# Why ChatGPT struggles with scanned PDFs: the hard part is the layout, not the translation

> Published 2026-10-28 · https://www.promptzone.com/li_lu_8925b64a9636130a26d/why-chatgpt-struggles-with-scanned-pdfs-the-hard-part-is-the-layout-not-the-translation-3m1h

Give ChatGPT or Gemini a scanned birth certificate and ask for an English translation, and the translation itself is usually fine. What you get back is a block of text: the table is gone, the stamp is gone, the signature line is gone. If you only need to know what the document says, that's enough. If you need a translation that still looks like the original, you're only halfway there.

## Why this happens

A scan has no text in it, only an image. The model can read the image, but what it gives back is text. It doesn't draw the words back where they were. Translation is the part that already works well. The hard part comes after: putting the translation back on the page.

## Putting it back takes three steps

1. Find where every line is. The text alone isn't enough; you need positions.
2. Erase the original text from the image without touching table lines, stamps or photos.
3. Set the translation back in place, at a size close to the original. Translations often run longer (English to German can be about 30% longer), and they still have to fit in the original box.

## How far a prompt alone gets you

You can ask the model for each line's text and coordinates:

```plaintext
Read this scanned page. Return JSON only: one object per text line,
{"text": "...", "bbox": [x0, y0, x1, y1]} in pixels of the image.
Do not translate. Do not merge lines.
```

Then a script pastes the translation back by those coordinates, and it works. Two problems show up quickly, though. General-purpose models aren't precise with coordinates, and a small offset puts the text right on top of a table line. And the erasing and re-layout take far more code than the translation does. For a page or two that's fine. For a tool, you end up moving to a dedicated OCR model and a real typesetting step.

## What I built

I built Reglyph to do those three steps for you: upload a scanned PDF, pick a language, and download a translated PDF laid out like the original. It supports 19 languages, new accounts get 5 free pages, and downloads have no watermark.

(Disclosure: I'm the maker.)

https://tryreglyph.com