The left column was read top to bottom before the right column, then both were flattened into one Markdown stream.
Document title
Kept
The main title became a level-one Markdown heading with the words and capitalization preserved.
Subtitle divider
Changed
Deep Read inserted a horizontal rule after the subtitle, even though the source only had a drawn divider near the top.
Bottom table
Kept
The sampling schedule became a valid five-column pipe table with Station, Date, Lead, Result, and Action headers.
Footer
Kept
The page footer was included as ordinary Markdown text instead of being recognized as removable page furniture.
Visual columns
Lost
Column geometry, table ruling, and the page border were not recreated because the output favors Markdown semantics.
Runtime
Kept
The 1500 by 1800 PNG completed in 7.8 seconds and used one credit for the single rendered page.
Two-column reading order Changed
The left column was read top to bottom before the right column, then both were flattened into one Markdown stream.
Document title Kept
The main title became a level-one Markdown heading with the words and capitalization preserved.
Subtitle divider Changed
Deep Read inserted a horizontal rule after the subtitle, even though the source only had a drawn divider near the top.
Bottom table Kept
The sampling schedule became a valid five-column pipe table with Station, Date, Lead, Result, and Action headers.
Footer Kept
The page footer was included as ordinary Markdown text instead of being recognized as removable page furniture.
Visual columns Lost
Column geometry, table ruling, and the page border were not recreated because the output favors Markdown semantics.
Runtime Kept
The 1500 by 1800 PNG completed in 7.8 seconds and used one credit for the single rendered page.
Sample conversions
Municipal briefing page
Input
PNG with a title, subtitle, left and right text columns, a small sampling schedule, and 'packet page 4 of 12' footer.
Output
Markdown kept the title, read the columns in a sensible order, built a five-column pipe table, and kept the footer as body text.
Screenshot note card
Input
PNG screenshot with one heading, a paragraph, three bullets, and a compact owner/date/status area.
Output
The Markdown-style transcript uses a heading and list markers. The colored owner strip becomes text lines rather than a callout block.
A real PNG → Markdown run
We converted our PNG sample (219 KB) with AI Deep Read under AI Deep Read on 2026-09-11. It took 2.8 s and produced a 499 B Markdown file, 100% smaller than the input. The engine raised 1 warning, listed below.
# Feedback Atlas — Q3 activity trail
Prepared for: Board of Directors · 12 Sep 2026
## Headline *numbers*
| Metric | Q2 | Q3 |
|---|---:|---:|
| ARR | $4.1M | $4.9M |
| Net revenue retention | 108% | 113% |
| Support CSAT | 93% | 94% |
| Refunds | (12,400) | (9,850) |
## What went well
- Enterprise pipeline doubled after the SOC 2 report shipped.
- Support CSAT held at 94% while ticket volume rose 31%.
- Two renewals slipped into October. Owner: VP Sales.
…
Is your file a scan?
PNG scan detection is straightforward because the file is pixels only. Deep Read costs 1 credit for a single PNG and adds structure such as headings and GFM tables, but it may keep footers that a human would discard.
Options and when to change them
extractTablesAsGfm
Keep this enabled when a PNG includes schedules, ledgers, or forms that should become Markdown pipe tables.
preserveLayout
Enable it if visual column placement matters more than the default clean reading order.
languageHint
Set the expected language when headings or table labels are short enough to be ambiguous.
Do it yourself
Tesseract + manual Markdown cleanup
bash
tesseract page.png stdout --psm 6 > page.txt
# Add Markdown headings, lists, and table pipes during review.
The local OCR command extracts words but does not infer the GFM table or column reading order that Deep Read produced.
PNG to Markdown by API or MCP
The HTTP API and the MCP server run the exact engine this page uses (AI Deep Read); the response carries the same warnings the web converter shows. Check support first with GET /api/v1/pairs?from=png&to=md.
If the Markdown is only an intermediate step for Word editing, use /convert/png-to-docx and review the DOCX directly.
If the PNG is a receipt or table intended for formulas, use /convert/png-to-xlsx instead of Markdown.
If you need a text-only log with no table syntax, /convert/png-to-txt is easier to diff.
PNG to Markdown questions
How does PNG to Markdown handle two columns?
The tested page was read down the left column and then down the right column. That is good for prose, but layouts with sidebars should be checked for interruptions.
Will tables become real Markdown tables?
The sampling schedule did become a valid pipe table with five headers. Dense tables or merged headings can still require manual pipe alignment.
Why did the footer appear in the Markdown?
The footer was visible text, so Deep Read retained it. Delete repeated page furniture after conversion when you are preparing a clean article or README.