LaTeX OCR

The model takes an image of a mathematical expression (handwritten or typeset) and produces the corresponding LaTeX code, enabling equations to be immediately usable in LaTeX documents, markdown, or any system that renders LaTeX math without manual transcription.

The system uses a transformer-based architecture (ViT encoder combined with an autoregressive decoder) trained on a large dataset of paired equation images and LaTeX code, enabling it to handle complex mathematical expressions including integrals, summations, matrices, Greek letters, fractions, and multi-line expressions.

The model is available as a Python library, a command-line tool, and a browser extension that enables right-click OCR of equations directly from web pagesall producing LaTeX that can be pasted directly into documents.

Researchers digitizing equations from papers they're reading, students transcribing textbook equations into their notes, anyone working with mathematical documents who needs to edit or repurpose equations that exist only as images use LaTeX OCR to avoid the tedious and error-prone process of manually transcribing mathematical notation.

The tool is particularly valuable for complex equations where the LaTeX syntax is non-obviousgetting the structure of a multi-level fraction or tensor expression right requires either deep LaTeX expertise or a reliable recognition tool.