a tag to compare

@charlesapetersen charlesapetersen released this 18 Aug 03:10

The two ways Automatic could destroy what was on a page — both listed as known in 1.12.0's notes — are fixed. If you kept a document in Grayscale to stay clear of them, you no longer need to.

A photograph could come out as a solid black shape. A continuous-tone plate covering roughly a fifth of a sheet was thresholded to pure black and white, taking the line of text above it with it. It slipped past two separate checks at once, because a picture that large dilutes its own tone across the page and both limits had been calibrated on pages a picture dominates. The page's tone is now measured inside each large mark rather than across the whole sheet — the same test, asked about the right region. Found on real pages, not in theory: a credited art photograph in one book was rendering as a blob, and one thesis in the test library had 87 more.

A pale line drawing could be erased. Not blurred — gone, with the text beside it untouched, which is what made the result look fine. Faded ink, light pencil, a pale wash. Four attempts to fix this by brightness were built and refused, because the range of tones too light to be ink and too light to be a photograph holds three different things: drawings you want kept, decorative shading you do not mind losing, and show-through from the back of the sheet you actively want gone. No brightness test separates those. What separates them is where the mark is — show-through lands on top of the page's own type, because it is type seen from behind, while a figure sits where the type is not.

Pages no longer borrow their resolution from a neighbour. On documents whose pages share one internal resource list, a page could be rebuilt at whatever resolution the largest image anywhere in that list implied — so in one journal article every page of type was rendered at 370 DPI because page 41 has a plate on it. The app now looks at what each page actually draws. 130 pages across the test library changed, every one of them measured, and the affected documents come out smaller with the same words recovered.

Three smaller things you might have noticed:

  • A line of text you could not highlight — a whole line could be drawn at 5% of its width.
  • Words welding together when one part of a line was much shorter than the next.
  • A rebuilt copy showing something the original had hidden at the edge of the sheet.

Requires macOS 13 or later. Unsigned by choice: System Settings ▸ Privacy & Security ▸ Open Anyway after the first launch attempt.


How this one was checked

Every document in the test library went through the whole app before this shipped: 233 of 233 succeeded, 0 failed, 16,987 pages, 34.2 million characters. Every output was then compared against its original pixel by pixel, not by its text — a check added after an earlier release gate proved it could pass a page whose image had been destroyed.

Forty pages came back flagged and none of them was a fault. Thirty-five are pages that are near-blank in the original and near-blank in the copy. Five were flagged for looking unlike their originals, and three of those were rendered and looked at by eye rather than trusted to the score: a dedication page reading "For Rosie", an Internet Archive boilerplate page, and a two-page spread of body type. In all three every word survives, including one copyright page's fine print, and what fell away was a grey scan background or a pale decorative watermark. That is the score's shape, not a defect — on a page whose entire content is two words, cleaning the paper changes almost every pixel while losing nothing.

The defect register behind this release is empty for the first time: 0 open of 163 entries. Every fix above is written up with its evidence in BUGS.md, including the fixes that were built, measured and refused.

Vision OCR 1.12.0 — big scans come out small again

Choose a tag to compare

@charlesapetersen charlesapetersen released this 14 Aug 01:56

A 568-page library scan went in at 31 MB and came out at 437 MB. It now comes out at 35 MB, with exactly the same text.

size
the original scan 31 MB
1.11.0 437 MB (14.0x)
1.12.0 35 MB (1.13x)

Two things were wrong, and the second was the interesting one.

Colour pages could not be stored in layers. Pages that mix text with a picture are normally split into a sharp 1-bit layer for the text and a compressed one for the picture — but colour pages were the single kind that could not be, so a scan the app read as coloured throughout kept every page as one full-resolution colour photograph of itself. 669 KB a page where three layers cost 110.

And even layered, the layers were far more detailed than paper needs. The 1-bit layer carrying the text already matched the Internet Archive's own scan of the same book to within 2%. The other layers cost 40.7 MB against their 4.3. Pages whose ink is all text now store their paper far more coarsely — and the app works that out per page, so there is nothing to set and nothing to think about. A page carrying a picture is left exactly as it was.

On the 232-document test corpus: 721 MB against 792, every document either smaller or byte-for-byte identical, not one larger, and the recognised text unmoved. Every photograph-heavy book in it came out identical to the byte. Scans of faint typescripts and manuscripts — the ones this app was mistaking for photographs — came down hardest: one 1941 speech to a fifth of its size.

If you keep Settings ▸ Photo detail at Maximum, nothing about your photographs changes at all; that setting is now honoured in full.

Also fixed: the updates section of Settings appeared twice, and the Start from preset buttons gave no sign they had done anything — they now say what changed, and say so out loud for VoiceOver instead of hiding it in a tooltip you need a mouse to find.

Requires macOS 13 or later. Unsigned by choice: System Settings ▸ Privacy & Security ▸ Open Anyway after the first launch attempt.

Every defect above is written up with its evidence in BUGS.md — R49 through R53, including a fix that was built, measured and refused, and three defects found by reviewing this release's own code before it shipped.


Known, and found after this release went out

Two ways Automatic can route a page wrongly were found on 2026-08-13 while measuring something else, and both lose what is on the page rather than merely degrading it. Neither is caused by anything in 1.12.0 — both are in 1.11.0 and earlier too.

  • A pale drawing can be erased. Faded ink, light pencil, a pale wash: if a mark is too light to count as ink and too light to count as continuous tone, the page is treated as plain text and the mark is rendered as blank paper. The text beside it is unaffected, which is what makes the result look fine.
  • A continuous-tone plate covering roughly a fifth of a page can come out solid black, taking the line of text above it with it.

If a document matters and carries pale artwork, compare the output against the original — or use Grayscale, which sends nothing down the 1-bit route and cannot do either of these things.

Both are written up with rendered before-and-after evidence as R56 and R57 in BUGS.md, including a luminance-based fix that was built, measured over four rounds and refused — the blind zone turns out to hold pale drawings you want kept, decorative shading you do not mind losing, and show-through from the back of the sheet you actively want gone, and no brightness test separates those three.

Vision OCR 1.11.0 — nothing to install, and large batches are fast again

Choose a tag to compare

@charlesapetersen charlesapetersen released this 13 Aug 15:00

Nothing to install any more. Vision OCR used to carry a copy of a command-line program called mac-ocr inside it to do the recognition — 2.4 MB of someone else's binary, which earlier versions asked you to install yourself in a Terminal. It now calls Apple's Vision framework directly. If you installed anything for an older version, you can remove it.

This is mostly invisible, and deliberately so: the recognised text is the same text. Both routes were run over 52 documents and 4,140 pages before the change was made — 9,211,704 characters against 9,254,956, a difference of +0.47% in favour of the new one.

And large batches are faster than they have ever been. The 232-document release gate:

1.10.1 1.11.0
documents succeeded 232 / 232 232 / 232
characters recovered 34,148,681 34,204,948
output size 792 MB 792 MB
minutes 75 48

Recognition now runs in a small helper program of Vision OCR's own, one per file being processed. It is not the old dependency coming back: it ships inside the app with nothing to install, and it is handed page images this app has already drawn rather than a PDF for something else to re-render. If it is ever missing or goes wrong, the app simply recognises the document itself and says so in the log — nothing fails and nothing is lost.

That number was measured on a machine that was busy with other work, so it is a floor rather than a best case.

Also fixed in this release

  • A very large sheet no longer needs a special case. The 200-megapixel limit belonged to the program that has just been removed, not to Vision — which reads a 216-megapixel page without complaint.
  • The page that is read is now exactly the page that is written. The app used to render each page, save it into a PDF, and hand that PDF to another program, which rendered it again at a resolution of its own choosing. That round trip no longer happens.
  • A photograph that says it is sideways is now read as sideways. Image files carry a rotation flag that was being ignored when extracting text; the text came out right but the word positions did not.
  • Language detection works as intended again when you have not named a language, which is the default.

799 checks in the suite, and the release gate ran over all 232 test documents. Written up as BUGS.md R40, with R41–R47 from reviewing that work afterwards.

Two features were investigated and declined on measurement rather than shipped: correcting crooked scans loses text rather than recovering it, and multi-column pages already come out in reading order. FEATURES.md carries both sets of numbers.

Installing. Vision OCR is signed but not notarized, so the first launch is blocked. Open it once, then go to System Settings ▸ Privacy & Security and click Open Anyway. Requires macOS 13 or later on Apple silicon. Nothing else to install.

Vision OCR 1.10.1 — a very large sheet could fail instead of being read

Choose a tag to compare

@charlesapetersen charlesapetersen released this 12 Aug 19:34

A very large sheet could fail instead of being read. Vision refuses to render a page above 200 megapixels, so Vision OCR works out the highest resolution each document can safely be recognised at and asks for that, rather than letting the whole file fail. On Automatic that safeguard could not do its job — it was comparing against the wrong number, and a page whose safe limit landed between 300 DPI and its own resolution was handed over unprotected. A 20 × 30 inch sheet scanned at 600 DPI came back as an error instead of as text.

Now fixed. Nothing else changes: no document in the 232-item test library is affected, because the safeguard only speaks up for sheets far larger than a book page. Every file this app has processed comes out exactly as it did.

A correction to 1.10.0's release note. It suggested setting Page DPI to 300 for very high-resolution scans. Measured properly afterwards — 52 documents, 4,140 pages, every rendering resolution from 150 to 400 — Automatic recovers more text than any fixed value. An explicit 300 costs 0.17% overall and makes 26 of 52 documents worse, and the gap is widest on exactly the high-resolution scans the advice was aimed at. It helps a particular document here and there and hurts more than it helps.

Leave Page DPI on Automatic unless you have measured your own material.


765 checks in the suite. Written up as BUGS.md R39, including the measurement that refuted the fix originally proposed for it.

Installing. Vision OCR is signed but not notarized, so the first launch is blocked. Open it once, then go to System Settings ▸ Privacy & Security and click Open Anyway. Requires macOS 13 or later on Apple silicon. Nothing else to install.

Vision OCR 1.10.0 — scans come out smaller than they went in

Choose a tag to compare

@charlesapetersen charlesapetersen released this 12 Aug 17:32

Files that used to come out bigger now come out smaller. Some scans grew when this app processed them — one book went from 16 MB to 156 MB, more than nine times its original size, and it was not alone. Across the whole 232-document test library the output was 1,039 MB against 1,198 MB going in; it is now 792 MB.

The cause was a page of dense small type being mistaken for a photograph. One of the three things the app looks at is how much ink is on a page, and a broadsheet of eight-point classifieds is a great deal of ink — so those pages were given a photographic layer on top of the sharp black-and-white one, carrying nothing the sharp version did not already have. Heavy ink now has to be corroborated by real photographic tone. Actual pictures were checked page by page at full size and are unaffected.

One honest caveat, found straight after release. Across the library the recognised text changed by −0.05%, and most documents gained. But one book lost 2.7% of its text, concentrated on about a dozen pages — and the cause turned out not to be the change above. On some high-resolution scans Vision recovers less text than it would at a lower rendering resolution, and Automatic leaves that choice to the recogniser. Pages that moved to black-and-white met that behaviour more often than they used to.

Correction (1.10.1). This note originally advised setting Page DPI to 300 for very high-resolution scans. Measured properly afterwards — 52 documents, 4,140 pages — Automatic recovers more text than any fixed value, and 300 makes more than half of those documents worse. That advice is withdrawn. Leave Page DPI on Automatic.

A written record of every batch. The results pane always lost its record when the window closed. Each finished run now leaves a report in ~/Library/Logs/VisionOCR: every input, where its output went, what happened to it, the settings that produced it, how long it took. For an overnight run over material that may not be re-scannable, that is the difference between "something failed last night" and knowing which document and why. On by default; Settings ▸ Behaviour has a Show Reports button. Your password is never written into it.

Retry just the files that failed. Four failures out of seventy-eight used to mean finding those four and dragging them in again. The results pane now offers Retry N Failed.

The language list comes from your Mac. An unrecognised BCP-47 code does not quietly do nothing — it fails every file in the batch. There is now an Add menu listing what this Mac actually supports, and a warning naming any code that will fail before you start.

Most importantly: "Fast" supports far fewer languages than the normal recogniser — six against thirty on macOS 26.6. Turning it on with Japanese, Russian, Chinese, Korean, Arabic or twenty others selected turned a working setup into a run where nothing succeeded, with no indication why. It says so now, the moment you tick the box.


232 of 232 documents through the full-corpus gate, 34.1 million characters recovered, no failures. 758 checks in the suite.

Installing. Vision OCR is signed but not notarized, so the first launch is blocked. Open it once, then go to System Settings ▸ Privacy & Security and click Open Anyway. Requires macOS 13 or later on Apple silicon. Nothing else to install.