Kivoza

Guide 06 / local OCR

Turn printed image text into editable words.

Choose one clear JPG, PNG, or static WebP image, select a language, and let the browser run the OCR worker locally. The result is a starting point you can review and edit.

By Kivoza · Updated

Privacy check: the image stays in the browser. The first recognition may request the selected local model and core resources from the same Kivoza origin; those resources are not an image upload or OCR API.

Use the OCR tool

01 / CHOOSE

Open Kivoza OCR, browse, drop, or paste one static JPG, PNG, or WebP image. The page checks the file bytes and decoded dimensions before loading OCR resources.

02 / LANGUAGE

Select English, 简体中文, or English + 简体中文. Choose the model that matches the printed text; mixed pages can need more review.

03 / RECOGNIZE

Start recognition and watch the honest stages: local engine loading, language data loading, worker preparation, and printed-text recognition. The page does not turn a progress number into an accuracy claim.

04 / REFINE

Edit the textarea when punctuation, names, line breaks, or characters need correction. Copy the text or download a plain TXT file.

First load and later runs

The worker and selected language model load only after you start. The browser can cache the model under the tool's OCR cache name, so a repeated run may avoid downloading the same language data again. A transient model failure settles to a visible error and leaves Recognize again available.

Copying on different origins

Clipboard writing is available in a secure context when the browser permits it. On an insecure preview address or when permission is blocked, Kivoza selects the text and tells you to use the device copy command. Download TXT remains available.

Accuracy boundaries

The first version targets printed text. Blur, tilt, glare, small type, unusual fonts, low contrast, and complex backgrounds can reduce accuracy. Review the editable result. Handwriting, tables, receipt fields, PDF OCR, layout reconstruction, and perfect transcription are outside this version.

Limits and failures

  • One image per run, up to 20 MB and 40 megapixels after decoding.
  • Animated PNG and animated WebP are rejected rather than reduced to one frame.
  • A blank result reports that no printed text was recognized.
  • If a model request fails, use Recognize again after the resource is available; clearing or replacing the image cancels the old run.