tflite_flutter_gdx_plus #

A community-maintained Flutter plugin that provides a flexible, low-latency Dart API for TensorFlow Lite inference. It binds directly to the TensorFlow Lite C API through Dart FFI and follows the structure of the native Java and Swift APIs.

The plugin supports CPU inference, multithreading, background-isolate execution, and platform acceleration through NNAPI and GPU delegates on Android, Metal and Core ML delegates on iOS, and XNNPACK on desktop.

Features #

• Run any compatible .tflite model from an asset, file, or byte buffer.
• Use single-input or multi-input/multi-output inference.
• Keep the UI responsive with IsolateInterpreter.
• Configure interpreter threads and hardware delegates.
• Target Android, iOS, Linux, macOS, and Windows.
• Use the included classification, detection, segmentation, pose, style transfer, super-resolution, question-answering, and reinforcement-learning examples.

Compatibility #

The package requires Dart 3.3 or newer. Use a compatible stable Flutter SDK.

Installation #

Add the maintained package:

flutter pub add tflite_flutter_gdx_plus

Or add it directly to your application's pubspec.yaml:

dependencies:
  tflite_flutter_gdx_plus: ^0.12.3

Then import the public library:

import 'package:tflite_flutter_gdx_plus/tflite_flutter_gdx_plus.dart';

Platform setup #

Android #

Android dependencies are downloaded by Gradle. Build and install on a connected device:

flutter build apk
flutter install

The current native Android libraries require API level 26 or newer at runtime.

iOS #

iOS dependencies are downloaded by Swift Package Manager on Flutter 3.44 and newer. CocoaPods remains supported for older or opted-out projects. Build and install from the example application's directory:

flutter build ios
flutter install

TensorFlow Lite may not work in every iOS simulator configuration, so testing on a physical device is recommended.

When creating an IPA, Xcode can strip symbols required by Dart FFI and report Failed to lookup symbol ... symbol not found. In Xcode, open Runner > Build Settings > Strip Style and change All Symbols to Non-Global Symbols.

macOS #

The plugin bundles a universal TensorFlow Lite C dynamic library for arm64 and x86_64 macOS builds. Flutter 3.44 and newer links it with Swift Package Manager; CocoaPods remains supported for older or opted-out projects.

Linux #

Build the TensorFlow Lite C .so, create a blobs directory at the application root, and copy the library there as libtensorflowlite_c-linux.so. Add this to the application's linux/CMakeLists.txt:

install(
  FILES ${PROJECT_BUILD_DIR}/../blobs/libtensorflowlite_c-linux.so
  DESTINATION ${INSTALL_BUNDLE_DATA_DIR}/../blobs/
)

Windows #

Build the TensorFlow Lite C DLL, create a blobs directory at the application root, and copy the library there as libtensorflowlite_c-win.dll. Add this to the application's windows/CMakeLists.txt:

install(
  FILES ${PROJECT_BUILD_DIR}/../blobs/libtensorflowlite_c-win.dll
  DESTINATION ${INSTALL_BUNDLE_DATA_DIR}/../blobs/
)

Usage #

Create an interpreter from an asset #

Place the model in your application, declare it under flutter.assets in pubspec.yaml, and load it:

final interpreter =
    await Interpreter.fromAsset('assets/your_model.tflite');

The API also supports creating an interpreter from a file or buffer. See the API reference for the available constructors and options.

Run inference #

For one input and one output:

final input = [
  [1.23, 6.54, 7.81, 3.21, 2.22],
];
final output = List<double>.filled(2, 0).reshape([1, 2]);

interpreter.run(input, output);
print(output);

For multiple inputs and outputs:

final inputs = [
  [1.23],
  [2.43],
];
final outputs = <int, Object>{
  0: List<double>.filled(1, 0),
  1: List<double>.filled(1, 0),
};

interpreter.runForMultipleInputs(inputs, outputs);
print(outputs);

Always release native resources when inference is complete:

interpreter.close();

Run inference in a background isolate #

Create the regular interpreter, then wrap its native address:

final interpreter =
    await Interpreter.fromAsset('assets/your_model.tflite');
final isolateInterpreter =
    await IsolateInterpreter.create(address: interpreter.address);

await isolateInterpreter.run(input, output);
await isolateInterpreter.runForMultipleInputs(inputs, outputs);

await isolateInterpreter.close();
interpreter.close();

IsolateInterpreter performs inference away from the main isolate to avoid blocking UI work.

Examples #

The example directory contains complete applications for:

• audio, digit, gesture, image, and text classification;
• BERT question answering;
• image segmentation and pose estimation;
• SSD MobileNet object detection;
• style transfer and ESRGAN super resolution;
• reinforcement learning.

Several examples download model files through their own scripts directory. Read the example's README before building it.

TFLite Flutter Helper Library #

The former helper library is deprecated. For higher-level vision and media tasks, evaluate MediaPipe for Flutter.

Contributing #

Read CONTRIBUTING.md before opening a change. This repository uses Melos:

dart pub global activate melos
melos bootstrap
flutter test
flutter analyze

FFI bindings are generated with ffigen:

melos run ffigen

Do not hand-edit lib/src/bindings/tensorflow_lite_bindings_generated.dart.

Maintained Package #

This package is community maintained because the upstream project is no longer actively maintained. This repository continues dropout/flutter-tflite and the TensorFlow flutter-tflite project on which it is based. TensorFlow's repository is itself a managed fork of Amish Garg's original tflite_flutter_plugin.

Original project credits, copyright notices, contributor attribution, and license terms remain intact. See AUTHORS, NOTICE, THIRD_PARTY_NOTICES, the repository history, and LICENSE.

Feature Requests #

Feature requests and Pull Requests are always welcome.

• Request a feature
• Report or review bugs
• Open or review Pull Requests

License and acknowledgements #

Licensed under the Apache License, Version 2.0. See LICENSE.

The original authors and contributors are credited in AUTHORS. Special thanks remain due to Amish Garg, the original author and Google Summer of Code participant, and to the TensorFlow maintainers and all contributors whose work forms the foundation of this package.

Inherited BSD-licensed source notices and terms are preserved in THIRD_PARTY_NOTICES.

Need help or a custom solution? #

Need help integrating this package, maintaining an existing project, or building a custom web, mobile, or AI solution? Get in touch to discuss your requirements:

• Contact Gurwinder DevX
• Hire me on Upwork

Author and support #

Maintained by Gurwinder Singh, a full-stack web and mobile application developer and founder of Gurwinder DevX.

• GitHub
• LinkedIn
• Upwork
• Buy Me a Coffee

If this package helps your project, consider supporting its continued development through Buy Me a Coffee.