Convert screenshots of math formulas to LaTeX code. 100% local AI — no upload, no API key. Powered by WebGPU.
LaTeX OCR uses the Texify2 model (ONNX, ~305MB, q8 quantized) from Xenova/texify2, running locally via Transformers.js / ONNX Runtime Web in your browser. The model downloads on first use and is cached afterward. Multi-CDN fallback ensures reliable downloads worldwide. All processing is 100% local — your images never leave your device.
This tool recognizes math formulas using Texify2, a purpose-built image-to-text model (ONNX, ~305MB) running locally through Transformers.js. The formula screenshot is loaded as a raw image — via Blob, not a fetch, to stay compatible with the page's cross-origin isolation — and passed to the model, which decodes the pixels into a LaTeX token sequence with a larger token budget than the model default so long equations aren't cut off. The raw output then goes through a cleanup pass: HTML tags and outer $ delimiters are stripped, redundant \mathrm and \text wrappers are removed, common misreads like \chi being mistaken for an italic x are corrected, and spacing around operators and parentheses is normalized. The result is clean LaTeX you can paste straight into Overleaf, Notion, or any editor.
A formula image is fed to the local Texify2 ONNX model, which decodes it into raw LaTeX, and a cleanup pass strips wrappers and fixes common misreads.
Dr. Alvarez screenshots a definite integral from a physics textbook and drops the PNG into the tool. The model downloads on first use (~305MB) and stays cached.
Convert screenshots of math formulas to LaTeX code using local AI. Everything runs locally; nothing is uploaded.
FreeToolHub LaTeX OCR is a free browser-based tool — convert screenshots of math formulas to LaTeX code using local AI. No signup, no upload; everything runs locally in your browser.
Convert screenshots of math formulas to LaTeX code using local AI. No upload, no API key, no subscription. 100% private, free.
The LaTeX OCR tool is a free, browser-based AI tool that recognizes mathematical formulas from screenshots and converts them to LaTeX code. It uses the texify-small model running locally via Transformers.js with WebGPU acceleration. Upload a screenshot of any formula — from a PDF, textbook, whiteboard, or handwritten note — and the tool outputs clean LaTeX code that you can paste into Overleaf, Notion, or any LaTeX editor. All processing happens in your browser — your images never leave your device.
Students who need to digitize formulas from textbooks or lecture notes. Researchers extracting equations from academic papers. Teachers creating LaTeX documents from whiteboard photos. Anyone who works with math but doesn't want to manually type complex LaTeX expressions. The tool is especially useful when you have a formula as an image and need to edit it in LaTeX.
Take a screenshot of the formula you want to convert. Upload it to the tool. Click Recognize Formula. The AI model downloads on first use (~40MB) and runs locally. The recognized LaTeX code appears in the output panel. Copy it to your clipboard and paste into any LaTeX editor. For best results, crop the screenshot to contain only the formula.
The model is trained primarily on typeset formulas (printed or digital). Handwritten formulas may work if clearly written, but accuracy will be lower than for printed formulas. For best results, use screenshots from PDFs, websites, or typed documents.
No. The AI model runs entirely in your browser using WebGPU. The first time you use the tool, the model (~40MB) is downloaded and cached. After that, it works offline with no internet connection.
On clean printed formulas — textbook scans, typeset PDFs, screenshots from papers — recognition is usually character-perfect including nested fractions, subscripts, and Greek letters. Handwriting is less consistent: legible block handwriting works, cursive and cramped notation often drops symbols. The confidence score alongside each result tells you when to double-check; anything below roughly 0.85 deserves a manual review against the source image.
Aim for a tight crop with the formula filling the frame, even lighting, and contrast between ink and background. Phone photos at a slight angle introduce perspective skew that hurts accuracy — scan or screenshot when possible. Very large images are downscaled automatically, so a 4000px photo gains nothing over a sharp 800px crop. Remove surrounding prose; the model wastes attention on text tokens it will discard.
Single-line formulas are most reliable. Multi-line aligned environments (aligned, cases, piecewise) are recognized, but the model may flatten line breaks or lose alignment markers — the LaTeX still compiles, just with \ rather than a proper environment. For long derivations, recognize each line separately and assemble the environment yourself; the per-line results combine cleanly.
The output is editable text, so common fixes are quick: mismatched braces (count \{ and \}), a misrecognized \xi versus \zeta or \ell versus 1, and missing \left/\right sizing pairs. Render the corrected string in any LaTeX editor to verify. Saved snapshots let you keep recognized results locally so you can revisit and refine a formula later without re-running recognition.
This tool is also known by these tasks — each link opens the same tool with a focused guide:
Why did the cookie go to the doctor?
No signups, no data sold. Every tool is free to use — the free tier allows 5 downloads or saves per day, and the optional Pro plan ($7/mo, $59/yr) adds unlimited downloads, batch processing, white-label exports and an ad-free experience.
☕Support me on Ko-fi— support the free tier100% of proceeds go towards hosting & building more free tools.