Hacker News new | ask | show | jobs
by ComputerGuru 597 days ago
I’d be very interested in the opposite! Lots of scanned or legacy images that would be nice to convert to LaTeX, or to create a robust PDF ingestion pipeline.
4 comments
Vetch 597 days ago
In addition to the already mentioned https://huggingface.co/facebook/nougat-base, I also highly recommend https://huggingface.co/stepfun-ai/GOT-OCR2_0. It might even be better.
link
hextex 597 days ago
Facebook's Nougat [1] should work with this, but not sure how much preprocessing is needed to yield good results with scanned copies of physical documents. Note that it outputs .mmd files (MultiMarkDown), but the equations and tables should (iirc) output plain LaTeX.

1: https://github.com/facebookresearch/nougat

link
Wdorf 596 days ago
This looks really interesting! I will definitely have a look.
link
sorenjan 597 days ago
Like Mathpix?

https://mathpix.com/

link
BeetleB 597 days ago
Mathpix is awesome. So far it has never gotten the output wrong. I even have it integrated into Emacs/org-mode.
link
Wdorf 597 days ago
I could add a GPT based pdf to latex functionality in the future.
link