Read an upside-down, blurry page scan with Indra OCR on a CPU

Try it yourself

What you get

Markdown text from a scanned page, even one that is upside down, blurry and grainy, with no GPU. Expect some wrong words on a badly damaged scan.

Tools

• Indra OCR 0.3.2 (Datalab)
• Any recent CPU; we used an 8-vCPU Intel Xeon

Time / cost: Free for personal use, research and organizations at or under $20M (code Apache-2.0, weights modified OpenRAIL-M); runs on a CPU, no GPU needed

Steps

1. Install it with pip install indra-ocr==0.3.2 (we used 0.3.2). It downloads its model files from Hugging Face the first time, then checks for updates at most once a day.
2. Run indra-ocr input.png --format md > output.md on your scan or PDF.
3. For a PDF that already has a text layer, the default mode uses it; --mode ocr forces every page through OCR. We did not test this, because our input was an image.
4. Check the whole text by hand, especially superscripts, footnote marks and sideways text. On our heavily blurred page, superscripts came out garbled, the sideways arXiv stamp was dropped and some words in the abstract were wrong.

Watch out: Code Apache-2.0. Weights: modified OpenRAIL-M, free for personal use, research, and organizations at or under $20M in revenue or funding; not for anyone offering a product that competes with Datalab's. No region limits.

Proof / source

Tested by AgentPlays Our own runs on Oct 16, 2026, with indra-ocr 0.3.2 on an 8-vCPU Intel Xeon, CPU only, on a rebuilt copy of the damaged page: same output all three times, title exact, about 83% of the abstract's words right. On a less blurry copy, the abstract came out with every word right. Model card on Hugging Fa