Models in the Inference tab
The machine-learning models you can use in the viewer's Inference tab (9 models). Each one is downloaded to your browser and runs there. Images are not sent anywhere.
How to use: open an image in the viewer, then download a model and run it in the Inference tab on the right. A downloaded model is kept in the browser, so it is ready next time. The results are shown on the image as annotations and can be saved as JSON.
Models run fast in browsers with WebGPU (Chrome, Edge and others). Without it they run on WebAssembly, which takes longer.
NDLOCR-Lite
Reads the text of printed Japanese material from the Meiji era on (vertical or horizontal), line by line.
- How to run
- A button (reads the whole image)
- Download size
- 154 MB
- Image size read
- Up to 2,048 px on the long side
- Runtime
- ONNX Runtime Web
- License
- CC BY 4.0
NDL Kotenseki OCR-Lite
Reads the text of pre-modern Japanese books (including kuzushiji), line by line.
- How to run
- A button (reads the whole image)
- Download size
- 83 MB
- Image size read
- Up to 1,600 px on the long side
- Runtime
- ONNX Runtime Web
- License
- CC BY 4.0
Files (2)
- detector40 MB
- recognizer42 MB
PP-OCRv5 (Chinese, English, Japanese)
Reads the text of printed material (mainly horizontal), line by line. Chinese (simplified and traditional), English and Japanese.
- How to run
- A button (reads the whole image)
- Download size
- 21 MB
- Image size read
- Up to 2,048 px on the long side
- Runtime
- ONNX Runtime Web
- License
- Apache-2.0
Files (2)
- detector5 MB
- recognizer17 MB
PP-OCRv5 (Latin script)
Reads printed text in Latin-script languages (French, German, Latin and others), line by line.
- How to run
- A button (reads the whole image)
- Download size
- 13 MB
- Image size read
- Up to 2,048 px on the long side
- Runtime
- ONNX Runtime Web
- License
- Apache-2.0
Files (2)
- detector5 MB
- recognizer8 MB
Faces in Japanese art
Finds the faces drawn in ukiyo-e, picture scrolls, modern bijin-ga and other Japanese art, and marks them with boxes.
- How to run
- A button (reads the whole image)
- Download size
- 114 MB
- Image size read
- Up to 2,560 px on the long side
- Runtime
- ONNX Runtime Web
- License
- AGPL-3.0
Files (1)
- detector114 MB
Face outlines in Japanese art
Finds the faces, then traces each one and marks it with its outline (a polygon). Slower than boxes only.
- How to run
- A button (reads the whole image)
- Download size
- 192 MB
- Image size read
- Up to 2,560 px on the long side
- Runtime
- ONNX Runtime Web
- License
- AGPL-3.0 (face detection) / Apache-2.0 (SAM 2.1)
Files (5)
Click to outline
Click on something in the image to trace its shape and mark it with an outline. Works on people, buildings, objects and more.
- How to run
- Click on the image
- Download size
- 78 MB
- Image size read
- Up to 3,072 px on the long side
- Runtime
- ONNX Runtime Web
- License
- Apache-2.0
- Made by
- Meta (SAM 2.1)
Files (4)
Describe the picture (Florence-2)
Writes, in English, what the image shows. Best on photographs and modern pictures; it may miss on old paintings.
- How to run
- A button (reads the whole image)
- Download size
- 360 MB
- Image size read
- Up to 1,024 px on the long side
- Runtime
- transformers.js
- License
- MIT
Objects (Florence-2)
Finds people, animals, buildings, vehicles and other things, and marks them with boxes named in English.
- How to run
- A button (reads the whole image)
- Download size
- 360 MB
- Image size read
- Up to 1,024 px on the long side
- Runtime
- transformers.js
- License
- MIT