TexImageScan

Turns images of equations into editable LaTeX, Markdown, PNG, or SVG.
Recognition runs on your device. No accounts. No cloud OCR.

Features

TexImageScan runs in a Chromium browser side panel next to the page you are reading. Capture an equation, matrix, or aligned equations, recognize them locally, then choose the export format you want.

1. Capture an equation	2. Review and copy the result

• Capture equations with a rectangle or freehand lasso, or upload a PNG, JPEG, or WebP image.
• Recognize locally with PaddleOCR-VL-1.6 running on your device via WebGPU.
• Review and edit the output source, copy LaTeX, Markdown with math, plain text, or export rendered PNG or SVG images.
• Reopen saved scans and keep editing. Images and edits stay in local history until you delete them.

Install

Requires Node.js 24 or later and pnpm 12.15.0:

pnpm install
pnpm build

1. Open chrome://extensions in your Chromium browser, enable Developer mode, choose Load unpacked, and select .output/chrome-mv3.
2. Click the TexImageScan toolbar icon to open the side panel.
3. Choose Download model to obtain PaddleOCR-VL-1.6 FP16. The first download is about 2.04 GB from Hugging Face and stays cached for later use.

Use desktop Chrome or a compatible Chromium browser with hardware acceleration and a GPU supporting WebGPU and shader-f16. There is no CPU fallback.

Optional Capture across tabs access lets you capture on other websites without clicking the toolbar icon each time. Manage it in Settings. Local files require Allow access to file URLs in Chrome's extension settings.

Testing

Install the test browser once, then run type checks, unit tests, the build, browser tests, and package checks:

pnpm exec playwright install chromium
pnpm verify

Privacy

• Images and recognized text stay on your device. They are saved locally and kept until manually deleted in the history tab.
• Model files use a separate cache, which can be deleted in the settings tab. No telemetry, analytics, or remote OCR API.
• Required website access covers model downloads. Capturing across websites is optional and starts only when you ask.

Limitations

• Recognition can omit symbols or matrix columns, or change indices. Compare the rendered result with the source before copying it. Valid LaTeX can still be wrong.
• Capture the complete equation without surrounding paragraphs. Zoom in before capturing small notation. Enlarging an existing screenshot cannot recover missing detail.
• Handwriting and full-page document recognition are outside the current focus.
• Uploads are limited to 12 MB. Uploaded images and page selections are limited to 8 million pixels.
• If output is cut off, increase the output token limit in the settings tab.

License

TexImageScan uses the MIT License. Bundled dependencies keep their own licenses, recorded in the third-party notices and attribution sources, which also ship with the extension. Model weights are downloaded separately from Hugging Face under Apache-2.0. See the PaddleOCR model and runtime notice.