Pictures survive an export

A Notion, Confluence or Obsidian export kept its pictures in the .zip and never unpacked them, so the document arrived with every image pointing at a file that was not there. A PowerPoint deck lost its slides’ pictures. A Word file lost its pictures having already handed every one of them over. All four now carry them, inside the document itself.

This was the loudest complaint about two of these formats, and this converter had it too.

What was happening

An export writes its pictures into folders and links to them relatively: Notion beside each page, Confluence under attachments/, Obsidian wherever you filed them. Only the text was ever read out of the archive, so the Markdown came out still saying ![](some/path.png) about files that were never unpacked — twenty broken images in a converted wiki, which reads as something this lost rather than something it never had.

PowerPoint was worse in a quieter way: a slide that was only a diagram said so and moved on. And Word was the strangest of the four, because mammoth reads a .docx's own pictures out of the file and hands over every one of them — and a single line, set years ago and never commented, threw them away.

They are carried, not uploaded

A picture becomes part of the document, as its own bytes. That is the constraint rather than a shortcut: every conversion here promises the file stays in your browser, and putting the images in storage on the way through would make that untrue. It also means they travel — into the downloaded Markdown, into the standalone HTML, into the Word export, into a shared link.

There is a ceiling, and it is not ours

A document has to fit in 4 MB to be saved to an account, so pictures have half of that between them and a megabyte each. Past it, a picture from an archive keeps the link it already had — nothing is lost that was not lost before — and one with no file behind it is replaced by its own description.

The part that changed the design

The obvious way to keep Word's pictures was to stop discarding them. Measured on a 3.5 MB document holding three photographs, that produced 4.7 MB of HTML in 224 ms and then did not finish: an <img> whose address is two million characters long is not what an HTML parser is built for. So the bytes never go near it now — each picture is weighed as it is read, held aside under a number, and put back once the Markdown is small again. Sixty milliseconds.