// the find
RQLuo/MixTeX-Latex-OCR
MixTeX multimodal LaTeX, ZhEn, and, Table OCR. It performs efficient CPU-based inference in a local offline on Windows.
A Windows desktop app that OCRs mixed LaTeX formulas, Chinese/English text, and tables from clipboard or screenshots, running entirely on CPU with no internet connection. It's aimed at students and researchers typing up math-heavy documents who want a fast, offline formula-to-LaTeX pipeline without GPU hardware or a cloud API.
CPU-only local inference is the actual selling point here — no GPU, no API key, no data leaving the machine, which matters for anyone handling unpublished research. It's backed by a real paper (arxiv 2406.17148) rather than just vibes, and there's a Colab notebook plus a HF Space so you can test accuracy on your own formulas before installing anything. The clipboard/screenshot-to-LaTeX workflow (Win+V, screenshot tool, F2 to pause) is a sensible, low-friction UX for a utility that people will use dozens of times a day.
Windows-only — the README's build step literally asks the community to "please submit the packaged programs from other platforms," meaning there's no real macOS/Linux story. Training data is admittedly synthetic ("somewhat rough, uniform, and limited" per their own Limitations section), so expect degraded accuracy on anything outside clean printed text — handwriting and complex tables are explicitly unsupported. There's no test suite or CI in the tree, and repo hygiene is rough: the changelog lives as strikethrough edits inside the README instead of tags/releases, and there's a stray file literally named `mixtexgui/conda create -nmixtex python=3.14` checked into the repo, which doesn't inspire confidence in the packaging process.