baidu unlimited-ocr meets claude opus 4.8: 1998 scan, clean table

baidu's (@Baidu_Inc) unlimited-ocr is currently #2 trending on hugging face, with 70k+ downloads. we ran a test

we downloaded it, ran it locally, and pointed it at a usa today "nasdaq & amex" stock page – the matt mahoney ocr benchmark: a 200dpi grayscale scan saved down to 75% jpeg, so the artifacts are real. it contains serif and sans-serif fonts, bold text, tightly packed columns of tickers and numbers, and a logo

we then took unlimited-ocr's output, including the coordinates it produces, and passed it to claude opus 4.8 (high) with one instruction: rebuild the layout from those coordinates

the result was a clean reconstruction of a dense, multi-column financial table extracted from a degraded 1998-era newspaper scan

the pipeline is straightforward: ocr + coordinates, then llm layout reconstruction, then structured editable output

possible applications:

– old newspaper & archive scans into searchable structured data
– receipts & invoices into correctly placed line items rather than a text blob
– complex scanned tables into clean spreadsheets or html
– legacy pdfs into editable docs that match the original
– magazines & catalogs re-laid-out for translation or reprint

same two steps each time: ocr captures the pixels and their positions, the llm rebuilds the structure