ffocr

Browser OCR for frontend apps using PaddleOCR models converted to ONNX.

npm: https://www.npmjs.com/package/ffocr

Live Demo: https://zxc88645.github.io/ffocr/demo/

Install

npm install ffocr

Quick start

If you want the built-in hosted model:

import { createDefaultPPOcrV5 } from "ffocr";

const ocr = createDefaultPPOcrV5();
const result = await ocr.ocr(fileOrBlobOrUrl);
console.log(result.text);

If you want to host the model files yourself:

import { createPPOcrV5 } from "ffocr";

const ocr = createPPOcrV5({
  baseUrl: "https://your-cdn.example.com/models/pp-ocrv5"
});

const result = await ocr.ocr(fileOrBlobOrUrl);
console.log(result.text);

Model variant

PP-OCRv5 ships two model variants: mobile (default) and server. The mobile variant is smaller and faster, while server offers higher accuracy at a larger size.

// Uses the default mobile model
const ocr = createDefaultPPOcrV5();

// Switch to the server model
const ocr = createDefaultPPOcrV5({ modelVariant: "server" });

The same option works with a custom baseUrl:

const ocr = createPPOcrV5({
  baseUrl: "https://your-cdn.example.com/models/pp-ocrv5",
  modelVariant: "server"
});

Model caching

Enable cacheModels to persist downloaded models in the browser's Cache API. Subsequent page loads will read from cache instead of re-downloading:

const ocr = createDefaultPPOcrV5({ cacheModels: true });

Progress callback

Track loading, download, and inference progress via onProgress:

const result = await ocr.ocr(file, {
  onProgress({ phase, current, total, loaded, totalBytes }) {
    // phase: "loading_dictionary" | "loading_detection_model"
    //      | "loading_recognition_model" | "warmup"
    //      | "preprocessing" | "detecting" | "recognizing"
    if (loaded != null && totalBytes != null) {
      const pct = Math.round((loaded / totalBytes) * 100);
      console.log(`${phase}: ${pct}%`);
    } else if (phase === "recognizing") {
      console.log(`Recognizing ${current}/${total}`);
    } else {
      console.log(phase);
    }
  }
});

Main API

createDefaultPPOcrV5(options?)

Creates an OCR instance using the package default hosted model URL.

ocrWithDefaultPPOcrV5(source, options?, ocrOptions?)

One-shot OCR helper using the package default hosted model URL.

createPPOcrV5({ baseUrl, ...options })

Recommended when you want full control over model hosting.

ocrWithPPOcrV5(source, options, ocrOptions?)

One-shot OCR helper for a custom baseUrl.

createPaddleOcr({ manifest, ...options })

Lower-level API for custom manifests or model paths.

Required model files

Your baseUrl should point to a folder that contains the detection and recognition ONNX pair for the variant you use (mobile by default):

• Mobile: det_mobile.onnx, rec_mobile.onnx
• Server: det_server.onnx, rec_server.onnx

Override filenames with detectionModelPath / recognitionModelPath if your layout differs.

Dictionary: By default the dictionary is loaded from the PaddleOCR project (not from baseUrl). To self-host it, set dictionaryUrl to your ppocrv5_dict.txt URL.

Default model URL

createDefaultPPOcrV5 loads weights from:

https://zxc88645.github.io/ffocr/models/pp-ocrv5

Use createPPOcrV5 with your own baseUrl when you need to self-host or pin assets.

The bundled conversion flow patches the generated ONNX files so the default PP-OCRv5 models stay compatible with ONNX Runtime WebGPU.

Supported inputs

• ImageData
• ImageBitmap
• HTMLImageElement
• HTMLCanvasElement
• OffscreenCanvas
• Blob
• string URL

Runtime

• Detects available ONNX Runtime Web execution providers including webgpu, webnn, webgl, and wasm
• In auto mode, picks the first available provider in this order: webgpu -> webnn -> webgl -> wasm

Extra docs

• Model conversion: docs/MODEL_CONVERSION.md
• Example app: examples/vite-demo