> Converts an image to a single-page PDF with a hidden text layer using Tesseract. This is the 'State Preservation' step.
Does this mean the text only pdf page is transformed into an image that covers the full page, but the text is still under there. So, any machine based extraction would still get the text, but would probably loose all the bounding box information and regular users cannot just use their mouse to select text anymore?
OP here, I added a sample PDF output in the project assets and put screenshots in the ReadMe. The text is selectable after rehydration. would this work with your app?
It might be feasible to have an intermediate AI call take the generated image and slice it into individual text and image elements that it would then render into the pdf page
Behold, the might of LLMs! Instead of ushering the age of AGI as advertised 6 months ago, now it cleans your PDFs for you.
Many thanks to humanity for failing to standardise PDF and this project for paying interest on that tech debt with datacenter levels of energy consumption.
Very nice! I wonder whether that could be used to get LLMs to annotate pdfs. Say an "agentic" CLI like Claude Code or Gemini-cli reviews a pdf and finds typos, could it use this to annotate the pdf like underlining them in red or something of that sort? That could be nice.
Do you mean example commands? we see those examples on the githib README, yes,
But people here are probably also looking for example input and output PDFs (or images/screenshots) showing the actual work done to get a sense of what to expect.
I like the example of updating latest market data. Updating a deck one-off is tedious. Keeping it updated long-term was never going to happen. But now it can
Please don't add an animated gif to your README. Nothing worse than an autoplaying video with no controls, that has 10 frames but takes 5.4MB to download. Github supports normal video files. It allows the user to rewind or pause, and it results in a much smaller file size.
Generally agreed! though fun point of info: you can use the .avif format to get something that behaves just like a gif (auto-playing, no sound, no controls) but supports modern features (HDR/transparency channel) and is compressed as well as a modern video is, since its just AV1. And it's supported in most all modern browsers these days: https://caniuse.com/?search=avif
Webm is better in many ways, but it doesn't give you gif-like behavior I think. As in, you can't just include it in an <img> tag and a get an autoplaying looping video. Though you can simulate it with <video>.
Basically, .avif is an "animated image" format, like .gif, but .webm is only a video format.
edit: just realized .webp i think can be an animated image! So that seems like the alternative
AFAIK -- even the "Designer" feature of Microsoft Powerpoint (now folded under Copilot license I believe) gives slide deigns with shapes etc that are not editable. Thankfully the text remains editable. But if we wnat to ever so slightly modify the suggested design my removing or reshaping some if the shapes ... nopes. Feels like they are worried about humans with taste ripping-off the AI output :D
I am disappointed that this doesn't modify the underlying pdf structure (which is a horror show, I know) but instead relies on fairly lossy OCR back&fourths.
I wish an agent with a validation and rendering tools could instead manipulate the structure to accomplish those edits way less destructively, checking its progress with the tools.
it will be more excited if i can use this feature in application with GUI, it’s now convenient to check the result after edit the PDF, i need to transfer between CLI and PDF reader
I recently tried to change a single word in a PDF and nearly tore my hair out (thank you LibreOffice) I'll definitely keep this in mind for next time, thank you.
PS: in my quick test of editing a PDF text -- the output PDF had weirdly added an extra "&" symbol at the end of every existing line of text. will try out more to see if it was something in the input PDF that was causing it.
> Converts an image to a single-page PDF with a hidden text layer using Tesseract. This is the 'State Preservation' step.
Does this mean the text only pdf page is transformed into an image that covers the full page, but the text is still under there. So, any machine based extraction would still get the text, but would probably loose all the bounding box information and regular users cannot just use their mouse to select text anymore?
Seems true and really wish the project included some sample PDF output.
My Text to Speech app uses bounding box to display what text in PDF is being read and would not work well PDF's from this project.
OP here, I added a sample PDF output in the project assets and put screenshots in the ReadMe. The text is selectable after rehydration. would this work with your app?
Amazing. Worked really well. Thank you.
Wait! what? This is incredible. Amazing work.
This is nuts and I absolutely love this. So you convert the PDF into image, edit the image, then convert the image back into a PDF.
This is the usual workflow dealing with pdfs (unfortunately)
No, it's not, unless you are dealing with scans. Lots of apps let you edit PDFs.
A side effect of replacing entire pages with images is that the file size will expand dramatically. Most PDFs only contain a couple of images
It might be feasible to have an intermediate AI call take the generated image and slice it into individual text and image elements that it would then render into the pdf page
Behold, the might of LLMs! Instead of ushering the age of AGI as advertised 6 months ago, now it cleans your PDFs for you.
Many thanks to humanity for failing to standardise PDF and this project for paying interest on that tech debt with datacenter levels of energy consumption.
I'd love to see clearer examples: a video, or original pdf / command / result pdf. Very cool!
It's incredible how many hacks there are to make PDFs semi-usable.
Very nice! I wonder whether that could be used to get LLMs to annotate pdfs. Say an "agentic" CLI like Claude Code or Gemini-cli reviews a pdf and finds typos, could it use this to annotate the pdf like underlining them in red or something of that sort? That could be nice.
The lack of examples makes me very reluctant to commit any time to trying this out - despite it being something that I’m interested in.
Has anyone given any it a go? Does it work?
What? There are examples in the repo and even in OP’s post.
I haven’t tried it, but there are plenty of examples.
Do you mean example commands? we see those examples on the githib README, yes,
But people here are probably also looking for example input and output PDFs (or images/screenshots) showing the actual work done to get a sense of what to expect.
Exactly - if these examples work really well, then include some screenshots.
I like the example of updating latest market data. Updating a deck one-off is tedious. Keeping it updated long-term was never going to happen. But now it can
Nice - but consider adding an animated screengrap like: https://github.com/pythops/oryx
Please don't add an animated gif to your README. Nothing worse than an autoplaying video with no controls, that has 10 frames but takes 5.4MB to download. Github supports normal video files. It allows the user to rewind or pause, and it results in a much smaller file size.
Generally agreed! though fun point of info: you can use the .avif format to get something that behaves just like a gif (auto-playing, no sound, no controls) but supports modern features (HDR/transparency channel) and is compressed as well as a modern video is, since its just AV1. And it's supported in most all modern browsers these days: https://caniuse.com/?search=avif
I tend to use webm but I’m curious, is avif better (performance, size) for gif?
Webm is better in many ways, but it doesn't give you gif-like behavior I think. As in, you can't just include it in an <img> tag and a get an autoplaying looping video. Though you can simulate it with <video>.
Basically, .avif is an "animated image" format, like .gif, but .webm is only a video format.
edit: just realized .webp i think can be an animated image! So that seems like the alternative
Thanks
I thought it was kinda funny that Google Slide’s own built in “beautify this slide” button converts the whole slide into an uneditable image.
AFAIK -- even the "Designer" feature of Microsoft Powerpoint (now folded under Copilot license I believe) gives slide deigns with shapes etc that are not editable. Thankfully the text remains editable. But if we wnat to ever so slightly modify the suggested design my removing or reshaping some if the shapes ... nopes. Feels like they are worried about humans with taste ripping-off the AI output :D
Somewhat unrelated but can anyone recommend a way to edit the text of a PDF using LLM? Something like AI + acrobat pro?
Congratulations on the release; that's a really good job.
I am disappointed that this doesn't modify the underlying pdf structure (which is a horror show, I know) but instead relies on fairly lossy OCR back&fourths.
I wish an agent with a validation and rendering tools could instead manipulate the structure to accomplish those edits way less destructively, checking its progress with the tools.
it will be more excited if i can use this feature in application with GUI, it’s now convenient to check the result after edit the PDF, i need to transfer between CLI and PDF reader
Love this.
After several iterations of edits, would the image quality decrease?
I recently tried to change a single word in a PDF and nearly tore my hair out (thank you LibreOffice) I'll definitely keep this in mind for next time, thank you.
Try photopea next time
Wow - didnt know about this tool for PDF editing - thanks!
https://www.photopea.com/
PS: in my quick test of editing a PDF text -- the output PDF had weirdly added an extra "&" symbol at the end of every existing line of text. will try out more to see if it was something in the input PDF that was causing it.
What is photopea built on?