Screenshot OCR: Tesseract, Apple Vision and RapidOCR, tested
95.1%
words right with Tesseract, up from 86.4%
Apple Vision got the most words right
We drew 16 screens with text we know letter for letter, from chat to code.
| Run | 1x | 2x |
|---|---|---|
| Apple Vision, no language correction | 97.1% | 97.8% |
| Apple Vision | 96.9% | 97.1% |
| RapidOCR | 96.0% | 97.4% |
| Tesseract, our command | 95.1% | 95.3% |
| Our Actor's code | 94.4% | 91.8% |
| Tesseract defaults | 86.4% | 90.2% |
| Tesseract, best model | 85.0% | 89.6% |
| Apple Vision, fast | 71.3% | 90.0% |
1x is a screen at 100% scale. 2x has twice the pixels each way, like a Retina shot.
Words right means exact words, case and punctuation kept, in any order.
The command: resize, gray, psm 6
magick shot.png -resize 200% -colorspace gray big.png
tesseract big.png - --psm 6
ImageMagick 7 and Tesseract run on Windows, Mac and Linux. On ImageMagick 6, use convert in place of magick.
On the 2x shots, we left out -resize 200%: 95.3% of words right.
| Tesseract | Words right | Letters in order |
|---|---|---|
| Defaults | 86.4% | 88.0% |
| + 2x resize | 89.9% | 88.6% |
| + gray copy | 95.2% | 93.4% |
| + --psm 6 | 95.1% | 99.3% |
--psm 6 reads the shot as one block. It kept our sheet in rows; the default mode read it column by column.
Tesseract skipped the blue chat bubbles
We drew 11 blue bubbles and 10 gray ones, each at 1x and 2x.
At its defaults, Tesseract read 2 of the 22 blue ones, white text on blue. It read all 20 gray ones.
With a gray copy first, it read all 22. Apple Vision, RapidOCR and our Actor's code read every bubble too.
Chat and terminal gained the most
| Screen | Defaults | Command |
|---|---|---|
| Chat | 52.9% | 95.1% |
| Terminal, 13 px | 73.7% | 95.8% |
| Code, 13 px | 78.5% | 82.5% |
| Sheet, 13 px | 83.8% | 98.1% |
| Small UI text, 12 to 14 px | 93.2% | 97.7% |
| Dark mode, 16 px | 95.0% | 95.8% |
| Web page, 16 px | 97.5% | 96.8% |
| E-book, 18 px serif | 98.7% | 99.4% |
The web page lost a little. Every other screen gained.
Code: turn off language correction
Apple Vision corrects words by default. On our code shots at 1x, it got 81.7% of words right with it on, and 89.4% with it off.
RapidOCR read code best: 95.5% at both scales.
ocr.swift: Apple Vision on the command line
import Foundation
import Vision
let req = VNRecognizeTextRequest()
req.recognitionLevel = .accurate
req.recognitionLanguages = ["en-US"]
req.usesLanguageCorrection = false
try VNImageRequestHandler(url: URL(fileURLWithPath: CommandLine.arguments[1])).perform([req])
for r in req.results ?? [] { print(r.topCandidates(1).first?.string ?? "") }
Run it with swift ocr.swift shot.png. Mac only.
swift comes with the Xcode Command Line Tools (xcode-select --install).
RapidOCR in Python
python3 -m venv .venv && . .venv/bin/activate
pip install rapidocr==3.10.0 onnxruntime
python3 -c 'from rapidocr import RapidOCR; print("\n".join(RapidOCR()("shot.png").txts or []))'
What to watch for
- The best model did not help: 85.0% on 1x shots, against 86.4% for the default model.
- Inverting dark mode first barely moved Tesseract: 95.0% to 95.3% on the dark mode shots at 1x.
- Apple Vision's fast level got 71.3% on 1x shots.
- RapidOCR missed much of the 1x terminal: 67.4%, against 94.7% at 2x.
- Our own Actor got 60.4% of words right on the 2x sheet, against 100% at 1x. It is on our fix list.
How we ran it
- Text from two 1892 Sherlock Holmes stories on English Wikisource, public domain in the US. Code from two of our own scripts. Terminal output from our Mac. The sheet holds random numbers.
- Headless Chromium drew each screen at 1x and 2x: 32 PNGs. A script checked that each shot's visible text matched its truth: 32 of 32.
- Words right: words split on spaces, case and punctuation kept, curly quotes as straight, matched in any order. Letters in order: the longest common subsequence of all characters but spaces.
- A chat bubble counts as read when half or more of its words are right.
- Tesseract 5.5.4 with eng from tessdata_fast 4.1.0 (best: tessdata_best 4.1.0). ImageMagick 7.1.2. Apple Vision: VNRecognizeTextRequest, accurate, en-US, macOS 26.6.
- RapidOCR 3.10.0 with its PP-OCRv6 small models. Our Actor's code ran on the Mac with its defaults, except plain text and tab-separated table rows.
- Text edges are gray, as on a Mac. Windows ClearType adds color fringes, and we did not test that, or Windows' own OCR.
- Licenses: Tesseract, tessdata and RapidOCR are Apache-2.0. We ran them and did not ship any of them.
- We tried 18 runs in all. Tesseract, RapidOCR and the Actor ran on one thread each, 2 Tesseract shots at a time, at low priority.
Many screenshots? No install
Run it yourself. Our Apify Actor reads PNG, JPEG, WEBP and TIFF images and scanned PDFs into Markdown or text. $0.006 an image.
Or send us the scan. Tack, our AI, does the job for $29, up to 50 pages. First page free.
Questions people ask
What is the best free screenshot OCR?
On our 16 screens at 1x, Apple Vision with language correction off got 97.1% of words right, RapidOCR 96.0%, and Tesseract 95.1% after a resize, a gray copy and --psm 6.
Why does Tesseract miss text in chat screenshots?
At its defaults it read 2 of our 22 blue bubbles, white text on blue. With a gray copy first, it read all 22.
Should I upscale a screenshot before OCR?
For Tesseract on a 1x screenshot, yes: a 2x resize took words right from 86.4% to 89.9%. A 3x resize did no better. We did not resize the 2x shots: as they were, Tesseract got 90.2%.
Can Tesseract read dark mode screenshots?
Yes. It got 95.0% of words right on our dark shots at 1x. Inverting the image first made it 95.3%.
How do I OCR a screenshot on a Mac for free?
Apple Vision comes with macOS. Our short Swift script runs it: swift ocr.swift shot.png. With language correction off, it got 97.1% of words right on our 1x shots.
Related guides
- Free TextSniper alternatives on Mac, tested on 32 screenshots
- Apple Vision OCR vs Tesseract, tested on 523 scans
- RapidOCR in Python: tested vs Tesseract and PaddleOCR
- Image preprocessing for OCR: what helps Tesseract, tested
- Tesseract psm modes: which one to use, tested (2026)
- Seven segment OCR, tested: Tesseract and ssocr on 47 displays
- Tibetan OCR, tested: Tesseract on 27 pecha and 11 book pages