I Found a Really Good use for AI

[Quoth]

Originally Posted by JSWolf

If I can find a book in PDF and an ePub version, I could try a PDF > ePub and then do a compare of the two ePub versions.

It has to be one that the LLMs can't know about! Lack of close match means nothing either as LLMs chop-up & reassemble the inputs.

[salamanderjuice]

Originally Posted by DNSB

In my experience, most text layers in PDF files seem to be the result of an unedited OCR conversion.

Probably because you're reading scans. PDFs made digitally with tools like LaTeX and Word typically have good text layers, sometimes there are still issues because text can't handle things like equations and tables that well, but generally the text is good. Since I mainly read PDFs of journal articles from the last 25 years they all tend to have good text layers.

[Loosheesh]

Originally Posted by Quoth

It has to be one that the LLMs can't know about!

Your recent posts seem to me to be implying that the AI will generate text from its general knowledge for the conversion; please correct me if I've misunderstood. I've used Claude to convert several PDFs to EPUBs, and this is not what is happening.

I feed Claude the PDF and give it a simple prompt (most times "Convert the attached PDF to a well formatted EPUB with a working TOC."). I see the commands ran as Claude is working (not that I sit there and watch the whole thing ). I get progress reports (if it hits a tool-use limit) and, when completed, I get a final report on how it worked through the conversion. I can also look through all the commands ran during the conversion.

Claude is working off the actual PDF I fed it. I attached a few images from a recent conversion; some PDFs are more gunky and stubborn than others, and this was one of them (it hit the tool-use limit twice).

[graycyn]

Originally Posted by JSWolf

Thanks. The new cover looks really good vs the one I started with. Sometimes with older books, it can be rather difficult to find a proper cover image. So I gave AI a try and it worked.

The other thing I tried with AI that worked well. I took a frame from Star Trek: The Next Generation and a frame from The Simpsons. Both in 4:3. I asked AI to change it to 16:9 by adding to the left/right sides and both worked very well. You would not know they they were not shot/drawn that way.

Another thing AI can do well is upscale TV programs/movies that are still popular enough in SD to full HD while keeping the original aspect ratio. Ones that were not shot on film.

Do you just ask AI to enlarge the image? What do your instructions look like? Feel free to get specific on how to tackle this! I'm kind of clueless with AI.

I've got a 1929 USA public domain title, text is done, epub is made, needs a final proofread, but the cover is a hold up. I did find ONE, and only one low-res image online. Getting a physical copy is a total no go. There's only two copies for sale presently, and the one with a dust jacket (source of my low res image) is like $450. Text was available via Google scan, fortunately.

Been painfully cleaning up the low res image in Photoshop, hoping I can enlarge it once done and have something decently passable. BUT ... my computer is dying, and could go at any point. I'll lose my Photoshop CS5 when I get a new computer, as I'm pretty sure Adobe has shut down the activation servers for that by now.

If AI can help with this stuff, I'm all for it. It's not the only 1920s book I'm working on. You aren't kidding about how it can be hard to find cover images for older books!

This is the book in question:

https://www.abebooks.com/first-editi...31332733643/bd

I can get it done by hand, but if I can learn faster ways, I'm listening! And yes, I know this title can't go up on MobileRead. I'll put the epub on the Internet Archive when finished.


Sent from my iPad using Tapatalk

[nana77]

Originally Posted by graycyn

Do you just ask AI to enlarge the image? What do your instructions look like? Feel free to get specific on how to tackle this! I'm kind of clueless with AI.

I've got a 1929 USA public domain title, text is done, epub is made, needs a final proofread, but the cover is a hold up. I did find ONE, and only one low-res image online. Getting a physical copy is a total no go. There's only two copies for sale presently, and the one with a dust jacket (source of my low res image) is like $450. Text was available via Google scan, fortunately.

Been painfully cleaning up the low res image in Photoshop, hoping I can enlarge it once done and have something decently passable. BUT ... my computer is dying, and could go at any point. I'll lose my Photoshop CS5 when I get a new computer, as I'm pretty sure Adobe has shut down the activation servers for that by now.

If AI can help with this stuff, I'm all for it. It's not the only 1920s book I'm working on. You aren't kidding about how it can be hard to find cover images for older books!

This is the book in question:

https://www.abebooks.com/first-editi...31332733643/bd

I can get it done by hand, but if I can learn faster ways, I'm listening! And yes, I know this title can't go up on MobileRead. I'll put the epub on the Internet Archive when finished.


Sent from my iPad using Tapatalk

If you want to give them a try, these programs are fine (imho):
https://www.kornelix.net/fotocx/fotocx.html
https://rawtherapee.com/