OCR to LaTeX

Stream: Machine Learning for Theorem Proving

[Heather Macbeth]

This is machine learning for math rather than machine learning for theorem proving, but I imagine the readership of this stream will be well-informed: what are the best tools currently available for the task of converting an image of handwritten text to LaTeX? Is there a way to create a pipeline in which a first approximation is compiled to DVI/PDF and then compared optically with the original image for feedback and adjustment?

[Newell Jensen]

I have used https://mathpix.com/image-to-latex in the past, worked fairly well.

[Newell Jensen]

Definitely saves a bunch of time even though it might not get it right 100% of the time.

[Heather Macbeth]

Interesting, I'll try it. Is there anything public about how it works?

[Newell Jensen]

Not sure, didn't look into it but others here may know more.

[Siddhartha Gadgil]

Mathpix seems to work well.

[Zhangir Azerbayev]

There is also this open source project: pix2tex. It's a vision transformer trained on wikipedia, arxiv, and the im2latex-100k dataset.