Style Transfer

Style Transfer is an application developed with Flutter that allows users to apply artistic style transformations to images using advanced style transfer models created by Google, as described in their publication Supercharging Style Transfer.

Technical Features

• Local Inference on Mobile: In the mobile versions, TensorFlow Lite is used to load the models and perform inference directly on the device. This ensures fast performance and does not rely on an internet connection.

• Isolates on Mobile: To avoid blocking the main thread responsible for rendering the user interface (UI), inference and image processing are performed in the background using isolates. This enhances the user experience by keeping the UI responsive.

• Server-side Inference for Web: In the web version, inference is handled by sending requests to a server running FastAPI and using TensorFlow to execute the models.

• Web Workers on Web: Similar to isolates in the mobile versions, web workers are used in the web version to perform inference and image processing in the background, ensuring the UI remains smooth and responsive.

Screenshots

Mobile

Web

Local deployment using Docker Compose

docker compose up

Build for web

Setup service workers api url

./build_service_workers.sh http://localhost:8080

Build web

flutter build web

Run local API

Download tensorflow lite models to root folder

curl https://storage.googleapis.com/download.tensorflow.org/models/tflite/task_library/style_transfer/android/magenta_arbitrary-image-stylization-v1-256_int8_prediction_1.tflite -o magenta_arbitrary-image-stylization-v1-256_int8_prediction_1.tflite

curl https://storage.googleapis.com/download.tensorflow.org/models/tflite/task_library/style_transfer/android/magenta_arbitrary-image-stylization-v1-256_int8_transfer_1.tflite -o magenta_arbitrary-image-stylization-v1-256_int8_transfer_1.tflite

python -m venv venv

source venv/bin/activate

pip install -r requirements.txt

uvicorn tflite_server:app --host 0.0.0.0 --port 8080

Build for android

Download tensorflow lite models to assets folder

curl https://storage.googleapis.com/download.tensorflow.org/models/tflite/task_library/style_transfer/android/magenta_arbitrary-image-stylization-v1-256_int8_prediction_1.tflite -o assets/models/magenta_arbitrary-image-stylization-v1-256_int8_prediction_1.tflite

curl https://storage.googleapis.com/download.tensorflow.org/models/tflite/task_library/style_transfer/android/magenta_arbitrary-image-stylization-v1-256_int8_transfer_1.tflite -o assets/models/magenta_arbitrary-image-stylization-v1-256_int8_transfer_1.tflite

Build

flutter build apk

Build for iOS (Development)

You need to have a macOS device with Xcode to build for iOS.

Download tensorflow lite models to assets folder

curl https://storage.googleapis.com/download.tensorflow.org/models/tflite/task_library/style_transfer/android/magenta_arbitrary-image-stylization-v1-256_int8_prediction_1.tflite -o assets/models/magenta_arbitrary-image-stylization-v1-256_int8_prediction_1.tflite

curl https://storage.googleapis.com/download.tensorflow.org/models/tflite/task_library/style_transfer/android/magenta_arbitrary-image-stylization-v1-256_int8_transfer_1.tflite -o assets/models/magenta_arbitrary-image-stylization-v1-256_int8_transfer_1.tflite

Build

flutter build ios

flutter install

Follow the instructions to run the app on your iOS device.

For more information, see the iOS | Flutter documentation.