📷 Drop image to recognize

Configuration

The defaults work well for most images. Adjust only what you need, then apply your changes.

Library

Models Pick the model tier for your workload

Pick a preset, or leave everything blank to use the built-in PP-OCR defaults.

Model preset Any URL filled in below overrides that part of the preset.

Warm up after applying (hidden inference so the first recognize is fast)

Detection model URL

Recognition model URL

Character dictionary URL

Recognition Control text reading speed and accuracy

Reading strategy Balanced is the default: line context improves short labels at equal speed.

Minimum confidence Items below this confidence are dropped before results return. 0 disables.

Recognition image height

Cross-line width factor Used only by the fast cross-line strategy.

Detection Find and crop text regions

Image processing size Larger values can detect smaller text but use more memory and time.

Minimum text area Increase this to ignore tiny noise and artifacts.

Crop padding

Vertical

Horizontal

Adds breathing room around detected text.

Advanced image normalization

Mean - R, G, B

Standard deviation - R, G, B

Change these only when using a model trained with different normalization.

Runtime & debugging Hardware acceleration and console logs

Processing engine Automatic provides the best compatibility across browsers.

Show detailed logs in the browser console

Using recommended defaults

Image Input

Upload an image to begin

JSON Output

// Inference output will appear here

ppu-paddle-ocr

Preparing...

Which model and config should I use?

Defaults are PP-OCRv6 tiny with the Balanced strategy - fast and accurate for screenshots and clean photos. Click a preset to select it, then Apply changes.

Your input	Recommended	Preset
App or web screenshot (dark or light theme)	Defaults as-is
Form or sparse labels	Defaults as-is
Receipt or document photo	Defaults as-is; small for hard shots
Dense document page (paragraphs, tables)	Small + the Fast (cross-line) strategy
Low contrast or blurry photo	Medium (server-grade); raise processing size
Rare CJK, kana, or full 50+ language coverage	Small or medium (full dictionary)
One known language only	A v5 language preset (English shown)
Tilted or rotated input	Deskew the image first, then defaults	-
Tiny text that still is not detected	Raise Image processing size (e.g. 1280+)	-