About this tool
One formula at a time
The math OCR works best on exactly one formula or one aligned block: paste a screenshot, drop a photo, or open an image and drag a box around the equation you want. The crop is padded to a square and scaled to the model's input size, so a tight box around the formula gives better results than a whole page with text around it. A page of lecture notes is not the target — for mixed text and maths, box each formula separately. The formulas you convert during a visit are listed on the page until you close it; nothing is saved.
Edit the LaTeX with the preview open
The recognised LaTeX lands in an editable field with a KaTeX rendering next to it, and the preview re-renders as you type. A parse error shows in the preview area rather than blanking it, so a missing brace is visible immediately. The page also tidies the model's output — collapsing stray spaces, turning \operatorname { s i n } back into \sin — and warns when braces do not balance. Spaces inside \text{…}, \mathrm{…} and \operatorname{…} are removed too, because the model spells letters one at a time and the page cannot tell where a word space was, so put those back by hand. Then copy it in the form you need: raw LaTeX for a document, $…$ for inline maths in Markdown or a Jupyter cell, $$…$$ for a display equation, or MathML for pasting into a word processor.
How accurate the image to LaTeX conversion is
The model is pix2text-mfr, an open-source formula recogniser under the MIT licence, about 118 MB downloaded once from huggingface.co and then cached by the browser, while the ONNX Runtime script that runs it loads from cdn.jsdelivr.net. We rendered 20 printed formulas with KaTeX at two sizes and ran all 40 through it; 39 came back exactly right. That test covers clean printed input only. Handwritten formulas have not been tested here yet, and formula recognisers are usually much worse on handwriting, so expect to correct more. Low-resolution screenshots and photos taken at an angle also lower the odds; a larger, straighter crop helps.
Same rendering, different LaTeX
A formula that renders identically can be written several ways: \frac{a}{b} and a/b, \left( and plain (, x^{2} and x^2. The recogniser returns the markup it thinks matches the picture, which may not be the markup the source used. If you are matching an existing document, read the LaTeX as well as the preview. And this is a transcription tool only — it turns the image into markup and does not simplify, solve or check the mathematics.
When a paid service is overkill
Paid services such as Mathpix do the same conversion with extra features on top. For the occasional screenshot in a paper, a homework write-up or a slide, this page covers the common case with no account and no upload: your images stay in the browser, the model runs on your GPU through WebGPU when the browser has one or on the CPU when it does not, and the result is the same either way. For a whole scanned document you want the OCR tool for text and this one for the formulas you box out of it.