lukas-blecher/LaTeX-OCR

Combines a Vision Transformer encoder with a Transformer decoder to predict LaTeX from equation images, incorporating an auxiliary resolution-prediction network that automatically preprocesses input to match training data characteristics. Supports multiple deployment modes including command-line, desktop GUI with screenshot capture, REST API, and direct Python integration, with MathJax rendering for real-time validation. Includes end-to-end training infrastructure with custom tokenizer generation and dataset preparation from rendered LaTeX via XeLaTeX and ImageMagick.

16,245 stars and 10,850 monthly downloads. No commits in the last 6 months. Available on PyPI.