Image to Text (OCR)
Extract Korean text from images locally with WebGPU acceleration.
Image to Text (OCR)
Extract Korean, English, and number text from photos, scans, and screenshots without sending the image to a server.
How to use
Runtime
A Korean PP-OCRv5 detection and recognition pair runs through PaddleOCR.js and ONNX Runtime Web. WebGPU is selected automatically when the browser supports it; the version-pinned WASM backend is used as a fallback.
The first run downloads about 18 MB of model assets. Keep the page open while the models are cached. Accuracy depends on image focus, lighting, and text layout.
Input limits and expected results
- One PNG, JPG, WebP, or GIF image up to 10 MB.
- Images above the safe pixel budget are reduced before inference; extreme dimensions or prepared images above the internal budget are rejected.
- The recognition model is primarily for Korean, English, and numbers. Handwriting, decorative fonts, vertical text, tables, formulas, and unsupported scripts may be missed or placed in the wrong reading order.
OCR output is a draft, not a verified transcript. Compare names, dates, totals, account numbers, medical text, and legal wording with the source image before relying on it.
Troubleshooting
Crop away unrelated areas, rotate the image upright, increase contrast without clipping characters, and use the original resolution when possible. If the model cannot load, check that content blockers or a restricted network are not blocking the documented model host, then reload once. WebGPU is an acceleration path, not a requirement; the WASM fallback can be slower.
Privacy
OCR runs in a dedicated browser worker. Image pixels are not sent to an OCR API or Aiviko processing server. Model and runtime assets are downloaded separately.